tor-browser

The Tor Browser
git clone https://git.dasho.dev/tor-browser.git
Log | Files | Refs | README | LICENSE

warp_plane_avx2.c (53875B)


      1 /*
      2 * Copyright (c) 2019, Alliance for Open Media. All rights reserved.
      3 *
      4 * This source code is subject to the terms of the BSD 2 Clause License and
      5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
      6 * was not distributed with this source code in the LICENSE file, you can
      7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
      8 * Media Patent License 1.0 was not distributed with this source code in the
      9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
     10 */
     11 
     12 #include <immintrin.h>
     13 #include "config/av1_rtcd.h"
     14 #include "av1/common/warped_motion.h"
     15 #include "aom_dsp/x86/synonyms.h"
     16 
     17 #if !CONFIG_HIGHWAY
     18 
     19 DECLARE_ALIGNED(32, static const uint8_t, shuffle_alpha0_mask01_avx2[32]) = {
     20  0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
     21  0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1
     22 };
     23 
     24 DECLARE_ALIGNED(32, static const uint8_t, shuffle_alpha0_mask23_avx2[32]) = {
     25  2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
     26  2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3
     27 };
     28 
     29 DECLARE_ALIGNED(32, static const uint8_t, shuffle_alpha0_mask45_avx2[32]) = {
     30  4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
     31  4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5
     32 };
     33 
     34 DECLARE_ALIGNED(32, static const uint8_t, shuffle_alpha0_mask67_avx2[32]) = {
     35  6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
     36  6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7
     37 };
     38 
     39 DECLARE_ALIGNED(32, static const uint8_t, shuffle_gamma0_mask0_avx2[32]) = {
     40  0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3,
     41  0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3
     42 };
     43 
     44 DECLARE_ALIGNED(32, static const uint8_t, shuffle_gamma0_mask1_avx2[32]) = {
     45  4, 5, 6, 7, 4, 5, 6, 7, 4, 5, 6, 7, 4, 5, 6, 7,
     46  4, 5, 6, 7, 4, 5, 6, 7, 4, 5, 6, 7, 4, 5, 6, 7
     47 };
     48 
     49 DECLARE_ALIGNED(32, static const uint8_t, shuffle_gamma0_mask2_avx2[32]) = {
     50  8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11,
     51  8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11
     52 };
     53 
     54 DECLARE_ALIGNED(32, static const uint8_t, shuffle_gamma0_mask3_avx2[32]) = {
     55  12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15,
     56  12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15
     57 };
     58 
     59 DECLARE_ALIGNED(32, static const uint8_t,
     60                shuffle_src0[32]) = { 0, 2, 2, 4, 4, 6, 6, 8, 1, 3, 3,
     61                                      5, 5, 7, 7, 9, 0, 2, 2, 4, 4, 6,
     62                                      6, 8, 1, 3, 3, 5, 5, 7, 7, 9 };
     63 
     64 DECLARE_ALIGNED(32, static const uint8_t,
     65                shuffle_src1[32]) = { 4,  6,  6,  8,  8,  10, 10, 12, 5,  7, 7,
     66                                      9,  9,  11, 11, 13, 4,  6,  6,  8,  8, 10,
     67                                      10, 12, 5,  7,  7,  9,  9,  11, 11, 13 };
     68 
     69 DECLARE_ALIGNED(32, static const uint8_t,
     70                shuffle_src2[32]) = { 1, 3, 3, 5, 5,  7, 7, 9, 2, 4, 4,
     71                                      6, 6, 8, 8, 10, 1, 3, 3, 5, 5, 7,
     72                                      7, 9, 2, 4, 4,  6, 6, 8, 8, 10 };
     73 
     74 DECLARE_ALIGNED(32, static const uint8_t,
     75                shuffle_src3[32]) = { 5,  7,  7,  9,  9,  11, 11, 13, 6,  8, 8,
     76                                      10, 10, 12, 12, 14, 5,  7,  7,  9,  9, 11,
     77                                      11, 13, 6,  8,  8,  10, 10, 12, 12, 14 };
     78 
     79 static inline void filter_src_pixels_avx2(const __m256i src, __m256i *horz_out,
     80                                          __m256i *coeff,
     81                                          const __m256i *shuffle_src,
     82                                          const __m256i *round_const,
     83                                          const __m128i *shift, int row) {
     84  const __m256i src_0 = _mm256_shuffle_epi8(src, shuffle_src[0]);
     85  const __m256i src_1 = _mm256_shuffle_epi8(src, shuffle_src[1]);
     86  const __m256i src_2 = _mm256_shuffle_epi8(src, shuffle_src[2]);
     87  const __m256i src_3 = _mm256_shuffle_epi8(src, shuffle_src[3]);
     88 
     89  const __m256i res_02 = _mm256_maddubs_epi16(src_0, coeff[0]);
     90  const __m256i res_46 = _mm256_maddubs_epi16(src_1, coeff[1]);
     91  const __m256i res_13 = _mm256_maddubs_epi16(src_2, coeff[2]);
     92  const __m256i res_57 = _mm256_maddubs_epi16(src_3, coeff[3]);
     93 
     94  const __m256i res_even = _mm256_add_epi16(res_02, res_46);
     95  const __m256i res_odd = _mm256_add_epi16(res_13, res_57);
     96  const __m256i res =
     97      _mm256_add_epi16(_mm256_add_epi16(res_even, res_odd), *round_const);
     98  horz_out[row] = _mm256_srl_epi16(res, *shift);
     99 }
    100 
    101 static inline void prepare_horizontal_filter_coeff_avx2(int alpha, int beta,
    102                                                        int sx,
    103                                                        __m256i *coeff) {
    104  __m128i tmp_0 = _mm_loadl_epi64(
    105      (__m128i *)&av1_filter_8bit[((unsigned)(sx + 0 * alpha)) >>
    106                                  WARPEDDIFF_PREC_BITS]);
    107  __m128i tmp_1 = _mm_loadl_epi64(
    108      (__m128i *)&av1_filter_8bit[((unsigned)(sx + 1 * alpha)) >>
    109                                  WARPEDDIFF_PREC_BITS]);
    110  __m128i tmp_2 = _mm_loadl_epi64(
    111      (__m128i *)&av1_filter_8bit[((unsigned)(sx + 2 * alpha)) >>
    112                                  WARPEDDIFF_PREC_BITS]);
    113  __m128i tmp_3 = _mm_loadl_epi64(
    114      (__m128i *)&av1_filter_8bit[((unsigned)(sx + 3 * alpha)) >>
    115                                  WARPEDDIFF_PREC_BITS]);
    116 
    117  __m128i tmp_4 = _mm_loadl_epi64(
    118      (__m128i *)&av1_filter_8bit[((unsigned)(sx + 4 * alpha)) >>
    119                                  WARPEDDIFF_PREC_BITS]);
    120  __m128i tmp_5 = _mm_loadl_epi64(
    121      (__m128i *)&av1_filter_8bit[((unsigned)(sx + 5 * alpha)) >>
    122                                  WARPEDDIFF_PREC_BITS]);
    123  __m128i tmp_6 = _mm_loadl_epi64(
    124      (__m128i *)&av1_filter_8bit[((unsigned)(sx + 6 * alpha)) >>
    125                                  WARPEDDIFF_PREC_BITS]);
    126  __m128i tmp_7 = _mm_loadl_epi64(
    127      (__m128i *)&av1_filter_8bit[((unsigned)(sx + 7 * alpha)) >>
    128                                  WARPEDDIFF_PREC_BITS]);
    129 
    130  __m256i tmp0_256 = _mm256_castsi128_si256(tmp_0);
    131  __m256i tmp2_256 = _mm256_castsi128_si256(tmp_2);
    132  __m256i tmp1_256 = _mm256_castsi128_si256(tmp_1);
    133  __m256i tmp3_256 = _mm256_castsi128_si256(tmp_3);
    134 
    135  __m256i tmp4_256 = _mm256_castsi128_si256(tmp_4);
    136  __m256i tmp6_256 = _mm256_castsi128_si256(tmp_6);
    137  __m256i tmp5_256 = _mm256_castsi128_si256(tmp_5);
    138  __m256i tmp7_256 = _mm256_castsi128_si256(tmp_7);
    139 
    140  __m128i tmp_8 = _mm_loadl_epi64(
    141      (__m128i *)&av1_filter_8bit[(unsigned)((sx + beta) + 0 * alpha) >>
    142                                  WARPEDDIFF_PREC_BITS]);
    143  tmp0_256 = _mm256_inserti128_si256(tmp0_256, tmp_8, 1);
    144 
    145  __m128i tmp_9 = _mm_loadl_epi64(
    146      (__m128i *)&av1_filter_8bit[(unsigned)((sx + beta) + 1 * alpha) >>
