cui-llama.rn 1.4.3 → 1.4.4

package/cpp/ggml-cpu-quants.c

@@ -297,6 +297,90 @@ static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
 static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
 #endif
 
+#if defined(__loongarch_sx)
+
+static __m128i lsx_packs_w(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_w(a, 15);
+    tmp1 = __lsx_vsat_w(b, 15);
+    return __lsx_vpickev_h(tmp1, tmp);
+}
+
+static __m128i lsx_packs_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_h(a, 7);
+    tmp1 = __lsx_vsat_h(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_packus_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_hu(a, 7);
+    tmp1 = __lsx_vsat_hu(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_h_b(a, b);
+    tmp2 = __lsx_vmulwod_h_b(a, b);
+    return __lsx_vsadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_madd_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_w_h(a, b);
+    tmp2 = __lsx_vmulwod_w_h(a, b);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
+    v4i32 __ret = {d, c, b, a};
+    return (__m128i)__ret;
+}
+
+static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
+    __m128i mask_f, zero, tmp0, tmp2, mask;
+    int f = 0x8f;
+    mask_f = __lsx_vreplgr2vr_b(f);
+    zero = __lsx_vldi(0);
+    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
+    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
+    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
+    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
+    return __lsx_vshuf_b(a, zero, tmp2);
+}
+
+static __m128i lsx_hadd_h(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_h(b, a);
+    __m128i tmp2 = __lsx_vpickod_h(b, a);
+    return __lsx_vadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_hadd_w(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_w(b, a);
+    __m128i tmp2 = __lsx_vpickod_w(b, a);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+static __m128 lsx_hadd_s(__m128 a, __m128 b) {
+    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
+    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
+
+    return __lsx_vfadd_s(tmp1, tmp2);
+}
+
+static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
+    __m128 res_0 =lsx_hadd_s(a, b);
+    __m128 res_1 =lsx_hadd_s(c, d);
+    __m128 res =lsx_hadd_s(res_0, res_1);
+    res =lsx_hadd_s(res, res);
+    res =lsx_hadd_s(res, res);
+
+    return ((v4f32)res)[0];
+}
+#endif
+
 #if defined(__loongarch_asx)
 
 #ifdef __clang__
@@ -395,11 +479,6 @@ static __m256i lasx_set_w(int e7, int e6, int e5, int e4, int e3, int e2, int e1
     return (__m256i)__ret;
 }
 
-static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
-    v4i32 __ret = {d, c, b, a};
-    return (__m128i)__ret;
-}
-
 static __m256i lasx_set_d(int64_t a, int64_t b, int64_t c, int64_t d) {
     v4i64 __ret = {d, c, b, a};
     return (__m256i)__ret;
@@ -409,18 +488,6 @@ static __m256i lasx_insertf128( __m128i x, __m128i y) {
     return lasx_set_q(x, y);
 }
 
-static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
-    __m128i mask_f, zero, tmp0, tmp2, mask;
-    int f = 0x8f;
-    mask_f = __lsx_vreplgr2vr_b(f);
-    zero = __lsx_vldi(0);
-    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
-    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
-    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
-    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
-    return __lsx_vshuf_b(a, zero, tmp2);
-}
-
 static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
     __m256i mask_f, zero, tmp0, tmp2, mask;
     int f = 0x8f;
@@ -434,30 +501,15 @@ static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
 }
 
 static __m256i lasx_extu8_16(__m128i a) {
-    __m128i zero = __lsx_vldi(0);
-    __m128i vlo = __lsx_vilvl_b(zero, a);
-    __m128i vhi = __lsx_vilvh_b(zero, a);
-    return lasx_set_q(vhi, vlo);
+    return __lasx_vext2xv_hu_bu(____m256i(a));
 }
 
 static __m256i lasx_ext8_16(__m128i a) {
-     __m128i sign = __lsx_vslti_b(a, 0);
-     __m128i vlo = __lsx_vilvl_b(sign, a);
-     __m128i vhi = __lsx_vilvh_b(sign, a);
-     return lasx_set_q(vhi, vlo);
+    return __lasx_vext2xv_h_b(____m256i(a));
 }
 
 static __m256i lasx_ext16_32(__m128i a) {
-    __m256i tmp1;
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 0), 0);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 1), 1);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 2), 2);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 3), 3);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 4), 4);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 5), 5);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 6), 6);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 7), 7);
-    return tmp1;
+    return __lasx_vext2xv_w_h(____m256i(a));
 }
 
 static __m128i lasx_extracti128( __m256i a, int pos) {
@@ -482,25 +534,6 @@ static __m128 lasx_extractf128( __m256 a, int pos) {
     return ret;
 }
 
-static __m128i lsx_hadd_h(__m128i a, __m128i b) {
-    __m128i tmp1 = __lsx_vpickev_h(b, a);
-    __m128i tmp2 = __lsx_vpickod_h(b, a);
-    return __lsx_vadd_h(tmp1, tmp2);
-}
-
-static __m128i lsx_hadd_w(__m128i a, __m128i b) {
-    __m128i tmp1 = __lsx_vpickev_w(b, a);
-    __m128i tmp2 = __lsx_vpickod_w(b, a);
-    return __lsx_vadd_w(tmp1, tmp2);
-}
-
-static __m128 lsx_hadd_s(__m128 a, __m128 b) {
-    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
-    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
-
-    return __lsx_vfadd_s(tmp1, tmp2);
-}
-
 static __m256i lasx_maddubs_h(__m256i a, __m256i b) {
     __m256i tmp1, tmp2;
     tmp1 = __lasx_xvmulwev_h_b(a, b);
@@ -529,40 +562,39 @@ static __m256i lasx_packs_h(__m256i a, __m256i b) {
     return __lasx_xvpickev_b(tmp1, tmp);
 }
 
-static __m128i lsx_packs_w(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_w(a, 15);
-    tmp1 = __lsx_vsat_w(b, 15);
-    return __lsx_vpickev_h(tmp1, tmp);
-}
-
-static __m128i lsx_packs_h(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_h(a, 7);
-    tmp1 = __lsx_vsat_h(b, 7);
-    return __lsx_vpickev_b(tmp1, tmp);
-}
-
-static __m128i lsx_packus_h(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_hu(a, 7);
-    tmp1 = __lsx_vsat_hu(b, 7);
-    return __lsx_vpickev_b(tmp1, tmp);
+static inline __m256i lasx_madd_h_b(__m256i a, __m256i b) {
+    __m256i tmp1, tmp2;
+    tmp1 = __lasx_xvmulwev_h_b(a, b);
+    tmp2 = __lasx_xvmulwod_h_b(a, b);
+    return __lasx_xvadd_h(tmp1, tmp2);
 }
 
-
-static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
-    __m128i tmp1, tmp2;
-    tmp1 = __lsx_vmulwev_h_b(a, b);
-    tmp2 = __lsx_vmulwod_h_b(a, b);
-    return __lsx_vsadd_h(tmp1, tmp2);
+static inline __m256i lasx_xvrepl128vei_h(__m256i a, const unsigned int b) {
+    switch (b) {
+        case 0: return __lasx_xvrepl128vei_h(a, 0);
+        case 1: return __lasx_xvrepl128vei_h(a, 1);
+        case 2: return __lasx_xvrepl128vei_h(a, 2);
+        case 3: return __lasx_xvrepl128vei_h(a, 3);
+        case 4: return __lasx_xvrepl128vei_h(a, 4);
+        case 5: return __lasx_xvrepl128vei_h(a, 5);
+        case 6: return __lasx_xvrepl128vei_h(a, 6);
+        case 7: return __lasx_xvrepl128vei_h(a, 7);
+        default: __builtin_unreachable();
+    }
 }
 
-static __m128i lsx_madd_h(__m128i a, __m128i b) {
-    __m128i tmp1, tmp2;
-    tmp1 = __lsx_vmulwev_w_h(a, b);
-    tmp2 = __lsx_vmulwod_w_h(a, b);
-    return __lsx_vadd_w(tmp1, tmp2);
+static inline __m256i lasx_xvandi_b_bit(__m256i a, const unsigned int b) {
+    switch (b) {
+        case 0: return __lasx_xvandi_b(a, 1 << 0);
+        case 1: return __lasx_xvandi_b(a, 1 << 1);
+        case 2: return __lasx_xvandi_b(a, 1 << 2);
+        case 3: return __lasx_xvandi_b(a, 1 << 3);
+        case 4: return __lasx_xvandi_b(a, 1 << 4);
+        case 5: return __lasx_xvandi_b(a, 1 << 5);
+        case 6: return __lasx_xvandi_b(a, 1 << 6);
+        case 7: return __lasx_xvandi_b(a, 1 << 7);
+        default: __builtin_unreachable();
+    }
 }
 
