llama_cpp 0.16.1 → 0.16.2

data/vendor/tmp/llama.cpp/ggml-quants.c

@@ -4,8 +4,6 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 
-#define GGML_COMMON_IMPL_C
-#include "ggml-common.h"
 
 #include <math.h>
 #include <string.h>
@@ -1078,6 +1076,7 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
         }
         vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
         vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
+    }
 
 #elif defined(__loongarch_asx)
     for (int i = 0; i < nb; i++) {
@@ -1437,6 +1436,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
         accv = vec_add(accv, vec_sld(accv, accv, 4));
         accv = vec_add(accv, vec_sld(accv, accv, 8));
         y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
+    }
 
 #elif defined(__loongarch_asx)
     for (int i = 0; i < nb; i++) {
@@ -4113,12 +4113,13 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
     const vector signed char v8 = vec_splats((signed char)0x8);
 
     vector float vsumf0 = vec_splats(0.0f);
 
-#pragma GCC unroll 4
+#pragma GCC unroll 8
     for (int i = 0; i < nb; i++) {
         __builtin_prefetch(x[i].qs, 0, 1);
         __builtin_prefetch(y[i].qs, 0, 1);
@@ -4140,9 +4141,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
         vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
 
-        qv0 = vec_add(qv0, qv1);
+        vector signed int vsumi0 = v0;
 
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi0 = vec_sum4s(qv1, vsumi0);
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
     }
@@ -4516,6 +4518,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
@@ -4537,15 +4540,13 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         vector signed char q8y0 = vec_xl( 0, y[i].qs);
         vector signed char q8y1 = vec_xl(16, y[i].qs);
 
-        vector signed char q4x0 = vec_and(qxs, lowMask);
-        vector signed char q4x1 = vec_sr(qxs, v4);
+        vector unsigned char q4x0 = (vector unsigned char)vec_and(qxs, lowMask);
+        vector unsigned char q4x1 = (vector unsigned char)vec_sr(qxs, v4);
 
-        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+        vector signed int vsumi0 = v0;
 
-        qv0 = vec_add(qv0, qv1);
-
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+        vsumi0 = vec_msum(q8y0, q4x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q4x1, vsumi0);
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
     }
@@ -5247,6 +5248,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
@@ -5272,18 +5274,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
         vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs);
 
-        vector signed char q5x0 = vec_or(vec_and(qxs, lowMask), qh0);
-        vector signed char q5x1 = vec_or(vec_sr(qxs, v4), qh1);
+        vector unsigned char q5x0 = (vector unsigned char)vec_or(vec_and(qxs, lowMask), qh0);
+        vector unsigned char q5x1 = (vector unsigned char)vec_or(vec_sr(qxs, v4), qh1);
 
         vector signed char q8y0 = vec_xl(  0, y[i].qs);
         vector signed char q8y1 = vec_xl( 16, y[i].qs);
 
-        vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1));
+        vector signed int vsumi0 = v0;
 
-        qv0 = vec_add(qv0, qv1);
-
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+        vsumi0 = vec_msum(q8y0, q5x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q5x1, vsumi0);
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
     }
@@ -5523,9 +5523,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     *s = sumf;
 
 #elif defined(__POWER9_VECTOR__)
+    const vector signed int v0 = vec_splats((int32_t)0);
     vector float vsumf0 = vec_splats(0.0f);
 
-#pragma GCC unroll 4
+#pragma GCC unroll 8
     for (int i = 0; i < nb; i++) {
         __builtin_prefetch(x[i].qs, 0, 1);
         __builtin_prefetch(y[i].qs, 0, 1);
@@ -5544,13 +5545,13 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         vector signed short qv2 = vec_mule(q8x1, q8y1);
         vector signed short qv3 = vec_mulo(q8x1, q8y1);
 
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackh(qv1));
-        vector signed int vsumi1 = vec_add(vec_unpackl(qv0), vec_unpackl(qv1));
-        vector signed int vsumi2 = vec_add(vec_unpackh(qv2), vec_unpackh(qv3));
-        vector signed int vsumi3 = vec_add(vec_unpackl(qv2), vec_unpackl(qv3));
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
 
-        vsumi0 = vec_add(vsumi0, vsumi2);
-        vsumi1 = vec_add(vsumi1, vsumi3);
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
+        vsumi0 = vec_sum4s(qv2, vsumi0);
+        vsumi1 = vec_sum4s(qv3, vsumi1);
 
         vsumi0 = vec_add(vsumi0, vsumi1);
 
@@ -5938,6 +5939,7 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0x3);
     const vector signed char lowScaleMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector unsigned char v2 = vec_splats((unsigned char)0x2);
     const vector unsigned char v6 = vec_splats((unsigned char)0x6);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
@@ -5975,15 +5977,17 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
         vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
 
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
 
         for (int j = 0; j < QK_K/128; ++j) {
             __builtin_prefetch(q2, 0, 1);
@@ -5993,14 +5997,14 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char qxs1 = (vector signed char)vec_xl(16, q2);
             q2 += 32;
 