    147                                  WARPEDDIFF_PREC_BITS]);
    148  tmp1_256 = _mm256_inserti128_si256(tmp1_256, tmp_9, 1);
    149 
    150  __m128i tmp_10 = _mm_loadl_epi64(
    151      (__m128i *)&av1_filter_8bit[(unsigned)((sx + beta) + 2 * alpha) >>
    152                                  WARPEDDIFF_PREC_BITS]);
    153  tmp2_256 = _mm256_inserti128_si256(tmp2_256, tmp_10, 1);
    154 
    155  __m128i tmp_11 = _mm_loadl_epi64(
    156      (__m128i *)&av1_filter_8bit[(unsigned)((sx + beta) + 3 * alpha) >>
    157                                  WARPEDDIFF_PREC_BITS]);
    158  tmp3_256 = _mm256_inserti128_si256(tmp3_256, tmp_11, 1);
    159 
    160  tmp_2 = _mm_loadl_epi64(
    161      (__m128i *)&av1_filter_8bit[(unsigned)((sx + beta) + 4 * alpha) >>
    162                                  WARPEDDIFF_PREC_BITS]);
    163  tmp4_256 = _mm256_inserti128_si256(tmp4_256, tmp_2, 1);
    164 
    165  tmp_3 = _mm_loadl_epi64(
    166      (__m128i *)&av1_filter_8bit[(unsigned)((sx + beta) + 5 * alpha) >>
    167                                  WARPEDDIFF_PREC_BITS]);
    168  tmp5_256 = _mm256_inserti128_si256(tmp5_256, tmp_3, 1);
    169 
    170  tmp_6 = _mm_loadl_epi64(
    171      (__m128i *)&av1_filter_8bit[(unsigned)((sx + beta) + 6 * alpha) >>
    172                                  WARPEDDIFF_PREC_BITS]);
    173  tmp6_256 = _mm256_inserti128_si256(tmp6_256, tmp_6, 1);
    174 
    175  tmp_7 = _mm_loadl_epi64(
    176      (__m128i *)&av1_filter_8bit[(unsigned)((sx + beta) + 7 * alpha) >>
    177                                  WARPEDDIFF_PREC_BITS]);
    178  tmp7_256 = _mm256_inserti128_si256(tmp7_256, tmp_7, 1);
    179 
    180  const __m256i tmp_12 = _mm256_unpacklo_epi16(tmp0_256, tmp2_256);
    181  const __m256i tmp_13 = _mm256_unpacklo_epi16(tmp1_256, tmp3_256);
    182  const __m256i tmp_14 = _mm256_unpacklo_epi16(tmp4_256, tmp6_256);
    183  const __m256i tmp_15 = _mm256_unpacklo_epi16(tmp5_256, tmp7_256);
    184 
    185  const __m256i res_0 = _mm256_unpacklo_epi32(tmp_12, tmp_14);
    186  const __m256i res_1 = _mm256_unpackhi_epi32(tmp_12, tmp_14);
    187  const __m256i res_2 = _mm256_unpacklo_epi32(tmp_13, tmp_15);
    188  const __m256i res_3 = _mm256_unpackhi_epi32(tmp_13, tmp_15);
    189 
    190  coeff[0] = _mm256_unpacklo_epi64(res_0, res_2);
    191  coeff[1] = _mm256_unpackhi_epi64(res_0, res_2);
    192  coeff[2] = _mm256_unpacklo_epi64(res_1, res_3);
    193  coeff[3] = _mm256_unpackhi_epi64(res_1, res_3);
    194 }
    195 
    196 static inline void prepare_horizontal_filter_coeff_beta0_avx2(int alpha, int sx,
    197                                                              __m256i *coeff) {
    198  __m128i tmp_0 = _mm_loadl_epi64(
    199      (__m128i *)&av1_filter_8bit[(sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS]);
    200  __m128i tmp_1 = _mm_loadl_epi64(
    201      (__m128i *)&av1_filter_8bit[(sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS]);
    202  __m128i tmp_2 = _mm_loadl_epi64(
    203      (__m128i *)&av1_filter_8bit[(sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS]);
    204  __m128i tmp_3 = _mm_loadl_epi64(
    205      (__m128i *)&av1_filter_8bit[(sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS]);
    206  __m128i tmp_4 = _mm_loadl_epi64(
    207      (__m128i *)&av1_filter_8bit[(sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS]);
    208  __m128i tmp_5 = _mm_loadl_epi64(
    209      (__m128i *)&av1_filter_8bit[(sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS]);
    210  __m128i tmp_6 = _mm_loadl_epi64(
    211      (__m128i *)&av1_filter_8bit[(sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS]);
    212  __m128i tmp_7 = _mm_loadl_epi64(
    213      (__m128i *)&av1_filter_8bit[(sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS]);
    214 
    215  tmp_0 = _mm_unpacklo_epi16(tmp_0, tmp_2);
    216  tmp_1 = _mm_unpacklo_epi16(tmp_1, tmp_3);
    217  tmp_4 = _mm_unpacklo_epi16(tmp_4, tmp_6);
    218  tmp_5 = _mm_unpacklo_epi16(tmp_5, tmp_7);
    219 
    220  const __m256i tmp_12 = _mm256_broadcastsi128_si256(tmp_0);
    221  const __m256i tmp_13 = _mm256_broadcastsi128_si256(tmp_1);
    222  const __m256i tmp_14 = _mm256_broadcastsi128_si256(tmp_4);
    223  const __m256i tmp_15 = _mm256_broadcastsi128_si256(tmp_5);
    224 
    225  const __m256i res_0 = _mm256_unpacklo_epi32(tmp_12, tmp_14);
    226  const __m256i res_1 = _mm256_unpackhi_epi32(tmp_12, tmp_14);
    227  const __m256i res_2 = _mm256_unpacklo_epi32(tmp_13, tmp_15);
    228  const __m256i res_3 = _mm256_unpackhi_epi32(tmp_13, tmp_15);
    229 
    230  coeff[0] = _mm256_unpacklo_epi64(res_0, res_2);
    231  coeff[1] = _mm256_unpackhi_epi64(res_0, res_2);
    232  coeff[2] = _mm256_unpacklo_epi64(res_1, res_3);
    233  coeff[3] = _mm256_unpackhi_epi64(res_1, res_3);
    234 }
    235 
    236 static inline void prepare_horizontal_filter_coeff_alpha0_avx2(int beta, int sx,
    237                                                               __m256i *coeff) {
    238  const __m128i tmp_0 =
    239      _mm_loadl_epi64((__m128i *)&av1_filter_8bit[sx >> WARPEDDIFF_PREC_BITS]);
    240  const __m128i tmp_1 = _mm_loadl_epi64(
    241      (__m128i *)&av1_filter_8bit[(sx + beta) >> WARPEDDIFF_PREC_BITS]);
    242 
    243  const __m256i res_0 =
    244      _mm256_inserti128_si256(_mm256_castsi128_si256(tmp_0), tmp_1, 0x1);
    245 
    246  coeff[0] = _mm256_shuffle_epi8(
    247      res_0, _mm256_load_si256((__m256i *)shuffle_alpha0_mask01_avx2));
    248  coeff[1] = _mm256_shuffle_epi8(
    249      res_0, _mm256_load_si256((__m256i *)shuffle_alpha0_mask23_avx2));
    250  coeff[2] = _mm256_shuffle_epi8(
    251      res_0, _mm256_load_si256((__m256i *)shuffle_alpha0_mask45_avx2));
    252  coeff[3] = _mm256_shuffle_epi8(
    253      res_0, _mm256_load_si256((__m256i *)shuffle_alpha0_mask67_avx2));
    254 }
    255 
    256 static inline void horizontal_filter_avx2(const __m256i src, __m256i *horz_out,
    257                                          int sx, int alpha, int beta, int row,
    258                                          const __m256i *shuffle_src,
    259                                          const __m256i *round_const,
    260                                          const __m128i *shift) {
    261  __m256i coeff[4];
    262  prepare_horizontal_filter_coeff_avx2(alpha, beta, sx, coeff);
    263  filter_src_pixels_avx2(src, horz_out, coeff, shuffle_src, round_const, shift,
    264                         row);
    265 }
    266 static inline void prepare_horizontal_filter_coeff(int alpha, int sx,
    267                                                   __m256i *coeff) {
    268  const __m128i tmp_0 = _mm_loadl_epi64(
    269      (__m128i *)&av1_filter_8bit[(sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS]);
    270  const __m128i tmp_1 = _mm_loadl_epi64(
    271      (__m128i *)&av1_filter_8bit[(sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS]);
    272  const __m128i tmp_2 = _mm_loadl_epi64(
    273      (__m128i *)&av1_filter_8bit[(sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS]);
    274  const __m128i tmp_3 = _mm_loadl_epi64(
    275      (__m128i *)&av1_filter_8bit[(sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS]);
    276  const __m128i tmp_4 = _mm_loadl_epi64(
    277      (__m128i *)&av1_filter_8bit[(sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS]);
    278  const __m128i tmp_5 = _mm_loadl_epi64(
    279      (__m128i *)&av1_filter_8bit[(sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS]);
    280  const __m128i tmp_6 = _mm_loadl_epi64(
    281      (__m128i *)&av1_filter_8bit[(sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS]);
    282  const __m128i tmp_7 = _mm_loadl_epi64(
