@fugood/llama.node 0.3.13 → 0.3.14

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -719,28 +719,28 @@ static inline __m128i packNibbles( __m256i bytes ) {
 }
 #endif  //__loongarch_asx
 
-void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) {
+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 * restrict x, void * restrict y, int64_t 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 * restrict x, void * restrict y, int64_t 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 * restrict x, void * restrict y, int64_t 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(const float * restrict x, void * restrict vy, int64_t k) {
+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 * restrict y = vy;
+    block_q8_0 * GGML_RESTRICT y = vy;
 
 #if defined(__ARM_NEON)
     for (int i = 0; i < nb; i++) {
@@ -1011,6 +1011,38 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
         __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
 
     }
+#elif 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
@@ -1018,11 +1050,11 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
 #endif
 }
 
-void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) {
+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 * restrict y = vy;
+    block_q8_1 * GGML_RESTRICT y = vy;
 
 #if defined(__ARM_NEON)
     for (int i = 0; i < nb; i++) {
@@ -1337,6 +1369,44 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
         __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
         __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
     }
+#elif 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
@@ -1358,8 +1428,8 @@ static inline int nearest_int(float fval) {
     return (i & 0x007fffff) - 0x00400000;
 }
 
-static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type,
-        const float * restrict qw) {
+static float make_qx_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, int rmse_type,
+        const float * GGML_RESTRICT qw) {
     float max = 0;
     float amax = 0;
     for (int i = 0; i < n; ++i) {
@@ -1427,7 +1497,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
     return scale;
 }
 
-static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, bool do_rmse) {
+static float make_q3_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, bool do_rmse) {
     float max = 0;
     float amax = 0;
     for (int i = 0; i < n; ++i) {
@@ -1486,7 +1556,7 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t *
     return 1/iscale;
 }
 
-static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
+static float make_qkx1_quants(int n, int nmax, const float * GGML_RESTRICT x, uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min,
         int ntry, float alpha) {
     float min = x[0];
     float max = x[0];
@@ -1529,8 +1599,8 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
     return scale;
 }
 
-static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
-        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, const float * GGML_RESTRICT weights,
+        uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, uint8_t * GGML_RESTRICT Laux,
         float rmin, float rdelta, int nstep, bool use_mad) {
     float min = x[0];
     float max = x[0];
@@ -1610,7 +1680,7 @@ static float make_qkx2_quants(int n, int nmax, const float * restrict x, const f
     return scale;
 }
 
-static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
+static inline void get_scale_min_k4(int j, const uint8_t * GGML_RESTRICT q, uint8_t * GGML_RESTRICT d, uint8_t * GGML_RESTRICT m) {
     if (j < 4) {
         *d = q[j] & 63; *m = q[j + 4] & 63;
     } else {
@@ -1621,51 +1691,51 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t *
 
 //========================- 2-bit (de)-quantization
 
-void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) {
+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 * restrict x, void * restrict vy, int64_t k) {
+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 * restrict x, void * restrict vy, int64_t k) {
+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 * restrict y = vy;
+    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 * restrict x, void * restrict vy, int64_t k) {
+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 * restrict y = vy;
+    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 * restrict x, void * restrict vy, int64_t k) {
+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 * restrict y = vy;
+    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 * restrict x, void * restrict vy, int64_t k) {
+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 * restrict y = vy;
+    block_tq1_0 * GGML_RESTRICT y = vy;
     quantize_row_tq1_0_ref(x, y, k);
 }
 
-void quantize_row_tq2_0(const float * restrict x, void * restrict vy, int64_t 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 * restrict y = vy;
+    block_tq2_0 * GGML_RESTRICT y = vy;
     quantize_row_tq2_0_ref(x, y, k);
 }
 
@@ -1673,11 +1743,11 @@ static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -1
 
 //===================================== Q8_K ==============================================
 
-void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t 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 * restrict yc = y; // Cast to proper type
+    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;
@@ -1839,7 +1909,7 @@ static inline __m128i get_scale_shuffle(int i) {
 }
 #endif
 
-void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
 
@@ -1854,23 +1924,23 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q4_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
 
 #if defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
-        const block_q4_0 * restrict vx0 = vx;
-        const block_q4_0 * restrict vx1 = (const block_q4_0 *) ((const uint8_t*)vx + bx);
-        const block_q8_0 * restrict vy0 = vy;
-        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
+        const block_q4_0 * GGML_RESTRICT vx0 = vx;
+        const block_q4_0 * GGML_RESTRICT vx1 = (const block_q4_0 *) ((const uint8_t*)vx + bx);
+        const block_q8_0 * GGML_RESTRICT vy0 = vy;
+        const block_q8_0 * GGML_RESTRICT vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
 
         float32x4_t sumv0 = vdupq_n_f32(0.0f);
 
         for (int i = 0; i < nb; i++) {
-            const block_q4_0 * restrict b_x0 = &vx0[i];
-            const block_q4_0 * restrict b_x1 = &vx1[i];
-            const block_q8_0 * restrict b_y0 = &vy0[i];
-            const block_q8_0 * restrict b_y1 = &vy1[i];
+            const block_q4_0 * GGML_RESTRICT b_x0 = &vx0[i];
+            const block_q4_0 * GGML_RESTRICT b_x1 = &vx1[i];
+            const block_q8_0 * GGML_RESTRICT b_y0 = &vy0[i];
+            const block_q8_0 * GGML_RESTRICT b_y1 = &vy1[i];
 
             const uint8x16_t m4b = vdupq_n_u8(0x0F);
             const int8x16_t  s8b = vdupq_n_s8(0x8);
@@ -1947,10 +2017,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
                 const svbool_t ph4 = svptrue_pat_b32(SV_VL4);
 
                 for (; ib + 1 < nb; ib += 2) {
-                    const block_q4_0 * restrict x0 = &x[ib + 0];
-                    const block_q4_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
+                    const block_q4_0 * GGML_RESTRICT x0 = &x[ib + 0];
+                    const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
 
                     // load x
                     const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
@@ -1993,10 +2063,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
                 const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
 
                 for (; ib + 1 < nb; ib += 2) {
-                    const block_q4_0 * restrict x0 = &x[ib + 0];
-                    const block_q4_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
+                    const block_q4_0 * GGML_RESTRICT x0 = &x[ib + 0];
+                    const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
 
                     // load x
                     const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
@@ -2034,10 +2104,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
                 const svbool_t pl16 = svnot_b_z(ph32, ph16);
 
                 for (; ib + 1 < nb; ib += 2) {
-                    const block_q4_0 * restrict x0 = &x[ib + 0];
-                    const block_q4_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
+                    const block_q4_0 * GGML_RESTRICT x0 = &x[ib + 0];
+                    const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
 