-            vector signed char q2x00 = vec_and(qxs0, lowMask);
-            vector signed char q2x01 = vec_and(vec_sr(qxs0, v2), lowMask);
-            vector signed char q2x02 = vec_and(vec_sr(qxs0, v4), lowMask);
-            vector signed char q2x03 = vec_and(vec_sr(qxs0, v6), lowMask);
-            vector signed char q2x10 = vec_and(qxs1, lowMask);
-            vector signed char q2x11 = vec_and(vec_sr(qxs1, v2), lowMask);
-            vector signed char q2x12 = vec_and(vec_sr(qxs1, v4), lowMask);
-            vector signed char q2x13 = vec_and(vec_sr(qxs1, v6), lowMask);
+            vector unsigned char q2x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q2x01 = (vector unsigned char)vec_and(vec_sr(qxs0, v2), lowMask);
+            vector unsigned char q2x02 = (vector unsigned char)vec_and(vec_sr(qxs0, v4), lowMask);
+            vector unsigned char q2x03 = (vector unsigned char)vec_and(vec_sr(qxs0, v6), lowMask);
+            vector unsigned char q2x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q2x11 = (vector unsigned char)vec_and(vec_sr(qxs1, v2), lowMask);
+            vector unsigned char q2x12 = (vector unsigned char)vec_and(vec_sr(qxs1, v4), lowMask);
+            vector unsigned char q2x13 = (vector unsigned char)vec_and(vec_sr(qxs1, v6), lowMask);
 
             vector signed char q8y00 = vec_xl(  0, q8);
             vector signed char q8y10 = vec_xl( 16, q8);
@@ -6012,45 +6016,36 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char q8y13 = vec_xl(112, q8);
             q8 += 128;
 
-            vector signed short qv0 = vec_add(vec_mule(q2x00, q8y00), vec_mulo(q2x00, q8y00));
-            vector signed short qv1 = vec_add(vec_mule(q2x01, q8y01), vec_mulo(q2x01, q8y01));
-            vector signed short qv2 = vec_add(vec_mule(q2x02, q8y02), vec_mulo(q2x02, q8y02));
-            vector signed short qv3 = vec_add(vec_mule(q2x03, q8y03), vec_mulo(q2x03, q8y03));
-            vector signed short qv4 = vec_add(vec_mule(q2x10, q8y10), vec_mulo(q2x10, q8y10));
-            vector signed short qv5 = vec_add(vec_mule(q2x11, q8y11), vec_mulo(q2x11, q8y11));
-            vector signed short qv6 = vec_add(vec_mule(q2x12, q8y12), vec_mulo(q2x12, q8y12));
-            vector signed short qv7 = vec_add(vec_mule(q2x13, q8y13), vec_mulo(q2x13, q8y13));
-
-            vector signed short vscales_h = vec_unpackh(vscales);
-            vector signed short vs0 = vec_splat(vscales_h, 0);
-            vector signed short vs1 = vec_splat(vscales_h, 1);
-            vector signed short vs2 = vec_splat(vscales_h, 2);
-            vector signed short vs3 = vec_splat(vscales_h, 3);
-            vector signed short vs4 = vec_splat(vscales_h, 4);
-            vector signed short vs5 = vec_splat(vscales_h, 5);
-            vector signed short vs6 = vec_splat(vscales_h, 6);
-            vector signed short vs7 = vec_splat(vscales_h, 7);
+            vector signed int qv0 = vec_msum(q8y00, q2x00, v0);
+            vector signed int qv1 = vec_msum(q8y01, q2x01, v0);
+            vector signed int qv2 = vec_msum(q8y02, q2x02, v0);
+            vector signed int qv3 = vec_msum(q8y03, q2x03, v0);
+            vector signed int qv4 = vec_msum(q8y10, q2x10, v0);
+            vector signed int qv5 = vec_msum(q8y11, q2x11, v0);
+            vector signed int qv6 = vec_msum(q8y12, q2x12, v0);
+            vector signed int qv7 = vec_msum(q8y13, q2x13, v0);
+
+            vector signed short vscales_07 = vec_unpackh(vscales);
+            vector signed int vscales_03 = vec_unpackh(vscales_07);
+            vector signed int vscales_47 = vec_unpackl(vscales_07);
+            vector signed int vs0 = vec_splat(vscales_03, 0);
+            vector signed int vs1 = vec_splat(vscales_03, 1);
+            vector signed int vs2 = vec_splat(vscales_03, 2);
+            vector signed int vs3 = vec_splat(vscales_03, 3);
+            vector signed int vs4 = vec_splat(vscales_47, 0);
+            vector signed int vs5 = vec_splat(vscales_47, 1);
+            vector signed int vs6 = vec_splat(vscales_47, 2);
+            vector signed int vs7 = vec_splat(vscales_47, 3);
             vscales = vec_sld(vscales, vscales, 8);
 