    283      (__m128i *)&av1_filter_8bit[(sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS]);
    284 
    285  const __m128i tmp_8 = _mm_unpacklo_epi16(tmp_0, tmp_2);
    286  const __m128i tmp_9 = _mm_unpacklo_epi16(tmp_1, tmp_3);
    287  const __m128i tmp_10 = _mm_unpacklo_epi16(tmp_4, tmp_6);
    288  const __m128i tmp_11 = _mm_unpacklo_epi16(tmp_5, tmp_7);
    289 
    290  const __m128i tmp_12 = _mm_unpacklo_epi32(tmp_8, tmp_10);
    291  const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_8, tmp_10);
    292  const __m128i tmp_14 = _mm_unpacklo_epi32(tmp_9, tmp_11);
    293  const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_9, tmp_11);
    294 
    295  coeff[0] = _mm256_castsi128_si256(_mm_unpacklo_epi64(tmp_12, tmp_14));
    296  coeff[1] = _mm256_castsi128_si256(_mm_unpackhi_epi64(tmp_12, tmp_14));
    297  coeff[2] = _mm256_castsi128_si256(_mm_unpacklo_epi64(tmp_13, tmp_15));
    298  coeff[3] = _mm256_castsi128_si256(_mm_unpackhi_epi64(tmp_13, tmp_15));
    299 }
    300 
    301 static inline void warp_horizontal_filter_avx2(
    302    const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
    303    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
    304    const __m256i *round_const, const __m128i *shift,
    305    const __m256i *shuffle_src) {
    306  int k, iy, sx, row = 0;
    307  __m256i coeff[4];
    308  for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
    309    iy = iy4 + k;
    310    iy = clamp(iy, 0, height - 1);
    311    const __m128i src_0 =
    312        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
    313    iy = iy4 + k + 1;
    314    iy = clamp(iy, 0, height - 1);
    315    const __m128i src_1 =
    316        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
    317    const __m256i src_01 =
    318        _mm256_inserti128_si256(_mm256_castsi128_si256(src_0), src_1, 0x1);
    319    sx = sx4 + beta * (k + 4);
    320    horizontal_filter_avx2(src_01, horz_out, sx, alpha, beta, row, shuffle_src,
    321                           round_const, shift);
    322    row += 1;
    323  }
    324  iy = iy4 + k;
    325  iy = clamp(iy, 0, height - 1);
    326  const __m256i src_01 = _mm256_castsi128_si256(
    327      _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7)));
    328  sx = sx4 + beta * (k + 4);
    329  prepare_horizontal_filter_coeff(alpha, sx, coeff);
    330  filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
    331                         shift, row);
    332 }
    333 
    334 static inline void warp_horizontal_filter_alpha0_avx2(
    335    const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
    336    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
    337    const __m256i *round_const, const __m128i *shift,
    338    const __m256i *shuffle_src) {
    339  (void)alpha;
    340  int k, iy, sx, row = 0;
    341  __m256i coeff[4];
    342  for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
    343    iy = iy4 + k;
    344    iy = clamp(iy, 0, height - 1);
    345    const __m128i src_0 =
    346        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
    347    iy = iy4 + k + 1;
    348    iy = clamp(iy, 0, height - 1);
    349    const __m128i src_1 =
    350        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
    351    const __m256i src_01 =
    352        _mm256_inserti128_si256(_mm256_castsi128_si256(src_0), src_1, 0x1);
    353    sx = sx4 + beta * (k + 4);
    354    prepare_horizontal_filter_coeff_alpha0_avx2(beta, sx, coeff);
    355    filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
    356                           shift, row);
    357    row += 1;
    358  }
    359  iy = iy4 + k;
    360  iy = clamp(iy, 0, height - 1);
    361  const __m256i src_01 = _mm256_castsi128_si256(
    362      _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7)));
    363  sx = sx4 + beta * (k + 4);
    364  prepare_horizontal_filter_coeff_alpha0_avx2(beta, sx, coeff);
    365  filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
    366                         shift, row);
    367 }
    368 
    369 static inline void warp_horizontal_filter_beta0_avx2(
    370    const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
    371    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
    372    const __m256i *round_const, const __m128i *shift,
    373    const __m256i *shuffle_src) {
    374  (void)beta;
    375  int k, iy, row = 0;
    376  __m256i coeff[4];
    377  prepare_horizontal_filter_coeff_beta0_avx2(alpha, sx4, coeff);
    378  for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
    379    iy = iy4 + k;
    380    iy = clamp(iy, 0, height - 1);
    381    const __m128i src_0 =
    382        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
    383    iy = iy4 + k + 1;
    384    iy = clamp(iy, 0, height - 1);
    385    const __m128i src_1 =
    386        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
    387    const __m256i src_01 =
    388        _mm256_inserti128_si256(_mm256_castsi128_si256(src_0), src_1, 0x1);
    389    filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
    390                           shift, row);
    391    row += 1;
    392  }
    393  iy = iy4 + k;
    394  iy = clamp(iy, 0, height - 1);
    395  const __m256i src_01 = _mm256_castsi128_si256(
    396      _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7)));
    397  filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
    398                         shift, row);
    399 }
    400 
    401 static inline void warp_horizontal_filter_alpha0_beta0_avx2(
    402    const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
    403    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
    404    const __m256i *round_const, const __m128i *shift,
    405    const __m256i *shuffle_src) {
    406  (void)alpha;
    407  int k, iy, row = 0;
    408  __m256i coeff[4];
    409  prepare_horizontal_filter_coeff_alpha0_avx2(beta, sx4, coeff);
    410  for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
    411    iy = iy4 + k;
    412    iy = clamp(iy, 0, height - 1);
    413    const __m128i src0 =
    414        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
    415    iy = iy4 + k + 1;
    416    iy = clamp(iy, 0, height - 1);
    417    const __m128i src1 =
    418        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
    419    const __m256i src_01 =
    420        _mm256_inserti128_si256(_mm256_castsi128_si256(src0), src1, 0x1);
    421    filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
    422                           shift, row);
    423    row += 1;
    424  }
    425  iy = iy4 + k;
    426  iy = clamp(iy, 0, height - 1);
    427  const __m256i src_01 = _mm256_castsi128_si256(
    428      _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7)));
    429  filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
    430                         shift, row);
    431 }
    432 
    433 static inline void unpack_weights_and_set_round_const_avx2(
    434    ConvolveParams *conv_params, const int round_bits, const int offset_bits,
    435    __m256i *res_sub_const, __m256i *round_bits_const, __m256i *wt) {
    436  *res_sub_const =
    437      _mm256_set1_epi16(-(1 << (offset_bits - conv_params->round_1)) -
    438                        (1 << (offset_bits - conv_params->round_1 - 1)));
    439  *round_bits_const = _mm256_set1_epi16(((1 << round_bits) >> 1));
    440 
    441  const int w0 = conv_params->fwd_offset;
    442  const int w1 = conv_params->bck_offset;
    443  const __m256i wt0 = _mm256_set1_epi16((short)w0);
    444  const __m256i wt1 = _mm256_set1_epi16((short)w1);
    445  *wt = _mm256_unpacklo_epi16(wt0, wt1);
    446 }
    447 
    448 static inline void prepare_vertical_filter_coeffs_avx2(int gamma, int delta,