                     // load x
                     const svuint8_t qx0r = svld1rq_u8(ph32, x0->qs);
@@ -2074,10 +2144,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
     for (; ib + 1 < nb; ib += 2) {
-        const block_q4_0 * restrict x0 = &x[ib + 0];
-        const block_q4_0 * restrict x1 = &x[ib + 1];
-        const block_q8_0 * restrict y0 = &y[ib + 0];
-        const block_q8_0 * restrict y1 = &y[ib + 1];
+        const block_q4_0 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
         const int8x16_t  s8b = vdupq_n_s8(0x8);
@@ -2119,10 +2189,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     const v128_t s8b = wasm_i8x16_splat(0x8);
 
     for (; ib + 1 < nb; ib += 2) {
-        const block_q4_0 * restrict x0 = &x[ib];
-        const block_q4_0 * restrict x1 = &x[ib + 1];
-        const block_q8_0 * restrict y0 = &y[ib];
-        const block_q8_0 * restrict y1 = &y[ib + 1];
+        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);
@@ -2488,6 +2558,37 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     }
 
     sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#elif 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;
@@ -2508,7 +2609,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     *s = sumf;
 }
 
-void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     const int qk = QK8_1;
     const int nb = n / qk;
 
@@ -2523,24 +2624,24 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q4_1 * restrict x = vx;
-    const block_q8_1 * restrict y = vy;
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
 
 #if defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
-        const block_q4_1 * restrict vx0 = vx;
-        const block_q4_1 * restrict vx1 = (const block_q4_1 *) ((const uint8_t*)vx + bx);
-        const block_q8_1 * restrict vy0 = vy;
-        const block_q8_1 * restrict vy1 = (const block_q8_1 *) ((const uint8_t*)vy + by);
+        const block_q4_1 * GGML_RESTRICT vx0 = vx;
+        const block_q4_1 * GGML_RESTRICT vx1 = (const block_q4_1 *) ((const uint8_t*)vx + bx);
+        const block_q8_1 * GGML_RESTRICT vy0 = vy;
+        const block_q8_1 * GGML_RESTRICT vy1 = (const block_q8_1 *) ((const uint8_t*)vy + by);
 
         float32x4_t sumv0 = vdupq_n_f32(0.0f);
         float32x4_t summs0 = vdupq_n_f32(0.0f);
 
         for (int i = 0; i < nb; i++) {
-            const block_q4_1 * restrict b_x0 = &vx0[i];
-            const block_q4_1 * restrict b_x1 = &vx1[i];
-            const block_q8_1 * restrict b_y0 = &vy0[i];
-            const block_q8_1 * restrict b_y1 = &vy1[i];
+            const block_q4_1 * GGML_RESTRICT b_x0 = &vx0[i];
+            const block_q4_1 * GGML_RESTRICT b_x1 = &vx1[i];
+            const block_q8_1 * GGML_RESTRICT b_y0 = &vy0[i];
+            const block_q8_1 * GGML_RESTRICT b_y1 = &vy1[i];
 
             float32_t summs_t[4] = {
                 GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
@@ -2614,10 +2715,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     float summs = 0;
 
     for (; ib + 1 < nb; ib += 2) {
-        const block_q4_1 * restrict x0 = &x[ib + 0];
-        const block_q4_1 * restrict x1 = &x[ib + 1];
-        const block_q8_1 * restrict y0 = &y[ib + 0];
-        const block_q8_1 * restrict y1 = &y[ib + 1];
+        const block_q4_1 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q4_1 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
 
         summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
 
@@ -2781,6 +2882,35 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     }
 
     sumf = hsum_float_8(acc) + summs;
+#elif 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;
@@ -2801,7 +2931,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     *s = sumf;
 }
 
-void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     const int qk = QK8_0;
     const int nb = n / qk;
 
@@ -2816,8 +2946,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q5_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
 
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
@@ -2830,10 +2960,10 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     uint64_t tmp1[4];
 
     for (; ib + 1 < nb; ib += 2) {
-        const block_q5_0 * restrict x0 = &x[ib];
-        const block_q5_0 * restrict x1 = &x[ib + 1];
-        const block_q8_0 * restrict y0 = &y[ib];
-        const block_q8_0 * restrict y1 = &y[ib + 1];
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib];
+        const block_q5_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];
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -2894,8 +3024,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 
     // TODO: check if unrolling this is better
     for (; ib < nb; ++ib) {
-        const block_q5_0 * restrict x0 = &x[ib];
-        const block_q8_0 * restrict y0 = &y[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);
 
@@ -3156,7 +3286,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     *s = sumf;
 }
 
-void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     const int qk = QK8_1;
     const int nb = n / qk;
 
@@ -3171,8 +3301,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q5_1 * restrict x = vx;
-    const block_q8_1 * restrict y = vy;
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
 
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
@@ -3188,10 +3318,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     uint64_t tmp1[4];
 
     for (; ib + 1 < nb; ib += 2) {
-        const block_q5_1 * restrict x0 = &x[ib];
-        const block_q5_1 * restrict x1 = &x[ib + 1];
-        const block_q8_1 * restrict y0 = &y[ib];
-        const block_q8_1 * restrict y1 = &y[ib + 1];
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
+        const block_q5_1 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
+        const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -3257,8 +3387,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
     // TODO: check if unrolling this is better
     for (; ib < nb; ++ib) {
-        const block_q5_1 * restrict x0 = &x[ib];
-        const block_q8_1 * restrict y0 = &y[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);
 
@@ -3530,7 +3660,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     *s = sumf;
 }
 
-void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     const int qk = QK8_0;
     const int nb = n / qk;
 
@@ -3545,24 +3675,24 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q8_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
 
 #if defined(__ARM_FEATURE_MATMUL_INT8)
     if (nrc == 2) {
-        const block_q8_0 * restrict vx0 = vx;
-        const block_q8_0 * restrict vx1 = (const block_q8_0 *) ((const uint8_t*)vx + bx);
-        const block_q8_0 * restrict vy0 = vy;
-        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
+        const block_q8_0 * GGML_RESTRICT vx0 = vx;
+        const block_q8_0 * GGML_RESTRICT vx1 = (const block_q8_0 *) ((const uint8_t*)vx + bx);
+        const block_q8_0 * GGML_RESTRICT vy0 = vy;
+        const block_q8_0 * GGML_RESTRICT vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
 
         float32x4_t sumv0 = vdupq_n_f32(0.0f);
 
         for (int i = 0; i < nb; i++) {
-            const block_q8_0 * restrict b_x0 = &vx0[i];
-            const block_q8_0 * restrict b_y0 = &vy0[i];
+            const block_q8_0 * GGML_RESTRICT b_x0 = &vx0[i];
+            const block_q8_0 * GGML_RESTRICT b_y0 = &vy0[i];
 
