diff --git a/StgSharp.Native/sn-src/mk.fma.x86.c b/StgSharp.Native/sn-src/mk.fma.x86.c
index 36cb31c8..c5a0459f 100644
--- a/StgSharp.Native/sn-src/mk.fma.x86.c
+++ b/StgSharp.Native/sn-src/mk.fma.x86.c
@@ -107,34 +107,36 @@ SN_MK_PROC_DECL_STD(float, avx, , fma)
 SN_MK_PROC_DECL_STD(float, avx, _fma, fma)
 {
         __mk_param_std(TILE, LEFT_RIGHT_ANS, float);
+        SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y);
+        register __m256 c0 = _mm256_setzero_ps();
+        register __m256 c1 = _mm256_setzero_ps();
+        register __m256 c2 = _mm256_setzero_ps();
+        register __m256 c3 = _mm256_setzero_ps();
 
-        SN_MAT_KERNEL_LOCAL *restrict ans_cur =
-                ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe
+        for (size_t z = 0; z < common_z; z+=2) {
+                const SN_MAT_KERNEL_LOCAL *a_base = left + z;
+                const SN_MAT_KERNEL_LOCAL *b_base = right + z;
 
-        register __m256 c01 = _mm256_load_ps((float *)&ans_cur->y[0]);
-        register __m256 c23 = _mm256_load_ps((float *)&ans_cur->y[1]);
+                register __m256 b0 = _mm256_load_ps((float *)&b_base[0].z[0]);
+                register __m256 b1 = _mm256_load_ps((float *)&b_base[1].z[0]);
+                register __m256 b2 = _mm256_load_ps((float *)&b_base[2].z[0]);
+                register __m256 b3 = _mm256_load_ps((float *)&b_base[3].z[0]);
 
-        for (size_t z = 0; z < common_z; z++) {
-                SN_MAT_KERNEL_LOCAL const *right_cur = right + GET_INDEX(common_z, x, z);
-                SN_MAT_KERNEL_LOCAL const *left_cur = left + GET_INDEX(common_y, z, y);
-                register __m256 b01 = _mm256_load_ps((float *)&right_cur->y[0]);
-                register __m256 b23 = _mm256_load_ps((float *)&right_cur->y[1]);
+                register __m256 a0 = _mm256_load_ps((float *)&a_base[0].z[0]);
+                register __m256 a1 = _mm256_load_ps((float *)&a_base[1].z[0]);
+                register __m256 a2 = _mm256_load_ps((float *)&a_base[2].z[0]);
+                register __m256 a3 = _mm256_load_ps((float *)&a_base[3].z[0]);
+
+                SN_MK_FMA_F32_AVX2_CYCLE(0);
+                SN_MK_FMA_F32_AVX2_CYCLE(1);
+                SN_MK_FMA_F32_AVX2_CYCLE(2);
+                SN_MK_FMA_F32_AVX2_CYCLE(3);
 
-                register __m256 a0 = _mm256_broadcast_ps(&left_cur->x[0]);
-                c01 = _mm256_fmadd_ps(a0, b01, c01);
-                c23 = _mm256_fmadd_ps(a0, b23, c23);
-                register __m256 a1 = _mm256_broadcast_ps(&left_cur->x[1]);
-                c01 = _mm256_fmadd_ps(a1, b01, c01);
-                c23 = _mm256_fmadd_ps(a1, b23, c23);
-                register __m256 a2 = _mm256_broadcast_ps(&left_cur->x[2]);
-                c01 = _mm256_fmadd_ps(a2, b01, c01);
-                c23 = _mm256_fmadd_ps(a2, b23, c23);
-                register __m256 a3 = _mm256_broadcast_ps(&left_cur->x[3]);
-                c01 = _mm256_fmadd_ps(a3, b01, c01);
-                c23 = _mm256_fmadd_ps(a3, b23, c23);
         }
-        _mm256_store_ps((float *)&ans_cur->y[0], c01);
-        _mm256_store_ps((float *)&ans_cur->y[1], c23);
+        SN_FMA_AVX2_PACK_COLUMN(c0, (float*)&(ans_cur->x[0]));
+        SN_FMA_AVX2_PACK_COLUMN(c1, (float*)&(ans_cur->x[1]));
+        SN_FMA_AVX2_PACK_COLUMN(c2, (float*)&(ans_cur->x[2]));
+        SN_FMA_AVX2_PACK_COLUMN(c3, (float*)&(ans_cur->x[3]));
 }
 
 
@@ -167,10 +169,10 @@ SN_MK_PROC_DECL_STD(float, 512, , fma)
                 SN_MK_FMA_F32_512_CYCLE(3);
 
         }
-        SN_FMA_PACK_COLUMN(c0, (float*)&(ans_cur->x[0]));
-        SN_FMA_PACK_COLUMN(c1, (float*)&(ans_cur->x[1]));
-        SN_FMA_PACK_COLUMN(c2, (float*)&(ans_cur->x[2]));
-        SN_FMA_PACK_COLUMN(c3, (float*)&(ans_cur->x[3]));
+        SN_FMA_512_PACK_COLUMN(c0, (float*)&(ans_cur->x[0]));
+        SN_FMA_512_PACK_COLUMN(c1, (float*)&(ans_cur->x[1]));
+        SN_FMA_512_PACK_COLUMN(c2, (float*)&(ans_cur->x[2]));
+        SN_FMA_512_PACK_COLUMN(c3, (float*)&(ans_cur->x[3]));
 }
 
 
diff --git a/StgSharp.Native/sn-template/sn_template.fma.x86.h b/StgSharp.Native/sn-template/sn_template.fma.x86.h
index 9267f06a..8dfce50a 100644
--- a/StgSharp.Native/sn-template/sn_template.fma.x86.h
+++ b/StgSharp.Native/sn-template/sn_template.fma.x86.h
@@ -10,7 +10,15 @@
                 c3 = _mm512_fmadd_ps(c3, a3, _mm512_permute_ps(b3, _FMA_SELECT(i))); \
         } while (0)
 
-#define SN_FMA_PACK_COLUMN(col, target)                                   \
+#define SN_MK_FMA_F32_AVX2_CYCLE(i)                                                       \
+        do {                                                                             \
+                c0 = _mm256_fmadd_ps(c0, a0, _mm256_permute_ps(b0, _FMA_SELECT(i))); \
+                c1 = _mm256_fmadd_ps(c1, a1, _mm256_permute_ps(b1, _FMA_SELECT(i))); \
+                c2 = _mm256_fmadd_ps(c2, a2, _mm256_permute_ps(b2, _FMA_SELECT(i))); \
+                c3 = _mm256_fmadd_ps(c3, a3, _mm256_permute_ps(b3, _FMA_SELECT(i))); \
+        } while (0)
+
+#define SN_FMA_512_PACK_COLUMN(col, target)                                   \
         do {                                                              \
                 register __m256 low = _mm512_castps512_ps256(col);        \
                 register __m256 high = _mm512_extractf32x8_ps(col, 1);    \
@@ -21,6 +29,14 @@
                 _mm_store_ps(target, half128);                            \
         } while (0)
 
+#define SN_FMA_AVX2_PACK_COLUMN(col, target)                                   \
+        do {                                                              \
+                register __m128 high128 = _mm256_extractf32x4_ps(col, 1); \
+                register __m128 low128 = _mm256_castps256_ps128(col);    \
+                register __m128 half128 = _mm_add_ps(high128, low128);    \
+                _mm_store_ps(target, half128);                            \
+        } while (0)
+
 // double
 #define SN_MK_FMA_F64_SSE_PHASE(i)                                               \
         do {                                                                     \