    449                                                       int sy,
    450                                                       __m256i *coeffs) {
    451  __m128i filt_00 =
    452      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    453                                  ((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
    454  __m128i filt_01 =
    455      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    456                                  ((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
    457  __m128i filt_02 =
    458      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    459                                  ((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
    460  __m128i filt_03 =
    461      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    462                                  ((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
    463 
    464  __m128i filt_10 = _mm_loadu_si128(
    465      (__m128i *)(av1_warped_filter +
    466                  (((sy + delta) + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
    467  __m128i filt_11 = _mm_loadu_si128(
    468      (__m128i *)(av1_warped_filter +
    469                  (((sy + delta) + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
    470  __m128i filt_12 = _mm_loadu_si128(
    471      (__m128i *)(av1_warped_filter +
    472                  (((sy + delta) + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
    473  __m128i filt_13 = _mm_loadu_si128(
    474      (__m128i *)(av1_warped_filter +
    475                  (((sy + delta) + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
    476 
    477  __m256i filt_0 =
    478      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_00), filt_10, 0x1);
    479  __m256i filt_1 =
    480      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_01), filt_11, 0x1);
    481  __m256i filt_2 =
    482      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_02), filt_12, 0x1);
    483  __m256i filt_3 =
    484      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_03), filt_13, 0x1);
    485 
    486  __m256i res_0 = _mm256_unpacklo_epi32(filt_0, filt_1);
    487  __m256i res_1 = _mm256_unpacklo_epi32(filt_2, filt_3);
    488  __m256i res_2 = _mm256_unpackhi_epi32(filt_0, filt_1);
    489  __m256i res_3 = _mm256_unpackhi_epi32(filt_2, filt_3);
    490 
    491  coeffs[0] = _mm256_unpacklo_epi64(res_0, res_1);
    492  coeffs[1] = _mm256_unpackhi_epi64(res_0, res_1);
    493  coeffs[2] = _mm256_unpacklo_epi64(res_2, res_3);
    494  coeffs[3] = _mm256_unpackhi_epi64(res_2, res_3);
    495 
    496  filt_00 =
    497      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    498                                  ((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
    499  filt_01 =
    500      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    501                                  ((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
    502  filt_02 =
    503      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    504                                  ((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
    505  filt_03 =
    506      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    507                                  ((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
    508 
    509  filt_10 = _mm_loadu_si128(
    510      (__m128i *)(av1_warped_filter +
    511                  (((sy + delta) + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
    512  filt_11 = _mm_loadu_si128(
    513      (__m128i *)(av1_warped_filter +
    514                  (((sy + delta) + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
    515  filt_12 = _mm_loadu_si128(
    516      (__m128i *)(av1_warped_filter +
    517                  (((sy + delta) + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
    518  filt_13 = _mm_loadu_si128(
    519      (__m128i *)(av1_warped_filter +
    520                  (((sy + delta) + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
    521 
    522  filt_0 =
    523      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_00), filt_10, 0x1);
    524  filt_1 =
    525      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_01), filt_11, 0x1);
    526  filt_2 =
    527      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_02), filt_12, 0x1);
    528  filt_3 =
    529      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_03), filt_13, 0x1);
    530 
    531  res_0 = _mm256_unpacklo_epi32(filt_0, filt_1);
    532  res_1 = _mm256_unpacklo_epi32(filt_2, filt_3);
    533  res_2 = _mm256_unpackhi_epi32(filt_0, filt_1);
    534  res_3 = _mm256_unpackhi_epi32(filt_2, filt_3);
    535 
    536  coeffs[4] = _mm256_unpacklo_epi64(res_0, res_1);
    537  coeffs[5] = _mm256_unpackhi_epi64(res_0, res_1);
    538  coeffs[6] = _mm256_unpacklo_epi64(res_2, res_3);
    539  coeffs[7] = _mm256_unpackhi_epi64(res_2, res_3);
    540 }
    541 
    542 static inline void prepare_vertical_filter_coeffs_delta0_avx2(int gamma, int sy,
    543                                                              __m256i *coeffs) {
    544  __m128i filt_00 =
    545      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    546                                  ((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
    547  __m128i filt_01 =
    548      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    549                                  ((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
    550  __m128i filt_02 =
    551      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    552                                  ((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
    553  __m128i filt_03 =
    554      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    555                                  ((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
    556 
    557  __m256i filt_0 = _mm256_broadcastsi128_si256(filt_00);
    558  __m256i filt_1 = _mm256_broadcastsi128_si256(filt_01);
    559  __m256i filt_2 = _mm256_broadcastsi128_si256(filt_02);
    560  __m256i filt_3 = _mm256_broadcastsi128_si256(filt_03);
    561 
    562  __m256i res_0 = _mm256_unpacklo_epi32(filt_0, filt_1);
    563  __m256i res_1 = _mm256_unpacklo_epi32(filt_2, filt_3);
    564  __m256i res_2 = _mm256_unpackhi_epi32(filt_0, filt_1);
    565  __m256i res_3 = _mm256_unpackhi_epi32(filt_2, filt_3);
    566 
    567  coeffs[0] = _mm256_unpacklo_epi64(res_0, res_1);
    568  coeffs[1] = _mm256_unpackhi_epi64(res_0, res_1);
    569  coeffs[2] = _mm256_unpacklo_epi64(res_2, res_3);
    570  coeffs[3] = _mm256_unpackhi_epi64(res_2, res_3);
    571 
    572  filt_00 =
    573      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    574                                  ((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
    575  filt_01 =
    576      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    577                                  ((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
    578  filt_02 =
    579      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    580                                  ((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
    581  filt_03 =
    582      _mm_loadu_si128((__m128i *)(av1_warped_filter +
    583                                  ((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
    584 
    585  filt_0 = _mm256_broadcastsi128_si256(filt_00);
    586  filt_1 = _mm256_broadcastsi128_si256(filt_01);
    587  filt_2 = _mm256_broadcastsi128_si256(filt_02);
    588  filt_3 = _mm256_broadcastsi128_si256(filt_03);
    589 
    590  res_0 = _mm256_unpacklo_epi32(filt_0, filt_1);
    591  res_1 = _mm256_unpacklo_epi32(filt_2, filt_3);
    592  res_2 = _mm256_unpackhi_epi32(filt_0, filt_1);