-            const block_q8_0 * restrict b_x1 = &vx1[i];
-            const block_q8_0 * restrict b_y1 = &vy1[i];
+            const block_q8_0 * GGML_RESTRICT b_x1 = &vx1[i];
+            const block_q8_0 * GGML_RESTRICT b_y1 = &vy1[i];
 
             const int8x16_t x0_l = vld1q_s8(b_x0->qs);
             const int8x16_t x0_h = vld1q_s8(b_x0->qs + 16);
@@ -3627,10 +3757,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
                 const svbool_t pl16 = svptrue_pat_b32(SV_VL4);
 
                 for (; ib + 1 < nb; ib += 2) {
-                    const block_q8_0 * restrict x0 = &x[ib + 0];
-                    const block_q8_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT x0 = &x[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT x1 = &x[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
 
                     // load x
                     const svint8_t qx0_0 = svld1_s8(ph16, x0->qs);
@@ -3658,10 +3788,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
             {
                 //printf("sve256");
                 for (; ib + 1 < nb; ib += 2) {
-                    const block_q8_0 * restrict x0 = &x[ib + 0];
-                    const block_q8_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT x0 = &x[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT x1 = &x[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
 
                     // load x
                     const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
@@ -3694,10 +3824,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
                 svfloat32_t sumv00 = svdup_n_f32(0.0f);
 
                 for (; ib + 1 < nb; ib += 2) {
-                    const block_q8_0 * restrict x0 = &x[ib + 0];
-                    const block_q8_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT x0 = &x[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT x1 = &x[ib + 1];
+                    const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+                    const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
 
                     //load 32 int8_t in first half of vector and put another 32 int8_t in second vector lower bits
                     // and add them to make one 64 element vector
@@ -3737,10 +3867,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
     for (; ib + 1 < nb; ib += 2) {
-        const block_q8_0 * restrict x0 = &x[ib + 0];
-        const block_q8_0 * restrict x1 = &x[ib + 1];
-        const block_q8_0 * restrict y0 = &y[ib + 0];
-        const block_q8_0 * restrict y1 = &y[ib + 1];
+        const block_q8_0 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q8_0 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
 
         const int8x16_t x0_0 = vld1q_s8(x0->qs);
         const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
@@ -3767,8 +3897,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     v128_t sumv = wasm_f32x4_splat(0.0f);
 
     for (; ib < nb; ++ib) {
-        const block_q8_0 * restrict x0 = &x[ib];
-        const block_q8_0 * restrict y0 = &y[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);
@@ -3915,6 +4045,27 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     }
 
     sumf = hsum_float_8(acc);
+#elif 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;
@@ -3929,15 +4080,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     *s = sumf;
 }
 
-void ggml_vec_dot_tq1_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
     UNUSED(by);
     UNUSED(bs);
 
-    const block_tq1_0 * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
+    const block_tq1_0 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -4252,15 +4403,15 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * restrict s, size_t bs, const void *
 #endif
 }
 
-void ggml_vec_dot_tq2_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
     UNUSED(by);
     UNUSED(bs);
 
-    const block_tq2_0 * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
+    const block_tq2_0 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -4424,19 +4575,264 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * restrict s, size_t bs, const void *
 #endif
 }
 