-            qv0 = vec_mul(qv0, vs0);
-            qv1 = vec_mul(qv1, vs2);
-            qv2 = vec_mul(qv2, vs4);
-            qv3 = vec_mul(qv3, vs6);
-
-            qv0 = vec_madd(qv4, vs1, qv0);
-            qv1 = vec_madd(qv5, vs3, qv1);
-            qv2 = vec_madd(qv6, vs5, qv2);
-            qv3 = vec_madd(qv7, vs7, qv3);
-
-            vsumi0 = vec_add(vec_unpackh(qv0), vsumi0);
-            vsumi1 = vec_add(vec_unpackh(qv1), vsumi1);
-            vsumi2 = vec_add(vec_unpackh(qv2), vsumi2);
-            vsumi3 = vec_add(vec_unpackh(qv3), vsumi3);
-
-            vsumi4 = vec_add(vec_unpackl(qv0), vsumi4);
-            vsumi5 = vec_add(vec_unpackl(qv1), vsumi5);
-            vsumi6 = vec_add(vec_unpackl(qv2), vsumi6);
-            vsumi7 = vec_add(vec_unpackl(qv3), vsumi7);
+            vsumi0 = vec_add(vec_mul(qv0, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv1, vs2), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv2, vs4), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv3, vs6), vsumi3);
+            vsumi4 = vec_add(vec_mul(qv4, vs1), vsumi4);
+            vsumi5 = vec_add(vec_mul(qv5, vs3), vsumi5);
+            vsumi6 = vec_add(vec_mul(qv6, vs5), vsumi6);
+            vsumi7 = vec_add(vec_mul(qv7, vs7), vsumi7);
         }
 
         vsumi0 = vec_add(vsumi0, vsumi4);
@@ -6641,6 +6636,9 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0x3);
+    const vector signed char lowMask1 = vec_splats((int8_t)0xf);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector signed char v1 = vec_splats((signed char)0x1);
     const vector unsigned char v2 = vec_splats((unsigned char)0x2);
     const vector unsigned char v3 = vec_splats((unsigned char)0x3);
@@ -6658,30 +6656,33 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        uint32_t aux[3];
-        uint32_t utmp[4];
+        UNUSED(kmask1);
+        UNUSED(kmask2);
 
-        memcpy(aux, x[i].scales, 12);
-        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
-        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
-        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
-        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(u0, lowMask1);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = (vector signed char)vec_mergeh((vector signed int)u2, (vector signed int)vec_sr(u2, v2));
+        vector signed char u30 = vec_sl(vec_and(u3, lowMask), v4);
+        vector signed char u31 = vec_and(u3, lowMask2);
 
-        vector signed char vscales = (vector signed char)vec_xl( 0, utmp);
+        u1 = vec_or(u1, u30);
+        u2 = vec_or(vec_sr(u0, v4), u31);
+
+        vector signed char vscales = (vector signed char)vec_mergeh((vector signed long long)u1, (vector signed long long)u2);
         vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].hmask);
         vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].hmask);
 
         vscales = vec_sub(vscales, off);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
-
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
 
         const uint8_t * restrict q3 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -6755,23 +6756,14 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed short qv12 = vec_add(vec_mule(q3x12, q8y12), vec_mulo(q3x12, q8y12));
             vector signed short qv13 = vec_add(vec_mule(q3x13, q8y13), vec_mulo(q3x13, q8y13));
 
-            vector signed int vsum0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0));
-            vector signed int vsum1 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2));
-            vector signed int vsum2 = vec_add(vec_mule(qv02, vs4), vec_mulo(qv02, vs4));
-            vector signed int vsum3 = vec_add(vec_mule(qv03, vs6), vec_mulo(qv03, vs6));
-            vector signed int vsum4 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1));
-            vector signed int vsum5 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3));
-            vector signed int vsum6 = vec_add(vec_mule(qv12, vs5), vec_mulo(qv12, vs5));
-            vector signed int vsum7 = vec_add(vec_mule(qv13, vs7), vec_mulo(qv13, vs7));
-
-            vsumi0 = vec_add(vsum0, vsumi0);
-            vsumi1 = vec_add(vsum1, vsumi1);
-            vsumi2 = vec_add(vsum2, vsumi2);
-            vsumi3 = vec_add(vsum3, vsumi3);
-            vsumi4 = vec_add(vsum4, vsumi4);
-            vsumi5 = vec_add(vsum5, vsumi5);
-            vsumi6 = vec_add(vsum6, vsumi6);
-            vsumi7 = vec_add(vsum7, vsumi7);
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs2, vsumi1);
+            vsumi2 = vec_msum(qv02, vs4, vsumi2);
+            vsumi3 = vec_msum(qv03, vs6, vsumi3);
+            vsumi4 = vec_msum(qv10, vs1, vsumi4);
+            vsumi5 = vec_msum(qv11, vs3, vsumi5);
+            vsumi6 = vec_msum(qv12, vs5, vsumi6);
+            vsumi7 = vec_msum(qv13, vs7, vsumi7);
         }
 
         vsumi0 = vec_add(vsumi0, vsumi4);
@@ -7270,6 +7262,10 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((uint8_t)2);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
@@ -7288,15 +7284,24 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
         vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
 
-        memcpy(utmp, x[i].scales, 12);
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
 
-        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;
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
 
-        vector signed char utmps = (vector signed char)vec_xl( 0, utmp);
         vector signed short vscales = vec_unpackh(utmps);
         vector signed short q4xmins = vec_unpackl(utmps);
         vector signed short q4xmins0 = vec_mergeh(q4xmins, q4xmins);
@@ -7312,14 +7317,10 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
         vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -7334,14 +7335,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char qxs3 = (vector signed char)vec_xl(48, q4);
             q4 += 64;
 