    593  res_3 = _mm256_unpackhi_epi32(filt_2, filt_3);
    594 
    595  coeffs[4] = _mm256_unpacklo_epi64(res_0, res_1);
    596  coeffs[5] = _mm256_unpackhi_epi64(res_0, res_1);
    597  coeffs[6] = _mm256_unpacklo_epi64(res_2, res_3);
    598  coeffs[7] = _mm256_unpackhi_epi64(res_2, res_3);
    599 }
    600 
    601 static inline void prepare_vertical_filter_coeffs_gamma0_avx2(int delta, int sy,
    602                                                              __m256i *coeffs) {
    603  const __m128i filt_0 = _mm_loadu_si128(
    604      (__m128i *)(av1_warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
    605  const __m128i filt_1 = _mm_loadu_si128(
    606      (__m128i *)(av1_warped_filter + ((sy + delta) >> WARPEDDIFF_PREC_BITS)));
    607 
    608  __m256i res_0 =
    609      _mm256_inserti128_si256(_mm256_castsi128_si256(filt_0), filt_1, 0x1);
    610 
    611  coeffs[0] = _mm256_shuffle_epi8(
    612      res_0, _mm256_load_si256((__m256i *)shuffle_gamma0_mask0_avx2));
    613  coeffs[1] = _mm256_shuffle_epi8(
    614      res_0, _mm256_load_si256((__m256i *)shuffle_gamma0_mask1_avx2));
    615  coeffs[2] = _mm256_shuffle_epi8(
    616      res_0, _mm256_load_si256((__m256i *)shuffle_gamma0_mask2_avx2));
    617  coeffs[3] = _mm256_shuffle_epi8(
    618      res_0, _mm256_load_si256((__m256i *)shuffle_gamma0_mask3_avx2));
    619 
    620  coeffs[4] = coeffs[0];
    621  coeffs[5] = coeffs[1];
    622  coeffs[6] = coeffs[2];
    623  coeffs[7] = coeffs[3];
    624 }
    625 
    626 static inline void filter_src_pixels_vertical_avx2(__m256i *horz_out,
    627                                                   __m256i *src,
    628                                                   __m256i *coeffs,
    629                                                   __m256i *res_lo,
    630                                                   __m256i *res_hi, int row) {
    631  const __m256i src_6 = horz_out[row + 3];
    632  const __m256i src_7 =
    633      _mm256_permute2x128_si256(horz_out[row + 3], horz_out[row + 4], 0x21);
    634 
    635  src[6] = _mm256_unpacklo_epi16(src_6, src_7);
    636 
    637  const __m256i res_0 = _mm256_madd_epi16(src[0], coeffs[0]);
    638  const __m256i res_2 = _mm256_madd_epi16(src[2], coeffs[1]);
    639  const __m256i res_4 = _mm256_madd_epi16(src[4], coeffs[2]);
    640  const __m256i res_6 = _mm256_madd_epi16(src[6], coeffs[3]);
    641 
    642  const __m256i res_even = _mm256_add_epi32(_mm256_add_epi32(res_0, res_2),
    643                                            _mm256_add_epi32(res_4, res_6));
    644 
    645  src[7] = _mm256_unpackhi_epi16(src_6, src_7);
    646 
    647  const __m256i res_1 = _mm256_madd_epi16(src[1], coeffs[4]);
    648  const __m256i res_3 = _mm256_madd_epi16(src[3], coeffs[5]);
    649  const __m256i res_5 = _mm256_madd_epi16(src[5], coeffs[6]);
    650  const __m256i res_7 = _mm256_madd_epi16(src[7], coeffs[7]);
    651 
    652  const __m256i res_odd = _mm256_add_epi32(_mm256_add_epi32(res_1, res_3),
    653                                           _mm256_add_epi32(res_5, res_7));
    654 
    655  // Rearrange pixels back into the order 0 ... 7
    656  *res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
    657  *res_hi = _mm256_unpackhi_epi32(res_even, res_odd);
    658 }
    659 
    660 static inline void store_vertical_filter_output_avx2(
    661    const __m256i *res_lo, const __m256i *res_hi, const __m256i *res_add_const,
    662    const __m256i *wt, const __m256i *res_sub_const,
    663    const __m256i *round_bits_const, uint8_t *pred, ConvolveParams *conv_params,
    664    int i, int j, int k, const int reduce_bits_vert, int p_stride, int p_width,
    665    const int round_bits) {
    666  __m256i res_lo_1 = *res_lo;
    667  __m256i res_hi_1 = *res_hi;
    668 
    669  if (conv_params->is_compound) {
    670    __m128i *const p_0 =
    671        (__m128i *)&conv_params->dst[(i + k + 4) * conv_params->dst_stride + j];
    672    __m128i *const p_1 =
    673        (__m128i *)&conv_params
    674            ->dst[(i + (k + 1) + 4) * conv_params->dst_stride + j];
    675 
    676    res_lo_1 = _mm256_srai_epi32(_mm256_add_epi32(res_lo_1, *res_add_const),
    677                                 reduce_bits_vert);
    678 
    679    const __m256i temp_lo_16 = _mm256_packus_epi32(res_lo_1, res_lo_1);
    680    __m256i res_lo_16;
    681    if (conv_params->do_average) {
    682      __m128i *const dst8_0 = (__m128i *)&pred[(i + k + 4) * p_stride + j];
    683      __m128i *const dst8_1 =
    684          (__m128i *)&pred[(i + (k + 1) + 4) * p_stride + j];
    685      const __m128i p_16_0 = _mm_loadl_epi64(p_0);
    686      const __m128i p_16_1 = _mm_loadl_epi64(p_1);
    687      const __m256i p_16 =
    688          _mm256_inserti128_si256(_mm256_castsi128_si256(p_16_0), p_16_1, 1);
    689      if (conv_params->use_dist_wtd_comp_avg) {
    690        const __m256i p_16_lo = _mm256_unpacklo_epi16(p_16, temp_lo_16);
    691        const __m256i wt_res_lo = _mm256_madd_epi16(p_16_lo, *wt);
    692        const __m256i shifted_32 =
    693            _mm256_srai_epi32(wt_res_lo, DIST_PRECISION_BITS);
    694        res_lo_16 = _mm256_packus_epi32(shifted_32, shifted_32);
    695      } else {
    696        res_lo_16 = _mm256_srai_epi16(_mm256_add_epi16(p_16, temp_lo_16), 1);
    697      }
    698      res_lo_16 = _mm256_add_epi16(res_lo_16, *res_sub_const);
    699      res_lo_16 = _mm256_srai_epi16(
    700          _mm256_add_epi16(res_lo_16, *round_bits_const), round_bits);
    701      const __m256i res_8_lo = _mm256_packus_epi16(res_lo_16, res_lo_16);
    702      const __m128i res_8_lo_0 = _mm256_castsi256_si128(res_8_lo);
    703      const __m128i res_8_lo_1 = _mm256_extracti128_si256(res_8_lo, 1);
    704      *(int *)dst8_0 = _mm_cvtsi128_si32(res_8_lo_0);
    705      *(int *)dst8_1 = _mm_cvtsi128_si32(res_8_lo_1);
    706    } else {
    707      const __m128i temp_lo_16_0 = _mm256_castsi256_si128(temp_lo_16);
    708      const __m128i temp_lo_16_1 = _mm256_extracti128_si256(temp_lo_16, 1);
    709      _mm_storel_epi64(p_0, temp_lo_16_0);
    710      _mm_storel_epi64(p_1, temp_lo_16_1);
    711    }
    712    if (p_width > 4) {
    713      __m128i *const p4_0 =
    714          (__m128i *)&conv_params
    715              ->dst[(i + k + 4) * conv_params->dst_stride + j + 4];
    716      __m128i *const p4_1 =
    717          (__m128i *)&conv_params
    718              ->dst[(i + (k + 1) + 4) * conv_params->dst_stride + j + 4];
    719      res_hi_1 = _mm256_srai_epi32(_mm256_add_epi32(res_hi_1, *res_add_const),
    720                                   reduce_bits_vert);
    721      const __m256i temp_hi_16 = _mm256_packus_epi32(res_hi_1, res_hi_1);
    722      __m256i res_hi_16;
    723      if (conv_params->do_average) {
    724        __m128i *const dst8_4_0 =
    725            (__m128i *)&pred[(i + k + 4) * p_stride + j + 4];
    726        __m128i *const dst8_4_1 =
    727            (__m128i *)&pred[(i + (k + 1) + 4) * p_stride + j + 4];
    728        const __m128i p4_16_0 = _mm_loadl_epi64(p4_0);
    729        const __m128i p4_16_1 = _mm_loadl_epi64(p4_1);
    730        const __m256i p4_16 = _mm256_inserti128_si256(
    731            _mm256_castsi128_si256(p4_16_0), p4_16_1, 1);
    732        if (conv_params->use_dist_wtd_comp_avg) {
    733          const __m256i p_16_hi = _mm256_unpacklo_epi16(p4_16, temp_hi_16);
    734          const __m256i wt_res_hi = _mm256_madd_epi16(p_16_hi, *wt);
    735          const __m256i shifted_32 =
    736              _mm256_srai_epi32(wt_res_hi, DIST_PRECISION_BITS);
    737          res_hi_16 = _mm256_packus_epi32(shifted_32, shifted_32);
    738        } else {
    739          res_hi_16 = _mm256_srai_epi16(_mm256_add_epi16(p4_16, temp_hi_16), 1);
