diff --git a/aom_dsp/x86/convolve_avx2.h b/aom_dsp/x86/convolve_avx2.h
index e041a62..644f3dd 100644
--- a/aom_dsp/x86/convolve_avx2.h
+++ b/aom_dsp/x86/convolve_avx2.h
@@ -34,31 +34,6 @@
   2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8,
 };
 
-DECLARE_ALIGNED(32, static const uint8_t, filt_center_global_avx2[32]) = {
-  3, 255, 4, 255, 5, 255, 6, 255, 7, 255, 8, 255, 9, 255, 10, 255,
-  3, 255, 4, 255, 5, 255, 6, 255, 7, 255, 8, 255, 9, 255, 10, 255
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
-  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
-  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt2_global_avx2[32]) = {
-  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10,
-  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt3_global_avx2[32]) = {
-  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
-  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
-  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14,
-  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
-};
-
 static INLINE void prepare_coeffs_lowbd(
     const InterpFilterParams *const filter_params, const int subpel_q4,
     __m256i *const coeffs /* [4] */) {
diff --git a/av1/common/x86/wiener_convolve_avx2.c b/av1/common/x86/wiener_convolve_avx2.c
index 1f13e2f..87a6e12 100644
--- a/av1/common/x86/wiener_convolve_avx2.c
+++ b/av1/common/x86/wiener_convolve_avx2.c
@@ -17,7 +17,6 @@
 #include "av1/common/convolve.h"
 #include "aom_dsp/aom_dsp_common.h"
 #include "aom_dsp/aom_filter.h"
-#include "aom_dsp/x86/convolve_avx2.h"
 #include "aom_dsp/x86/synonyms.h"
 #include "aom_dsp/x86/synonyms_avx2.h"
 
@@ -26,207 +25,236 @@
 // on the left.
 // A row of, say, 8-bit pixels with values p0, p1, p2, ..., p30, p31 will be
 // loaded and stored as [ p31 ... p17 p16 ][ p15 ... p1 p0 ].
-
-// Exploiting the range of wiener filter coefficients,
-// horizontal filtering can be done in 16 bit intermediate precision.
-// The details are as follows :
-// Consider the horizontal wiener filter coefficients of the following form :
-//      [C0, C1, C2, 2^(FILTER_BITS) -2 * (C0 + C1 + C2), C2, C1, C0]
-// Subtracting  2^(FILTER_BITS) from the centre tap we get the following  :
-//      [C0, C1, C2,     -2 * (C0 + C1 + C2),             C2, C1, C0]
-// The sum of the product "C0 * p0 + C1 * p1 + C2 * p2 -2 * (C0 + C1 + C2) * p3
-// + C2 * p4 + C1 * p5 + C0 * p6" would be in the range of signed 16 bit
-// precision. Finally, after rounding the above result by round_0, we multiply
-// the centre pixel by 2^(FILTER_BITS - round_0) and add it to get the
-// horizontal filter output.
-
 void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride,
                                       uint8_t *dst, ptrdiff_t dst_stride,
                                       const int16_t *filter_x, int x_step_q4,
                                       const int16_t *filter_y, int y_step_q4,
                                       int w, int h,
                                       const ConvolveParams *conv_params) {
+  const int bd = 8;
   assert(x_step_q4 == 16 && y_step_q4 == 16);
   assert(!(w & 7));
   (void)x_step_q4;
   (void)y_step_q4;
 
-  DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + SUBPEL_TAPS) * 8]);
-  int im_h = h + SUBPEL_TAPS - 2;
-  int im_stride = 8;
-  memset(im_block + (im_h * im_stride), 0, MAX_SB_SIZE);
-  int i, j;
-  const int center_tap = (SUBPEL_TAPS - 1) / 2;
+  DECLARE_ALIGNED(32, uint16_t,
+                  temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
+  int intermediate_height = h + SUBPEL_TAPS - 2;
+  memset(temp + (intermediate_height * MAX_SB_SIZE), 0, MAX_SB_SIZE);
+  const int center_tap = ((SUBPEL_TAPS - 1) / 2);
   const uint8_t *const src_ptr = src - center_tap * src_stride - center_tap;
 