 // multiply int8_t, add results pairwise twice
@@ -580,12 +612,10 @@ static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
 // horizontally add 8 floats
 static inline float hsum_float_8(const __m256 x) {
     __m128 res = lasx_extractf128(x, 1);
-    ft_union tmp;
     res = __lsx_vfadd_s(res, lasx_extractf128(x, 0));
     res = __lsx_vfadd_s(res, (__m128)__lsx_vpickod_d((__m128i)res, (__m128i)res));
     res = __lsx_vfadd_s(res, (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w(res, 1), 0));
-    tmp.i = __lsx_vpickve2gr_w(res, 0);
-    return tmp.f;
+    return ((v4f32)res)[0];
 }
 
 // horizontally add 8 int32_t
@@ -661,13 +691,8 @@ static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy)
 
 // multiply int8_t, add results pairwise twice and return as float vector
 static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
-
-    // Get absolute values of x vectors
-    const __m256i ax = __lasx_xvsigncov_b(x, x);
-    // Sign the values of the y vectors
-    const __m256i sy = __lasx_xvsigncov_b(x, y);
-
-    return mul_sum_us8_pairs_float(ax, sy);
+    const __m256i dot = lasx_madd_h_b(x, y);
+    return sum_i16_pairs_float(dot);
 }
 
 static inline __m128i packNibbles( __m256i bytes ) {
@@ -747,7 +772,7 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
             y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
         }
     }
-#elif defined(__wasm_simd128__)
+#elif defined __wasm_simd128__
     for (int i = 0; i < nb; i++) {
         v128_t srcv [8];
         v128_t asrcv[8];
@@ -927,7 +952,6 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
 
 #elif defined(__loongarch_asx)
     for (int i = 0; i < nb; i++) {
-        ft_union fi;
         __m256 v0 = (__m256)__lasx_xvld( x , 0);
         __m256 v1 = (__m256)__lasx_xvld( x , 32);
         __m256 v2 = (__m256)__lasx_xvld( x , 64);
@@ -945,8 +969,7 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vinsgr2vr_w(tmp, __lsx_vpickve2gr_w( max4, 1 ), 0 ));
-        fi.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
-        const float max_scalar = fi.f;
+        const float max_scalar = ((v4f32)max4)[0];
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
@@ -1037,7 +1060,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
 
         y[i].s = LM_GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
     }
-#elif defined(__wasm_simd128__)
+#elif defined __wasm_simd128__
     for (int i = 0; i < nb; i++) {
         v128_t srcv [8];
         v128_t asrcv[8];
@@ -1251,7 +1274,6 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
 
 #elif defined(__loongarch_asx)
     for (int i = 0; i < nb; i++) {
-        ft_union ft;
         __m256 v0 = (__m256)__lasx_xvld( x , 0 );
         __m256 v1 = (__m256)__lasx_xvld( x , 32 );
         __m256 v2 = (__m256)__lasx_xvld( x , 64 );
@@ -1269,8 +1291,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
-        ft.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
-        const float max_scalar = ft.f;
+        const float max_scalar = ((v4f32)max4)[0];
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
@@ -1653,7 +1674,87 @@ 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) {
+#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
+
+    for (int i = 0; i < nb; i++) {
+        const float * x_block = x + i * QK_K;
+
+        v128_t min_vec = wasm_v128_load(x_block);
+        v128_t max_vec = min_vec;
+
+        for (int j = 4; j < QK_K; j += 4) {
+            v128_t x_vec = wasm_v128_load(x_block + j);
+            max_vec = wasm_f32x4_pmax(max_vec, x_vec);
+            min_vec = wasm_f32x4_pmin(min_vec, x_vec);
+        }
+        max_vec = wasm_f32x4_pmax(max_vec, wasm_i32x4_shuffle(max_vec, max_vec, 2, 3, 0, 1));
+        max_vec = wasm_f32x4_pmax(max_vec, wasm_i32x4_shuffle(max_vec, max_vec, 1, 0, 3, 2));
+        min_vec = wasm_f32x4_pmin(min_vec, wasm_i32x4_shuffle(min_vec, min_vec, 2, 3, 0, 1));
+        min_vec = wasm_f32x4_pmin(min_vec, wasm_i32x4_shuffle(min_vec, min_vec, 1, 0, 3, 2));
+        float max = wasm_f32x4_extract_lane(max_vec, 0);
+        float min = wasm_f32x4_extract_lane(min_vec, 0);
+        float amax = -min > max ? min : max;
+
+        if (amax == 0.0f) {
+            yc[i].d = 0.0f;
+            const v128_t zero = wasm_i8x16_splat(0);
+            for (int j = 0; j < QK_K; j += 16) {
+                wasm_v128_store(yc[i].qs + j, zero);
+            }
+            continue;
+        }
+
+        const float iscale = -127.0f / amax;
+        const v128_t scale_vec = wasm_f32x4_splat(iscale);
+
+        // Process 16 elements per iteration
+        for (int j = 0, jb = 0; j < QK_K; j += 16, jb++) {
+            // Load and quantize 16 floats
+            v128_t x0 = wasm_v128_load(x_block + j);
+            v128_t x1 = wasm_v128_load(x_block + j + 4);
+            v128_t x2 = wasm_v128_load(x_block + j + 8);
+            v128_t x3 = wasm_v128_load(x_block + j + 12);
+
+            v128_t q0 = wasm_f32x4_nearest(wasm_f32x4_mul(x0, scale_vec));
+            v128_t q1 = wasm_f32x4_nearest(wasm_f32x4_mul(x1, scale_vec));
+            v128_t q2 = wasm_f32x4_nearest(wasm_f32x4_mul(x2, scale_vec));
+            v128_t q3 = wasm_f32x4_nearest(wasm_f32x4_mul(x3, scale_vec));
+
+            // Convert to i32 with saturation
+            v128_t i0 = wasm_i32x4_trunc_sat_f32x4(q0);
+            v128_t i1 = wasm_i32x4_trunc_sat_f32x4(q1);
+            v128_t i2 = wasm_i32x4_trunc_sat_f32x4(q2);
+            v128_t i3 = wasm_i32x4_trunc_sat_f32x4(q3);
+
+            // Pack into 16 i8 values
+            v128_t i8 = wasm_i8x16_narrow_i16x8(
+                wasm_i16x8_narrow_i32x4(i0, i1),
+                wasm_i16x8_narrow_i32x4(i2, i3)
+            );
+            wasm_v128_store(yc[i].qs + j, i8);
+
+            // Calculate bsums using SIMD
+            v128_t sum16 = wasm_i16x8_add(
+                wasm_i16x8_extend_low_i8x16(i8),
+                wasm_i16x8_extend_high_i8x16(i8)
+            );
+            v128_t sum32 = wasm_i32x4_add(
+                wasm_i32x4_extend_low_i16x8(sum16),
+                wasm_i32x4_extend_high_i16x8(sum16)
+            );
+            sum32 = wasm_i32x4_add(sum32, wasm_i32x4_shuffle(sum32, sum32, 2, 3, 0, 1));
+            sum32 = wasm_i32x4_add(sum32, wasm_i32x4_shuffle(sum32, sum32, 1, 0, 3, 2));
+            yc[i].bsums[jb] = wasm_i32x4_extract_lane(sum32, 0);
+        }
+
+        yc[i].d = 1.0f / iscale;
+    }
+#else
     quantize_row_q8_K_ref(x, y, k);
+#endif
 }
 