    740        }
    741        res_hi_16 = _mm256_add_epi16(res_hi_16, *res_sub_const);
    742        res_hi_16 = _mm256_srai_epi16(
    743            _mm256_add_epi16(res_hi_16, *round_bits_const), round_bits);
    744        __m256i res_8_hi = _mm256_packus_epi16(res_hi_16, res_hi_16);
    745        const __m128i res_8_hi_0 = _mm256_castsi256_si128(res_8_hi);
    746        const __m128i res_8_hi_1 = _mm256_extracti128_si256(res_8_hi, 1);
    747        *(int *)dst8_4_0 = _mm_cvtsi128_si32(res_8_hi_0);
    748        *(int *)dst8_4_1 = _mm_cvtsi128_si32(res_8_hi_1);
    749      } else {
    750        const __m128i temp_hi_16_0 = _mm256_castsi256_si128(temp_hi_16);
    751        const __m128i temp_hi_16_1 = _mm256_extracti128_si256(temp_hi_16, 1);
    752        _mm_storel_epi64(p4_0, temp_hi_16_0);
    753        _mm_storel_epi64(p4_1, temp_hi_16_1);
    754      }
    755    }
    756  } else {
    757    const __m256i res_lo_round = _mm256_srai_epi32(
    758        _mm256_add_epi32(res_lo_1, *res_add_const), reduce_bits_vert);
    759    const __m256i res_hi_round = _mm256_srai_epi32(
    760        _mm256_add_epi32(res_hi_1, *res_add_const), reduce_bits_vert);
    761 
    762    const __m256i res_16bit = _mm256_packs_epi32(res_lo_round, res_hi_round);
    763    const __m256i res_8bit = _mm256_packus_epi16(res_16bit, res_16bit);
    764    const __m128i res_8bit0 = _mm256_castsi256_si128(res_8bit);
    765    const __m128i res_8bit1 = _mm256_extracti128_si256(res_8bit, 1);
    766 
    767    // Store, blending with 'pred' if needed
    768    __m128i *const p = (__m128i *)&pred[(i + k + 4) * p_stride + j];
    769    __m128i *const p1 = (__m128i *)&pred[(i + (k + 1) + 4) * p_stride + j];
    770 
    771    if (p_width == 4) {
    772      *(int *)p = _mm_cvtsi128_si32(res_8bit0);
    773      *(int *)p1 = _mm_cvtsi128_si32(res_8bit1);
    774    } else {
    775      _mm_storel_epi64(p, res_8bit0);
    776      _mm_storel_epi64(p1, res_8bit1);
    777    }
    778  }
    779 }
    780 
    781 static inline void warp_vertical_filter_avx2(
    782    uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
    783    int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
    784    int i, int j, int sy4, const int reduce_bits_vert,
    785    const __m256i *res_add_const, const int round_bits,
    786    const __m256i *res_sub_const, const __m256i *round_bits_const,
    787    const __m256i *wt) {
    788  int k, row = 0;
    789  __m256i src[8];
    790  const __m256i src_0 = horz_out[0];
    791  const __m256i src_1 =
    792      _mm256_permute2x128_si256(horz_out[0], horz_out[1], 0x21);
    793  const __m256i src_2 = horz_out[1];
    794  const __m256i src_3 =
    795      _mm256_permute2x128_si256(horz_out[1], horz_out[2], 0x21);
    796  const __m256i src_4 = horz_out[2];
    797  const __m256i src_5 =
    798      _mm256_permute2x128_si256(horz_out[2], horz_out[3], 0x21);
    799 
    800  src[0] = _mm256_unpacklo_epi16(src_0, src_1);
    801  src[2] = _mm256_unpacklo_epi16(src_2, src_3);
    802  src[4] = _mm256_unpacklo_epi16(src_4, src_5);
    803 
    804  src[1] = _mm256_unpackhi_epi16(src_0, src_1);
    805  src[3] = _mm256_unpackhi_epi16(src_2, src_3);
    806  src[5] = _mm256_unpackhi_epi16(src_4, src_5);
    807 
    808  for (k = -4; k < AOMMIN(4, p_height - i - 4); k += 2) {
    809    int sy = sy4 + delta * (k + 4);
    810    __m256i coeffs[8];
    811    prepare_vertical_filter_coeffs_avx2(gamma, delta, sy, coeffs);
    812    __m256i res_lo, res_hi;
    813    filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi,
    814                                    row);
    815    store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt,
    816                                      res_sub_const, round_bits_const, pred,
    817                                      conv_params, i, j, k, reduce_bits_vert,
    818                                      p_stride, p_width, round_bits);
    819    src[0] = src[2];
    820    src[2] = src[4];
    821    src[4] = src[6];
    822    src[1] = src[3];
    823    src[3] = src[5];
    824    src[5] = src[7];
    825 
    826    row += 1;
    827  }
    828 }
    829 
    830 static inline void warp_vertical_filter_gamma0_avx2(
    831    uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
    832    int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
    833    int i, int j, int sy4, const int reduce_bits_vert,
    834    const __m256i *res_add_const, const int round_bits,
    835    const __m256i *res_sub_const, const __m256i *round_bits_const,
    836    const __m256i *wt) {
    837  (void)gamma;
    838  int k, row = 0;
    839  __m256i src[8];
    840  const __m256i src_0 = horz_out[0];
    841  const __m256i src_1 =
    842      _mm256_permute2x128_si256(horz_out[0], horz_out[1], 0x21);
    843  const __m256i src_2 = horz_out[1];
    844  const __m256i src_3 =
    845      _mm256_permute2x128_si256(horz_out[1], horz_out[2], 0x21);
    846  const __m256i src_4 = horz_out[2];
    847  const __m256i src_5 =
    848      _mm256_permute2x128_si256(horz_out[2], horz_out[3], 0x21);
    849 
    850  src[0] = _mm256_unpacklo_epi16(src_0, src_1);
    851  src[2] = _mm256_unpacklo_epi16(src_2, src_3);
    852  src[4] = _mm256_unpacklo_epi16(src_4, src_5);
    853 
    854  src[1] = _mm256_unpackhi_epi16(src_0, src_1);
    855  src[3] = _mm256_unpackhi_epi16(src_2, src_3);
    856  src[5] = _mm256_unpackhi_epi16(src_4, src_5);
    857 
    858  for (k = -4; k < AOMMIN(4, p_height - i - 4); k += 2) {
    859    int sy = sy4 + delta * (k + 4);
    860    __m256i coeffs[8];
    861    prepare_vertical_filter_coeffs_gamma0_avx2(delta, sy, coeffs);
    862    __m256i res_lo, res_hi;
    863    filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi,
    864                                    row);
    865    store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt,
    866                                      res_sub_const, round_bits_const, pred,
    867                                      conv_params, i, j, k, reduce_bits_vert,
    868                                      p_stride, p_width, round_bits);
    869    src[0] = src[2];
    870    src[2] = src[4];
    871    src[4] = src[6];
    872    src[1] = src[3];
    873    src[3] = src[5];
    874    src[5] = src[7];
    875    row += 1;
    876  }
    877 }
    878 
    879 static inline void warp_vertical_filter_delta0_avx2(
    880    uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
    881    int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
    882    int i, int j, int sy4, const int reduce_bits_vert,
    883    const __m256i *res_add_const, const int round_bits,
    884    const __m256i *res_sub_const, const __m256i *round_bits_const,
    885    const __m256i *wt) {
    886  (void)delta;
    887  int k, row = 0;
    888  __m256i src[8], coeffs[8];
    889  const __m256i src_0 = horz_out[0];
    890  const __m256i src_1 =
    891      _mm256_permute2x128_si256(horz_out[0], horz_out[1], 0x21);
    892  const __m256i src_2 = horz_out[1];
    893  const __m256i src_3 =
    894      _mm256_permute2x128_si256(horz_out[1], horz_out[2], 0x21);
    895  const __m256i src_4 = horz_out[2];
    896  const __m256i src_5 =
    897      _mm256_permute2x128_si256(horz_out[2], horz_out[3], 0x21);
    898 
    899  src[0] = _mm256_unpacklo_epi16(src_0, src_1);
    900  src[2] = _mm256_unpacklo_epi16(src_2, src_3);
    901  src[4] = _mm256_unpacklo_epi16(src_4, src_5);
    902 
    903  src[1] = _mm256_unpackhi_epi16(src_0, src_1);
    904  src[3] = _mm256_unpackhi_epi16(src_2, src_3);
    905  src[5] = _mm256_unpackhi_epi16(src_4, src_5);
    906 
    907  prepare_vertical_filter_coeffs_delta0_avx2(gamma, sy4, coeffs);
    908 
    909  for (k = -4; k < AOMMIN(4, p_height - i - 4); k += 2) {
