whispercpp 1.3.3 → 1.3.5

data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -24,6 +24,31 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(__aarch64__) && defined(__ARM_NEON) && (defined(__ARM_FEATURE_MATMUL_INT8) || defined(__ARM_FEATURE_DOTPROD))
+static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
+                                             int16x8_t *     out_mins,
+                                             int8_t *        out_scales) {
+    constexpr uint32_t kmask1 = 0x3f3f3f3f;
+    constexpr uint32_t kmask2 = 0x0f0f0f0f;
+    constexpr uint32_t kmask3 = 0x03030303;
+    constexpr uint8_t  scales_size = 12;
+
+    uint32_t sm[3];
+    memcpy(sm, scales_in, scales_size);
+
+    const uint32_t   mins_0_3 = sm[1] & kmask1;
+    const uint32_t   mins_4_7 = ((sm[2] >> 4) & kmask2) | (((sm[1] >> 6) & kmask3) << 4);
+    const uint32x2_t mins_u32 = { mins_0_3, mins_4_7 };
+
+    *out_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins_u32)));
+
+    uint32_t scales_u32[2];
+    scales_u32[0] = sm[0] & kmask1;
+    scales_u32[1] = (sm[2] & kmask2) | (((sm[0] >> 6) & kmask3) << 4);
+    memcpy(out_scales, scales_u32, 8);
+}
+#endif
+
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -86,35 +111,9 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR
         }
     }
 #else
-    // scalar
-    const int blck_size_interleave = 4;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x4_generic(x, vy, k);
 #endif
 }
 
@@ -205,35 +204,9 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
     }
 
 #else
-    // scalar
-    const int blck_size_interleave = 8;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
-            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
-            src_offset += (j % blck_size_interleave);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
 #endif
 }
 
@@ -295,29 +268,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    float sumf[4];
-    int sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
+    ggml_gemv_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -383,29 +334,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    float sumf[4];
-    int sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
+    ggml_gemv_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -497,31 +426,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif // #if defined(__ARM_FEATURE_SVE)
 
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    {
-        float sumf[8];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -591,31 +496,421 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4];
-        int sumi;
+    ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
 
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
 
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_groups = ncols_interleaved / 4; // 0123 and 4567
+    const uint8x16_t m4b        = vdupq_n_u8(0x0f);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[col_groups];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < col_groups; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q4_d_0        = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q4_d_1        = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d          = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0123 = vmulq_f32(q4_d_0, q8_d);
+            float32x4_t sb_scale_4567 = vmulq_f32(q4_d_1, q8_d);
+            float32x4_t q4_dmin_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q4_dmin_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0123   = vmulq_f32(q4_dmin_0, q8_d);
+            float32x4_t sb_min_4567   = vmulq_f32(q4_dmin_1, q8_d);
+
+            // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567
+            int32x4_t bias_acc[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
+            int32x4_t acc_lo[col_groups];
+            int32x4_t acc_hi[col_groups];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_groups; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q4sb_mins[2];
+                int16x8_t q4sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q4sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                }
+
+                int8x16_t q8_qs[64 / 16];
+                for (int i = 0; i < 64 / 16; i++) {
+                    q8_qs[i] = vld1q_s8(q8_ptr[b].qs + sb * 64 + i * 16);
+                }
+
+                for (int c = 0; c < col_groups; c++) {
+                    uint8x16_t q4_cols[8];
+                    for (int i = 0; i < 8; i++) {
+                        q4_cols[i] = vld1q_u8(q4_ptr[b].qs + sb * QK_K + i * 32 + 16 * c);
                     }
+
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[0], m4b)), q8_qs[0], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[1], m4b)), q8_qs[0], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[2], m4b)), q8_qs[0], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[3], m4b)), q8_qs[0], 3);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[4], m4b)), q8_qs[1], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[5], m4b)), q8_qs[1], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[6], m4b)), q8_qs[1], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], vreinterpretq_s8_u8(vandq_u8(q4_cols[7], m4b)), q8_qs[1], 3);
