diff --git a/src/strategies/avx2/picture-avx2.c b/src/strategies/avx2/picture-avx2.c
index a98f373b..3aad0ee9 100644
--- a/src/strategies/avx2/picture-avx2.c
+++ b/src/strategies/avx2/picture-avx2.c
@@ -1309,8 +1309,8 @@ static uint32_t hor_sad_avx2(const kvz_pixel *pic_data, const kvz_pixel *ref_dat
     return hor_sad_sse41_w16(pic_data, ref_data, width, height,
                              pic_stride, ref_stride, left, right);
   if (width == 32)
-    return hor_sad_sse41_w32(pic_data, ref_data, width, height,
-                             pic_stride, ref_stride, left, right);
+    return hor_sad_avx2_w32(pic_data, ref_data, width, height,
+                            pic_stride, ref_stride, left, right);
   else
     return hor_sad_sse41_arbitrary(pic_data, ref_data, width, height,
                                    pic_stride, ref_stride, left, right);
diff --git a/src/strategies/avx2/reg_sad_pow2_widths-avx2.h b/src/strategies/avx2/reg_sad_pow2_widths-avx2.h
index 68e7604a..b3d9224e 100644
--- a/src/strategies/avx2/reg_sad_pow2_widths-avx2.h
+++ b/src/strategies/avx2/reg_sad_pow2_widths-avx2.h
@@ -109,4 +109,86 @@ static INLINE uint32_t reg_sad_w64(const kvz_pixel * const data1, const kvz_pixe
   return _mm_cvtsi128_si32(sad);
 }
 
+static uint32_t hor_sad_avx2_w32(const kvz_pixel *pic_data, const kvz_pixel *ref_data,
+                                 int32_t width, int32_t height, uint32_t pic_stride,
+                                 uint32_t ref_stride, uint32_t left, uint32_t right)
+{
+  const int32_t height_fourline_groups = height & ~3;
+  const int32_t height_residual_lines  = height &  3;
+
+  __m256i ns = _mm256_setr_epi8(0,  1,  2,  3,  4,  5,  6,  7,
+                                8,  9,  10, 11, 12, 13, 14, 15,
+                                16, 17, 18, 19, 20, 21, 22, 23,
+                                24, 25, 26, 27, 28, 29, 30, 31);
+  __m256i epol_mask;
+  int32_t border_pix_off;
+
+  if (left) {
+    border_pix_off          = left;
+    __m256i first_valid_idx = _mm256_set1_epi8(left);
+
+    epol_mask = _mm256_cmpgt_epi8(first_valid_idx, ns);
+  } else {
+    border_pix_off          = width - (right + 1);
+    __m256i last_valid_idx  = _mm256_set1_epi8(border_pix_off);
+
+    epol_mask = _mm256_cmpgt_epi8(ns, last_valid_idx);
+  }
+
+  __m256i avx_inc = _mm256_setzero_si256();
+  int32_t y;
+  for (y = 0; y < height_fourline_groups; y += 4) {
+    __m256i a = _mm256_loadu_si256((__m256i *)(pic_data + (y + 0) * pic_stride));
+    __m256i b = _mm256_loadu_si256((__m256i *)(ref_data + (y + 0) * ref_stride));
+    __m256i c = _mm256_loadu_si256((__m256i *)(pic_data + (y + 1) * pic_stride));
+    __m256i d = _mm256_loadu_si256((__m256i *)(ref_data + (y + 1) * ref_stride));
+    __m256i e = _mm256_loadu_si256((__m256i *)(pic_data + (y + 2) * pic_stride));
+    __m256i f = _mm256_loadu_si256((__m256i *)(ref_data + (y + 2) * ref_stride));
+    __m256i g = _mm256_loadu_si256((__m256i *)(pic_data + (y + 3) * pic_stride));
+    __m256i h = _mm256_loadu_si256((__m256i *)(ref_data + (y + 3) * ref_stride));
+
+    __m256i border_px_b  = _mm256_set1_epi8   (*(uint8_t *)(ref_data + (y + 0) * ref_stride + border_pix_off));
+    __m256i border_px_d  = _mm256_set1_epi8   (*(uint8_t *)(ref_data + (y + 1) * ref_stride + border_pix_off));
+    __m256i border_px_f  = _mm256_set1_epi8   (*(uint8_t *)(ref_data + (y + 2) * ref_stride + border_pix_off));
+    __m256i border_px_h  = _mm256_set1_epi8   (*(uint8_t *)(ref_data + (y + 3) * ref_stride + border_pix_off));
+
+    __m256i b_epol       = _mm256_blendv_epi8(b, border_px_b, epol_mask);
+    __m256i d_epol       = _mm256_blendv_epi8(d, border_px_d, epol_mask);
+    __m256i f_epol       = _mm256_blendv_epi8(f, border_px_f, epol_mask);
+    __m256i h_epol       = _mm256_blendv_epi8(h, border_px_h, epol_mask);
+
+    __m256i curr_sads_ab = _mm256_sad_epu8(a, b_epol);
+    __m256i curr_sads_cd = _mm256_sad_epu8(c, d_epol);
+    __m256i curr_sads_ef = _mm256_sad_epu8(e, f_epol);
+    __m256i curr_sads_gh = _mm256_sad_epu8(g, h_epol);
+
+    avx_inc = _mm256_add_epi64(avx_inc, curr_sads_ab);
+    avx_inc = _mm256_add_epi64(avx_inc, curr_sads_cd);
+    avx_inc = _mm256_add_epi64(avx_inc, curr_sads_ef);
+    avx_inc = _mm256_add_epi64(avx_inc, curr_sads_gh);
+  }
+  if (height_residual_lines) {
+    for (; y < height; y++) {
+      __m256i a = _mm256_loadu_si256((__m256i *)(pic_data + y * pic_stride));
+      __m256i b = _mm256_loadu_si256((__m256i *)(ref_data + y * ref_stride));
+
+      __m256i border_px_b = _mm256_set1_epi8  (*(uint8_t *)(ref_data + y * ref_stride + border_pix_off));
+      __m256i b_epol      = _mm256_blendv_epi8(b, border_px_b, epol_mask);
+
+      __m256i curr_sads_ab = _mm256_sad_epu8(a, b_epol);
+
+      avx_inc = _mm256_add_epi64(avx_inc, curr_sads_ab);
+    }
+  }
+
+  __m128i inchi = _mm256_extracti128_si256(avx_inc, 1);
+  __m128i inclo = _mm256_castsi256_si128  (avx_inc);
+
+  __m128i sum_1 = _mm_add_epi64    (inclo, inchi);
+  __m128i sum_2 = _mm_shuffle_epi32(sum_1, _MM_SHUFFLE(1, 0, 3, 2));
+  __m128i sad   = _mm_add_epi64    (sum_1, sum_2);
+
+  return _mm_cvtsi128_si32(sad);
+}
+
 #endif