-void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q2_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
+    const block_q2_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
-#ifdef __ARM_NEON
+#ifdef __ARM_FEATURE_SVE
+    const int vector_length = svcntb()*8;
+    const svuint8_t m3s = svdup_n_u8(0x3);
+    const svuint32_t m4s = svdup_n_u32(0xF);
+    const svint32_t vzero_sv = svdup_n_s32(0);
+    svfloat32_t acc_sum = svdup_n_f32(0);
+    svbool_t pred_s32 = svptrue_pat_b32(SV_VL4);
+
+    switch (vector_length) {
+        case 128:
+            for (int i = 0; i < nb; ++i) {
+                const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+                svfloat32_t d_broad = svdup_n_f32((float32_t)d);
+                const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+                svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);
+
+                const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+                const int8_t  * GGML_RESTRICT q8_sv = y[i].qs;
+                const uint8_t * GGML_RESTRICT sc = x[i].scales;
+
+                svuint32_t mins_and_scales_sve = svld1ub_u32(svptrue_b32(), sc);
+                const svint32_t mins_sv_1 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_b32(), mins_and_scales_sve, 4));
+
+                mins_and_scales_sve = svld1ub_u32(svptrue_b32(), sc+4);
+                const svint32_t mins_sv_2 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_b32(), mins_and_scales_sve, 4));
+
+                svint32_t q8sums_sv_1 = svld1sh_s32(svptrue_b32(), y[i].bsums);
+                svint32_t q8sums_sv_2 = svld1sh_s32(svptrue_b32(), y[i].bsums+4);
+
+                const svint32_t s0 = svadd_s32_x(svptrue_b32(), svmul_s32_x(svptrue_b32(), mins_sv_1, q8sums_sv_1), svmul_s32_x(svptrue_b32(), mins_sv_2, q8sums_sv_2));
+
+                mins_and_scales_sve = svld1ub_u32(svptrue_b32(), sc+8);
+                const svint32_t mins_sv_3 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_b32(), mins_and_scales_sve, 4));
+
+                mins_and_scales_sve = svld1ub_u32(svptrue_b32(), sc+12);
+                const svint32_t mins_sv_4 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_b32(), mins_and_scales_sve, 4));
+
+                q8sums_sv_1 = svld1sh_s32(svptrue_b32(), y[i].bsums+8);
+                q8sums_sv_2 = svld1sh_s32(svptrue_b32(), y[i].bsums+12);
+
+                svint32_t s1 = svadd_s32_x(svptrue_b32(), svmul_s32_x(svptrue_b32(), mins_sv_3, q8sums_sv_1), svmul_s32_x(svptrue_b32(), mins_sv_4, q8sums_sv_2));
+
+                svfloat32_t temp = svcvt_f32_s32_x(svptrue_b32(), svadd_s32_x(svptrue_b32(), s0, s1));
+
+                acc_sum = svmla_f32_m(svptrue_b32(), acc_sum, temp, dmin_broad);
+
+                svint32_t sumi1 = svdup_n_s32(0);
+
+                {
+                    const svuint8_t q2bits_1 = svld1_u8(svptrue_b8(), q2);
+                    svint8_t q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), q2bits_1, m3s));
+                    svint8_t q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+                    const svint32_t scales_sv = svreinterpret_s32_u32(svand_u32_m(svptrue_b32(), svld1ub_u32(svptrue_b32(), sc), m4s));
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv, 0));
+
+                    const svuint8_t q2bits_3 = svld1_u8(svptrue_b8(), q2+16);
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), q2bits_3, m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv, 1));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_1, 2), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv, 2));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_3, 2), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv, 3));
+
+
+                    const svint32_t scales_sv_1 = svreinterpret_s32_u32(svand_u32_m(svptrue_b32(), svld1ub_u32(svptrue_b32(), sc+4), m4s));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_1, 4), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_1, 0));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_3, 4), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_1, 1));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_1, 6), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_1, 2));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_3, 6), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_1, 3));
+
+                    //-------------------------------
+
+                    q2 += 32;
+                    const svint32_t scales_sv_2 = svreinterpret_s32_u32(svand_u32_m(svptrue_b32(), svld1ub_u32(svptrue_b32(), sc+8), m4s));
+                    const svuint8_t q2bits_2 = svld1_u8(svptrue_b8(), q2);
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), q2bits_2, m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_2, 0));
+
+                    const svuint8_t q2bits_4 = svld1_u8(svptrue_b8(), q2+16);
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), q2bits_4, m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_2, 1));
+
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_2, 2), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_2, 2));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_4, 2), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_2, 3));
+
+
+                    const svint32_t scales_sv_3 = svreinterpret_s32_u32(svand_u32_m(svptrue_b32(), svld1ub_u32(svptrue_b32(), sc+12), m4s));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_2, 4), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_3, 0));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_4, 4), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_3, 1));
+
+
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_2, 6), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_3, 2));
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_4, 6), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                    sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_3, 3));
+                }
+                acc_sum = svmla_f32_m(svptrue_b32(), acc_sum, svcvt_f32_s32_x(svptrue_b32(), sumi1), d_broad);
+            }
+            *s = svaddv_f32(svptrue_b32(), acc_sum);
+            break;
+
+        case 256:
+        case 512:
+            for (int i = 0; i < nb; ++i) {
+                const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+                svfloat32_t d_broad = svdup_n_f32((float32_t)d);
+                const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+                svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);
+
+                const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+                const int8_t  * GGML_RESTRICT q8_sv = y[i].qs;
+                const uint8_t * GGML_RESTRICT sc = x[i].scales;
+
+                const svuint32_t mins_and_scales_sve = svld1ub_u32(svptrue_pat_b32(SV_VL8), sc); sc += 8;
+                const svint32_t scales_sv = svreinterpret_s32_u32(svand_u32_m(svptrue_pat_b32(SV_VL8), mins_and_scales_sve, m4s));
+                const svint32_t mins_sv_1 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_pat_b32(SV_VL8), mins_and_scales_sve, 4));
+                svint32_t q8sums_sv_1 = svld1sh_s32(svptrue_pat_b32(SV_VL8), y[i].bsums);
+
+                const svuint32_t mins_and_scales_sve_1 = svld1ub_u32(svptrue_pat_b32(SV_VL8), sc);
+                const svint32_t scales_sv_1 = svreinterpret_s32_u32(svand_u32_m(svptrue_pat_b32(SV_VL8), mins_and_scales_sve_1, m4s));
+                const svint32_t mins_sv_2 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_pat_b32(SV_VL8), mins_and_scales_sve_1, 4));
+
+                svint32_t q8sums_sv_2 = svld1sh_s32(svptrue_pat_b32(SV_VL8), y[i].bsums+8);
+
+                svfloat32_t temp = svcvt_f32_s32_x(svptrue_pat_b32(SV_VL8), svadd_s32_x(svptrue_pat_b32(SV_VL8), svmul_s32_x(svptrue_pat_b32(SV_VL8), mins_sv_1, q8sums_sv_1), svmul_s32_x(svptrue_pat_b32(SV_VL8), mins_sv_2, q8sums_sv_2)));
+
+                acc_sum = svmla_f32_m(svptrue_pat_b32(SV_VL8), acc_sum, temp, dmin_broad);
+
+                svint32_t sumi1 = svdup_n_s32(0);
+
+                {
+                    const svuint8_t q2bits_1 = svld1_u8(svptrue_pat_b8(SV_VL32), q2);
+                    svint8_t q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), q2bits_1, m3s));
+                    svint8_t q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                    svint32_t scale_1 = svsel(pred_s32, svdup_lane_s32(scales_sv, 0), svdup_lane_s32(scales_sv, 1));
+                    sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_1);
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_1, 2), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                    svint32_t scale_2 = svsel(pred_s32, svdup_lane_s32(scales_sv, 2), svdup_lane_s32(scales_sv, 3));
+                    sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(svdup_n_s32(0), q2bytes_sv, q8bytes_sv), scale_2);
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_1, 4), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                    scale_1 = svsel(pred_s32, svdup_lane_s32(scales_sv, 4), svdup_lane_s32(scales_sv, 5));
+                    sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_1);
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_1, 6), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                    scale_2 = svsel(pred_s32, svdup_lane_s32(scales_sv, 6), svdup_lane_s32(scales_sv, 7));
+                    sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_2);
+
+                    q2 += 32;
+
+                    const svuint8_t q2bits_2 = svld1_u8(svptrue_pat_b8(SV_VL32), q2);
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), q2bits_2, m3s));
+                    q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                    scale_1 = svsel(pred_s32, svdup_lane_s32(scales_sv_1, 0), svdup_lane_s32(scales_sv_1, 1));
+                    sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_1);
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_2, 2), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                    scale_2 = svsel(pred_s32, svdup_lane_s32(scales_sv_1, 2), svdup_lane_s32(scales_sv_1, 3));
+                    sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_2);
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_2, 4), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                    scale_1 = svsel(pred_s32, svdup_lane_s32(scales_sv_1, 4), svdup_lane_s32(scales_sv_1, 5));
+                    sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_1);
+
+                    q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_2, 6), m3s));
+                    q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                    scale_2 = svsel(pred_s32, svdup_lane_s32(scales_sv_1, 6), svdup_lane_s32(scales_sv_1, 7));
+                    sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_2);
+                }
+                acc_sum = svmla_f32_m(svptrue_pat_b32(SV_VL8), acc_sum, svcvt_f32_s32_x(svptrue_pat_b32(SV_VL8), sumi1), d_broad);
+            }
+            *s = svaddv_f32(svptrue_pat_b32(SV_VL8), acc_sum);
+            break;
+
+        default:
+            assert(false && "Unsupported vector length");
+            break;
+    }
+
+#elif __ARM_NEON
     const uint8x16_t m3 = vdupq_n_u8(0x3);
     const uint8x16_t m4 = vdupq_n_u8(0xF);
 
@@ -4451,9 +4847,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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 uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-        const uint8_t * restrict sc = x[i].scales;
+        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT sc = x[i].scales;
 
         const uint8x16_t mins_and_scales = vld1q_u8(sc);
         const uint8x16_t scales = vandq_u8(mins_and_scales, m4);
@@ -4516,8 +4912,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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 uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
         const __m128i scales8 = _mm_and_si128(mins_and_scales, m4);
@@ -4583,8 +4979,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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);
 
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         // load mins and scales from block_q2_K.scales[QK_K/16]
         const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
@@ -4910,8 +5306,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed int vsumi6 = v0;
         vector signed int vsumi7 = v0;
 
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/128; ++j) {
             __builtin_prefetch(q2, 0, 1);
@@ -5002,8 +5398,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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 uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         const __m128i mins_and_scales128 = __lsx_vld((const __m128i*)x[i].scales, 0);
         const __m128i scales128 = __lsx_vandi_b(mins_and_scales128, 0xf);
@@ -5096,7 +5492,7 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 #endif
 }
 