+
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[0], 4)), q8_qs[2], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[1], 4)), q8_qs[2], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[2], 4)), q8_qs[2], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[3], 4)), q8_qs[2], 3);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[4], 4)), q8_qs[3], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[5], 4)), q8_qs[3], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[6], 4)), q8_qs[3], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], vreinterpretq_s8_u8(vshrq_n_u8(q4_cols[7], 4)), q8_qs[3], 3);
                 }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+
+                // Scales
+                // row c0123 blk0 and blk1
+                const int16x4_t   sc_0123_lo = vget_low_s16(q4sb_scales[0]);
+                const int16x4_t   sc_0123_hi = vget_low_s16(q4sb_scales[1]);
+                const float32x4_t sumf_0123  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[0]),
+                                                                       vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[0])));
+                acc_f32[0]                   = vfmaq_f32(acc_f32[0], sb_scale_0123, sumf_0123);
+                // row c4567 blk0 and blk1
+                const int16x4_t   sc_4567_lo = vget_high_s16(q4sb_scales[0]);
+                const int16x4_t   sc_4567_hi = vget_high_s16(q4sb_scales[1]);
+                const float32x4_t sumf_4567  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[1]),
+                                                                       vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[1])));
+                acc_f32[1]                   = vfmaq_f32(acc_f32[1], sb_scale_4567, sumf_4567);
+
+                // Bias Correction
+                const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+            }  // for sb
+
+            acc_f32[0] = vmlsq_f32(acc_f32[0], vcvtq_f32_s32(bias_acc[0]), sb_min_0123);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_4567);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // #if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q4_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_q4_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_pairs = ncols_interleaved / 2;
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[ncols_interleaved / 4];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < ncols_interleaved / 4; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q4_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q4_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q4_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q4_d_1, q8_d);
+            float32x4_t q4_dmin_0  = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q4_dmin_1  = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0   = vmulq_f32(q4_dmin_0, q8_d);
+            float32x4_t sb_min_1   = vmulq_f32(q4_dmin_1, q8_d);
+
+            // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567
+            int32x4_t bias_acc[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
+            // 2 sb each iteration
+            int32x4_t acc_lo[col_pairs];
+            int32x4_t acc_hi[col_pairs];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_pairs; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
+                int16x8_t q4sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q4sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                }
+
+                const uint8_t * q4_base = q4_ptr[b].qs + sb * QK_K;
+
+                // Load the 64 quants from q8K duplicated to use vecdots with the interelaved columns
+                // but still need the qs to use the low and hi bits from q4
+                const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
+                int8x16_t      q8_qs[8];
+                for (int i = 0; i < 8; i++) {
+                    q8_qs[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base + i * 8));
+                }
+
+                // Q4s columns iterated in pairs (01, 23, 45, 67)
+                for (int cp = 0; cp < col_pairs; cp++) {
+                    uint8x16_t q4_qs_cp_0 = vld1q_u8(q4_base + 16 * cp);
+                    uint8x16_t q4_qs_cp_1 = vld1q_u8(q4_base + 16 * cp + 64);
+                    uint8x16_t q4_qs_cp_2 = vld1q_u8(q4_base + 16 * cp + 128);
+                    uint8x16_t q4_qs_cp_3 = vld1q_u8(q4_base + 16 * cp + 192);
+
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_0, m4b)), q8_qs[0]);  // 0 .. 7
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_1, m4b)), q8_qs[1]);  // 8 ..15
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_2, m4b)), q8_qs[2]);  // 16..23
+                    acc_lo[cp] =
+                        ggml_vdotq_s32(acc_lo[cp], vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_3, m4b)), q8_qs[3]);  // 24..31
+
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_0, 4)), q8_qs[4]);  // 32..39