-            vector signed char q4x00 = vec_and(qxs0, lowMask);
-            vector signed char q4x01 = vec_sr(qxs0, v4);
-            vector signed char q4x10 = vec_and(qxs1, lowMask);
-            vector signed char q4x11 = vec_sr(qxs1, v4);
-            vector signed char q4x20 = vec_and(qxs2, lowMask);
-            vector signed char q4x21 = vec_sr(qxs2, v4);
-            vector signed char q4x30 = vec_and(qxs3, lowMask);
-            vector signed char q4x31 = vec_sr(qxs3, v4);
+            vector unsigned char q4x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q4x01 = (vector unsigned char)vec_sr(qxs0, v4);
+            vector unsigned char q4x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q4x11 = (vector unsigned char)vec_sr(qxs1, v4);
+            vector unsigned char q4x20 = (vector unsigned char)vec_and(qxs2, lowMask);
+            vector unsigned char q4x21 = (vector unsigned char)vec_sr(qxs2, v4);
+            vector unsigned char q4x30 = (vector unsigned char)vec_and(qxs3, lowMask);
+            vector unsigned char q4x31 = (vector unsigned char)vec_sr(qxs3, v4);
 
             vector signed char q8y00 = vec_xl(  0, q8);
             vector signed char q8y10 = vec_xl( 16, q8);
@@ -7353,41 +7354,33 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char q8y31 = vec_xl(112, q8);
             q8 += 128;
 
-            vector signed short qv00 = vec_add(vec_mule(q4x00, q8y00), vec_mulo(q4x00, q8y00));
-            vector signed short qv01 = vec_add(vec_mule(q4x01, q8y01), vec_mulo(q4x01, q8y01));
-            vector signed short qv10 = vec_add(vec_mule(q4x10, q8y10), vec_mulo(q4x10, q8y10));
-            vector signed short qv11 = vec_add(vec_mule(q4x11, q8y11), vec_mulo(q4x11, q8y11));
-            vector signed short qv20 = vec_add(vec_mule(q4x20, q8y20), vec_mulo(q4x20, q8y20));
-            vector signed short qv21 = vec_add(vec_mule(q4x21, q8y21), vec_mulo(q4x21, q8y21));
-            vector signed short qv30 = vec_add(vec_mule(q4x30, q8y30), vec_mulo(q4x30, q8y30));
-            vector signed short qv31 = vec_add(vec_mule(q4x31, q8y31), vec_mulo(q4x31, q8y31));
-
-            vector signed short vs0 = vec_splat(vscales, 0);
-            vector signed short vs1 = vec_splat(vscales, 1);
-            vector signed short vs2 = vec_splat(vscales, 2);
-            vector signed short vs3 = vec_splat(vscales, 3);
+            vector signed int qv00 = vec_msum(q8y00, q4x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q4x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q4x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q4x11, v0);
+            vector signed int qv20 = vec_msum(q8y20, q4x20, v0);
+            vector signed int qv21 = vec_msum(q8y21, q4x21, v0);
+            vector signed int qv30 = vec_msum(q8y30, q4x30, v0);
+            vector signed int qv31 = vec_msum(q8y31, q4x31, v0);
+
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
+            vector signed int vs2 = vec_splat(vscales_h, 2);
+            vector signed int vs3 = vec_splat(vscales_h, 3);
             vscales = vec_sld(vscales, vscales, 8);
 
-            qv00 = vec_add(qv00, qv10);
-            qv10 = vec_add(qv01, qv11);
-            qv20 = vec_add(qv20, qv30);
-            qv30 = vec_add(qv21, qv31);
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv01, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv20, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv21, vs3), vsumi3);
 
-            vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv10, vs1), vsumi2);
-            vsumi3 = vec_add(vec_mulo(qv10, vs1), vsumi3);
-            vsumi4 = vec_add(vec_mule(qv20, vs2), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv20, vs2), vsumi5);
-            vsumi6 = vec_add(vec_mule(qv30, vs3), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv30, vs3), vsumi7);
+            vsumi0 = vec_add(vec_mul(qv10, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv11, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv30, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv31, vs3), vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -7889,6 +7882,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector unsigned char v1 = vec_splats((unsigned char)0x1);
     const vector unsigned char v2 = vec_splats((unsigned char)0x2);
     const vector unsigned char v3 = vec_splats((unsigned char)0x3);
@@ -7907,18 +7903,27 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
         vector float vdmin = vec_mul(vxmin, vyd);
 
-        memcpy(utmp, x[i].scales, 12);
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
 
-        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;
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
 
         vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
         vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
 
-        vector signed char utmps = (vector signed char)vec_xl( 0, utmp);
         vector signed short vscales = vec_unpackh(utmps);
 
         vector signed short q5xmins = vec_unpackl(utmps);
@@ -7938,10 +7943,10 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh);
         vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].qh);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint8_t * restrict q5 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -7966,10 +7971,10 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             qxhs0 = vec_sr(qxhs0, v2);
             qxhs1 = vec_sr(qxhs1, v2);
 