-void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -5107,12 +5503,187 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     const uint32_t kmask1 = 0x03030303;
     const uint32_t kmask2 = 0x0f0f0f0f;
 
-    const block_q3_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
-#ifdef __ARM_NEON
+#if defined(__ARM_FEATURE_SVE)
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    const int8_t m32 = 32;
+    const int vector_length = svcntb()*8;
+    const svuint8_t m3b_sv = svdup_n_u8(0x3);
+    const svint32_t vzero_sv = svdup_n_s32(0);
+
+    const svuint8_t m0_sv = svdup_n_u8(1);
+    const svuint8_t m1_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 1);
+    const svuint8_t m2_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 2);
+    const svuint8_t m3_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 3);
+
+    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 * GGML_RESTRICT q3_sv = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh_sv = x[i].hmask;
+        const int8_t  * GGML_RESTRICT q8_sv = y[i].qs;
+
+        // Set up scales
+        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] -= m32;
+
+        switch (vector_length) {
+            case 128:
+                {
+                    svuint8_t qhbits_sv_1 = svld1_u8(svptrue_b8(), qh_sv);
+                    svuint8_t qhbits_sv_2 = svld1_u8(svptrue_b8(), qh_sv+16);
+                    svuint8_t q3h_sv;
+
+                    svint32_t sumi1_1 = svdup_n_s32(0);
+                    svint8_t q3bytes_sv;
+
+                    for (int j = 0; j < QK_K/128; ++j) {
+
+                        const svuint8_t q3bits_sv = svld1_u8(svptrue_b8(), q3_sv); q3_sv += 16;
+                        const svuint8_t q3bits_sv_1 = svld1_u8(svptrue_b8(), q3_sv); q3_sv += 16;
+                        svint8_t q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+                        svint8_t q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                        q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m0_sv, qhbits_sv_1), 2);
+                        q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), q3bits_sv, m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[0]));
+
+                        q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m0_sv, qhbits_sv_2), 2);
+                        q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), q3bits_sv_1, m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[1]));
+
+                        q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+                        q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                        q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m1_sv, qhbits_sv_1), 1);
+                        q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[2]));
+
+                        q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m1_sv, qhbits_sv_2), 1);
+                        q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[3]));
+
+
+                        scale += 4;
+                        q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+                        q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                        q3h_sv = svbic_u8_x(svptrue_b8(), m2_sv, qhbits_sv_1);
+                        q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[0]));
+
+                        q3h_sv = svbic_u8_x(svptrue_b8(), m2_sv, qhbits_sv_2);
+                        q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[1]));
+
+
+                        q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+                        q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
+
+                        q3h_sv = svlsr_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m3_sv, qhbits_sv_1), 1);
+                        q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[2]));
+
+                        q3h_sv = svlsr_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m3_sv, qhbits_sv_2), 1);
+                        q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[3]));
+
+                        if (j == 0) {
+                            qhbits_sv_1 = svlsr_n_u8_x(svptrue_b8(), qhbits_sv_1, 4);
+                            qhbits_sv_2 = svlsr_n_u8_x(svptrue_b8(), qhbits_sv_2, 4);
+                        }
+
+                        scale += 4;
+                    }
+
+                    sum += d * (svaddv_s32(svptrue_b32(), sumi1_1));
+                } break;
+            case 256:
+            case 512:
+                {
+                    svuint8_t qhbits_sv = svld1_u8(svptrue_pat_b8(SV_VL32), qh_sv);
+                    svuint8_t q3h_sv;
+
+                    svint32_t sumi1_1 = svdup_n_s32(0);
+                    svint8_t q3bytes_sv;
+
+                    for (int j = 0; j < QK_K/128; ++j) {
+
+                        const svuint8_t q3bits_sv = svld1_u8(svptrue_pat_b8(SV_VL32), q3_sv); q3_sv += 32;
+                        svint8_t q8bytes_1_sv_1 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+                        svint8_t q8bytes_1_sv_2 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                        q3h_sv = svlsl_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m0_sv, qhbits_sv), 2);
+                        q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), q3bits_sv, m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+
+                        svint32_t scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[0]), svdup_n_s32((int32_t)scale[1]));
+                        sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), scale_1);
+
+                        q3h_sv = svlsl_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m1_sv, qhbits_sv), 1);
+                        q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[2]), svdup_n_s32((int32_t)scale[3]));
+                        sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), scale_1);
+
+                        scale += 4;
+                        q8bytes_1_sv_1 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+                        q8bytes_1_sv_2 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
+
+                        q3h_sv = svbic_u8_x(svptrue_pat_b8(SV_VL32), m2_sv, qhbits_sv);
+                        q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[0]), svdup_n_s32((int32_t)scale[1]));
+                        sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), scale_1);
+
+                        q3h_sv = svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m3_sv, qhbits_sv), 1);
+                        q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
+
+                        scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[2]), svdup_n_s32((int32_t)scale[3]));
+                        sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), scale_1);
+
+                        if (j == 0) {
+                            qhbits_sv = svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), qhbits_sv, 4);
+                        }
+
+                        scale += 4;
+                    }
+
+                    sum += d * (svaddv_s32(svptrue_pat_b32(SV_VL8), sumi1_1));
+                } break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        }
+    }
+    *s = sum;
+
+#elif __ARM_NEON
 
     uint32_t aux[3];
     uint32_t utmp[4];
@@ -5134,9 +5705,9 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
 
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict qh = x[i].hmask;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
 
         ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
 
@@ -5220,8 +5791,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
 
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         // Set up scales
         memcpy(aux, x[i].scales, 12);
@@ -5325,8 +5896,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
 
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         // Set up scales
         aux = (const uint32_t *)x[i].scales;
@@ -5459,9 +6030,9 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
@@ -5548,9 +6119,9 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     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;
+        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);
@@ -5690,8 +6261,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed int vsumi6 = v0;
         vector signed int vsumi7 = v0;
 
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/128; ++j) {
             __builtin_prefetch(q3, 0, 1);
@@ -5804,8 +6375,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     for (int i = 0; i < nb; ++i) {
 
         const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
         // Set up scales
         memcpy(aux, x[i].scales, 12);
         __m128i scales128 = lsx_set_w(
@@ -5890,11 +6461,11 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
-        const  int8_t * restrict q8 = y[i].qs;
+        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 * restrict a = aux8;
+        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;
@@ -5937,7 +6508,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 }
 
-void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -5945,8 +6516,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q4_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -5981,8 +6552,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const uint8_t * scales = (const uint8_t *)utmp;
 
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         const int vector_length = ggml_cpu_get_sve_cnt()*8;
         const svuint8_t m4b = svdup_n_u8(0xf);
@@ -6069,8 +6640,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const uint8_t * scales = (const uint8_t *)utmp;
 