-  __m256i filt[4], coeffs_h[4], coeffs_v[4], filt_center;
-
-  assert(conv_params->round_0 > 0);
-
-  filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
-  filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
-  filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
-  filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);
-
-  filt_center = _mm256_load_si256((__m256i const *)filt_center_global_avx2);
-
-  const __m128i coeffs_x = _mm_loadu_si128((__m128i *)filter_x);
-  const __m256i filter_coeffs_x = _mm256_broadcastsi128_si256(coeffs_x);
-
-  // coeffs 0 1 0 1 0 1 0 1
-  coeffs_h[0] =
-      _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0200u));
-  // coeffs 2 3 2 3 2 3 2 3
-  coeffs_h[1] =
-      _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0604u));
-  // coeffs 4 5 4 5 4 5 4 5
-  coeffs_h[2] =
-      _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0a08u));
-  // coeffs 6 7 6 7 6 7 6 7
-  coeffs_h[3] =
-      _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0e0cu));
-
-  const __m256i round_const_h =
-      _mm256_set1_epi16((1 << (conv_params->round_0 - 1)));
-  const __m128i round_shift_h = _mm_cvtsi32_si128(conv_params->round_0);
+  const __m128i zero_128 = _mm_setzero_si128();
+  const __m256i zero_256 = _mm256_setzero_si256();
 
   // Add an offset to account for the "add_src" part of the convolve function.
-  const __m128i zero_128 = _mm_setzero_si128();
-  const __m128i offset_0 = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3);
-  const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset_0);
+  const __m128i offset = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3);
 
-  const __m256i filter_coeffs_y = _mm256_broadcastsi128_si256(coeffs_y);
+  const __m256i clamp_low = zero_256;
+  const __m256i clamp_high =
+      _mm256_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1);
 
-  // coeffs 0 1 0 1 0 1 0 1
-  coeffs_v[0] = _mm256_shuffle_epi32(filter_coeffs_y, 0x00);
-  // coeffs 2 3 2 3 2 3 2 3
-  coeffs_v[1] = _mm256_shuffle_epi32(filter_coeffs_y, 0x55);
-  // coeffs 4 5 4 5 4 5 4 5
-  coeffs_v[2] = _mm256_shuffle_epi32(filter_coeffs_y, 0xaa);
-  // coeffs 6 7 6 7 6 7 6 7
-  coeffs_v[3] = _mm256_shuffle_epi32(filter_coeffs_y, 0xff);
+  /* Horizontal filter */
+  {
+    // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ]
+    const __m128i coeffs_x = _mm_add_epi16(xx_loadu_128(filter_x), offset);
 
-  const __m256i round_const_v =
-      _mm256_set1_epi32((1 << (conv_params->round_1 - 1)));
-  const __m128i round_shift_v = _mm_cvtsi32_si128(conv_params->round_1);
+    // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ]
+    const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ]
+    const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
 
-  for (j = 0; j < w; j += 8) {
-    for (i = 0; i < im_h; i += 2) {
-      __m256i data = _mm256_castsi128_si256(
-          _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));
+    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ]
+    const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ]
+    const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ]
+    const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
+    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ]
+    const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);
 
-      // Load the next line
-      if (i + 1 < im_h)
-        data = _mm256_inserti128_si256(
-            data,
-            _mm_loadu_si128(
-                (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]),
-            1);
+    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ]
+    const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
+    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ]
+    const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
+    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ]
+    const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
+    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ]
+    const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);
 
-      __m256i res = convolve_lowbd_x(data, coeffs_h, filt);
+    const __m256i round_const = _mm256_set1_epi32(
+        (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));
 
-      res =
-          _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h);
+    for (int i = 0; i < intermediate_height; ++i) {
+      for (int j = 0; j < w; j += 16) {
+        const uint8_t *data_ij = src_ptr + i * src_stride + j;
 
-      __m256i data_0 = _mm256_shuffle_epi8(data, filt_center);
+        // Load 8-bit src data
+        const __m128i data_0 = xx_loadu_128(data_ij + 0);
+        const __m128i data_1 = xx_loadu_128(data_ij + 1);
+        const __m128i data_2 = xx_loadu_128(data_ij + 2);
+        const __m128i data_3 = xx_loadu_128(data_ij + 3);
+        const __m128i data_4 = xx_loadu_128(data_ij + 4);
+        const __m128i data_5 = xx_loadu_128(data_ij + 5);
+        const __m128i data_6 = xx_loadu_128(data_ij + 6);
+        const __m128i data_7 = xx_loadu_128(data_ij + 7);
 