-            vector signed char q5x00 = vec_or(q5h00, qxs00);
-            vector signed char q5x01 = vec_or(q5h01, qxs01);
-            vector signed char q5x10 = vec_or(q5h10, qxs10);
-            vector signed char q5x11 = vec_or(q5h11, qxs11);
+            vector unsigned char q5x00 = (vector unsigned char)vec_or(q5h00, qxs00);
+            vector unsigned char q5x01 = (vector unsigned char)vec_or(q5h01, qxs01);
+            vector unsigned char q5x10 = (vector unsigned char)vec_or(q5h10, qxs10);
+            vector unsigned char q5x11 = (vector unsigned char)vec_or(q5h11, qxs11);
 
             vector signed char q8y00 = vec_xl( 0, q8);
             vector signed char q8y10 = vec_xl(16, q8);
@@ -7977,22 +7982,20 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed char q8y11 = vec_xl(48, q8);
             q8 += 64;
 
-            vector signed short qv00 = vec_add(vec_mule(q5x00, q8y00), vec_mulo(q5x00, q8y00));
-            vector signed short qv01 = vec_add(vec_mule(q5x01, q8y01), vec_mulo(q5x01, q8y01));
-            vector signed short qv10 = vec_add(vec_mule(q5x10, q8y10), vec_mulo(q5x10, q8y10));
-            vector signed short qv11 = vec_add(vec_mule(q5x11, q8y11), vec_mulo(q5x11, q8y11));
+            vector signed int qv00 = vec_msum(q8y00, q5x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q5x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q5x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q5x11, v0);
 
-            vector signed short vs0 = vec_splat(vscales, 0);
-            vector signed short vs1 = vec_splat(vscales, 1);
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
             vscales = vec_sld(vscales, vscales, 12);
 
-            qv00 = vec_add(qv00, qv10);
-            qv01 = vec_add(qv01, qv11);
-
-            vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv01, vs1), vsumi2);
-            vsumi3 = vec_add(vec_mulo(qv01, vs1), vsumi3);
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv10, vs0), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv01, vs1), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv11, vs1), vsumi3);
         }
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
@@ -8553,6 +8556,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector unsigned char v2 = vec_splats((unsigned char)0x2);
     const vector unsigned char v3 = vec_splats((unsigned char)0x3);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
@@ -8569,14 +8573,14 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
 
         const uint8_t * restrict q6 = x[i].ql;
         const uint8_t * restrict qh = x[i].qh;
@@ -8656,23 +8660,14 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             vector signed short vs6 = vec_splat(vscales, 6);
             vector signed short vs7 = vec_splat(vscales, 7);
 
-            vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv01, vs4), vsumi2);
-            vsumi3 = vec_add(vec_mulo(qv01, vs4), vsumi3);
-            vsumi4 = vec_add(vec_mule(qv10, vs1), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv10, vs1), vsumi5);
-            vsumi6 = vec_add(vec_mule(qv11, vs5), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv11, vs5), vsumi7);
-
-            vsumi0 = vec_add(vec_mule(qv20, vs2), vsumi0);
-            vsumi1 = vec_add(vec_mulo(qv20, vs2), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv21, vs6), vsumi2);
-            vsumi3 = vec_add(vec_mulo(qv21, vs6), vsumi3);
-            vsumi4 = vec_add(vec_mule(qv30, vs3), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv30, vs3), vsumi5);
-            vsumi6 = vec_add(vec_mule(qv31, vs7), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv31, vs7), vsumi7);
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs4, vsumi1);
+            vsumi2 = vec_msum(qv10, vs1, vsumi2);
+            vsumi3 = vec_msum(qv11, vs5, vsumi3);
+            vsumi4 = vec_msum(qv20, vs2, vsumi4);
+            vsumi5 = vec_msum(qv21, vs6, vsumi5);
+            vsumi6 = vec_msum(qv30, vs3, vsumi6);
+            vsumi7 = vec_msum(qv31, vs7, vsumi7);
         }
 
         vsumi0 = vec_add(vsumi0, vsumi4);
@@ -8819,7 +8814,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 #endif
 }
 
-#if defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx)
+#if defined (__AVX__) || defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx)
 static const int8_t keven_signs_q2xs[1024] = {
      1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
      1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
@@ -8952,7 +8947,63 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = 0.125f * hsum_float_8(accumf);
 
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xxs_grid[aux8[3]], iq2xxs_grid[aux8[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]]);
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
 #elif defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -8965,14 +9016,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint16_t * restrict q2 = x[i].qs;
         const int8_t  *  restrict q8 = y[i].qs;
@@ -9019,21 +9066,12 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
             vector signed short vscales01 = vec_splats((int16_t)(2*ls0+1));
             vector signed short vscales23 = vec_splats((int16_t)(2*ls1+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -9307,6 +9345,165 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
     }
 