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_1, 4)), q8_qs[5]);  // 40..47
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_2, 4)), q8_qs[6]);  // 48..55
+                    acc_hi[cp] =
+                        ggml_vdotq_s32(acc_hi[cp], vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_3, 4)), q8_qs[7]);  // 56..63
+                }
+
+                // Iterates over a pair of column pairs (4 columns) to use a single 128 register
+                // p = 0 -> 0123  p2 -> 4567
+                for (int i = 0, p = 0; p < col_pairs; i++, p += 2) {
+                    int16x4_t   group_scales_lo = p == 0 ? vget_low_s16(q4sb_scales[0]) : vget_high_s16(q4sb_scales[0]);
+                    int16x4_t   group_scales_hi = p == 0 ? vget_low_s16(q4sb_scales[1]) : vget_high_s16(q4sb_scales[1]);
+                    float32x4_t sb_scale        = p == 0 ? sb_scale_0 : sb_scale_1;
+
+                    // 0123 or 4567
+                    float32x4_t sumf_0 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_lo), vpaddq_s32(acc_lo[p], acc_lo[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_0);
+
+                    float32x4_t sumf_1 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_hi), vpaddq_s32(acc_hi[p], acc_hi[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_1);
+                }
+
+                // Multiply Acc bsum + mins
+                // Each pair of subblocks share the same bsums
+                // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                // cols 0-3 bias
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+
+                // cols 4-7 bias
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+            }  // for sb
+
+            acc_f32[0] = vmlsq_f32(acc_f32[0], vcvtq_f32_s32(bias_acc[0]), sb_min_0);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_1);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_q8_0_4x4_q8_0(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx;
+
+    for (int c = 0; c < nc; c += ncols_interleaved) {
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        float32x4_t        acc   = vdupq_n_f32(0);
+        for (int b = 0; b < nb; b++) {
+            int8x16x4_t b_low  = vld1q_s8_x4((const int8_t *) b_ptr->qs);
+            int8x16x4_t b_high = vld1q_s8_x4((const int8_t *) b_ptr->qs + 64);
+            float16x4_t bd     = vld1_f16((const __fp16 *) b_ptr->d);
+
+            int8x16x2_t a  = vld1q_s8_x2(a_ptr->qs);
+            float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
+
+            int32x4_t ret = vdupq_n_s32(0);
+
+            ret = vdotq_laneq_s32(ret, b_low.val[0], a.val[0], 0);
+            ret = vdotq_laneq_s32(ret, b_low.val[1], a.val[0], 1);
+            ret = vdotq_laneq_s32(ret, b_low.val[2], a.val[0], 2);
+            ret = vdotq_laneq_s32(ret, b_low.val[3], a.val[0], 3);
+
+            ret = vdotq_laneq_s32(ret, b_high.val[0], a.val[1], 0);
+            ret = vdotq_laneq_s32(ret, b_high.val[1], a.val[1], 1);
+            ret = vdotq_laneq_s32(ret, b_high.val[2], a.val[1], 2);
+            ret = vdotq_laneq_s32(ret, b_high.val[3], a.val[1], 3);
+
+            acc = vfmaq_f32(acc, vcvtq_f32_s32(ret), vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+            a_ptr++;
+            b_ptr++;
         }
+        vst1q_f32(s, acc);
+        s += ncols_interleaved;
     }
+    return;
+
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q8_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_q8_0_4x8_q8_0(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx;
+
+    for (int c = 0; c < nc; c += ncols_interleaved) {
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        float32x4_t        acc   = vdupq_n_f32(0);
+
+        for (int b = 0; b < nb; b++) {
+            int8x16x4_t b_low  = vld1q_s8_x4((const int8_t *) b_ptr->qs);
+            int8x16x4_t b_high = vld1q_s8_x4((const int8_t *) b_ptr->qs + 64);
+            float16x4_t bd     = vld1_f16((const __fp16 *) b_ptr->d);
+
+            int8x8x4_t  a_chunks = vld1_s8_x4(a_ptr->qs);
+            int8x16_t   a0       = vcombine_s8(a_chunks.val[0], a_chunks.val[0]);
+            int8x16_t   a1       = vcombine_s8(a_chunks.val[1], a_chunks.val[1]);
+            int8x16_t   a2       = vcombine_s8(a_chunks.val[2], a_chunks.val[2]);
+            int8x16_t   a3       = vcombine_s8(a_chunks.val[3], a_chunks.val[3]);
+            float16x4_t ad       = vld1_dup_f16((const __fp16 *) &a_ptr->d);
+
+            int32x4_t ret0 = vdupq_n_s32(0);
+            int32x4_t ret1 = vdupq_n_s32(0);
+
+            // 0..7
+            ret0 = vdotq_s32(ret0, b_low.val[0], a0);
+            ret1 = vdotq_s32(ret1, b_low.val[1], a0);
+            // 8..15
+            ret0 = vdotq_s32(ret0, b_low.val[2], a1);
+            ret1 = vdotq_s32(ret1, b_low.val[3], a1);
+            // 16..23
+            ret0 = vdotq_s32(ret0, b_high.val[0], a2);
+            ret1 = vdotq_s32(ret1, b_high.val[1], a2);
+            // 24..31
+            ret0 = vdotq_s32(ret0, b_high.val[2], a3);
+            ret1 = vdotq_s32(ret1, b_high.val[3], a3);
+
+            int32x4_t ret = vpaddq_s32(ret0, ret1);