 //===================================== Dot products =================================
@@ -2011,6 +2112,94 @@ void lm_ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined __wasm_simd128__
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    const v128_t m4b = wasm_i8x16_splat(0x0F);
+    const v128_t s8b = wasm_i8x16_splat(0x8);
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q4_0 * 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];
+
+        // Load and process x0
+        v128_t v0_0 = wasm_v128_load(x0->qs);
+        v128_t v0_0l = wasm_v128_and(v0_0, m4b);
+        v128_t v0_0h = wasm_u8x16_shr(v0_0, 4);
+        v128_t v0_0ls = wasm_i8x16_sub(v0_0l, s8b);
+        v128_t v0_0hs = wasm_i8x16_sub(v0_0h, s8b);
+
+        // Load y0 vectors
+        v128_t y0_l = wasm_v128_load(y0->qs);
+        v128_t y0_h = wasm_v128_load(y0->qs + 16);
+
+        // Extend to i16x8 and compute dot products
+        v128_t dx0l = wasm_i16x8_extend_low_i8x16(v0_0ls);
+        v128_t dx0h = wasm_i16x8_extend_high_i8x16(v0_0ls);
+        v128_t dx0hl = wasm_i16x8_extend_low_i8x16(v0_0hs);
+        v128_t dx0hh = wasm_i16x8_extend_high_i8x16(v0_0hs);
+
+        v128_t dy0ll = wasm_i16x8_extend_low_i8x16(y0_l);
+        v128_t dy0lh = wasm_i16x8_extend_high_i8x16(y0_l);
+        v128_t dy0hl = wasm_i16x8_extend_low_i8x16(y0_h);
+        v128_t dy0hh = wasm_i16x8_extend_high_i8x16(y0_h);
+
+        v128_t dp0 = wasm_i32x4_add(
+            wasm_i32x4_add(
+                wasm_i32x4_dot_i16x8(dx0l, dy0ll),
+                wasm_i32x4_dot_i16x8(dx0h, dy0lh)
+            ),
+            wasm_i32x4_add(
+                wasm_i32x4_dot_i16x8(dx0hl, dy0hl),
+                wasm_i32x4_dot_i16x8(dx0hh, dy0hh)
+            )
+        );
+
+        // Load and process x1
+        v128_t v0_1 = wasm_v128_load(x1->qs);
+        v128_t v0_1l = wasm_v128_and(v0_1, m4b);
+        v128_t v0_1h = wasm_u8x16_shr(v0_1, 4);
+        v128_t v0_1ls = wasm_i8x16_sub(v0_1l, s8b);
+        v128_t v0_1hs = wasm_i8x16_sub(v0_1h, s8b);
+
+        // Load y1 vectors
+        v128_t y1_l = wasm_v128_load(y1->qs);
+        v128_t y1_h = wasm_v128_load(y1->qs + 16);
+
+        // Extend to i16x8 and compute dot products
+        v128_t dx1l = wasm_i16x8_extend_low_i8x16(v0_1ls);
+        v128_t dx1h = wasm_i16x8_extend_high_i8x16(v0_1ls);
+        v128_t dx1hl = wasm_i16x8_extend_low_i8x16(v0_1hs);
+        v128_t dx1hh = wasm_i16x8_extend_high_i8x16(v0_1hs);
+
+        v128_t dy1ll = wasm_i16x8_extend_low_i8x16(y1_l);
+        v128_t dy1lh = wasm_i16x8_extend_high_i8x16(y1_l);
+        v128_t dy1hl = wasm_i16x8_extend_low_i8x16(y1_h);
+        v128_t dy1hh = wasm_i16x8_extend_high_i8x16(y1_h);
+
+        v128_t dp1 = wasm_i32x4_add(
+            wasm_i32x4_add(
+                wasm_i32x4_dot_i16x8(dx1l, dy1ll),
+                wasm_i32x4_dot_i16x8(dx1h, dy1lh)
+            ),
+            wasm_i32x4_add(
+                wasm_i32x4_dot_i16x8(dx1hl, dy1hl),
+                wasm_i32x4_dot_i16x8(dx1hh, dy1hh)
+            )
+        );
+
+        // Accumulate results with scaling
+        float scale0 = LM_GGML_FP16_TO_FP32(x0->d) * LM_GGML_FP16_TO_FP32(y0->d);
+        float scale1 = LM_GGML_FP16_TO_FP32(x1->d) * LM_GGML_FP16_TO_FP32(y1->d);
+
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0)));
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1)));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -2232,21 +2421,22 @@ void lm_ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void
     }
 
     sumf = hsum_float_8(acc);
+
 #elif defined(__loongarch_sx)
     // set constants
     const __m128i low_mask = __lsx_vreplgr2vr_b(0xF);
     const __m128i off = __lsx_vreplgr2vr_b(8);
 
     // Initialize accumulator with zeros
-    __m128 acc_0 = __lsx_vldi(0);
-    __m128 acc_1 = __lsx_vldi(0);
-    __m128 acc_2 = __lsx_vldi(0);
-    __m128 acc_3 = __lsx_vldi(0);
+    __m128 acc_0 = (__m128)__lsx_vldi(0);
+    __m128 acc_1 = (__m128)__lsx_vldi(0);
+    __m128 acc_2 = (__m128)__lsx_vldi(0);
+    __m128 acc_3 = (__m128)__lsx_vldi(0);
 
     for (; ib + 1 < nb; ib += 2) {
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = __lsx_vreplgr2vr_w( LM_GGML_FP16_TO_FP32(x[ib].d) * LM_GGML_FP16_TO_FP32(y[ib].d) );
+        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( LM_GGML_FP16_TO_FP32(x[ib].d) * LM_GGML_FP16_TO_FP32(y[ib].d) );
 
         const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
 
@@ -2264,7 +2454,7 @@ void lm_ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void
         //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = __lsx_vreplgr2vr_w( LM_GGML_FP16_TO_FP32(x[ib + 1].d) * LM_GGML_FP16_TO_FP32(y[ib + 1].d) );
+        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( LM_GGML_FP16_TO_FP32(x[ib + 1].d) * LM_GGML_FP16_TO_FP32(y[ib + 1].d) );
 
         const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
 
@@ -2696,10 +2886,10 @@ void lm_ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
-#elif defined(__wasm_simd128__)
+#elif defined __wasm_simd128__
     v128_t sumv = wasm_f32x4_splat(0.0f);
 
-    uint32_t qh;
+    uint32_t qh_;
     uint64_t tmp[4];
 
     // TODO: check if unrolling this is better
@@ -2710,12 +2900,12 @@ void lm_ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void
         const v128_t m4b  = wasm_i8x16_splat(0x0F);
 
         // extract the 5th bit
-        memcpy(&qh, x0->qh, sizeof(qh));
+        memcpy(&qh_, x0->qh, sizeof(qh_));
 
-        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
-        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
-        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
-        tmp[3] = table_b2b_1[(qh >> 24)       ];
+        tmp[0] = table_b2b_1[(qh_ >>  0) & 0xFF];
+        tmp[1] = table_b2b_1[(qh_ >>  8) & 0xFF];
+        tmp[2] = table_b2b_1[(qh_ >> 16) & 0xFF];
+        tmp[3] = table_b2b_1[(qh_ >> 24)       ];
 
         const v128_t qhl = wasm_v128_load(tmp + 0);
         const v128_t qhh = wasm_v128_load(tmp + 2);
@@ -3057,12 +3247,12 @@ void lm_ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
-#elif defined(__wasm_simd128__)
+#elif defined __wasm_simd128__
     v128_t sumv = wasm_f32x4_splat(0.0f);
 
     float summs = 0.0f;
 
-    uint32_t qh;
+    uint32_t qh_;
     uint64_t tmp[4];
 
     // TODO: check if unrolling this is better
@@ -3075,12 +3265,12 @@ void lm_ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void
         const v128_t m4b = wasm_i8x16_splat(0x0F);
 
         // extract the 5th bit
-        memcpy(&qh, x0->qh, sizeof(qh));
+        memcpy(&qh_, x0->qh, sizeof(qh_));
 
-        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
-        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
-        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
-        tmp[3] = table_b2b_0[(qh >> 24)       ];
+        tmp[0] = table_b2b_0[(qh_ >>  0) & 0xFF];
+        tmp[1] = table_b2b_0[(qh_ >>  8) & 0xFF];
+        tmp[2] = table_b2b_0[(qh_ >> 16) & 0xFF];
+        tmp[3] = table_b2b_0[(qh_ >> 24)       ];
 
         const v128_t qhl = wasm_v128_load(tmp + 0);
         const v128_t qhh = wasm_v128_load(tmp + 2);
@@ -3573,6 +3763,45 @@ void lm_ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined __wasm_simd128__
+    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 v128_t x0_0 = wasm_v128_load(x0->qs);
+        const v128_t x0_1 = wasm_v128_load(x0->qs + 16);
+        const v128_t y0_0 = wasm_v128_load(y0->qs);
+        const v128_t y0_1 = wasm_v128_load(y0->qs + 16);
+
+        // Extend 8-bit to 16-bit
+        const v128_t x0_0l = wasm_i16x8_extend_low_i8x16(x0_0);
+        const v128_t x0_0h = wasm_i16x8_extend_high_i8x16(x0_0);
+        const v128_t x0_1l = wasm_i16x8_extend_low_i8x16(x0_1);
+        const v128_t x0_1h = wasm_i16x8_extend_high_i8x16(x0_1);
+
+        const v128_t y0_0l = wasm_i16x8_extend_low_i8x16(y0_0);
+        const v128_t y0_0h = wasm_i16x8_extend_high_i8x16(y0_0);
+        const v128_t y0_1l = wasm_i16x8_extend_low_i8x16(y0_1);
+        const v128_t y0_1h = wasm_i16x8_extend_high_i8x16(y0_1);
+
+        // Compute dot products
+        const v128_t dx0_0 = wasm_i32x4_dot_i16x8(x0_0l, y0_0l);
+        const v128_t dx0_1 = wasm_i32x4_dot_i16x8(x0_0h, y0_0h);
+        const v128_t dx1_0 = wasm_i32x4_dot_i16x8(x0_1l, y0_1l);
+        const v128_t dx1_1 = wasm_i32x4_dot_i16x8(x0_1h, y0_1h);
+
+        // Sum all dot products
+        const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1));
+
+        // Convert to float and accumulate
+        const float scale = LM_GGML_FP16_TO_FP32(x0->d) * LM_GGML_FP16_TO_FP32(y0->d);
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale)));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -4447,6 +4676,106 @@ void lm_ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = hsum_float_8(acc);
 