     *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const __m128i mone = _mm_set1_epi8(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i bit_selector_mask_0 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes);
+    const __m128i bit_selector_mask_1 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes + 1);
+    const __m128i block_sign_shuffle_1_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1);
+    const __m128i block_sign_shuffle_1_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1 + 1);
+    const __m128i block_sign_shuffle_2_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2);
+    const __m128i block_sign_shuffle_2_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2 + 1);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m128i bit_helper_0 = _mm_loadu_si128((const __m128i*)k_bit_helper);
+    const __m128i bit_helper_1 = _mm_loadu_si128((const __m128i*)k_bit_helper + 1);
+    const __m128i m511 = _mm_set1_epi16(511);
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m128i q2_data_0 = _mm_loadu_si128((const __m128i*)q2);
+            const __m128i q2_data_1 = _mm_loadu_si128((const __m128i*)q2 + 1);  q2 += 16;
+            aux_gindex = MM256_SET_M128I(_mm_and_si128(q2_data_1, m511), _mm_and_si128(q2_data_0, m511));
+
+            const __m128i partial_sign_bits_0 = _mm_srli_epi16(q2_data_0, 9);
+            const __m128i partial_sign_bits_1 = _mm_srli_epi16(q2_data_1, 9);
+            const __m128i partial_sign_bits_upper_0 = _mm_srli_epi16(q2_data_0, 13);
+            const __m128i partial_sign_bits_upper_1 = _mm_srli_epi16(q2_data_1, 13);
+            const __m128i partial_sign_bits_for_counting_0 = _mm_xor_si128(partial_sign_bits_0, partial_sign_bits_upper_0);
+            const __m128i partial_sign_bits_for_counting_1 = _mm_xor_si128(partial_sign_bits_1, partial_sign_bits_upper_1);
+
+            const __m128i odd_bits_0 = _mm_shuffle_epi8(bit_helper_0, partial_sign_bits_for_counting_0);
+            const __m128i odd_bits_1 = _mm_shuffle_epi8(bit_helper_1, partial_sign_bits_for_counting_1);
+            const __m128i full_sign_bits_0 = _mm_or_si128(partial_sign_bits_0, odd_bits_0);
+            const __m128i full_sign_bits_1 = _mm_or_si128(partial_sign_bits_1, odd_bits_1);
+
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]]);
+            const __m128i q2_3_0 = _mm_set_epi64x(iq2xs_grid[gindex[9]], iq2xs_grid[gindex[8]]);
+            const __m128i q2_3_1 = _mm_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]]);
+            const __m128i q2_4_0 = _mm_set_epi64x(iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+            const __m128i q2_4_1 = _mm_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]]);
+
+            // AVX2 full_signs_1 is full_sign_bits_0 here
+            // AVX2 full_signs_2 is full_sign_bits_1 here
+            __m128i signs_0, signs_1;
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_3_0 = _mm_sign_epi8(q8_3_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_3_1 = _mm_sign_epi8(q8_3_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_4_0 = _mm_sign_epi8(q8_4_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_4_1 = _mm_sign_epi8(q8_4_1, _mm_or_si128(signs_1, mone));
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const __m128i dot3_0  = _mm_maddubs_epi16(q2_3_0, q8s_3_0);
+            const __m128i dot3_1  = _mm_maddubs_epi16(q2_3_1, q8s_3_1);
+            const __m128i dot4_0  = _mm_maddubs_epi16(q2_4_0, q8s_4_0);
+            const __m128i dot4_1  = _mm_maddubs_epi16(q2_4_1, q8s_4_1);
+
+            __m128i sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0));
+            const __m128i sc1_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc1_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1));
+            const __m128i sc2_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc2_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2));
+            const __m128i sc3_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc3_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3));
+            const __m128i sc4_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc4_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot1_0, sc1_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot1_1, sc1_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot2_0, sc2_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot2_1, sc2_1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot3_0, sc3_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot3_1, sc3_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot4_0, sc4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot4_1, sc4_1));
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
 #elif defined(__loongarch_asx)
 
     const __m256i mone = __lasx_xvreplgr2vr_b(1);
@@ -9425,6 +9622,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 
     *s = 0.125f * hsum_float_8(accumf);
 #elif defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -9437,14 +9635,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint16_t * restrict q2 = x[i].qs;
         const uint8_t  * restrict sc = x[i].scales;
@@ -9492,21 +9686,12 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
             vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
             vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales0), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales1), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales2), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales3), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales0), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales1), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales2), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales3), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -9722,6 +9907,98 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
 
     *s = 0.125f * hsum_float_8(accumf);
 
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    uint64_t aux64;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
+        const __m128i scales16_0 = _mm_cvtepi8_epi16(scales8);
+        const __m128i scales16_1 = _mm_cvtepi8_epi16(_mm_srli_si128(scales8, 8));
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                  iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                  iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                  iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                  iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)]);
+            qs += 8;
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 0)));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 1)));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 0)));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 1)));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
 #elif defined(__POWER9_VECTOR__)
     static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
                                         0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
@@ -9729,6 +10006,8 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
 
     static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
 
+    const vector int v0 = vec_splats((int32_t)0);
+
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -9743,14 +10022,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint8_t *  restrict q2 = x[i].qs;
         const uint8_t *  restrict qh = x[i].qh;
@@ -9810,21 +10085,12 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
             vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
             vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales0), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales1), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales2), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales3), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales0), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales1), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales2), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales3), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -10059,9 +10325,68 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
 
     *s = 0.25f * hsum_float_8(accumf);
 
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            const __m128i q2_2_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
 #elif defined(__POWER9_VECTOR__)
     const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
 