+
+            acc = vfmaq_f32(acc, vcvtq_f32_s32(ret), vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+            a_ptr++;
+            b_ptr++;
+        }
+        vst1q_f32(s, acc);
+        s += ncols_interleaved;
+    }
+    return;
+
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q8_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1096,40 +1391,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     );
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4][4];
-        int sumi;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-                }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                        for (int m = 0; m < 4; m++) {
-                            for (int j = 0; j < ncols_interleaved; j++) {
-                                sumi = 0;
-                                for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                                }
-                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                            }
-                        }
-                    }
-                }
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++)
-                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-                }
-            }
-        }
-    }
+    ggml_gemm_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1550,38 +1812,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     );
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
-    float sumf[4][4];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+    ggml_gemm_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2019,38 +2250,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
 
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-    float sumf[4][8];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++)
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-            }
-        }
-    }
+    ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2126,38 +2326,570 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     }
     return;
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4][4];
-        int sumi;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+    ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    acc_size  = 2 * 4;  // 2 row pairs × 4 col pairs
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[acc_size];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < acc_size; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // d4 0 1 2 3, 4 5 6 7
+                float32x4_t q4_d_0123    = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d));
+                float32x4_t q4_d_4567    = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].d + 4));
+                // d8 0 1 2 3
+                float32x4_t q8_d_0123    = vld1q_f32(q8_ptr[b].d);
+                // mins
+                float32x4_t q4_dmin_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin));
+                float32x4_t q4_dmin_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q4_ptr[b].dmin + 4));
+
+                // Precomputation of scales and mins
+                float32x4_t sbd_scale_0123[q8_k_blocklen];
+                float32x4_t sbd_scale_4567[q8_k_blocklen];
+                float32x4_t sbd_min_0123[q8_k_blocklen];
+                float32x4_t sbd_min_4567[q8_k_blocklen];
+
+                sbd_scale_0123[0] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 0);
+                sbd_scale_4567[0] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 0);
+                sbd_min_0123[0]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 0);
+                sbd_min_4567[0]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 0);
+
+                sbd_scale_0123[1] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 1);
+                sbd_scale_4567[1] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 1);
+                sbd_min_0123[1]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 1);
+                sbd_min_4567[1]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 1);
+
+                sbd_scale_0123[2] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 2);
+                sbd_scale_4567[2] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 2);
+                sbd_min_0123[2]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 2);
+                sbd_min_4567[2]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 2);
+
+                sbd_scale_0123[3] = vmulq_laneq_f32(q4_d_0123, q8_d_0123, 3);
+                sbd_scale_4567[3] = vmulq_laneq_f32(q4_d_4567, q8_d_0123, 3);
+                sbd_min_0123[3]   = vmulq_laneq_f32(q4_dmin_0123, q8_d_0123, 3);
+                sbd_min_4567[3]   = vmulq_laneq_f32(q4_dmin_4567, q8_d_0123, 3);
+
+                // Precomputation of bsums, each vpaddq calcs all the bsums for each row
+                const int16x8_t bsums[q8_k_blocklen] = {
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[QK_K / 64][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                // interleaved bias_acc: [0]->r0 0123, [1]->r1 0123, .., [4]->r0 4567, [5]->r1 4567 ..