+#elif defined __wasm_simd128__
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * q2 = x[i].qs;
+        const int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        // Vectorized summs calculation
+        v128_t summs_vec = wasm_i32x4_splat(0);
+        {
+            v128_t sc_vec = wasm_v128_load(sc);
+            v128_t sc_upper = wasm_u8x16_shr(sc_vec, 4);
+
+            v128_t sc_low = wasm_u16x8_extend_low_u8x16(sc_upper);
+            v128_t sc_high = wasm_u16x8_extend_high_u8x16(sc_upper);
+
+            v128_t bsums1 = wasm_v128_load(&y[i].bsums[0]);
+            v128_t bsums2 = wasm_v128_load(&y[i].bsums[8]);
+
+            summs_vec = wasm_i32x4_add(
+                wasm_i32x4_add(wasm_i32x4_dot_i16x8(sc_low, bsums1),
+                               wasm_i32x4_dot_i16x8(sc_high, bsums2)),
+                summs_vec
+            );
+
+            summs_vec = wasm_i32x4_add(summs_vec, wasm_i32x4_shuffle(summs_vec, summs_vec, 2, 3, 0, 1));
+            summs_vec = wasm_i32x4_add(summs_vec, wasm_i32x4_shuffle(summs_vec, summs_vec, 1, 0, 3, 2));
+        }
+        int32_t summs = wasm_i32x4_extract_lane(summs_vec, 0);
+
+        // Vectorized isum calculation
+        int32_t isum = 0;
+        const uint8_t * sc_ptr = sc;
+        const int k_iters = QK_K/128;
+
+        for (int k = 0; k < k_iters; ++k) {
+            v128_t isum_vec = wasm_i32x4_splat(0);
+            int shift = 0;
+
+            for (int j = 0; j < 4; ++j) {
+                const int d0 = (sc_ptr[0] & 0xF);
+                const int d1 = (sc_ptr[1] & 0xF);
+                sc_ptr += 2;
+
+                // Process first 16 elements
+                v128_t q2_0 = wasm_v128_load(q2);
+                v128_t q8_0 = wasm_v128_load(q8);
+                v128_t q2_shift_0 = wasm_u8x16_shr(q2_0, shift);
+                v128_t q2_bits_0 = wasm_v128_and(q2_shift_0, wasm_i8x16_splat(0x03));
+
+                // Process next 16 elements
+                v128_t q2_1 = wasm_v128_load(q2 + 16);
+                v128_t q8_1 = wasm_v128_load(q8 + 16);
+                v128_t q2_shift_1 = wasm_u8x16_shr(q2_1, shift);
+                v128_t q2_bits_1 = wasm_v128_and(q2_shift_1, wasm_i8x16_splat(0x03));
+
+                // Calculate dot products
+                v128_t p0 = wasm_i32x4_dot_i16x8(
+                    wasm_i16x8_extend_low_i8x16(q8_0),
+                    wasm_i16x8_extend_low_i8x16(q2_bits_0)
+                );
+                v128_t p1 = wasm_i32x4_dot_i16x8(
+                    wasm_i16x8_extend_high_i8x16(q8_0),
+                    wasm_i16x8_extend_high_i8x16(q2_bits_0)
+                );
+                v128_t p2 = wasm_i32x4_dot_i16x8(
+                    wasm_i16x8_extend_low_i8x16(q8_1),
+                    wasm_i16x8_extend_low_i8x16(q2_bits_1)
+                );
+                v128_t p3 = wasm_i32x4_dot_i16x8(
+                    wasm_i16x8_extend_high_i8x16(q8_1),
+                    wasm_i16x8_extend_high_i8x16(q2_bits_1)
+                );
+
+                // Accumulate scaled results
+                v128_t scaled = wasm_i32x4_add(
+                    wasm_i32x4_mul(wasm_i32x4_add(p0, p1), wasm_i32x4_splat(d0)),
+                    wasm_i32x4_mul(wasm_i32x4_add(p2, p3), wasm_i32x4_splat(d1))
+                );
+
+                isum_vec = wasm_i32x4_add(isum_vec, scaled);
+                q8 += 32;
+                shift += 2;
+            }
+            q2 += 32;
+
+            // Horizontal sum of isum_vec
+            isum_vec = wasm_i32x4_add(isum_vec, wasm_i32x4_shuffle(isum_vec, isum_vec, 2, 3, 0, 1));
+            isum_vec = wasm_i32x4_add(isum_vec, wasm_i32x4_shuffle(isum_vec, isum_vec, 1, 0, 3, 2));
+            isum += wasm_i32x4_extract_lane(isum_vec, 0);
+        }
+
+        const float dall = LM_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = LM_GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf += dall * isum - dmin * summs;
+    }
+
+    *s = sumf;
+
 #elif defined __riscv_v_intrinsic
 
     float sumf = 0;
@@ -4666,9 +4995,6 @@ void lm_ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void
 
 #elif defined __loongarch_asx
 
-    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
-    const __m128i m4 = __lsx_vreplgr2vr_b(0xF);
-
     __m256 acc = (__m256)__lasx_xvldi(0);
 
     for (int i = 0; i < nb; ++i) {
@@ -4679,18 +5005,15 @@ void lm_ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void
         const uint8_t * restrict q2 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0);
-        const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4);
-        const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4);
-        const __m256i mins = lasx_ext8_16(mins8);
+        const __m128i mins_and_scales128 = __lsx_vld((const __m128i*)x[i].scales, 0);
+        const __m128i scales128 = __lsx_vandi_b(mins_and_scales128, 0xf);
+        const __m256i mins = lasx_ext8_16(__lsx_vsrli_b(mins_and_scales128, 4));
         const __m256i prod = lasx_madd_h(mins, __lasx_xvld((const __m256i*)y[i].bsums, 0));
 
         acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(dmin), __lasx_xvffint_s_w(prod), acc);
 
-        const __m256i all_scales = lasx_ext8_16(scales8);
-        const __m128i l_scales = lasx_extracti128(all_scales, 0);
-        const __m128i h_scales = lasx_extracti128(all_scales, 1);
-        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
+        const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15};
+        const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask));
 
         __m256i sumi = __lasx_xvldi(0);
 
@@ -4703,20 +5026,20 @@ void lm_ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void
             const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
 
-            const __m256i q2_0 = __lasx_xvand_v(q2bits, m3);
-            const __m256i q2_1 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 2), m3);
-            const __m256i q2_2 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 4), m3);
-            const __m256i q2_3 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 6), m3);
+            const __m256i q2_0 = __lasx_xvandi_b(q2bits, 3);
+            const __m256i q2_1 = __lasx_xvandi_b(__lasx_xvsrli_b(q2bits, 2), 3);
+            const __m256i q2_2 = __lasx_xvandi_b(__lasx_xvsrli_b(q2bits, 4), 3);
+            const __m256i q2_3 = __lasx_xvsrli_b(q2bits, 6);
 
-            __m256i p0 = lasx_maddubs_h(q2_0, q8_0);
-            __m256i p1 = lasx_maddubs_h(q2_1, q8_1);
-            __m256i p2 = lasx_maddubs_h(q2_2, q8_2);
-            __m256i p3 = lasx_maddubs_h(q2_3, q8_3);
+            __m256i p0 = lasx_madd_h_b(q2_0, q8_0);
+            __m256i p1 = lasx_madd_h_b(q2_1, q8_1);
+            __m256i p2 = lasx_madd_h_b(q2_2, q8_2);
+            __m256i p3 = lasx_madd_h_b(q2_3, q8_3);
 