+    const vector int v0 = vec_splats((int32_t)0);
+
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -10072,14 +10397,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         const uint8_t * restrict q3 = x[i].qs;
         const uint32_t * restrict signs = (const uint32_t *)(x[i].qs + QK_K/4);
@@ -10124,21 +10445,12 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
             vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
             vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -10421,6 +10733,112 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 
     *s = hsum_float_8(accumf);
 
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    const __m128i idx_mul_0 = _mm_set_epi32(32, 64, 128, 256);
+    const __m128i idx_mul_1 = _mm_set_epi32(2, 4, 8, 16);
+    const __m128i idx_mask  = _mm_set1_epi32(256);
+
+    typedef union {
+        __m128i  vec[4];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i qs_tmp = _mm_loadu_si128((const __m128i *)qs);
+            const __m128i idx_l_0 = _mm_cvtepu8_epi16(qs_tmp);
+            const __m128i idx_l_1 = _mm_cvtepu8_epi16(_mm_srli_si128(qs_tmp, 8)); qs += 16;
+            idx.vec[0] = _mm_set1_epi32(qh[ib32+0]);
+            idx.vec[1] = idx.vec[0];
+            idx.vec[2] = _mm_set1_epi32(qh[ib32+1]);
+            idx.vec[3] = idx.vec[2];
+
+            idx.vec[0] = _mm_and_si128(_mm_mullo_epi32(idx.vec[0], idx_mul_0), idx_mask);
+            idx.vec[1] = _mm_and_si128(_mm_mullo_epi32(idx.vec[1], idx_mul_1), idx_mask);
+            idx.vec[2] = _mm_and_si128(_mm_mullo_epi32(idx.vec[2], idx_mul_0), idx_mask);
+            idx.vec[3] = _mm_and_si128(_mm_mullo_epi32(idx.vec[3], idx_mul_1), idx_mask);
+
+            idx.vec[0] = _mm_or_si128(idx.vec[0], _mm_cvtepi16_epi32(idx_l_0));
+            idx.vec[1] = _mm_or_si128(idx.vec[1], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_0, 8)));
+            idx.vec[2] = _mm_or_si128(idx.vec[2], _mm_cvtepi16_epi32(idx_l_1));
+            idx.vec[3] = _mm_or_si128(idx.vec[3], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_1, 8)));
+
+            const __m128i q2_1_0 = _mm_set_epi32(iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]]);
+            const __m128i q2_2_0 = _mm_set_epi32(iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[9]], iq3s_grid[idx.index[8]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]]);
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | (signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | (signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = hsum_float_8(accumf);
+
 #elif defined(__POWER9_VECTOR__)
     static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
                                         0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
@@ -10428,6 +10846,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 
     static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
 
+    const vector int v0 = vec_splats((int32_t)0);
+
     vector float vsumf0 = vec_splats(0.0f);
     vector float vsumf1 = vec_splats(0.0f);
     vector float vsumf2 = vec_splats(0.0f);
@@ -10448,14 +10868,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
         const uint8_t *  restrict sc = x[i].scales;
         const int8_t  *  restrict q8 = y[i].qs;
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         for (int j = 0; j < QK_K/32; j += 2) {
             __builtin_prefetch(q3, 0, 1);
@@ -10509,21 +10925,12 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
             vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
             vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -10669,6 +11076,14 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
 }
 
 
+#if defined(__AVX__)
+static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
+    const __m128i ax = _mm_sign_epi8(x, x);
+    const __m128i sy = _mm_sign_epi8(y, x);
+    return _mm_maddubs_epi16(ax, sy);
+}
+#endif
+
 #if defined(__AVX2__)
 static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
     const __m256i ax = _mm256_sign_epi8(x, x);
@@ -10786,6 +11201,54 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
 
     *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
 
+#elif defined __AVX__
+    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)]);
+            qs += 8;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(ls1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(ls1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(ls2));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(ls2));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
+        accum1 += d * sumi1;
+
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
 #elif defined(__POWER9_VECTOR__)
     const vector unsigned char v0 = vec_splats((unsigned char)0x0);
     const vector unsigned short vsign = vec_splats((unsigned short)0x8000);
@@ -10804,10 +11267,6 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
         vector signed int vsumi1 = vec_splats((int32_t)0);
         vector signed int vsumi2 = vec_splats((int32_t)0);
         vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
         vector signed int vsumi8 = vec_splats((int32_t)0);
 
         const uint8_t  * restrict q1 = x[i].qs;
@@ -10849,14 +11308,10 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
             vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
             vector signed short vscales = vec_sld(vscales23, vscales01, 8);
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
 
             vector signed short q8ysums = vec_xl_len(qs, 8);
             qs += 4;
@@ -10871,11 +11326,6 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
             vsumi8 = vec_add(vec_mule(q8ysum, vscales), vsumi8);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -11137,6 +11587,92 @@ void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void
 
     *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
 
+#elif defined __AVX__
+    const __m128i mask = _mm_set1_epi16(0x7);
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)]);
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+
+            const __m128i delta1_0 = _mm_set_epi64x(qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta1_1 = _mm_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_0 = _mm_set_epi64x(qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_1 = _mm_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+            const __m128i dot3_0 = mul_add_epi8_sse(delta1_0, q8b_1_0);
+            const __m128i dot3_1 = mul_add_epi8_sse(delta1_1, q8b_1_1);
+            const __m128i dot4_0 = mul_add_epi8_sse(delta2_0, q8b_2_0);
+            const __m128i dot4_1 = mul_add_epi8_sse(delta2_1, q8b_2_1);
+
+            __m128i scale1_0 = _mm_set1_epi16(sc[ib/2] >> 0);
+            __m128i scale1_1 = _mm_set1_epi16(sc[ib/2] >> 3);
+            __m128i scale2_0 = _mm_set1_epi16(sc[ib/2] >> 6);
+            __m128i scale2_1 = _mm_set1_epi16(sc[ib/2] >> 9);
+
+            scale1_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_0, mask), 1), mone);
+            scale1_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_1, mask), 1), mone);
+            scale2_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_0, mask), 1), mone);
+            scale2_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_1, mask), 1), mone);
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, scale1_0);
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, scale1_1);
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, scale2_0);
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, scale2_1);
+            const __m128i p3_0 = _mm_madd_epi16(dot3_0, scale1_0);
+            const __m128i p3_1 = _mm_madd_epi16(dot3_1, scale1_1);
+            const __m128i p4_0 = _mm_madd_epi16(dot4_0, scale2_0);
+            const __m128i p4_1 = _mm_madd_epi16(dot4_1, scale2_1);
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_add_epi32(p3_0, p4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_add_epi32(p3_1, p4_1));
+
+            qs += 8; qh += 4;
+        }
+
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+
+        accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
 #else
 
     int sum1[2], sum2[2], delta[4];
@@ -11267,8 +11803,47 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
 
     *s = hsum_float_8(_mm256_add_ps(accum1, accum2));
 
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int ib = 0; ib < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+        const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+        const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+        const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+        const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+        const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+        const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+        const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+        accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[0].d)*GGML_FP16_TO_FP32(x[0].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[1].d)*GGML_FP16_TO_FP32(x[1].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
+
+        y += 2;
+        x += 2;
+    }
+
+    *s = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
@@ -11299,8 +11874,11 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
         vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
         vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
 
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
-        vector signed int vsumi1 = vec_add(vec_unpackh(qv1), vec_unpackl(qv1));
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
 
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
@@ -11453,8 +12031,57 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 
     *s = hsum_float_8(accum);
 
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    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;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+            const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+            const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+            const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+            const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+            const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+            const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+            const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+            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;
+            const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, _mm_set1_epi16(ls1));
+            const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, _mm_set1_epi16(ls1));
+            const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, _mm_set1_epi16(ls2));
+            const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, _mm_set1_epi16(ls2));
+            sumi1_0 = _mm_add_epi32(p_1_0, sumi1_0);
+            sumi1_1 = _mm_add_epi32(p_1_1, sumi1_1);
+            sumi2_0 = _mm_add_epi32(p_2_0, sumi2_0);
+            sumi2_1 = _mm_add_epi32(p_2_1, sumi2_1);
+        }
+        __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
+        __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
 #elif defined(__POWER9_VECTOR__)
     const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
     const vector unsigned char v4 = vec_splats((unsigned char)0x4);
 
     vector float vsumf0 = vec_splats(0.0f);
@@ -11470,14 +12097,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
         vector float vyd = vec_splats(y[ibl].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi4 = vec_splats((int32_t)0);
-        vector signed int vsumi5 = vec_splats((int32_t)0);
-        vector signed int vsumi6 = vec_splats((int32_t)0);
-        vector signed int vsumi7 = vec_splats((int32_t)0);
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
 
         uint16_t h = x[ibl].scales_h;
 
@@ -11522,21 +12145,12 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
             vector signed short vscales01 = vec_splats((int16_t)ls0);
             vector signed short vscales23 = vec_splats((int16_t)ls1);
 
-            vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0);
-            vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1);
-            vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2);
-            vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3);
-            vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4);
-            vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5);
-            vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6);
-            vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7);
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
         }
 
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
         vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
         vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
         vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
@@ -13139,7 +13753,7 @@ static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const u
         const float * restrict xval, const float * restrict weight, float * scale, int8_t * restrict L, int ngrid) {
     int num_neighbors = neighbours[0];
     GGML_ASSERT(num_neighbors > 0);
-    float best_score = 0;
+    float best_score = -FLT_MAX;
     int grid_index = -1;
     for (int j = 1; j <= num_neighbors; ++j) {
         const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
@@ -13337,7 +13951,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
                     sumw[j+1] = sumw[j] + weight[i];
                 }
             }
-            float best_score = 0, scale = max;
+            float best_score = -FLT_MIN, scale = max;
             int besti1 = -1, besti2 = -1, best_shift = 0;
             for (int i1 = 0; i1 <= block_size; ++i1) {
                 for (int i2 = i1; i2 <= block_size; ++i2) {
@@ -13513,7 +14127,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
                 idx[2*j] = j;
             }
             qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
-            float best_score = 0, scale = max;
+            float best_score = -FLT_MIN, scale = max;
             int besti1 = -1, besti2 = -1, best_k = -1;
             // 0: +, +
             // 1: +, -