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         int32_t sumi1 = 0;
         int32_t sumi2 = 0;
@@ -6108,8 +6679,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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 * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        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);
@@ -6121,7 +6692,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         // Sum mins * q8sums
         int32_t sumi = 0;
-        const int16_t * restrict q8sums = y[i].bsums;
+        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];
@@ -6220,8 +6791,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         utmp[2] = uaux;
         utmp[0] &= kmask1;
 
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
 
@@ -6279,8 +6850,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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 uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -6380,8 +6951,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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 * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         vl = 32;
 
@@ -6482,8 +7053,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed int vsumi2 = v0;
         vector signed int vsumi3 = v0;
 
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/64; j+=2) {
             __builtin_prefetch(q4, 0, 1);
@@ -6574,8 +7145,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         utmp[2] = uaux;
         utmp[0] &= kmask1;
 
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         const __m128i mins_and_scales128 = lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]);
         const __m128i mins128 = __lsx_vexth_h_b(mins_and_scales128);
@@ -6622,6 +7193,77 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 
     *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
+#elif 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];
@@ -6635,10 +7277,10 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].qs;
-        const  int8_t * restrict q8 = y[i].qs;
+        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 * restrict a = aux8;
+        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;
@@ -6681,7 +7323,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 #endif
 }
 
-void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -6689,8 +7331,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q5_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -6732,9 +7374,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const uint8_t * scales = (const uint8_t *)utmp;
 
-        const uint8_t * restrict q5 = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
 
         ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
 
@@ -6779,8 +7421,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     float summs = 0.f;
 
     for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         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);
@@ -6863,8 +7505,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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 uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -6955,9 +7597,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         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 * restrict q5 = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
+        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);
@@ -6969,7 +7611,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         // Sum mins * q8sums
         int32_t sumi_mins = 0;
-        const int16_t * restrict q8sums = y[i].bsums;
+        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];
@@ -7073,9 +7715,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         vl = 8;
 
-        const uint8_t * restrict q5 = x[i].qs;
-        const uint8_t * restrict hm = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
+        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;
@@ -7214,8 +7856,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed int vsumi2 = v0;
         vector signed int vsumi3 = v0;
 
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/64; ++j) {
             __builtin_prefetch(q5, 0, 1);
@@ -7287,8 +7929,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     for (int i = 0; i < nb; ++i) {
 
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q5 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         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);
@@ -7351,7 +7993,94 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vbsrl_v(acc_m, 4));
 
     *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
+#elif 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];
@@ -7365,11 +8094,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].qs;
-        const uint8_t * restrict hm = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
+        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 * restrict a = aux8;
+        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);
@@ -7416,7 +8145,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 #endif
 }
 
-void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -7424,8 +8153,8 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     UNUSED(by);
     UNUSED(bs);
 
-    const block_q6_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -7445,11 +8174,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const float d_all = GGML_FP16_TO_FP32(x[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 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 * restrict scale = x[i].scales;
+        const int8_t * GGML_RESTRICT scale = x[i].scales;
 
         const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
         const int8x16_t scales = vld1q_s8(scale);
@@ -7536,9 +8265,9 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
 
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
 
         const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
 
@@ -7614,9 +8343,9 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
 
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
 
         // handle the q6_k -32 offset separately using bsums
         const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)y[i].bsums);
@@ -7715,8 +8444,8 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     for (int i = 0; i < nb; ++i) {
         // Unpack 6-bit quantized data into aux8 (unchanged)
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
+        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) {
@@ -7730,8 +8459,8 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             qh += 32;
         }
 
-        const int8_t * restrict a_ptr = aux8;
-        const int8_t * restrict q8 = y[i].qs;
+        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);
 
@@ -7794,11 +8523,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         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 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 * restrict scale = x[i].scales;
+        const int8_t * GGML_RESTRICT scale = x[i].scales;
 
         size_t vl;
 
@@ -7900,10 +8629,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed int vsumi6 = v0;
         vector signed int vsumi7 = v0;
 
-        const uint8_t * restrict q6 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict qs = x[i].scales;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q6 = x[i].ql;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const int8_t  * GGML_RESTRICT qs = x[i].scales;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/128; ++j) {
             __builtin_prefetch(q6, 0, 0);
@@ -8019,9 +8748,9 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
         const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
 
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
 
         const __m128i scales128 = __lsx_vld((const __m128i*)x[i].scales, 0);
         const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15};
@@ -8068,7 +8797,130 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     }
 
     *s = hsum_float_8(acc);
+#elif 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];
@@ -8079,11 +8931,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
+        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 * restrict a = aux8;
+        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;
@@ -8151,7 +9003,7 @@ static const int8_t keven_signs_q2xs[1024] = {
 };
 #endif
 
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -8159,8 +9011,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     UNUSED(by);
     UNUSED(bs);
 
-    const block_iq2_xxs * restrict x = vx;
-    const block_q8_K    * restrict y = vy;
+    const block_iq2_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -8178,8 +9030,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     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 * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        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; ib32 += 2) {
             q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
@@ -8215,8 +9067,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m256i sumi1 = _mm256_setzero_si256();
         __m256i sumi2 = _mm256_setzero_si256();
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
@@ -8256,8 +9108,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m128i sumi1_0 = _mm_setzero_si128();
         __m128i sumi1_1 = _mm_setzero_si128();
         __m128i sumi2_0 = _mm_setzero_si128();
@@ -8321,8 +9173,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
         vector signed int vsumi2 = v0;
         vector signed int vsumi3 = v0;
 
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t  *  restrict q8 = y[i].qs;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t  *  GGML_RESTRICT q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/32; j += 2) {
             __builtin_prefetch(q2, 0, 1);
@@ -8398,8 +9250,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m256i sumi1 = __lasx_xvldi(0);
         __m256i sumi2 = __lasx_xvldi(0);
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
@@ -8429,7 +9281,57 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     }
 
     *s = 0.125f * hsum_float_8(accumf);
-
+//#elif 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];
@@ -8438,8 +9340,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     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 * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        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));
@@ -8462,7 +9364,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
 #endif
 }
 
-void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -8470,8 +9372,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     UNUSED(by);
     UNUSED(bs);
 
-    const block_iq2_xs * restrict x = vx;
-    const block_q8_K   * restrict y = vy;
+    const block_iq2_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -8488,8 +9390,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     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 * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         const uint8x8_t scales8 = vld1_u8(x[i].scales);
         const uint8x8_t scales_l = vand_u8(scales8, vdup_n_u8(0xf));
         const uint8x8_t scales_h = vshr_n_u8(scales8, 4);
@@ -8566,8 +9468,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
         memcpy(&aux64, x[i].scales, 8);
         __m128i stmp = _mm_set1_epi64x(aux64);
@@ -8687,8 +9589,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
         memcpy(&aux64, x[i].scales, 8);
         __m128i stmp = _mm_set1_epi64x(aux64);
@@ -8842,8 +9744,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
+        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
         memcpy(&aux64, x[i].scales, 8);
         __m128i stmp = __lsx_vreplgr2vr_d(aux64);
@@ -8940,9 +9842,9 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
         vector signed int vsumi2 = v0;
         vector signed int vsumi3 = v0;
 