-      // multiply the center pixel by 2^(FILTER_BITS - round_0) and add it to
-      // the result
-      data_0 = _mm256_slli_epi16(data_0, FILTER_BITS - conv_params->round_0);
-      res = _mm256_add_epi16(res, data_0);
+        // (Zero-)Extend 8-bit data to 16-bit data
+        const __m256i src_0 = _mm256_cvtepu8_epi16(data_0);
+        const __m256i src_1 = _mm256_cvtepu8_epi16(data_1);
+        const __m256i src_2 = _mm256_cvtepu8_epi16(data_2);
+        const __m256i src_3 = _mm256_cvtepu8_epi16(data_3);
+        const __m256i src_4 = _mm256_cvtepu8_epi16(data_4);
+        const __m256i src_5 = _mm256_cvtepu8_epi16(data_5);
+        const __m256i src_6 = _mm256_cvtepu8_epi16(data_6);
+        const __m256i src_7 = _mm256_cvtepu8_epi16(data_7);
 
-      _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
-    }
+        // Multiply src data by filter coeffs and sum pairs
+        const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
+        const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
+        const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
+        const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
+        const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
+        const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
+        const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
+        const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);
 
-    /* Vertical filter */
-    {
-      __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));
-      __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));
-      __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));
-      __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));
-      __m256i src_4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride));
-      __m256i src_5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride));
+        // Calculate scalar product for even- and odd-indices separately,
+        // increasing to 32-bit precision
+        const __m256i res_even_sum = _mm256_add_epi32(
+            _mm256_add_epi32(res_0, res_4), _mm256_add_epi32(res_2, res_6));
+        const __m256i res_odd_sum = _mm256_add_epi32(
+            _mm256_add_epi32(res_1, res_5), _mm256_add_epi32(res_3, res_7));
 
-      __m256i s[8];
-      s[0] = _mm256_unpacklo_epi16(src_0, src_1);
-      s[1] = _mm256_unpacklo_epi16(src_2, src_3);
-      s[2] = _mm256_unpacklo_epi16(src_4, src_5);
+        const __m256i res_even = _mm256_srai_epi32(
+            _mm256_add_epi32(res_even_sum, round_const), conv_params->round_0);
+        const __m256i res_odd = _mm256_srai_epi32(
+            _mm256_add_epi32(res_odd_sum, round_const), conv_params->round_0);
 
-      s[4] = _mm256_unpackhi_epi16(src_0, src_1);
-      s[5] = _mm256_unpackhi_epi16(src_2, src_3);
-      s[6] = _mm256_unpackhi_epi16(src_4, src_5);
+        // Reduce to 16-bit precision and pack even- and odd-index results
+        // back into one register. The _mm256_packs_epi32 intrinsic returns
+        // a register with the pixels ordered as follows:
+        // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
+        const __m256i res = _mm256_packs_epi32(res_even, res_odd);
+        const __m256i res_clamped =
+            _mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high);
 
-      for (i = 0; i < h - 1; i += 2) {
-        const int16_t *data = &im_block[i * im_stride];
-
-        const __m256i s6 =
-            _mm256_loadu_si256((__m256i *)(data + 6 * im_stride));
-        const __m256i s7 =
-            _mm256_loadu_si256((__m256i *)(data + 7 * im_stride));
-
-        s[3] = _mm256_unpacklo_epi16(s6, s7);
-        s[7] = _mm256_unpackhi_epi16(s6, s7);
-
-        __m256i res_a = convolve(s, coeffs_v);
-        __m256i res_b = convolve(s + 4, coeffs_v);
-
-        const __m256i res_a_round = _mm256_sra_epi32(
-            _mm256_add_epi32(res_a, round_const_v), round_shift_v);
-        const __m256i res_b_round = _mm256_sra_epi32(
-            _mm256_add_epi32(res_b, round_const_v), round_shift_v);
-
-        /* rounding code */
-        // 16 bit conversion
-        const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
-        // 8 bit conversion and saturation to uint8
-        const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);
-
-        const __m128i res_0 = _mm256_castsi256_si128(res_8b);
-        const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);
-
-        // Store values into the destination buffer
-        __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];
-        __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];
-
-        _mm_storel_epi64(p_0, res_0);
-        _mm_storel_epi64(p_1, res_1);
-
-        s[0] = s[1];
-        s[1] = s[2];
-        s[2] = s[3];
-
-        s[4] = s[5];
-        s[5] = s[6];
-        s[6] = s[7];
+        // Store in a temporary array
+        yy_storeu_256(temp + i * MAX_SB_SIZE + j, res_clamped);
       }
-      if (h - i) {
-        s[0] = _mm256_permute2x128_si256(s[0], s[4], 0x20);
-        s[1] = _mm256_permute2x128_si256(s[1], s[5], 0x20);
-        s[2] = _mm256_permute2x128_si256(s[2], s[6], 0x20);
+    }
+  }
 