-            p0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(0)), p0);
-            p1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(1)), p1);
-            p2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(2)), p2);
-            p3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(3)), p3);
+            p0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p0);
+            p1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p1);
+            p2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p2);
+            p3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p3);
 
             p0 = __lasx_xvadd_w(p0, p1);
             p2 = __lasx_xvadd_w(p2, p3);
@@ -5129,6 +5452,94 @@ void lm_ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = hsum_float_8(acc);
 
+#elif defined __wasm_simd128__
+    int8_t  aux8[QK_K];
+    float   sums[8] = {0};
+    uint32_t auxs[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict hm = x[i].hmask;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // Process blocks with SIMD
+        int8_t * a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int shift = 0; shift <= 6; shift += 2) {
+                v128_t v_m = wasm_i8x16_splat(m);
+                for (int l = 0; l < 32; l += 16) {
+                    v128_t v_q3 = wasm_v128_load(q3 + l);
+                    v128_t v_shift = wasm_i8x16_shr(v_q3, shift);
+                    v128_t v_low2 = wasm_v128_and(v_shift, wasm_i8x16_splat(0x03));
+
+                    v128_t v_hm = wasm_v128_load(hm + l);
+                    v128_t v_mask = wasm_v128_and(v_hm, v_m);
+                    v_mask = wasm_i8x16_ne(v_mask, wasm_i8x16_splat(0));
+
+                    v_low2 = wasm_i8x16_sub(v_low2, wasm_v128_and(wasm_i8x16_splat(4), wasm_v128_not(v_mask)));
+                    wasm_v128_store(a + l, v_low2);
+                }
+                a += 32;
+                m <<= 1;
+            }
+            q3 += 32;
+        }
+
+        // Extract scales
+        memcpy(auxs, x[i].scales, 12);
+        uint32_t tmp = auxs[2];
+        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+        const int8_t * scales = (const int8_t *)auxs;
+
+        // SIMD dot product with register accumulators
+        v128_t v_acc0 = wasm_i32x4_splat(0);
+        v128_t v_acc1 = wasm_i32x4_splat(0);
+        a = aux8;
+        for (int j = 0; j < QK_K/16; ++j) {
+            const v128_t v_scale = wasm_i16x8_splat(scales[j] - 32);
+
+            // Process 16 elements per iteration
+            for (int k = 0; k < 2; ++k) {
+                const v128_t v_q8 = wasm_i16x8_load8x8(q8);
+                const v128_t v_a = wasm_i16x8_load8x8(a);
+
+                v128_t v_prod = wasm_i16x8_mul(v_q8, v_a);
+                v_prod = wasm_i16x8_mul(v_prod, v_scale);
+
+                v_acc0 = wasm_i32x4_add(v_acc0, wasm_i32x4_extend_low_i16x8(v_prod));
+                v_acc1 = wasm_i32x4_add(v_acc1, wasm_i32x4_extend_high_i16x8(v_prod));
+
+                q8 += 8;
+                a += 8;
+            }
+        }
+
+        // Accumulate results
+        const float d = LM_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const v128_t v_d = wasm_f32x4_splat(d);
+        v128_t v_sum = wasm_f32x4_add(
+            wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d),
+            wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc1), v_d)
+        );
+
+        // Accumulate into sums vector
+        wasm_v128_store(sums, wasm_f32x4_add(wasm_v128_load(sums), v_sum));
+    }
+
+    // Horizontal sum
+    v128_t v_sum = wasm_f32x4_add(wasm_v128_load(sums), wasm_v128_load(sums + 4));
+    sumf = wasm_f32x4_extract_lane(v_sum, 0) +
+           wasm_f32x4_extract_lane(v_sum, 1) +
+           wasm_f32x4_extract_lane(v_sum, 2) +
+           wasm_f32x4_extract_lane(v_sum, 3);
+
+    *s = sumf;
+
 #elif defined __riscv_v_intrinsic
 
     uint32_t aux[3];
@@ -5384,8 +5795,6 @@ void lm_ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void
 
 #elif defined __loongarch_asx
 
-    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
-    const __m256i mone = __lasx_xvreplgr2vr_b(1);
     const __m128i m32 = __lsx_vreplgr2vr_b(32);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
@@ -5405,10 +5814,9 @@ void lm_ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void
                 (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
                 (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
         scales128 = __lsx_vsub_b(scales128, m32);
-        const __m256i all_scales = lasx_ext8_16(scales128);
-        const __m128i l_scales = lasx_extracti128(all_scales, 0);
-        const __m128i h_scales = lasx_extracti128(all_scales, 1);
-        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
+
+        const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15};
+        const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask));
 
         // high bit
         const __m256i hbits = __lasx_xvld((const __m256i*)x[i].hmask, 0);
@@ -5416,35 +5824,23 @@ void lm_ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void
         // integer accumulator
         __m256i sumi = __lasx_xvldi(0);
 
-        int bit = 0;
-        int is  = 0;
-        __m256i xvbit;
-
-
         for (int j = 0; j < QK_K/128; ++j) {
             // load low 2 bits
             const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
 
-            xvbit = __lasx_xvreplgr2vr_h(bit);
             // prepare low and high bits
-            const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3);
-            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3);
-            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3);
-            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3);
-            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
+            const __m256i q3l_0 = __lasx_xvandi_b(q3bits, 3);
+            const __m256i q3l_1 = __lasx_xvandi_b(__lasx_xvsrli_b(q3bits, 2), 3);
+            const __m256i q3l_2 = __lasx_xvandi_b(__lasx_xvsrli_b(q3bits, 4), 3);
+            const __m256i q3l_3 = __lasx_xvsrli_b(q3bits, 6);
+            const __m256i q3h_0 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 0), 0), 2);
+            const __m256i q3h_1 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 1), 0), 2);
+            const __m256i q3h_2 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 2), 0), 2);
+            const __m256i q3h_3 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 3), 0), 2);
+            const __m256i q3_0 = __lasx_xvor_v(q3h_0, q3l_0);
+            const __m256i q3_1 = __lasx_xvor_v(q3h_1, q3l_1);
+            const __m256i q3_2 = __lasx_xvor_v(q3h_2, q3l_2);
+            const __m256i q3_3 = __lasx_xvor_v(q3h_3, q3l_3);
 
             // load Q8 quants
             const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
@@ -5452,29 +5848,16 @@ void lm_ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void
             const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
 
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use lasx_maddubs_h,
-            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-            // and 2 if the high bit was set)
-            __m256i q8s_0 = lasx_maddubs_h(q3h_0, q8_0);
-            __m256i q8s_1 = lasx_maddubs_h(q3h_1, q8_1);
-            __m256i q8s_2 = lasx_maddubs_h(q3h_2, q8_2);
-            __m256i q8s_3 = lasx_maddubs_h(q3h_3, q8_3);
-
-            __m256i p16_0 = lasx_maddubs_h(q3l_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q3l_1, q8_1);
-            __m256i p16_2 = lasx_maddubs_h(q3l_2, q8_2);
-            __m256i p16_3 = lasx_maddubs_h(q3l_3, q8_3);
-
-            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
-            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
-            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
-            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
+            __m256i p16_0 = lasx_madd_h_b(q8_0, q3_0);
+            __m256i p16_1 = lasx_madd_h_b(q8_1, q3_1);
+            __m256i p16_2 = lasx_madd_h_b(q8_2, q3_2);
+            __m256i p16_3 = lasx_madd_h_b(q8_3, q3_3);
 
             // multiply with scales
-            p16_0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
-            p16_1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
-            p16_2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
-            p16_3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
+            p16_0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p16_0);
+            p16_1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p16_1);
+            p16_2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p16_2);
+            p16_3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p16_3);
 
             // accumulate
             p16_0 = __lasx_xvadd_w(p16_0, p16_1);
@@ -5482,7 +5865,7 @@ void lm_ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void
             sumi  = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_2));
         }
         // multiply with block scale and accumulate
-        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);//FIXME
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
     }
 
     *s = hsum_float_8(acc);
@@ -5654,7 +6037,7 @@ void lm_ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void
         }
     }
     *s = sumf;
-#elif __ARM_NEON
+#elif defined __ARM_NEON
     const uint8x16_t m4b = vdupq_n_u8(0xf);
     const int32x4_t mzero = vdupq_n_s32(0);
 
@@ -5717,6 +6100,107 @@ void lm_ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = sumf;
 