    910    __m256i res_lo, res_hi;
    911    filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi,
    912                                    row);
    913    store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt,
    914                                      res_sub_const, round_bits_const, pred,
    915                                      conv_params, i, j, k, reduce_bits_vert,
    916                                      p_stride, p_width, round_bits);
    917    src[0] = src[2];
    918    src[2] = src[4];
    919    src[4] = src[6];
    920    src[1] = src[3];
    921    src[3] = src[5];
    922    src[5] = src[7];
    923    row += 1;
    924  }
    925 }
    926 
    927 static inline void warp_vertical_filter_gamma0_delta0_avx2(
    928    uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
    929    int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
    930    int i, int j, int sy4, const int reduce_bits_vert,
    931    const __m256i *res_add_const, const int round_bits,
    932    const __m256i *res_sub_const, const __m256i *round_bits_const,
    933    const __m256i *wt) {
    934  (void)gamma;
    935  int k, row = 0;
    936  __m256i src[8], coeffs[8];
    937  const __m256i src_0 = horz_out[0];
    938  const __m256i src_1 =
    939      _mm256_permute2x128_si256(horz_out[0], horz_out[1], 0x21);
    940  const __m256i src_2 = horz_out[1];
    941  const __m256i src_3 =
    942      _mm256_permute2x128_si256(horz_out[1], horz_out[2], 0x21);
    943  const __m256i src_4 = horz_out[2];
    944  const __m256i src_5 =
    945      _mm256_permute2x128_si256(horz_out[2], horz_out[3], 0x21);
    946 
    947  src[0] = _mm256_unpacklo_epi16(src_0, src_1);
    948  src[2] = _mm256_unpacklo_epi16(src_2, src_3);
    949  src[4] = _mm256_unpacklo_epi16(src_4, src_5);
    950 
    951  src[1] = _mm256_unpackhi_epi16(src_0, src_1);
    952  src[3] = _mm256_unpackhi_epi16(src_2, src_3);
    953  src[5] = _mm256_unpackhi_epi16(src_4, src_5);
    954 
    955  prepare_vertical_filter_coeffs_gamma0_avx2(delta, sy4, coeffs);
    956 
    957  for (k = -4; k < AOMMIN(4, p_height - i - 4); k += 2) {
    958    __m256i res_lo, res_hi;
    959    filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi,
    960                                    row);
    961    store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt,
    962                                      res_sub_const, round_bits_const, pred,
    963                                      conv_params, i, j, k, reduce_bits_vert,
    964                                      p_stride, p_width, round_bits);
    965    src[0] = src[2];
    966    src[2] = src[4];
    967    src[4] = src[6];
    968    src[1] = src[3];
    969    src[3] = src[5];
    970    src[5] = src[7];
    971    row += 1;
    972  }
    973 }
    974 
    975 static inline void prepare_warp_vertical_filter_avx2(
    976    uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
    977    int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
    978    int i, int j, int sy4, const int reduce_bits_vert,
    979    const __m256i *res_add_const, const int round_bits,
    980    const __m256i *res_sub_const, const __m256i *round_bits_const,
    981    const __m256i *wt) {
    982  if (gamma == 0 && delta == 0)
    983    warp_vertical_filter_gamma0_delta0_avx2(
    984        pred, horz_out, conv_params, gamma, delta, p_height, p_stride, p_width,
    985        i, j, sy4, reduce_bits_vert, res_add_const, round_bits, res_sub_const,
    986        round_bits_const, wt);
    987  else if (gamma == 0 && delta != 0)
    988    warp_vertical_filter_gamma0_avx2(
    989        pred, horz_out, conv_params, gamma, delta, p_height, p_stride, p_width,
    990        i, j, sy4, reduce_bits_vert, res_add_const, round_bits, res_sub_const,
    991        round_bits_const, wt);
    992  else if (gamma != 0 && delta == 0)
    993    warp_vertical_filter_delta0_avx2(
    994        pred, horz_out, conv_params, gamma, delta, p_height, p_stride, p_width,
    995        i, j, sy4, reduce_bits_vert, res_add_const, round_bits, res_sub_const,
    996        round_bits_const, wt);
    997  else
    998    warp_vertical_filter_avx2(pred, horz_out, conv_params, gamma, delta,
    999                              p_height, p_stride, p_width, i, j, sy4,
   1000                              reduce_bits_vert, res_add_const, round_bits,
   1001                              res_sub_const, round_bits_const, wt);
   1002 }
   1003 
   1004 static inline void prepare_warp_horizontal_filter_avx2(
   1005    const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
   1006    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
   1007    const __m256i *round_const, const __m128i *shift,
   1008    const __m256i *shuffle_src) {
   1009  if (alpha == 0 && beta == 0)
   1010    warp_horizontal_filter_alpha0_beta0_avx2(
   1011        ref, horz_out, stride, ix4, iy4, sx4, alpha, beta, p_height, height, i,
   1012        round_const, shift, shuffle_src);
   1013  else if (alpha == 0 && beta != 0)
   1014    warp_horizontal_filter_alpha0_avx2(ref, horz_out, stride, ix4, iy4, sx4,
   1015                                       alpha, beta, p_height, height, i,
   1016                                       round_const, shift, shuffle_src);
   1017  else if (alpha != 0 && beta == 0)
   1018    warp_horizontal_filter_beta0_avx2(ref, horz_out, stride, ix4, iy4, sx4,
   1019                                      alpha, beta, p_height, height, i,
   1020                                      round_const, shift, shuffle_src);
   1021  else
   1022    warp_horizontal_filter_avx2(ref, horz_out, stride, ix4, iy4, sx4, alpha,
   1023                                beta, p_height, height, i, round_const, shift,
   1024                                shuffle_src);
   1025 }
   1026 
   1027 void av1_warp_affine_avx2(const int32_t *mat, const uint8_t *ref, int width,
   1028                          int height, int stride, uint8_t *pred, int p_col,
   1029                          int p_row, int p_width, int p_height, int p_stride,
   1030                          int subsampling_x, int subsampling_y,
   1031                          ConvolveParams *conv_params, int16_t alpha,
   1032                          int16_t beta, int16_t gamma, int16_t delta) {
   1033  __m256i horz_out[8];
   1034  int i, j, k;
   1035  const int bd = 8;
   1036  const int reduce_bits_horiz = conv_params->round_0;
   1037  const int reduce_bits_vert = conv_params->is_compound
   1038                                   ? conv_params->round_1
   1039                                   : 2 * FILTER_BITS - reduce_bits_horiz;
   1040  const int offset_bits_horiz = bd + FILTER_BITS - 1;
   1041  assert(IMPLIES(conv_params->is_compound, conv_params->dst != NULL));
   1042 
   1043  const int offset_bits_vert = bd + 2 * FILTER_BITS - reduce_bits_horiz;
   1044  const __m256i reduce_bits_vert_const =
   1045      _mm256_set1_epi32(((1 << reduce_bits_vert) >> 1));
   1046  const __m256i res_add_const = _mm256_set1_epi32(1 << offset_bits_vert);
   1047  const int round_bits =
   1048      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
   1049  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
   1050  assert(IMPLIES(conv_params->do_average, conv_params->is_compound));
   1051 
   1052  const __m256i round_const = _mm256_set1_epi16(
   1053      (1 << offset_bits_horiz) + ((1 << reduce_bits_horiz) >> 1));
   1054  const __m128i shift = _mm_cvtsi32_si128(reduce_bits_horiz);
   1055 
   1056  __m256i res_sub_const, round_bits_const, wt;
   1057  unpack_weights_and_set_round_const_avx2(conv_params, round_bits, offset_bits,
   1058                                          &res_sub_const, &round_bits_const,