+                int32x4_t bias_acc[acc_size];
+                for (int i = 0; i < acc_size; i++) {
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Int accumulators for qs vecdot (4 row x 2 col quartets)
+                    int32x4_t acc_lo[acc_size];
+                    int32x4_t acc_hi[acc_size];
+                    for (int i = 0; i < acc_size; i++) {
+                        acc_lo[i] = vdupq_n_s32(0);
+                        acc_hi[i] = vdupq_n_s32(0);
+                    }
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int16x8_t q4sb_scales[2];
+                    int16x8_t q4sb_mins[2];
+                    for (int i = 0; i < 2; i++) {
+                        int8_t    aux_q4sb[8];
+                        const int offset = sb * 24 + i * 12;
+                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                        q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
+                    }
+
+                    constexpr int reads_per_sb = 8;  // 8 * 16 bytes each => 32 qs * 4 rows
+                    for (int k = 0; k < reads_per_sb; k++) {
+                        const int8x16_t q8_blk0 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k);
+                        const int8x16_t q8_blk1 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k + 128);
+
+                        // 0..3 & 32..35
+                        const uint8x16_t q4_0123 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 32 * k);
+                        const uint8x16_t q4_4567 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 32 * k + 16);
+
+                        const int8x16_t q4_0123_lo = vreinterpretq_s8_u8(vandq_u8(q4_0123, m4b));
+                        const int8x16_t q4_0123_hi = vreinterpretq_s8_u8(vshrq_n_u8(q4_0123, 4));
+
+                        acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q4_0123_lo, q8_blk0, 0);  //  0..3  r0 c0123
+                        acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q4_0123_lo, q8_blk0, 1);  //  0..3  r1 c0123
+                        acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q4_0123_lo, q8_blk0, 2);  //  0..3  r2 c0123
+                        acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q4_0123_lo, q8_blk0, 3);  //  0..3  r3 c0123
+
+                        acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q4_0123_hi, q8_blk1, 0);  // 32..35 r0 c0123
+                        acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q4_0123_hi, q8_blk1, 1);  // 32..35 r1 c0123
+                        acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q4_0123_hi, q8_blk1, 2);  // 32..35 r2 c0123
+                        acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q4_0123_hi, q8_blk1, 3);  // 32..35 r3 c0123
+
+                        const int8x16_t q4_4567_lo = vreinterpretq_s8_u8(vandq_u8(q4_4567, m4b));
+                        const int8x16_t q4_4567_hi = vreinterpretq_s8_u8(vshrq_n_u8(q4_4567, 4));
+
+                        acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q4_4567_lo, q8_blk0, 0);  //  0..3  r0 c4567
+                        acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q4_4567_lo, q8_blk0, 1);  //  0..3  r1 c4567
+                        acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q4_4567_lo, q8_blk0, 2);  //  0..3  r2 c4567
+                        acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q4_4567_lo, q8_blk0, 3);  //  0..3  r3 c4567
+
+                        acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q4_4567_hi, q8_blk1, 0);  // 32..35 r0 c4567
+                        acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q4_4567_hi, q8_blk1, 1);  // 32..35 r1 c4567
+                        acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q4_4567_hi, q8_blk1, 2);  // 32..35 r2 c4567
+                        acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q4_4567_hi, q8_blk1, 3);  // 32..35 r3 c4567
+                    }
+
+                    // Scale and bias application
+                    // acc is stored interleaved to match output layout
+                    const int16x4_t sc_0123_lo = vget_low_s16(q4sb_scales[0]);
+                    const int16x4_t sc_4567_lo = vget_high_s16(q4sb_scales[0]);
+                    const int16x4_t sc_0123_hi = vget_low_s16(q4sb_scales[1]);
+                    const int16x4_t sc_4567_hi = vget_high_s16(q4sb_scales[1]);
+                    for (int row = 0; row < q8_k_blocklen; row++) {
+                        // Bias correction
+                        // row c0123 blk0 and blk1
+                        const float32x4_t sumf_0123 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[row]),
+                                                    vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[row])));
+                        acc_f32[2 * row] = vfmaq_f32(acc_f32[2 * row], sbd_scale_0123[row], sumf_0123);
+
+                        // row c4567 blk0 and blk1
+                        const float32x4_t sumf_4567 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[row + 4]),
+                                                    vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[row + 4])));
+                        acc_f32[2 * row + 1] = vfmaq_f32(acc_f32[2 * row + 1], sbd_scale_4567[row], sumf_4567);
+
+                        // Bias
+                        const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][row * 2]);
+                        const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][row * 2 + 1]);
+
+                        // row c0123 blk0 and blk1
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+
+                        // row c4567 blk0 and blk1
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+                    }
+                }  // for sb
+
+                for (int row = 0; row < q8_k_blocklen; row++) {
+                    acc_f32[2 * row] = vmlsq_f32(acc_f32[2 * row], vcvtq_f32_s32(bias_acc[2 * row]), sbd_min_0123[row]);
+                    acc_f32[2 * row + 1] =
+                        vmlsq_f32(acc_f32[2 * row + 1], vcvtq_f32_s32(bias_acc[2 * row + 1]), sbd_min_4567[row]);
+                }
+            }  // for b
+
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemm_q4_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemm_q4_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    constexpr int    q8_k_blocklen = 4;
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[blocklen];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < blocklen; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // bsums pairs belongs to the same q8_k subblock
+                const int16x8_t bsums[4]{
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[4][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
                 }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                        for (int m = 0; m < 4; m++) {
-                            for (int j = 0; j < ncols_interleaved; j++) {
-                                sumi = 0;
-                                for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+
+                int32x4_t sb_acc[4];    // Aux accumulators to store subblock (partial) results
+                int32x4_t acc[8];       // rows 01 stored in [0][1][2][3] rows 23 stored in [4][5][6][7]
+                int32x4_t bias_acc[8];  // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567, [2]->r1 0123 ...