+#elif defined __wasm_simd128__
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * LM_GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * LM_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;
+
+        // Process scales and mins
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        // Sum mins * q8sums
+        int32_t sumi = 0;
+        const int16_t * restrict q8sums = y[i].bsums;
+        const uint8_t * m = (const uint8_t *)&utmp[2];
+        for (int j = 0; j < 16; j += 2) {
+            sumi += (q8sums[j] + q8sums[j+1]) * m[j/2];
+        }
+        sumf -= dmin * sumi;
+
+        int32_t sumi1 = 0;
+        int32_t sumi2 = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // Load 64 4-bit weights (32 bytes)
+            const v128_t q4x0 = wasm_v128_load(q4);
+            const v128_t q4x1 = wasm_v128_load(q4 + 16);
+            q4 += 32;
+
+            // Split into low/high nibbles
+            const v128_t q4l0 = wasm_v128_and(q4x0, wasm_i8x16_splat(0x0F));
+            const v128_t q4h0 = wasm_u8x16_shr(q4x0, 4);
+            const v128_t q4l1 = wasm_v128_and(q4x1, wasm_i8x16_splat(0x0F));
+            const v128_t q4h1 = wasm_u8x16_shr(q4x1, 4);
+
+            // Load 64 8-bit values (64 bytes)
+            const v128_t q8x0 = wasm_v128_load(q8);
+            const v128_t q8x1 = wasm_v128_load(q8 + 16);
+            const v128_t q8x2 = wasm_v128_load(q8 + 32);
+            const v128_t q8x3 = wasm_v128_load(q8 + 48);
+            q8 += 64;
+
+            // Low nibble products
+            v128_t vacc1 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q4l0),
+                wasm_i16x8_extend_low_i8x16(q8x0)
+            );
+            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q4l0),
+                wasm_i16x8_extend_high_i8x16(q8x0)
+            ));
+            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q4l1),
+                wasm_i16x8_extend_low_i8x16(q8x1)
+            ));
+            vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q4l1),
+                wasm_i16x8_extend_high_i8x16(q8x1)
+            ));
+
+            // High nibble products
+            v128_t vacc2 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q4h0),
+                wasm_i16x8_extend_low_i8x16(q8x2)
+            );
+            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q4h0),
+                wasm_i16x8_extend_high_i8x16(q8x2)
+            ));
+            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q4h1),
+                wasm_i16x8_extend_low_i8x16(q8x3)
+            ));
+            vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q4h1),
+                wasm_i16x8_extend_high_i8x16(q8x3)
+            ));
+
+            // Accumulate scaled results
+            int32_t vacc1_sum = wasm_i32x4_extract_lane(vacc1, 0) + wasm_i32x4_extract_lane(vacc1, 1) +
+                                wasm_i32x4_extract_lane(vacc1, 2) + wasm_i32x4_extract_lane(vacc1, 3);
+            sumi1 += vacc1_sum * scales[2*j];
+
+            int32_t vacc2_sum = wasm_i32x4_extract_lane(vacc2, 0) + wasm_i32x4_extract_lane(vacc2, 1) +
+                                wasm_i32x4_extract_lane(vacc2, 2) + wasm_i32x4_extract_lane(vacc2, 3);
+            sumi2 += vacc2_sum * scales[2*j+1];
+        }
+
+        sumf += d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
+
 #elif defined __AVX2__
 
     const __m256i m4 = _mm256_set1_epi8(0xF);
@@ -6074,11 +6558,6 @@ void lm_ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void
     *s = vec_extract(vsumf0, 0);
 
 #elif defined __loongarch_asx
-    LM_GGML_UNUSED(kmask1);
-    LM_GGML_UNUSED(kmask2);
-    LM_GGML_UNUSED(kmask3);
-
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
     __m128 acc_m = (__m128)__lsx_vldi(0);
@@ -6098,33 +6577,34 @@ void lm_ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+        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);
+        const __m128i scales128 = __lsx_vsllwil_h_b(mins_and_scales128, 0);
 
         const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
         const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
-        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
+        const __m128i prod = lsx_madd_h(mins128, q8s);
         acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
 
-        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
-        const __m256i scales = lasx_insertf128(sc128, sc128);
+        const __m256i scales = lasx_insertf128(scales128, scales128);
 
         __m256i sumi = __lasx_xvldi(0);
 
         for (int j = 0; j < QK_K/64; ++j) {
 
-            const __m256i scale_l = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_h = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+            const __m256i scale_l = lasx_xvrepl128vei_h(scales, 2 * j + 0);
+            const __m256i scale_h = lasx_xvrepl128vei_h(scales, 2 * j + 1);
 
             const __m256i q4bits = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
-            const __m256i q4l = __lasx_xvand_v(q4bits, m4);
-            const __m256i q4h = __lasx_xvand_v(__lasx_xvsrli_h(q4bits, 4), m4);
+            const __m256i q4l = __lasx_xvandi_b(q4bits, 0xf);
+            const __m256i q4h = __lasx_xvsrli_b(q4bits, 4);
 
             const __m256i q8l = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            __m256i p16l = lasx_maddubs_h(q4l, q8l);
+            __m256i p16l = lasx_madd_h_b(q4l, q8l);
             p16l = lasx_madd_h(scale_l, p16l);
 
             const __m256i q8h = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            __m256i p16h = lasx_maddubs_h(q4h, q8h);
+            __m256i p16h = lasx_madd_h_b(q4h, q8h);
             p16h = lasx_madd_h(scale_h, p16h);
             const __m256i sumj = __lasx_xvadd_w(p16l, p16h);
 
@@ -6141,9 +6621,7 @@ void lm_ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void
     acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1);
 
 
-    ft_union fi;
-    fi.i = __lsx_vpickve2gr_w(acc_m, 0);
-    *s = hsum_float_8(acc) + fi.f ;
+    *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
 #else
 
     const uint8_t * scales = (const uint8_t*)&utmp[0];
@@ -6469,6 +6947,118 @@ void lm_ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = hsum_float_8(acc) + summs;
 
+#elif defined __wasm_simd128__
+    //const uint8_t * scales = (const uint8_t*)&utmp[0];
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * LM_GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * LM_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;
+
+        // Process scales and mins
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        // Sum mins * q8sums
+        int32_t sumi_mins = 0;
+        const int16_t * restrict q8sums = y[i].bsums;
+        const uint8_t * m = (const uint8_t *)&utmp[2];
+        for (int j = 0; j < 16; j += 2) {
+            sumi_mins += (q8sums[j] + q8sums[j+1]) * m[j/2];
+        }
+        sumf -= dmin * sumi_mins; // Correct subtraction
+
+        v128_t qh0 = wasm_v128_load(qh);
+        v128_t qh1 = wasm_v128_load(qh + 16);
+        const uint8_t * sc = (const uint8_t *)utmp;
+
+        int32_t sumi = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            const int shift = j * 2;
+            v128_t qh_shift0 = wasm_u8x16_shr(qh0, shift);
+            v128_t qh_shift1 = wasm_u8x16_shr(qh1, shift);
+
+            v128_t qh_low0 = wasm_i8x16_shl(wasm_v128_and(qh_shift0, wasm_i8x16_splat(0x01)), 4);
+            v128_t qh_high0 = wasm_i8x16_shl(wasm_v128_and(qh_shift0, wasm_i8x16_splat(0x02)), 3);
+            v128_t qh_low1 = wasm_i8x16_shl(wasm_v128_and(qh_shift1, wasm_i8x16_splat(0x01)), 4);
+            v128_t qh_high1 = wasm_i8x16_shl(wasm_v128_and(qh_shift1, wasm_i8x16_splat(0x02)), 3);
+
+            v128_t q5_0 = wasm_v128_load(q5);
+            v128_t q5_1 = wasm_v128_load(q5 + 16);
+            q5 += 32;
+
+            v128_t q5l_0 = wasm_v128_or(wasm_v128_and(q5_0, wasm_i8x16_splat(0x0F)), qh_low0);
+            v128_t q5h_0 = wasm_v128_or(wasm_u8x16_shr(q5_0, 4), qh_high0);
+            v128_t q5l_1 = wasm_v128_or(wasm_v128_and(q5_1, wasm_i8x16_splat(0x0F)), qh_low1);
+            v128_t q5h_1 = wasm_v128_or(wasm_u8x16_shr(q5_1, 4), qh_high1);
+
+            v128_t q8_0 = wasm_v128_load(q8);
+            v128_t q8_1 = wasm_v128_load(q8 + 16);
+            v128_t q8_2 = wasm_v128_load(q8 + 32);
+            v128_t q8_3 = wasm_v128_load(q8 + 48);
+            q8 += 64;
+
+            // Process low quants
+            v128_t pl0 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q5l_0),
+                wasm_i16x8_extend_low_i8x16(q8_0)
+            );
+            pl0 = wasm_i32x4_add(pl0, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q5l_0),
+                wasm_i16x8_extend_high_i8x16(q8_0)
+            ));
+            v128_t pl1 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q5l_1),
+                wasm_i16x8_extend_low_i8x16(q8_1)
+            );
+            pl1 = wasm_i32x4_add(pl1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q5l_1),
+                wasm_i16x8_extend_high_i8x16(q8_1)
+            ));
+            v128_t sum_low = wasm_i32x4_add(pl0, pl1);
+
+            // Process high quants
+            v128_t ph0 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q5h_0),
+                wasm_i16x8_extend_low_i8x16(q8_2)
+            );
+            ph0 = wasm_i32x4_add(ph0, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q5h_0),
+                wasm_i16x8_extend_high_i8x16(q8_2)
+            ));
+            v128_t ph1 = wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_low_i8x16(q5h_1),
+                wasm_i16x8_extend_low_i8x16(q8_3)
+            );
+            ph1 = wasm_i32x4_add(ph1, wasm_i32x4_dot_i16x8(
+                wasm_i16x8_extend_high_i8x16(q5h_1),
+                wasm_i16x8_extend_high_i8x16(q8_3)
+            ));
+            v128_t sum_high = wasm_i32x4_add(ph0, ph1);
+
+            // Accumulate with scale factors
+            int32_t sl = wasm_i32x4_extract_lane(sum_low, 0) + wasm_i32x4_extract_lane(sum_low, 1) +
+                        wasm_i32x4_extract_lane(sum_low, 2) + wasm_i32x4_extract_lane(sum_low, 3);
+            int32_t sh = wasm_i32x4_extract_lane(sum_high, 0) + wasm_i32x4_extract_lane(sum_high, 1) +
+                        wasm_i32x4_extract_lane(sum_high, 2) + wasm_i32x4_extract_lane(sum_high, 3);
+
+            sumi += sl * sc[2*j] + sh * sc[2*j+1];
+        }
+
+        sumf += d * sumi;
+    }
+
+    *s = sumf;
+
 #elif defined __riscv_v_intrinsic
 