-        const uint16_t * restrict q2 = x[i].qs;
-        const uint8_t  * restrict sc = x[i].scales;
-        const int8_t  *  restrict q8 = y[i].qs;
+        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;
 
         for (int j = 0; j < QK_K/64; ++j) {
             __builtin_prefetch(q2, 0, 1);
@@ -9012,9 +9914,9 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     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 * restrict q2 = x[i].qs;
-        const uint8_t  * restrict sc = x[i].scales;
-        const int8_t   * restrict q8 = y[i].qs;
+        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;
@@ -9047,7 +9949,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 #endif
 }
 
-void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -9055,8 +9957,8 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
     UNUSED(by);
     UNUSED(bs);
 
-    const block_iq2_s * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -9082,10 +9984,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
 
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
 
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         int sumi1 = 0, sumi2 = 0;
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
@@ -9156,10 +10058,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         memcpy(&aux64, x[i].scales, 8);
         const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
@@ -9229,10 +10131,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         memcpy(&aux64, x[i].scales, 8);
         const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
@@ -9327,11 +10229,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
         vector signed int vsumi2 = v0;
         vector signed int vsumi3 = v0;
 
-        const uint8_t *  restrict q2 = x[i].qs;
-        const uint8_t *  restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const uint8_t *  restrict sc = x[i].scales;
-        const int8_t  *  restrict q8 = y[i].qs;
+        const uint8_t *  GGML_RESTRICT q2 = x[i].qs;
+        const uint8_t *  GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const uint8_t *  GGML_RESTRICT sc = x[i].scales;
+        const int8_t  *  GGML_RESTRICT q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/32; j += 2) {
             __builtin_prefetch(q2, 0, 1);
@@ -9428,10 +10330,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
     __m256 accumf = (__m256)__lasx_xvldi(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 qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
         __m128i tmp1;
         memcpy(&aux64, x[i].scales, 8);
@@ -9525,7 +10427,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
 
 }
 
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -9533,8 +10435,8 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     UNUSED(by);
     UNUSED(bs);
 
-    const block_iq3_xxs * restrict x = vx;
-    const block_q8_K    * restrict y = vy;
+    const block_iq3_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -9550,9 +10452,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     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 q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t   * restrict q8 = y[i].qs;
+        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;
         float sumf1 = 0, sumf2 = 0;
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
             q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
@@ -9588,9 +10490,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
         __m256i sumi1 = _mm256_setzero_si256();
         __m256i sumi2 = _mm256_setzero_si256();
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
@@ -9633,9 +10535,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
         __m128i sumi1_0 = _mm_setzero_si128();
         __m128i sumi1_1 = _mm_setzero_si128();
         __m128i sumi2_0 = _mm_setzero_si128();
@@ -9702,9 +10604,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
         vector signed int vsumi2 = v0;
         vector signed int vsumi3 = v0;
 
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint32_t * restrict signs = (const uint32_t *)(x[i].qs + QK_K/4);
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+        const uint32_t * GGML_RESTRICT signs = (const uint32_t *)(x[i].qs + QK_K/4);
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
 #pragma GCC unroll 1
         for (int j = 0; j < QK_K/32; j += 2) {
@@ -9776,9 +10678,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     __m256 accumf = (__m256)__lasx_xvldi(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 q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
         __m256i sumi1 = __lasx_xvldi(0);
         __m256i sumi2 = __lasx_xvldi(0);
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
@@ -9821,9 +10723,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
     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 * restrict q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t  * restrict q8 = y[i].qs;
+        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);
@@ -9848,7 +10750,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
 #endif
 }
 
-void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -9856,8 +10758,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
     UNUSED(by);
     UNUSED(bs);
 
-    const block_iq3_s * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -9894,10 +10796,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
     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 qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
-        const int8_t   * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
         memcpy(scales32, x[i].scales, 4);
         scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101;
@@ -9976,10 +10878,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
         __m256i sumi1 = _mm256_setzero_si256();
         __m256i sumi2 = _mm256_setzero_si256();
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
@@ -10061,10 +10963,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
         __m128i sumi1_0 = _mm_setzero_si128();
         __m128i sumi1_1 = _mm_setzero_si128();
         __m128i sumi2_0 = _mm_setzero_si128();
@@ -10162,11 +11064,11 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        const uint8_t *  restrict q3 = x[i].qs;
-        const uint8_t *  restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].signs);
-        const uint8_t *  restrict sc = x[i].scales;
-        const int8_t  *  restrict q8 = y[i].qs;
+        const uint8_t *  GGML_RESTRICT q3 = x[i].qs;
+        const uint8_t *  GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].signs);
+        const uint8_t *  GGML_RESTRICT sc = x[i].scales;
+        const int8_t  *  GGML_RESTRICT q8 = y[i].qs;
 
         vector signed int vsumi0 = v0;
         vector signed int vsumi1 = v0;
@@ -10273,10 +11175,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
     __m256 accumf = (__m256)__lasx_xvldi(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 qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
-        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
         __m256i sumi1 = __lasx_xvldi(0);
         __m256i sumi2 = __lasx_xvldi(0);
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
@@ -10334,10 +11236,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
     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 * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint8_t * restrict signs = x[i].signs;
-        const int8_t  * restrict q8 = y[i].qs;
+        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;
@@ -10389,7 +11291,7 @@ static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
 }
 #endif
 
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -10397,8 +11299,8 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
     UNUSED(by);
     UNUSED(bs);
 
-    const block_iq1_s * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -10460,10 +11362,19 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
         __m256i sumi = _mm256_setzero_si256();
         int sumi1 = 0;
         for (int ib = 0; ib < QK_K/32; ib += 2) {
+#ifdef __BMI2__
+            const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib], 0x700070007000700ULL);
+            const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib + 1], 0x700070007000700ULL);
+            const uint16_t *idx1 = (const uint16_t *)(&packed_idx1);
+            const uint16_t *idx2 = (const uint16_t *)(&packed_idx2);
+            const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[idx1[3]], iq1s_grid[idx1[2]], iq1s_grid[idx1[1]], iq1s_grid[idx1[0]]);
+            const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[idx2[3]], iq1s_grid[idx2[2]], iq1s_grid[idx2[1]], iq1s_grid[idx2[0]]);
+#else
             const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)],
                                                     iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
             const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)],
                                                     iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+#endif
             qs += 8;
             const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
@@ -10556,10 +11467,10 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
         vector signed int vsumi3 = vec_splats((int32_t)0);
         vector signed int vsumi8 = vec_splats((int32_t)0);
 