+                for (int i = 0; i < 8; i++) {
+                    acc[i]      = vdupq_n_s32(0);
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int8_t    q4sb_scales[2][8];
+                    int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
+                    for (int i = 0; i < 2; i++) {
+                        const int offset = sb * 24 + i * 12;
+                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], q4sb_scales[i]);
+                    }
+
+                    // q8_ptr[b].qs has interleaved Q8 rows (01, 23)
+                    const int8_t * q8_base = q8_ptr[b].qs + sb * 256;
+
+                    int8x16_t q8_qs_01[8];
+                    int8x16_t q8_qs_23[8];
+
+                    // Load 32-byte per row pair, 1 subblock each time
+                    for (int i = 0; i < 8; i++) {
+                        const int offset = i * 32;  // 16 for row 01, 16 for row 23
+                        q8_qs_01[i]      = vld1q_s8(q8_base + offset);
+                        q8_qs_23[i]      = vld1q_s8(q8_base + offset + 16);
+                    }
+
+                    const int8x16_t q8s[2][8] = {
+                        { q8_qs_01[0], q8_qs_01[1], q8_qs_01[2], q8_qs_01[3],
+                          q8_qs_01[4], q8_qs_01[5], q8_qs_01[6], q8_qs_01[7] },
+                        { q8_qs_23[0], q8_qs_23[1], q8_qs_23[2], q8_qs_23[3],
+                          q8_qs_23[4], q8_qs_23[5], q8_qs_23[6], q8_qs_23[7] },
+                    };
+
+                    // Q4s columns iterated in pairs (01, 23, 45, 67)
+                    for (int cp = 0; cp < ncols_interleaved / 2; cp++) {
+                        for (int i = 0; i < 4; i++) {
+                            sb_acc[i] = vdupq_n_s32(0);
+                        }
+
+                        uint8x16_t q4_qs_cp_0 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 0);    // 0 .. 7 & 32..39
+                        uint8x16_t q4_qs_cp_1 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 64);   // 8 ..15 & 40..47
+                        uint8x16_t q4_qs_cp_2 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 128);  // 16..23 & 48..55
+                        uint8x16_t q4_qs_cp_3 = vld1q_u8(q4_ptr[b].qs + sb * QK_K + 16 * cp + 192);  // 24..31 & 56..63
+                        const int8x16_t q4_nibbles[2][4] = {
+                            {
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_0, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_1, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_2, m4b)),
+                                vreinterpretq_s8_u8(vandq_u8(q4_qs_cp_3, m4b)),
+                            },
+                            {
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_0, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_1, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_2, 4)),
+                                vreinterpretq_s8_u8(vshrq_n_u8(q4_qs_cp_3, 4)),
+                            }
+                        };
+
+                        // Calculates the Qs muladd of every row pair (rp) rows 01 and 23 of q8
+                        // for each of the internal 32 qs subblock (blk)
+                        for (int rp = 0; rp < 2; rp++) {
+                            for (int blk = 0; blk < 2; blk++) {
+                                const int8x16_t * q8  = &q8s[rp][4 * blk];
+                                const int8x16_t * q4  = q4_nibbles[blk];
+                                int32x4_t         acc = sb_acc[2 * rp + blk];
+                                // mul add for each qs in the same subblock
+                                for (int qs_offset = 0; qs_offset < 4; qs_offset++) {
+                                    acc = vmmlaq_s32(acc, q4[qs_offset], q8[qs_offset]);
                                 }