   1059                                          &wt);
   1060 
   1061  __m256i res_add_const_1;
   1062  if (conv_params->is_compound == 1) {
   1063    res_add_const_1 = _mm256_add_epi32(reduce_bits_vert_const, res_add_const);
   1064  } else {
   1065    res_add_const_1 = _mm256_set1_epi32(-(1 << (bd + reduce_bits_vert - 1)) +
   1066                                        ((1 << reduce_bits_vert) >> 1));
   1067  }
   1068  const int32_t const1 = alpha * (-4) + beta * (-4) +
   1069                         (1 << (WARPEDDIFF_PREC_BITS - 1)) +
   1070                         (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
   1071  const int32_t const2 = gamma * (-4) + delta * (-4) +
   1072                         (1 << (WARPEDDIFF_PREC_BITS - 1)) +
   1073                         (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
   1074  const int32_t const3 = ((1 << WARP_PARAM_REDUCE_BITS) - 1);
   1075  const int16_t const4 = (1 << (bd + FILTER_BITS - reduce_bits_horiz - 1));
   1076  const int16_t const5 = (1 << (FILTER_BITS - reduce_bits_horiz));
   1077 
   1078  __m256i shuffle_src[4];
   1079  shuffle_src[0] = _mm256_load_si256((__m256i *)shuffle_src0);
   1080  shuffle_src[1] = _mm256_load_si256((__m256i *)shuffle_src1);
   1081  shuffle_src[2] = _mm256_load_si256((__m256i *)shuffle_src2);
   1082  shuffle_src[3] = _mm256_load_si256((__m256i *)shuffle_src3);
   1083 
   1084  for (i = 0; i < p_height; i += 8) {
   1085    for (j = 0; j < p_width; j += 8) {
   1086      const int32_t src_x = (p_col + j + 4) << subsampling_x;
   1087      const int32_t src_y = (p_row + i + 4) << subsampling_y;
   1088      const int64_t dst_x =
   1089          (int64_t)mat[2] * src_x + (int64_t)mat[3] * src_y + (int64_t)mat[0];
   1090      const int64_t dst_y =
   1091          (int64_t)mat[4] * src_x + (int64_t)mat[5] * src_y + (int64_t)mat[1];
   1092      const int64_t x4 = dst_x >> subsampling_x;
   1093      const int64_t y4 = dst_y >> subsampling_y;
   1094 
   1095      int32_t ix4 = (int32_t)(x4 >> WARPEDMODEL_PREC_BITS);
   1096      int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
   1097      int32_t iy4 = (int32_t)(y4 >> WARPEDMODEL_PREC_BITS);
   1098      int32_t sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
   1099 
   1100      // Add in all the constant terms, including rounding and offset
   1101      sx4 += const1;
   1102      sy4 += const2;
   1103 
   1104      sx4 &= ~const3;
   1105      sy4 &= ~const3;
   1106 
   1107      // Horizontal filter
   1108      // If the block is aligned such that, after clamping, every sample
   1109      // would be taken from the leftmost/rightmost column, then we can
   1110      // skip the expensive horizontal filter.
   1111 
   1112      if (ix4 <= -7) {
   1113        int iy, row = 0;
   1114        for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
   1115          iy = iy4 + k;
   1116          iy = clamp(iy, 0, height - 1);
   1117          const __m256i temp_0 =
   1118              _mm256_set1_epi16(const4 + ref[iy * stride] * const5);
   1119          iy = iy4 + k + 1;
   1120          iy = clamp(iy, 0, height - 1);
   1121          const __m256i temp_1 =
   1122              _mm256_set1_epi16(const4 + ref[iy * stride] * const5);
   1123          horz_out[row] = _mm256_blend_epi32(temp_0, temp_1, 0xf0);
   1124          row += 1;
   1125        }
   1126        iy = iy4 + k;
   1127        iy = clamp(iy, 0, height - 1);
   1128        horz_out[row] = _mm256_set1_epi16(const4 + ref[iy * stride] * const5);
   1129      } else if (ix4 >= width + 6) {
   1130        int iy, row = 0;
   1131        for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
   1132          iy = iy4 + k;
   1133          iy = clamp(iy, 0, height - 1);
   1134          const __m256i temp_0 = _mm256_set1_epi16(
   1135              const4 + ref[iy * stride + (width - 1)] * const5);
   1136          iy = iy4 + k + 1;
   1137          iy = clamp(iy, 0, height - 1);
   1138          const __m256i temp_1 = _mm256_set1_epi16(
   1139              const4 + ref[iy * stride + (width - 1)] * const5);
   1140          horz_out[row] = _mm256_blend_epi32(temp_0, temp_1, 0xf0);
   1141          row += 1;
   1142        }
   1143        iy = iy4 + k;
   1144        iy = clamp(iy, 0, height - 1);
   1145        horz_out[row] =
   1146            _mm256_set1_epi16(const4 + ref[iy * stride + (width - 1)] * const5);
   1147      } else if (((ix4 - 7) < 0) || ((ix4 + 9) > width)) {
   1148        const int out_of_boundary_left = -(ix4 - 6);
   1149        const int out_of_boundary_right = (ix4 + 8) - width;
   1150        int iy, sx, row = 0;
   1151        for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
   1152          iy = iy4 + k;
   1153          iy = clamp(iy, 0, height - 1);
   1154          __m128i src0 =
   1155              _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
   1156          iy = iy4 + k + 1;
   1157          iy = clamp(iy, 0, height - 1);
   1158          __m128i src1 =
   1159              _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
   1160 
   1161          if (out_of_boundary_left >= 0) {
   1162            const __m128i shuffle_reg_left =
   1163                _mm_loadu_si128((__m128i *)warp_pad_left[out_of_boundary_left]);
   1164            src0 = _mm_shuffle_epi8(src0, shuffle_reg_left);
   1165            src1 = _mm_shuffle_epi8(src1, shuffle_reg_left);
   1166          }
   1167          if (out_of_boundary_right >= 0) {
   1168            const __m128i shuffle_reg_right = _mm_loadu_si128(
   1169                (__m128i *)warp_pad_right[out_of_boundary_right]);
   1170            src0 = _mm_shuffle_epi8(src0, shuffle_reg_right);
   1171            src1 = _mm_shuffle_epi8(src1, shuffle_reg_right);
   1172          }
   1173          sx = sx4 + beta * (k + 4);
   1174          const __m256i src_01 =
   1175              _mm256_inserti128_si256(_mm256_castsi128_si256(src0), src1, 0x1);
   1176          horizontal_filter_avx2(src_01, horz_out, sx, alpha, beta, row,
   1177                                 shuffle_src, &round_const, &shift);
   1178          row += 1;
   1179        }
   1180        iy = iy4 + k;
   1181        iy = clamp(iy, 0, height - 1);
   1182        __m128i src = _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
   1183        if (out_of_boundary_left >= 0) {
   1184          const __m128i shuffle_reg_left =
   1185              _mm_loadu_si128((__m128i *)warp_pad_left[out_of_boundary_left]);
   1186          src = _mm_shuffle_epi8(src, shuffle_reg_left);
   1187        }
   1188        if (out_of_boundary_right >= 0) {
   1189          const __m128i shuffle_reg_right =
   1190              _mm_loadu_si128((__m128i *)warp_pad_right[out_of_boundary_right]);
   1191          src = _mm_shuffle_epi8(src, shuffle_reg_right);
   1192        }
   1193        sx = sx4 + beta * (k + 4);
   1194        const __m256i src_01 = _mm256_castsi128_si256(src);
   1195        __m256i coeff[4];
   1196        prepare_horizontal_filter_coeff(alpha, sx, coeff);
   1197        filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src,
   1198                               &round_const, &shift, row);
   1199      } else {
   1200        prepare_warp_horizontal_filter_avx2(
   1201            ref, horz_out, stride, ix4, iy4, sx4, alpha, beta, p_height, height,
   1202            i, &round_const, &shift, shuffle_src);
   1203      }
   1204 
   1205      // Vertical filter
   1206      prepare_warp_vertical_filter_avx2(
   1207          pred, horz_out, conv_params, gamma, delta, p_height, p_stride,
   1208          p_width, i, j, sy4, reduce_bits_vert, &res_add_const_1, round_bits,
   1209          &res_sub_const, &round_bits_const, &wt);
   1210    }
   1211  }
   1212 }
   1213 
   1214 #endif  // !CONFIG_HIGHWAY