     const uint8_t * scales = (const uint8_t*)&utmp[0];
@@ -6691,19 +7281,11 @@ void lm_ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void
     *s = vec_extract(vsumf0, 0);
 
 #elif defined __loongarch_asx
-    LM_GGML_UNUSED(kmask1);
-    LM_GGML_UNUSED(kmask2);
-    LM_GGML_UNUSED(kmask3);
-
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
-    const __m128i mzero = __lsx_vldi(0);
-    const __m256i mone  = __lasx_xvreplgr2vr_b(1);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
+    __m128 acc_m = (__m128)__lsx_vldi(0);
 
-    float summs = 0.f;
-
-   for (int i = 0; i < nb; ++i) {
+    for (int i = 0; i < nb; ++i) {
 
         const uint8_t * restrict q5 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -6718,49 +7300,40 @@ void lm_ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void
         utmp[2] = uaux;
         utmp[0] &= kmask1;
 
-        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+        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);
+        const __m128i scales128 = __lsx_vsllwil_h_b(mins_and_scales128, 0);
 
         const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
         const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
-        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
-        const __m128i hsum = lsx_hadd_w(lsx_hadd_w(prod, mzero), mzero);
-        summs += dmin * __lsx_vpickve2gr_w(hsum, 0);    //TODO check
+        const __m128i prod = lsx_madd_h(mins128, q8s);
+        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
 
-        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
-        const __m256i scales = lasx_insertf128(sc128, sc128);
+        const __m256i scales = lasx_insertf128(scales128, scales128);
 
         const __m256i hbits = __lasx_xvld((const __m256i*)x[i].qh, 0);
-        __m256i hmask = mone;
 
         __m256i sumi = __lasx_xvldi(0);
 
-        int bit = 0;
-        __m256i xvbit;
-
         for (int j = 0; j < QK_K/64; ++j) {
 
-            const __m256i scale_0 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_1 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+            const __m256i scale_0 = lasx_xvrepl128vei_h(scales, 2 * j + 0);
+            const __m256i scale_1 = lasx_xvrepl128vei_h(scales, 2 * j + 1);
 
             const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32;
 
-            xvbit = __lasx_xvreplgr2vr_h(bit++);
-            const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4);
-            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
-            const __m256i q5_0  = __lasx_xvadd_b(q5l_0, q5h_0);
-            hmask = __lasx_xvslli_h(hmask, 1);
-
-            xvbit = __lasx_xvreplgr2vr_h(bit++);
-            const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4);
-            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
-            const __m256i q5_1  = __lasx_xvadd_b(q5l_1, q5h_1);
-            hmask = __lasx_xvslli_h(hmask, 1);
+            const __m256i q5l_0 = __lasx_xvandi_b(q5bits, 0xf);
+            const __m256i q5l_1 = __lasx_xvsrli_b(q5bits, 4);
+            const __m256i q5h_0 = __lasx_xvnori_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 2 * j + 0), 0), 0xef);
+            const __m256i q5h_1 = __lasx_xvnori_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 2 * j + 1), 0), 0xef);
+            const __m256i q5_0  = __lasx_xvor_v(q5l_0, q5h_0);
+            const __m256i q5_1  = __lasx_xvor_v(q5l_1, q5h_1);
 
             const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
 
-            __m256i p16_0 = lasx_maddubs_h(q5_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q5_1, q8_1);
+            __m256i p16_0 = lasx_madd_h_b(q5_0, q8_0);
+            __m256i p16_1 = lasx_madd_h_b(q5_1, q8_1);
 
             p16_0 = lasx_madd_h(scale_0, p16_0);
             p16_1 = lasx_madd_h(scale_1, p16_1);
@@ -6774,7 +7347,10 @@ void lm_ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void
 
     }
 
-    *s = hsum_float_8(acc) + summs;
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vbsrl_v(acc_m, 8));
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vbsrl_v(acc_m, 4));
+
+    *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
 
 #else
 
@@ -7132,6 +7708,85 @@ void lm_ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = hsum_float_8(acc);
 
+#elif defined __wasm_simd128__
+    int8_t aux8[QK_K] __attribute__((aligned(16)));
+    int32_t aux32[8] __attribute__((aligned(16))) = {0};
+    float sums[8] __attribute__((aligned(16))) = {0};
+
+    for (int i = 0; i < nb; ++i) {
+        // Unpack 6-bit quantized data into aux8 (unchanged)
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        int8_t * a = aux8;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            }
+            a += 128;
+            q4 += 64;
+            qh += 32;
+        }
+
+        const int8_t * restrict a_ptr = aux8;
+        const int8_t * restrict q8 = y[i].qs;
+        v128_t acc0 = wasm_i32x4_splat(0);
+        v128_t acc1 = wasm_i32x4_splat(0);
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            const int scale = x[i].scales[j];
+            const v128_t vscale = wasm_i32x4_splat(scale);
+
+            // Load 16 elements from a and q8
+            const v128_t a_vec = wasm_v128_load(a_ptr);
+            const v128_t q8_vec = wasm_v128_load(q8);
+
+            // Process low 8 elements
+            v128_t a_low = wasm_i16x8_extend_low_i8x16(a_vec);
+            v128_t q8_low = wasm_i16x8_extend_low_i8x16(q8_vec);
+            v128_t prod_low = wasm_i16x8_mul(a_low, q8_low);
+            v128_t prod_lo_lo = wasm_i32x4_extend_low_i16x8(prod_low);
+            v128_t prod_lo_hi = wasm_i32x4_extend_high_i16x8(prod_low);
+
+            // Process high 8 elements
+            v128_t a_high = wasm_i16x8_extend_high_i8x16(a_vec);
+            v128_t q8_high = wasm_i16x8_extend_high_i8x16(q8_vec);
+            v128_t prod_high = wasm_i16x8_mul(a_high, q8_high);
+            v128_t prod_hi_lo = wasm_i32x4_extend_low_i16x8(prod_high);
+            v128_t prod_hi_hi = wasm_i32x4_extend_high_i16x8(prod_high);
+
+            // Scale and accumulate
+            prod_lo_lo = wasm_i32x4_mul(prod_lo_lo, vscale);
+            prod_lo_hi = wasm_i32x4_mul(prod_lo_hi, vscale);
+            prod_hi_lo = wasm_i32x4_mul(prod_hi_lo, vscale);
+            prod_hi_hi = wasm_i32x4_mul(prod_hi_hi, vscale);
+
+            acc0 = wasm_i32x4_add(acc0, wasm_i32x4_add(prod_lo_lo, prod_hi_lo));
+            acc1 = wasm_i32x4_add(acc1, wasm_i32x4_add(prod_lo_hi, prod_hi_hi));
+
+            a_ptr += 16;
+            q8 += 16;
+        }
+
+        // Store accumulated results
+        wasm_v128_store(&aux32[0], acc0);
+        wasm_v128_store(&aux32[4], acc1);
+
+        const float d = LM_GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) {
+            sums[l] += d * aux32[l];
+        }
+    }
+
+    // Sum final results
+    float sumf = 0;
+    for (int l = 0; l < 8; ++l) {
+        sumf += sums[l];
+    }
+    *s = sumf;
+
 #elif defined __riscv_v_intrinsic
 
     float sumf = 0;
@@ -7356,8 +8011,6 @@ void lm_ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void
 