-                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sb_acc[2 * rp + blk] = acc;
                             }
                         }
+
+                        // Scales[i] corresponds to column i
+                        const int scale_offset = cp * 2;
+                        for (int blk = 0; blk < 2; blk++) {
+                            const int32x4_t block_scale = {
+                                (int32_t) q4sb_scales[blk][scale_offset],
+                                (int32_t) q4sb_scales[blk][scale_offset],
+                                (int32_t) q4sb_scales[blk][scale_offset + 1],
+                                (int32_t) q4sb_scales[blk][scale_offset + 1],
+                            };
+                            acc[cp]     = vmlaq_s32(acc[cp], sb_acc[blk], block_scale);
+                            acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[blk + 2], block_scale);
+                        }
+                    }
+
+                    // Multiply Acc bsum + mins
+                    for (int q8_row = 0; q8_row < 4; q8_row++) {
+                        // Each pair of subblocks share the same bsums
+                        // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                        int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][q8_row * 2]);
+                        int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][q8_row * 2 + 1]);
+
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_lo, vget_low_s16(q4sb_mins[0]));
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_hi, vget_low_s16(q4sb_mins[1]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_lo, vget_high_s16(q4sb_mins[0]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_hi, vget_high_s16(q4sb_mins[1]));
+                    }
+                }  // for sb
+
+                // Reorder of i8mm output with bias and output layout
+                for (int i = 0; i < 8; i++) {
+                    int32x2x2_t aux = vzip_s32(vget_low_s32(acc[i]), vget_high_s32(acc[i]));
+                    acc[i]          = vcombine_s32(aux.val[0], aux.val[1]);
+                }
+                int32x4_t reorder_acc[8] = {
+                    vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1])),
+                    vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3])),
+                    vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1])),
+                    vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3])),
+                    vcombine_s32(vget_low_s32(acc[4]), vget_low_s32(acc[5])),
+                    vcombine_s32(vget_low_s32(acc[6]), vget_low_s32(acc[7])),
+                    vcombine_s32(vget_high_s32(acc[4]), vget_high_s32(acc[5])),
+                    vcombine_s32(vget_high_s32(acc[6]), vget_high_s32(acc[7])),
+                };
+
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    for (int j = 0; j < 2; j++) {
+                        float32x4_t       q8_d    = vdupq_n_f32(q8_ptr[b].d[i]);
+                        float32x4_t       q4_dmin = vcvt_f32_f16(vld1_f16((const __fp16 *) (q4_ptr[b].dmin + j * 4)));
+                        const float32x4_t dmins   = vmulq_f32(q4_dmin, q8_d);
+
+                        float32x4_t       q4_d  = vcvt_f32_f16(vld1_f16((const __fp16 *) (q4_ptr[b].d + j * 4)));
+                        const float32x4_t scale = vmulq_f32(q4_d, q8_d);
+
+                        acc_f32[2 * i + j] = vmlsq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(bias_acc[2 * i + j]), dmins);
+                        acc_f32[2 * i + j] =
+                            vmlaq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(reorder_acc[2 * i + j]), scale);
+                    }
+                }
+            }  // for b
+
+            // With the previous reorder, the tile is already in the correct memory layout.