-        const int16_t *data = &im_block[i * im_stride];
-        const __m128i s6_ = _mm_loadu_si128((__m128i *)(data + 6 * im_stride));
-        const __m128i s7_ = _mm_loadu_si128((__m128i *)(data + 7 * im_stride));
+  /* Vertical filter */
+  {
+    // coeffs [ g7 g6 g5 g4 g3 g2 g1 g0 ]
+    const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset);
 
-        __m128i s3 = _mm_unpacklo_epi16(s6_, s7_);
-        __m128i s7 = _mm_unpackhi_epi16(s6_, s7_);
+    // coeffs [ g3 g2 g3 g2 g1 g0 g1 g0 ]
+    const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs [ g7 g6 g7 g6 g5 g4 g5 g4 ]
+    const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
 
-        s[3] = _mm256_inserti128_si256(_mm256_castsi128_si256(s3), s7, 1);
-        __m256i convolveres = convolve(s, coeffs_v);
+    // coeffs [ g1 g0 g1 g0 g1 g0 g1 g0 ]
+    const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ g3 g2 g3 g2 g3 g2 g3 g2 ]
+    const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ g5 g4 g5 g4 g5 g4 g5 g4 ]
+    const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
+    // coeffs [ g7 g6 g7 g6 g7 g6 g7 g6 ]
+    const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);
 
-        const __m256i res_round = _mm256_sra_epi32(
-            _mm256_add_epi32(convolveres, round_const_v), round_shift_v);
+    // coeffs [ g1 g0 g1 g0 g1 g0 g1 g0 ][ g1 g0 g1 g0 g1 g0 g1 g0 ]
+    const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
+    // coeffs [ g3 g2 g3 g2 g3 g2 g3 g2 ][ g3 g2 g3 g2 g3 g2 g3 g2 ]
+    const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
+    // coeffs [ g5 g4 g5 g4 g5 g4 g5 g4 ][ g5 g4 g5 g4 g5 g4 g5 g4 ]
+    const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
+    // coeffs [ g7 g6 g7 g6 g7 g6 g7 g6 ][ g7 g6 g7 g6 g7 g6 g7 g6 ]
+    const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);
 
-        /* rounding code */
-        // 16 bit conversion
-        __m128i reslo = _mm256_castsi256_si128(res_round);
-        __m128i reshi = _mm256_extracti128_si256(res_round, 1);
-        const __m128i res_16bit = _mm_packus_epi32(reslo, reshi);
+    const __m256i round_const =
+        _mm256_set1_epi32((1 << (conv_params->round_1 - 1)) -
+                          (1 << (bd + conv_params->round_1 - 1)));
 