-        const uint8_t  * restrict q1 = x[i].qs;
-        const uint16_t * restrict qh = x[i].qh;
-        const int8_t   * restrict q8 = y[i].qs;
-        const int16_t  * restrict qs = y[i].bsums;
+        const uint8_t  * GGML_RESTRICT q1 = x[i].qs;
+        const uint16_t * GGML_RESTRICT qh = x[i].qh;
+        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
+        const int16_t  * GGML_RESTRICT qs = y[i].bsums;
 
         for (int j = 0; j < QK_K/32; j += 2) {
             __builtin_prefetch(q1, 0, 1);
@@ -10720,7 +11631,7 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
 #endif
 }
 
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
     UNUSED(nrc);
@@ -10728,8 +11639,8 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
     UNUSED(by);
     UNUSED(bs);
 
-    const block_iq1_m * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
+    const block_iq1_m * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -10809,6 +11720,10 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
 
     const __m256i mask = _mm256_set1_epi16(0x7);
     const __m256i mone = _mm256_set1_epi16(1);
+    const __m256i mone8 = _mm256_set1_epi8(1);
+    const __m256i mtwo8 = _mm256_set1_epi8(2);
+    // VPSHUFB cannot cross 128-bit lanes so odd shifts go to upper half.
+    const __m256i scales_shift = _mm256_set_epi64x(9, 3, 6, 0);
 
     __m256 accum1 = _mm256_setzero_ps();
     __m256 accum2 = _mm256_setzero_ps();
@@ -10820,10 +11735,33 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
         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);
+        // Extract 3-bit scales (16 values)
+        __m256i scales = _mm256_set1_epi64x(*(const uint64_t*)sc);
+        scales = _mm256_srlv_epi64(scales, scales_shift);
+        scales = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scales, mask), 1), mone);
+
+        // Indices to repeat each scale 8 times.
+        __m256i scales_idx1 = _mm256_set1_epi16(0x0100);
+        __m256i scales_idx2 = _mm256_add_epi8(scales_idx1, _mm256_set1_epi8(8));
 
         __m256i sumi1 = _mm256_setzero_si256();
         __m256i sumi2 = _mm256_setzero_si256();
         for (int ib = 0; ib < QK_K/32; ib += 2) {
+#ifdef __BMI2__
+            const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL)
+                                       | _pdep_u64(*(const uint16_t*)(qh) & 0x7777, 0xf000f000f000f00ULL);
+            const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL)
+                                       | _pdep_u64(*(const uint16_t*)(qh + 2) & 0x7777, 0xf000f000f000f00ULL);
+            const uint16_t *idx1 = (const uint16_t *)(&packed_idx1);
+            const uint16_t *idx2 = (const uint16_t *)(&packed_idx2);
+            const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[idx1[3]], iq1s_grid[idx1[2]], iq1s_grid[idx1[1]], iq1s_grid[idx1[0]]);
+            const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[idx2[3]], iq1s_grid[idx2[2]], iq1s_grid[idx2[1]], iq1s_grid[idx2[0]]);
+
+            // Convert signs to bytes 0x81 (negative) or 0x01 (positive)
+            const uint64_t delta_sign = _pdep_u64(*(const uint32_t*)(qh) & 0x88888888, 0xf0f0f0f0f0f0f0f0ULL);
+            const __m256i delta1 = _mm256_or_si256(mone8, _mm256_cvtepi8_epi64(_mm_set1_epi32(delta_sign)));
+            const __m256i delta2 = _mm256_or_si256(mone8, _mm256_cvtepi8_epi64(_mm_set1_epi32(delta_sign >> 32)));
+#else
             const __m256i q1b_1 = _mm256_set_epi64x(
                     iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)],
                     iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]
@@ -10832,11 +11770,6 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
                     iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)],
                     iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]
             );
-            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
-            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
 
             const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
                                                      qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
@@ -10846,15 +11779,21 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
                                                      qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
                                                      qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
                                                      qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+#endif
+            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
 
-            const __m256i dot3 = mul_add_epi8(delta1, q8b_1);
-            const __m256i dot4 = mul_add_epi8(delta2, q8b_2);
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+            const __m256i dot3 = _mm256_maddubs_epi16(mone8, _mm256_sign_epi8(q8b_1, delta1));
+            const __m256i dot4 = _mm256_maddubs_epi16(mone8, _mm256_sign_epi8(q8b_2, delta2));
+
+            __m256i scale1 = _mm256_shuffle_epi8(scales, scales_idx1);
+            __m256i scale2 = _mm256_shuffle_epi8(scales, scales_idx2);
 
-            __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0));
-            __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6));
+            scales_idx1 = _mm256_add_epi8(scales_idx1, mtwo8);
+            scales_idx2 = _mm256_add_epi8(scales_idx2, mtwo8);
 
-            scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone);
-            scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone);
             const __m256i p1 = _mm256_madd_epi16(dot1, scale1);
             const __m256i p2 = _mm256_madd_epi16(dot2, scale2);
             const __m256i p3 = _mm256_madd_epi16(dot3, scale1);
@@ -11010,7 +11949,7 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
 #endif
 }
 
-void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
@@ -11019,8 +11958,8 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
     assert(n % QK4_NL == 0);
     static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
 
-    const block_iq4_nl * restrict x = vx;
-    const block_q8_0   * restrict y = vy;
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
 
     const int nb = n / QK4_NL;
 
@@ -11190,6 +12129,27 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
 
     sumf = hsum_float_8(__lasx_xvfadd_s(accum1, accum2));
 
+#elif 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);
@@ -11203,7 +12163,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
     *s = sumf;
 }
 
-void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * 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) {
     assert(nrc == 1);
     UNUSED(nrc);
     UNUSED(bx);
@@ -11211,8 +12171,8 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     UNUSED(bs);
     assert(n % QK_K == 0);
 
-    const block_iq4_xs * restrict x = vx;
-    const block_q8_K   * restrict y = vy;
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
 
     const int nb = n / QK_K;
 
@@ -11369,9 +12329,9 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 
         uint16_t h = x[ibl].scales_h;
 
-        const uint8_t * restrict q4 = x[ibl].qs;
-        const uint8_t * restrict sc = x[ibl].scales_l;
-        const int8_t  * restrict q8 = y[ibl].qs;
+        const uint8_t * GGML_RESTRICT q4 = x[ibl].qs;
+        const uint8_t * GGML_RESTRICT sc = x[ibl].scales_l;
+        const int8_t  * GGML_RESTRICT q8 = y[ibl].qs;
 
         for (int ib = 0; ib < QK_K/64; ib ++ ) {
             __builtin_prefetch(q4, 0, 1);
@@ -11468,6 +12428,56 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     }
 
     *s = hsum_float_8(accum);
+#elif 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;
@@ -11506,12 +12516,12 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 
 // ============================ 4-bit non-linear quants
 
-void quantize_row_iq4_nl(const float * restrict x, void * restrict y, int64_t k) {
+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 * restrict x, void * restrict y, int64_t 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);
 }