 #elif defined __loongarch_asx
 
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
-    const __m256i m2 = __lasx_xvreplgr2vr_b(3);
     const __m256i m32s = __lasx_xvreplgr2vr_b(32);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
@@ -7370,58 +8023,42 @@ void lm_ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void
         const uint8_t * restrict qh = x[i].qh;
         const int8_t  * restrict q8 = y[i].qs;
 
-        const __m128i scales = __lsx_vld((const __m128i*)x[i].scales, 0);
+        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};
+        const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask));
 
         __m256i sumi = __lasx_xvldi(0);
 
-        int is = 0;
-
         for (int j = 0; j < QK_K/128; ++j) {
 
-            const __m128i scale_0 = lsx_shuffle_b(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = lsx_shuffle_b(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = lsx_shuffle_b(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = lsx_shuffle_b(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
             const __m256i q4bits1 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
             const __m256i q4bits2 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
             const __m256i q4bitsH = __lasx_xvld((const __m256i*)qh, 0); qh += 32;
 
-            const __m256i q4h_0 = __lasx_xvslli_h(__lasx_xvand_v(q4bitsH, m2), 4);
-            const __m256i q4h_1 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 2), m2), 4);
-            const __m256i q4h_2 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 4), m2), 4);
-            const __m256i q4h_3 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 6), m2), 4);
+            const __m256i q4h_0 = __lasx_xvslli_b(__lasx_xvandi_b(q4bitsH, 3), 4);
+            const __m256i q4h_1 = __lasx_xvslli_b(__lasx_xvandi_b(q4bitsH, 3 << 2), 2);
+            const __m256i q4h_2 = __lasx_xvandi_b(q4bitsH, 3 << 4);
+            const __m256i q4h_3 = __lasx_xvsrli_b(__lasx_xvandi_b(q4bitsH, 3 << 6), 2);
 
-            const __m256i q4_0 = __lasx_xvor_v(__lasx_xvand_v(q4bits1, m4), q4h_0);
-            const __m256i q4_1 = __lasx_xvor_v(__lasx_xvand_v(q4bits2, m4), q4h_1);
-            const __m256i q4_2 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits1, 4), m4), q4h_2);
-            const __m256i q4_3 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits2, 4), m4), q4h_3);
+            const __m256i q4_0 = __lasx_xvor_v(__lasx_xvandi_b(q4bits1, 0xf), q4h_0);
+            const __m256i q4_1 = __lasx_xvor_v(__lasx_xvandi_b(q4bits2, 0xf), q4h_1);
+            const __m256i q4_2 = __lasx_xvor_v(__lasx_xvsrli_b(q4bits1, 4), q4h_2);
+            const __m256i q4_3 = __lasx_xvor_v(__lasx_xvsrli_b(q4bits2, 4), q4h_3);
 
             const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
 
-            __m256i q8s_0 = lasx_maddubs_h(m32s, q8_0);
-            __m256i q8s_1 = lasx_maddubs_h(m32s, q8_1);
-            __m256i q8s_2 = lasx_maddubs_h(m32s, q8_2);
-            __m256i q8s_3 = lasx_maddubs_h(m32s, q8_3);
+            __m256i p16_0 = lasx_madd_h_b(__lasx_xvsub_b(q4_0, m32s), q8_0);
+            __m256i p16_1 = lasx_madd_h_b(__lasx_xvsub_b(q4_1, m32s), q8_1);
+            __m256i p16_2 = lasx_madd_h_b(__lasx_xvsub_b(q4_2, m32s), q8_2);
+            __m256i p16_3 = lasx_madd_h_b(__lasx_xvsub_b(q4_3, m32s), q8_3);
 
-            __m256i p16_0 = lasx_maddubs_h(q4_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q4_1, q8_1);
-            __m256i p16_2 = lasx_maddubs_h(q4_2, q8_2);
-            __m256i p16_3 = lasx_maddubs_h(q4_3, q8_3);
-
-            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
-            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
-            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
-            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
-
-            p16_0 = lasx_madd_h(lasx_ext8_16(scale_0), p16_0);
-            p16_1 = lasx_madd_h(lasx_ext8_16(scale_1), p16_1);
-            p16_2 = lasx_madd_h(lasx_ext8_16(scale_2), p16_2);
-            p16_3 = lasx_madd_h(lasx_ext8_16(scale_3), p16_3);
+            p16_0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p16_0);
+            p16_1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p16_1);
+            p16_2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p16_2);
+            p16_3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p16_3);
 
             sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
             sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_2, p16_3));
@@ -9746,13 +10383,9 @@ static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
 }
 #elif defined(__loongarch_asx)
 static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
-    const __m256i ax = __lasx_xvsigncov_b(x, x);
-    const __m256i sy = __lasx_xvsigncov_b(x, y);
-    __m256i tmp1, tmp2, tmp3;
-    tmp1 = __lasx_xvmulwev_h_bu_b(ax, sy);
-    tmp2 = __lasx_xvmulwod_h_bu_b(ax, sy);
-    tmp3 = __lasx_xvadd_h(tmp1, tmp2);
-    return __lasx_xvsat_h(tmp3, 15);
+    const __m256i a = __lasx_xvmulwev_h_b(x, y);
+    const __m256i b = __lasx_xvmulwod_h_b(x, y);
+    return __lasx_xvadd_h(a, b);
 }
 #endif
 
@@ -10802,67 +11435,31 @@ void lm_ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const voi
 #elif defined(__loongarch_asx)
 
     const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
-    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
 
     __m256 accum = (__m256)__lasx_xvldi(0);
-    __m256i tmp1;
-    __m128i tmp0, tmp2, tmp3, tmp4, mask_8f, mask;
 
-    mask_8f = __lsx_vreplgr2vr_b(0x8f);
     for (int ibl = 0; ibl < nb; ++ibl) {
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;
         uint16_t sh = x[ibl].scales_h;
         __m256i sumi1 = __lasx_xvldi(0);
         __m256i sumi2 = __lasx_xvldi(0);
-        __m128i zero = __lsx_vldi(0);
         for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
-            const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
+            const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0); qs += 16;
+            const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0); qs += 16;
             const __m256i q8b_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
             const __m256i q8b_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp3 = __lsx_vand_v(tmp0, mask);
-            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_1, m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp4 = __lsx_vand_v(tmp0, mask);
-            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
-
-            const __m256i q4b_1 = lasx_insertf128(tmp3, tmp4);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp3 = __lsx_vand_v(tmp0, mask);
-            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_2, m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp4 = __lsx_vand_v(tmp0, mask);
-            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
-
-            const __m256i q4b_2 = lasx_insertf128(tmp3, tmp4);
-
+            const __m256i q4b_1 = lasx_insertf128(__lsx_vshuf_b(values128, values128, __lsx_vsrli_b(q4bits_1, 4)),
+                                                  __lsx_vshuf_b(values128, values128, __lsx_vandi_b(q4bits_1, 0xf)));
+            const __m256i q4b_2 = lasx_insertf128(__lsx_vshuf_b(values128, values128, __lsx_vsrli_b(q4bits_2, 4)),
+                                                  __lsx_vshuf_b(values128, values128, __lsx_vandi_b(q4bits_2, 0xf)));
             const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
             const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
             const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
             const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
             sh >>= 4;
-            __m256i tmp5, tmp6;
-            tmp1 = __lasx_xvreplgr2vr_h(ls1);
-            tmp5 = __lasx_xvmulwev_w_h(p16_1, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(p16_1, tmp1);
-            const __m256i p_1 = __lasx_xvadd_w(tmp5, tmp6);
-            tmp1 = __lasx_xvreplgr2vr_h(ls2);
-            tmp5 = __lasx_xvmulwev_w_h(p16_2, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(p16_2, tmp1);
-            const __m256i p_2 = __lasx_xvadd_w(tmp5, tmp6);
+            const __m256i p_1 = lasx_madd_h(p16_1, __lasx_xvreplgr2vr_h(ls1));
+            const __m256i p_2 = lasx_madd_h(p16_2, __lasx_xvreplgr2vr_h(ls2));
             sumi1 = __lasx_xvadd_w(p_1, sumi1);
             sumi2 = __lasx_xvadd_w(p_2, sumi2);
         }