-        // 8 bit conversion and saturation to uint8
-        const __m128i res_8b = _mm_packus_epi16(res_16bit, res_16bit);
-        __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];
-        _mm_storel_epi64(p_0, res_8b);
+    for (int i = 0; i < h; ++i) {
+      for (int j = 0; j < w; j += 16) {
+        const uint16_t *data_ij = temp + i * MAX_SB_SIZE + j;
+
+        // Load 16-bit data from the output of the horizontal filter in
+        // which the pixels are ordered as follows:
+        // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
+        const __m256i data_0 = yy_loadu_256(data_ij + 0 * MAX_SB_SIZE);
+        const __m256i data_1 = yy_loadu_256(data_ij + 1 * MAX_SB_SIZE);
+        const __m256i data_2 = yy_loadu_256(data_ij + 2 * MAX_SB_SIZE);
+        const __m256i data_3 = yy_loadu_256(data_ij + 3 * MAX_SB_SIZE);
+        const __m256i data_4 = yy_loadu_256(data_ij + 4 * MAX_SB_SIZE);
+        const __m256i data_5 = yy_loadu_256(data_ij + 5 * MAX_SB_SIZE);
+        const __m256i data_6 = yy_loadu_256(data_ij + 6 * MAX_SB_SIZE);
+        const __m256i data_7 = yy_loadu_256(data_ij + 7 * MAX_SB_SIZE);
+
+        // Filter the even-indices, increasing to 32-bit precision
+        const __m256i src_0 = _mm256_unpacklo_epi16(data_0, data_1);
+        const __m256i src_2 = _mm256_unpacklo_epi16(data_2, data_3);
+        const __m256i src_4 = _mm256_unpacklo_epi16(data_4, data_5);
+        const __m256i src_6 = _mm256_unpacklo_epi16(data_6, data_7);
+
+        const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
+        const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
+        const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
+        const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
+
+        const __m256i res_even = _mm256_add_epi32(
+            _mm256_add_epi32(res_0, res_2), _mm256_add_epi32(res_4, res_6));
+
+        // Filter the odd-indices, increasing to 32-bit precision
+        const __m256i src_1 = _mm256_unpackhi_epi16(data_0, data_1);
+        const __m256i src_3 = _mm256_unpackhi_epi16(data_2, data_3);
+        const __m256i src_5 = _mm256_unpackhi_epi16(data_4, data_5);
+        const __m256i src_7 = _mm256_unpackhi_epi16(data_6, data_7);
+
+        const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
+        const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
+        const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
+        const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);
+
+        const __m256i res_odd = _mm256_add_epi32(
+            _mm256_add_epi32(res_1, res_3), _mm256_add_epi32(res_5, res_7));
+
+        // Pixels are currently in the following order:
+        // res_even order: [ 14 12 10 8 ] [ 6 4 2 0 ]
+        // res_odd order:  [ 15 13 11 9 ] [ 7 5 3 1 ]
+        //
+        // Rearrange the pixels into the following order:
+        // res_lo order: [ 11 10  9  8 ] [ 3 2 1 0 ]
+        // res_hi order: [ 15 14 13 12 ] [ 7 6 5 4 ]
+        const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
+        const __m256i res_hi = _mm256_unpackhi_epi32(res_even, res_odd);
+
+        const __m256i res_lo_round = _mm256_srai_epi32(
+            _mm256_add_epi32(res_lo, round_const), conv_params->round_1);
+        const __m256i res_hi_round = _mm256_srai_epi32(
+            _mm256_add_epi32(res_hi, round_const), conv_params->round_1);
+
+        // Reduce to 16-bit precision and pack into the correct order:
+        // [ 15 14 13 12 11 10 9 8 ][ 7 6 5 4 3 2 1 0 ]
+        const __m256i res_16bit =
+            _mm256_packs_epi32(res_lo_round, res_hi_round);
+
+        // Reduce to 8-bit precision. This messes up the order:
+        // [ - - - - - - - - 15 14 13 12 11 10 9 8 ]
+        // [ - - - - - - - - 7 6 5 4 3 2 1 0 ]
+        const __m256i res_8bit =
+            _mm256_packus_epi16(res_16bit, zero_256 /* don't care value */);
+
+        // Swap the two central 32-bit values to get the order:
+        // [ - - - - - - - - - - - - - - - - ]
+        // [ 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 ]
+        const __m256i res_8bit2 = _mm256_permute4x64_epi64(res_8bit, 0xd8);
+
+        // Store the lower 128-bit lane in the dst array
+        xx_storeu_128(dst + i * dst_stride + j,
+                      _mm256_castsi256_si128(res_8bit2));
       }
     }
   }
diff --git a/test/hiprec_convolve_test_util.cc b/test/hiprec_convolve_test_util.cc
index f5bf56e..2672bce 100644
--- a/test/hiprec_convolve_test_util.cc
+++ b/test/hiprec_convolve_test_util.cc
@@ -31,7 +31,7 @@
   hkernel[2] = hkernel[4] =
       WIENER_FILT_TAP2_MINV +
       rnd->PseudoUniform(WIENER_FILT_TAP2_MAXV + 1 - WIENER_FILT_TAP2_MINV);
-  hkernel[3] = -2 * (hkernel[0] + hkernel[1] + hkernel[2]);
+  hkernel[3] = -(hkernel[0] + hkernel[1] + hkernel[2]);
   hkernel[7] = 0;
 
   vkernel[0] = vkernel[6] =
@@ -43,7 +43,7 @@
   vkernel[2] = vkernel[4] =
       WIENER_FILT_TAP2_MINV +
       rnd->PseudoUniform(WIENER_FILT_TAP2_MAXV + 1 - WIENER_FILT_TAP2_MINV);
-  vkernel[3] = -2 * (vkernel[0] + vkernel[1] + vkernel[2]);
+  vkernel[3] = -(vkernel[0] + vkernel[1] + vkernel[2]);
   vkernel[7] = 0;
 }