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+
+void ggml_gemm_q8_0_4x4_q8_0(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb);
+
+            float32x4_t sumf[4];
+            for (int m = 0; m < 4; m++) {
+                sumf[m] = vdupq_n_f32(0);
+            }
+
+            for (int l = 0; l < nb; l++) {
+                float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *) a_ptr[l].d));
+                float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *) b_ptr[l].d));
+
+                int32x4_t sumi_0 = vdupq_n_s32(0);
+                int32x4_t sumi_1 = vdupq_n_s32(0);
+                int32x4_t sumi_2 = vdupq_n_s32(0);
+                int32x4_t sumi_3 = vdupq_n_s32(0);
+
+                for (int k_group = 0; k_group < 8; k_group += 4) {
+                    int8x16x4_t a = vld1q_s8_x4(a_ptr[l].qs + 16 * k_group);
+                    int8x16x4_t b = vld1q_s8_x4(b_ptr[l].qs + 16 * k_group);
+
+                    for (int k = 0; k < 4; k++) {
+                        sumi_0 = vdotq_laneq_s32(sumi_0, b.val[k], a.val[k], 0);
+                        sumi_1 = vdotq_laneq_s32(sumi_1, b.val[k], a.val[k], 1);
+                        sumi_2 = vdotq_laneq_s32(sumi_2, b.val[k], a.val[k], 2);
+                        sumi_3 = vdotq_laneq_s32(sumi_3, b.val[k], a.val[k], 3);
                     }
                 }
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++)
-                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+
+                sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0));
+                sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1));
+                sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2));
+                sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3));
+            }
+
+            for (int m = 0; m < 4; m++) {
+                vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]);
+            }
+        }
+    }
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemm_q8_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemm_q8_0_4x8_q8_0(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    const int qk                = QK8_0;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    const block_q8_0x4 * b_ptr_base = (const block_q8_0x4 *) vx;
+
+    for (int y = 0; y < nr; y += 4) {
+        const block_q8_0x4 * a_ptr_base = (const block_q8_0x4 *) vy + (y / 4) * nb;
+
+        for (int x = 0; x < nc; x += ncols_interleaved) {
+            const block_q8_0x4 * b_ptr = b_ptr_base + (x / 4) * nb;
+            const block_q8_0x4 * a_ptr = a_ptr_base;
+
+            float32x4_t acc_f32[4];
+            for (int i = 0; i < 4; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                int32x4_t acc[4];
+                for (int i = 0; i < 4; i++) {
+                    acc[i] = vdupq_n_s32(0);
                 }
+
+                // Process 4 chunks of 8 positions each
+                for (int chunk = 0; chunk < 4; chunk++) {
+                    int8x16_t a01 = vld1q_s8(a_ptr->qs + chunk * 32);
+                    int8x16_t a23 = vld1q_s8(a_ptr->qs + chunk * 32 + 16);
+                    int8x16_t b01 = vld1q_s8(b_ptr->qs + chunk * 32);
+                    int8x16_t b23 = vld1q_s8(b_ptr->qs + chunk * 32 + 16);
+
+                    acc[0] = vmmlaq_s32(acc[0], a01, b01);
+                    acc[1] = vmmlaq_s32(acc[1], a01, b23);
+                    acc[2] = vmmlaq_s32(acc[2], a23, b01);
+                    acc[3] = vmmlaq_s32(acc[3], a23, b23);
+                }
+
+                // Reorder outputs from 2×2 tiles to row-major
+                // acc[0] = [r0c0, r0c1, r1c0, r1c1]
+                // acc[1] = [r0c2, r0c3, r1c2, r1c3]
+                // acc[2] = [r2c0, r2c1, r3c0, r3c1]
+                // acc[3] = [r2c2, r2c3, r3c2, r3c3]
+                int32x4_t row0 = vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1]));
+                int32x4_t row1 = vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1]));
+                int32x4_t row2 = vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3]));
+                int32x4_t row3 = vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3]));
+
+                // Scales
+                float32x4_t a_d = vcvt_f32_f16(vld1_f16((const __fp16 *) a_ptr->d));
+                float32x4_t b_d = vcvt_f32_f16(vld1_f16((const __fp16 *) b_ptr->d));
+
+                acc_f32[0] = vfmaq_f32(acc_f32[0], vcvtq_f32_s32(row0), vmulq_laneq_f32(b_d, a_d, 0));
+                acc_f32[1] = vfmaq_f32(acc_f32[1], vcvtq_f32_s32(row1), vmulq_laneq_f32(b_d, a_d, 1));
+                acc_f32[2] = vfmaq_f32(acc_f32[2], vcvtq_f32_s32(row2), vmulq_laneq_f32(b_d, a_d, 2));
+                acc_f32[3] = vfmaq_f32(acc_f32[3], vcvtq_f32_s32(row3), vmulq_laneq_f32(b_d, a_d, 3));
+
+                a_ptr++;
+                b_ptr++;
+            }
+
+            for (int row = 0; row < 4; row++) {
+                vst1q_f32(s + (y + row) * bs + x, acc_f32[row]);
             }
         }
     }
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_gemm_q8_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }