Faster R4 quants on Zen4 (#139)

* q3_k_r4: faster Zen4

* q3_k_r4: faster Zen4

256.2 -> 272.7 t/s for PP-512

* q6_k_r4: faster Zen4

243.2 -> 261.3 t/s for PP-512

* q4_k_r4: slightly faster Zen4

262.4 t/s -> 268.1 t/s

* q5_k_r4: slightly faster Zen4

248.3 t/s -> 256.7 t/s

* iq4_xs_r4: slightly faster Zen4

256.8 t/s -> 272.0 t/s

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

1 files changed, 68 insertions, 51 deletions

diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
index 84a4b619..3f448275 100644
--- a/ggml/src/iqk/iqk_mul_mat.cpp
+++ b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -3011,7 +3011,7 @@ static void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const Data
     using helper_t = union { __m512i vec; uint32_t val[16]; };
     helper_t h;
     __m512  acc[nrc_y] = {};
-    __m512  d4s[nrc_y];
+    __m512i isum[nrc_y] = {};
     __m512i qx[4];
     for (int ix = 0; ix < nrc_x; ix += 8) {
         const block_iq4_xs_r4 * iq4l = (const block_iq4_xs_r4 *)((const char *)vx + (ix+0)*bx);
@@ -3020,10 +3020,7 @@ static void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const Data
             auto dl = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq4l[ibl].d));
             auto dh = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq4h[ibl].d));
             auto d4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_set_m128(dl, dl)), _mm256_set_m128(dh, dh), 1);
-            for (int iy = 0; iy < nrc_y; ++iy) {
-                d4s[iy] = _mm512_mul_ps(d4, _mm512_set1_ps(q8.scale(iy, ibl)));
-            }
-            d4 = _mm512_mul_ps(d4, _mm512_set1_ps(-64.f));
+            auto d4x64 = _mm512_mul_ps(d4, _mm512_set1_ps(-64.f));
             auto slbits_l = _mm_loadu_si128((const __m128i *)iq4l[ibl].scales_l);
             auto shbits_l = _mm_loadu_si128((const __m128i *)iq4h[ibl].scales_l);
             auto sl_l = MM256_SET_M128I(_mm_srli_epi16(slbits_l, 4), slbits_l);
@@ -3040,7 +3037,7 @@ static void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const Data
             for (int ib = 0; ib < QK_K/32; ++ib) {
                 auto iscales = _mm512_cvtepi8_epi32(_mm_blend_epi32(_mm_set1_epi32(h.val[ib+0]), _mm_set1_epi32(h.val[ib+8]), 0x0c));
                 auto scales  = _mm512_cvtepi32_ps(iscales);
-                auto scales_m = _mm512_mul_ps(scales, d4);
+                auto scales_m = _mm512_mul_ps(scales, d4x64);
                 auto bits1 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l[ibl].qs+2*ib+0)),
                                                                        _mm256_loadu_si256((const __m256i *)iq4h[ibl].qs+2*ib+0), 1);
                 auto bits2 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l[ibl].qs+2*ib+1)),
@@ -3057,11 +3054,15 @@ static void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const Data
                     sumi = _mm512_dpbusd_epi32(sumi, qx[1], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x55)));
                     sumi = _mm512_dpbusd_epi32(sumi, qx[2], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xaa)));
                     sumi = _mm512_dpbusd_epi32(sumi, qx[3], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xff)));
+                    isum[iy] = _mm512_add_epi32(isum[iy], _mm512_mullo_epi32(iscales, sumi));
                     float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib];
-                    acc[iy] = _mm512_fmadd_ps(_mm512_mul_ps(scales, d4s[iy]), _mm512_cvtepi32_ps(sumi), acc[iy]);
                     acc[iy] = _mm512_fmadd_ps(scales_m, _mm512_set1_ps(m8), acc[iy]);
                 }
             }
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                acc[iy] = _mm512_fmadd_ps(_mm512_mul_ps(d4, _mm512_set1_ps(q8.scale(iy, ibl))), _mm512_cvtepi32_ps(isum[iy]), acc[iy]);
+                isum[iy] = _mm512_setzero_si512();
+            }
         }
         for (int iy = 0; iy < nrc_y; ++iy) {
             auto sum1 = _mm_add_ps(_mm512_extractf32x4_ps(acc[iy], 0), _mm512_extractf32x4_ps(acc[iy], 1));
@@ -3176,7 +3177,7 @@ static void mul_mat_q4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
     using helper_t = union { __m512i vec; uint32_t val[16]; };
     helper_t hd, hm;
     __m512  acc[nrc_y] = {};
-    __m512  d4s[nrc_y];
+    __m512i isum[nrc_y] = {};
     __m512i qx[4];
     for (int ix = 0; ix < nrc_x; ix += 8) {
         const block_q4_k_r4 * iq4l = (const block_q4_k_r4 *)((const char *)vx + (ix+0)*bx);
@@ -3189,9 +3190,6 @@ static void mul_mat_q4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
             auto dh = _mm256_castps256_ps128(d2);
             auto mh = _mm256_extractf128_ps(d2, 1);
             auto d4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_set_m128(dl, dl)), _mm256_set_m128(dh, dh), 1);
-            for (int iy = 0; iy < nrc_y; ++iy) {
-                d4s[iy] = _mm512_mul_ps(d4, _mm512_set1_ps(q8.scale(iy, ibl)));
-            }
             auto m4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_set_m128(ml, ml)), _mm256_set_m128(mh, mh), 1);
             m4 = _mm512_mul_ps(m4, _mm512_set1_ps(-0.5f));
             auto slbits_l = _mm256_loadu_si256((const __m256i *)iq4l[ibl].scales_l);
@@ -3212,11 +3210,10 @@ static void mul_mat_q4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                 auto scales1 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hd.val[ib+0]));
                 auto scales2 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hd.val[ib+8]));
                 auto iscales = _mm512_inserti32x8(_mm512_castsi256_si512(scales1), scales2, 1);
-                auto scales  = _mm512_cvtepi32_ps(iscales);
                 scales1 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hm.val[ib+0]));
                 scales2 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hm.val[ib+8]));
-                iscales = _mm512_inserti32x8(_mm512_castsi256_si512(scales1), scales2, 1);
-                auto scales_m  = _mm512_mul_ps(m4, _mm512_cvtepi32_ps(iscales));
+                auto iscales_m = _mm512_inserti32x8(_mm512_castsi256_si512(scales1), scales2, 1);
+                auto scales_m  = _mm512_mul_ps(m4, _mm512_cvtepi32_ps(iscales_m));
                 auto bits1 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l[ibl].qs+2*ib+0)),
                                                                        _mm256_loadu_si256((const __m256i *)iq4h[ibl].qs+2*ib+0), 1);
                 auto bits2 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l[ibl].qs+2*ib+1)),
@@ -3233,11 +3230,15 @@ static void mul_mat_q4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                     sumi = _mm512_dpbusd_epi32(sumi, qx[1], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x55)));
                     sumi = _mm512_dpbusd_epi32(sumi, qx[2], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xaa)));
                     sumi = _mm512_dpbusd_epi32(sumi, qx[3], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xff)));
-                    acc[iy] = _mm512_fmadd_ps(_mm512_mul_ps(scales, d4s[iy]), _mm512_cvtepi32_ps(sumi), acc[iy]);
+                    isum[iy] = _mm512_add_epi32(isum[iy], _mm512_mullo_epi32(iscales, sumi));
                     float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib];
                     acc[iy] = _mm512_fmadd_ps(scales_m, _mm512_set1_ps(m8), acc[iy]);
                 }
             }
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                acc[iy] = _mm512_fmadd_ps(_mm512_mul_ps(d4, _mm512_set1_ps(q8.scale(iy, ibl))), _mm512_cvtepi32_ps(isum[iy]), acc[iy]);
+                isum[iy] = _mm512_setzero_si512();
+            }
         }
         for (int iy = 0; iy < nrc_y; ++iy) {
             auto sum1 = _mm_add_ps(_mm512_extractf32x4_ps(acc[iy], 0), _mm512_extractf32x4_ps(acc[iy], 1));
@@ -3355,7 +3356,7 @@ static void mul_mat_q5_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
     using helper_t = union { __m512i vec; uint32_t val[16]; };
     helper_t hd, hm;
     __m512  acc[nrc_y] = {};
-    __m512  d4s[nrc_y];
+    __m512i isum[nrc_y] = {};
     __m512i qx[4];
     for (int ix = 0; ix < nrc_x; ix += 8) {
         const block_q5_k_r4 * iq5l = (const block_q5_k_r4 *)((const char *)vx + (ix+0)*bx);
@@ -3368,9 +3369,6 @@ static void mul_mat_q5_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
             auto dh = _mm256_castps256_ps128(d2);
             auto mh = _mm256_extractf128_ps(d2, 1);
             auto d4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_set_m128(dl, dl)), _mm256_set_m128(dh, dh), 1);
-            for (int iy = 0; iy < nrc_y; ++iy) {
-                d4s[iy] = _mm512_mul_ps(d4, _mm512_set1_ps(q8.scale(iy, ibl)));
-            }
             auto m4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_set_m128(ml, ml)), _mm256_set_m128(mh, mh), 1);
             m4 = _mm512_mul_ps(m4, _mm512_set1_ps(-0.5f));
             auto slbits_l = _mm256_loadu_si256((const __m256i *)iq5l[ibl].scales_l);
@@ -3391,11 +3389,10 @@ static void mul_mat_q5_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                 auto scales1 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hd.val[ib+0]));
                 auto scales2 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hd.val[ib+8]));
                 auto iscales = _mm512_inserti32x8(_mm512_castsi256_si512(scales1), scales2, 1);
-                auto scales  = _mm512_cvtepi32_ps(iscales);
                 scales1 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hm.val[ib+0]));
                 scales2 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hm.val[ib+8]));
-                iscales = _mm512_inserti32x8(_mm512_castsi256_si512(scales1), scales2, 1);
-                auto scales_m  = _mm512_mul_ps(m4, _mm512_cvtepi32_ps(iscales));
+                auto iscales_m = _mm512_inserti32x8(_mm512_castsi256_si512(scales1), scales2, 1);
+                auto scales_m  = _mm512_mul_ps(m4, _mm512_cvtepi32_ps(iscales_m));
                 auto lbits1 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq5l[ibl].qs+2*ib+0)),
                                                                         _mm256_loadu_si256((const __m256i *)iq5h[ibl].qs+2*ib+0), 1);
                 auto lbits2 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq5l[ibl].qs+2*ib+1)),
@@ -3417,11 +3414,15 @@ static void mul_mat_q5_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                     sumi = _mm512_dpbusd_epi32(sumi, qx[1], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x55)));
                     sumi = _mm512_dpbusd_epi32(sumi, qx[2], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xaa)));
                     sumi = _mm512_dpbusd_epi32(sumi, qx[3], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xff)));
-                    acc[iy] = _mm512_fmadd_ps(_mm512_mul_ps(scales, d4s[iy]), _mm512_cvtepi32_ps(sumi), acc[iy]);
+                    isum[iy] = _mm512_add_epi32(isum[iy], _mm512_mullo_epi32(iscales, sumi));
                     float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib];
                     acc[iy] = _mm512_fmadd_ps(scales_m, _mm512_set1_ps(m8), acc[iy]);
                 }
             }
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                acc[iy] = _mm512_fmadd_ps(_mm512_mul_ps(d4, _mm512_set1_ps(q8.scale(iy, ibl))), _mm512_cvtepi32_ps(isum[iy]), acc[iy]);
+                isum[iy] = _mm512_setzero_si512();
+            }
         }
         for (int iy = 0; iy < nrc_y; ++iy) {
             auto sum1 = _mm_add_ps(_mm512_extractf32x4_ps(acc[iy], 0), _mm512_extractf32x4_ps(acc[iy], 1));
@@ -3449,7 +3450,7 @@ static void mul_mat_q2_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
     static const uint8_t k_shuff[32] = {0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
     auto shuff = _mm256_loadu_si256((const __m256i *)k_shuff);
 #ifdef HAVE_FANCY_SIMD
-    __m256 d4s[nrc_y];
+    __m256i isum[nrc_y] = {};
 #else
     auto m1 = _mm256_set1_epi16(1);
 #endif
@@ -3487,7 +3488,6 @@ static void mul_mat_q2_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                     sumi = _mm256_dpwssd_epi32(sumi, s4, _mm256_shuffle_epi32(bsums, 0xff));
                     auto d8 = _mm256_set1_ps(q8.scale(iy, ibl));
                     acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(m4, d8), _mm256_cvtepi32_ps(sumi), acc[iy]);
-                    d4s[iy] = _mm256_mul_ps(d4, d8);
 #else
                     sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s1, _mm256_shuffle_epi32(bsums, 0x00)));
                     sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s2, _mm256_shuffle_epi32(bsums, 0x55)));
@@ -3507,9 +3507,7 @@ static void mul_mat_q2_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
             _mm256_storeu_si256((__m256i *)scales+1, all_scales2);
             for (int ib = 0; ib < QK_K/32; ++ib) {
                 auto iscales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(scales + 8*ib)));
-#ifdef HAVE_FANCY_SIMD
-                auto scales  = _mm256_cvtepi32_ps(iscales);
-#else
+#ifndef HAVE_FANCY_SIMD
                 auto scales  = _mm256_mul_ps(d4, _mm256_cvtepi32_ps(iscales));
 #endif
                 auto lb = _mm256_loadu_si256((const __m256i *)iq2[ibl].qs+ib);
@@ -3525,7 +3523,7 @@ static void mul_mat_q2_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                     sumi = _mm256_dpbusd_epi32(sumi, qx[1], _mm256_shuffle_epi32(y, 0x55));
                     sumi = _mm256_dpbusd_epi32(sumi, qx[2], _mm256_shuffle_epi32(y, 0xaa));
                     sumi = _mm256_dpbusd_epi32(sumi, qx[3], _mm256_shuffle_epi32(y, 0xff));
-                    acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, d4s[iy]), _mm256_cvtepi32_ps(sumi), acc[iy]);
+                    isum[iy] = _mm256_add_epi32(isum[iy], _mm256_mullo_epi32(iscales, sumi));
 #else
                     auto sumi1 = _mm256_add_epi16(_mm256_maddubs_epi16(qx[0], _mm256_shuffle_epi32(y, 0x00)),
                                                   _mm256_maddubs_epi16(qx[1], _mm256_shuffle_epi32(y, 0x55)));
@@ -3541,6 +3539,13 @@ static void mul_mat_q2_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
 #endif
                 }
             }
+#ifdef HAVE_FANCY_SIMD
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                auto d4y = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl)));
+                acc[iy] = _mm256_fmadd_ps(d4y, _mm256_cvtepi32_ps(isum[iy]), acc[iy]);
+                isum[iy] = _mm256_setzero_si256();
+            }
+#endif
         }
         for (int iy = 0; iy < nrc_y; ++iy) {
             auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));
@@ -3562,7 +3567,7 @@ static void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
     static const uint8_t k_shuff[32] = {0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
     auto shuff = _mm256_loadu_si256((const __m256i *)k_shuff);
 #ifdef HAVE_FANCY_SIMD
-    __m256 d4s[nrc_y];
+    __m256i isum[nrc_y];
 #else
     auto m1 = _mm256_set1_epi16(1);
 #endif
@@ -3575,11 +3580,7 @@ static void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
         for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256
             auto dl = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq3[ibl].d));
             auto d4 = _mm256_set_m128(dl, dl);
-#ifdef HAVE_FANCY_SIMD
-            for (int iy = 0; iy < nrc_y; ++iy) {
-                d4s[iy] = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl)));
-            }
-#else
+#ifndef HAVE_FANCY_SIMD
             if constexpr (nrc_y == 1) {
                 d4 = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(0, ibl)));
             }
@@ -3603,6 +3604,12 @@ static void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                 auto s2 = MM256_SET_M128I(_mm256_extracti128_si256(t3, 1), _mm256_extracti128_si256(t1, 1)); // blocks 2, 3, 10, 11
                 auto s3 = MM256_SET_M128I(_mm256_extracti128_si256(t4, 0), _mm256_extracti128_si256(t2, 0)); // blocks 4, 5, 12, 13
                 auto s4 = MM256_SET_M128I(_mm256_extracti128_si256(t4, 1), _mm256_extracti128_si256(t2, 1)); // blocks 6, 7, 14, 15
+#ifdef HAVE_FANCY_SIMD
+                s1 = _mm256_mullo_epi16(s1, _mm256_set1_epi16(-4));
+                s2 = _mm256_mullo_epi16(s2, _mm256_set1_epi16(-4));
+                s3 = _mm256_mullo_epi16(s3, _mm256_set1_epi16(-4));
+                s4 = _mm256_mullo_epi16(s4, _mm256_set1_epi16(-4));
+#endif
                 for (int iy = 0; iy < nrc_y; ++iy) {
                     auto bsums = q8.load_bsums(iy, ibl);
                     auto sumi = _mm256_setzero_si256();
@@ -3611,7 +3618,7 @@ static void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                     sumi = _mm256_dpwssd_epi32(sumi, s2, _mm256_shuffle_epi32(bsums, 0x55));
                     sumi = _mm256_dpwssd_epi32(sumi, s3, _mm256_shuffle_epi32(bsums, 0xaa));
                     sumi = _mm256_dpwssd_epi32(sumi, s4, _mm256_shuffle_epi32(bsums, 0xff));
-                    acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4s[iy], _mm256_set1_ps(-4.f)), _mm256_cvtepi32_ps(sumi), acc[iy]);
+                    isum[iy] = sumi;
 #else
                     sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s1, _mm256_shuffle_epi32(bsums, 0x00)));
                     sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s2, _mm256_shuffle_epi32(bsums, 0x55)));
@@ -3627,9 +3634,7 @@ static void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
             }
             for (int ib = 0; ib < QK_K/32; ++ib) {
                 auto iscales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(scales + 8*ib)));
-#ifdef HAVE_FANCY_SIMD
-                auto scales  = _mm256_cvtepi32_ps(iscales);
-#else
+#ifndef HAVE_FANCY_SIMD
                 auto scales  = _mm256_mul_ps(d4, _mm256_cvtepi32_ps(iscales));
 #endif
                 auto lb = _mm256_loadu_si256((const __m256i *)iq3[ibl].qs+ib);
@@ -3647,7 +3652,7 @@ static void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                     sumi = _mm256_dpbusd_epi32(sumi, qx[1], _mm256_shuffle_epi32(y, 0x55));
                     sumi = _mm256_dpbusd_epi32(sumi, qx[2], _mm256_shuffle_epi32(y, 0xaa));
                     sumi = _mm256_dpbusd_epi32(sumi, qx[3], _mm256_shuffle_epi32(y, 0xff));
-                    acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, d4s[iy]), _mm256_cvtepi32_ps(sumi), acc[iy]);
+                    isum[iy] = _mm256_add_epi32(isum[iy], _mm256_mullo_epi32(iscales, sumi));
 #else
                     auto sumi1 = _mm256_add_epi16(_mm256_maddubs_epi16(qx[0], _mm256_shuffle_epi32(y, 0x00)),
                                                   _mm256_maddubs_epi16(qx[1], _mm256_shuffle_epi32(y, 0x55)));
@@ -3664,6 +3669,12 @@ static void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
 
                 }
             }
+#ifdef HAVE_FANCY_SIMD
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                auto d4y = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl)));
+                acc[iy] = _mm256_fmadd_ps(d4y, _mm256_cvtepi32_ps(isum[iy]), acc[iy]);
+            }
+#endif
         }
         for (int iy = 0; iy < nrc_y; ++iy) {
             auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));
@@ -3682,7 +3693,7 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
     static const uint8_t k_shuff[32] = {0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
     auto shuff = _mm256_loadu_si256((const __m256i *)k_shuff);
 #ifdef HAVE_FANCY_SIMD
-    __m256 d4s[nrc_y];
+    __m256i isum[nrc_y];
 #else
     auto m1 = _mm256_set1_epi16(1);
 #endif
@@ -3694,11 +3705,7 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
         for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256
             auto dl = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq6[ibl].d));
             auto d4 = _mm256_set_m128(dl, dl);
-#ifdef HAVE_FANCY_SIMD
-            for (int iy = 0; iy < nrc_y; ++iy) {
-                d4s[iy] = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl)));
-            }
-#else
+#ifndef HAVE_FANCY_SIMD
             if constexpr (nrc_y == 1) {
                 d4 = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(0, ibl)));
             }
@@ -3715,6 +3722,12 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                 auto s2 = MM256_SET_M128I(_mm256_extracti128_si256(t3, 1), _mm256_extracti128_si256(t1, 1)); // blocks 2, 3, 10, 11
                 auto s3 = MM256_SET_M128I(_mm256_extracti128_si256(t4, 0), _mm256_extracti128_si256(t2, 0)); // blocks 4, 5, 12, 13
                 auto s4 = MM256_SET_M128I(_mm256_extracti128_si256(t4, 1), _mm256_extracti128_si256(t2, 1)); // blocks 6, 7, 14, 15
+#ifdef HAVE_FANCY_SIMD
+                s1 = _mm256_mullo_epi16(s1, _mm256_set1_epi16(-32));
+                s2 = _mm256_mullo_epi16(s2, _mm256_set1_epi16(-32));
+                s3 = _mm256_mullo_epi16(s3, _mm256_set1_epi16(-32));
+                s4 = _mm256_mullo_epi16(s4, _mm256_set1_epi16(-32));
+#endif
                 for (int iy = 0; iy < nrc_y; ++iy) {
                     auto bsums = q8.load_bsums(iy, ibl);
                     auto sumi = _mm256_setzero_si256();
@@ -3723,7 +3736,7 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                     sumi = _mm256_dpwssd_epi32(sumi, s2, _mm256_shuffle_epi32(bsums, 0x55));
                     sumi = _mm256_dpwssd_epi32(sumi, s3, _mm256_shuffle_epi32(bsums, 0xaa));
                     sumi = _mm256_dpwssd_epi32(sumi, s4, _mm256_shuffle_epi32(bsums, 0xff));
-                    acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4s[iy], _mm256_set1_ps(-32.f)), _mm256_cvtepi32_ps(sumi), acc[iy]);
+                    isum[iy] = sumi;
 #else
                     sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s1, _mm256_shuffle_epi32(bsums, 0x00)));
                     sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s2, _mm256_shuffle_epi32(bsums, 0x55)));
@@ -3740,9 +3753,7 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
             const uint32_t * scales = (const uint32_t *)iq6[ibl].scales;
             for (int ib = 0; ib < QK_K/32; ++ib) {
                 auto iscales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(scales + 2*ib)));
-#ifdef HAVE_FANCY_SIMD
-                auto scales  = _mm256_cvtepi32_ps(iscales);
-#else
+#ifndef HAVE_FANCY_SIMD
                 auto scales  = _mm256_mul_ps(d4, _mm256_cvtepi32_ps(iscales));
 #endif
                 auto lbits1 = _mm256_loadu_si256((const __m256i *)iq6[ibl].ql+2*ib+0);
@@ -3760,7 +3771,7 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
                     sumi = _mm256_dpbusd_epi32(sumi, qx[1], _mm256_shuffle_epi32(y, 0x55));
                     sumi = _mm256_dpbusd_epi32(sumi, qx[2], _mm256_shuffle_epi32(y, 0xaa));
                     sumi = _mm256_dpbusd_epi32(sumi, qx[3], _mm256_shuffle_epi32(y, 0xff));
-                    acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, d4s[iy]), _mm256_cvtepi32_ps(sumi), acc[iy]);
+                    isum[iy] = _mm256_add_epi32(isum[iy], _mm256_mullo_epi32(iscales, sumi));
 #else
                     auto sumi1 = _mm256_add_epi16(_mm256_maddubs_epi16(qx[0], _mm256_shuffle_epi32(y, 0x00)),
                                                   _mm256_maddubs_epi16(qx[1], _mm256_shuffle_epi32(y, 0x55)));
@@ -3776,6 +3787,12 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
 #endif
                 }
             }
+#ifdef HAVE_FANCY_SIMD
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                auto d4y = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl)));
+                acc[iy] = _mm256_fmadd_ps(d4y, _mm256_cvtepi32_ps(isum[iy]), acc[iy]);
+            }
+#endif
         }
         for (int iy = 0; iy < nrc_y; ++iy) {
             auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));