tor-browser

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

cdef_block_simd.h (37698B)


      1 /*
      2 * Copyright (c) 2016, 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 #ifndef AOM_AV1_COMMON_CDEF_BLOCK_SIMD_H_
     13 #define AOM_AV1_COMMON_CDEF_BLOCK_SIMD_H_
     14 
     15 #include "config/aom_config.h"
     16 #include "config/av1_rtcd.h"
     17 
     18 #include "av1/common/cdef_block.h"
     19 
     20 /* partial A is a 16-bit vector of the form:
     21   [x8 x7 x6 x5 x4 x3 x2 x1] and partial B has the form:
     22   [0  y1 y2 y3 y4 y5 y6 y7].
     23   This function computes (x1^2+y1^2)*C1 + (x2^2+y2^2)*C2 + ...
     24   (x7^2+y2^7)*C7 + (x8^2+0^2)*C8 where the C1..C8 constants are in const1
     25   and const2. */
     26 static inline v128 fold_mul_and_sum(v128 partiala, v128 partialb, v128 const1,
     27                                    v128 const2) {
     28  v128 tmp;
     29  /* Reverse partial B. */
     30  partialb = v128_shuffle_8(
     31      partialb, v128_from_32(0x0f0e0100, 0x03020504, 0x07060908, 0x0b0a0d0c));
     32  /* Interleave the x and y values of identical indices and pair x8 with 0. */
     33  tmp = partiala;
     34  partiala = v128_ziplo_16(partialb, partiala);
     35  partialb = v128_ziphi_16(partialb, tmp);
     36  /* Square and add the corresponding x and y values. */
     37  partiala = v128_madd_s16(partiala, partiala);
     38  partialb = v128_madd_s16(partialb, partialb);
     39  /* Multiply by constant. */
     40  partiala = v128_mullo_s32(partiala, const1);
     41  partialb = v128_mullo_s32(partialb, const2);
     42  /* Sum all results. */
     43  partiala = v128_add_32(partiala, partialb);
     44  return partiala;
     45 }
     46 
     47 static inline v128 hsum4(v128 x0, v128 x1, v128 x2, v128 x3) {
     48  v128 t0, t1, t2, t3;
     49  t0 = v128_ziplo_32(x1, x0);
     50  t1 = v128_ziplo_32(x3, x2);
     51  t2 = v128_ziphi_32(x1, x0);
     52  t3 = v128_ziphi_32(x3, x2);
     53  x0 = v128_ziplo_64(t1, t0);
     54  x1 = v128_ziphi_64(t1, t0);
     55  x2 = v128_ziplo_64(t3, t2);
     56  x3 = v128_ziphi_64(t3, t2);
     57  return v128_add_32(v128_add_32(x0, x1), v128_add_32(x2, x3));
     58 }
     59 
     60 /* Computes cost for directions 0, 5, 6 and 7. We can call this function again
     61   to compute the remaining directions. */
     62 static inline v128 compute_directions(v128 lines[8], int32_t tmp_cost1[4]) {
     63  v128 partial4a, partial4b, partial5a, partial5b, partial7a, partial7b;
     64  v128 partial6;
     65  v128 tmp;
     66  /* Partial sums for lines 0 and 1. */
     67  partial4a = v128_shl_n_byte(lines[0], 14);
     68  partial4b = v128_shr_n_byte(lines[0], 2);
     69  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[1], 12));
     70  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[1], 4));
     71  tmp = v128_add_16(lines[0], lines[1]);
     72  partial5a = v128_shl_n_byte(tmp, 10);
     73  partial5b = v128_shr_n_byte(tmp, 6);
     74  partial7a = v128_shl_n_byte(tmp, 4);
     75  partial7b = v128_shr_n_byte(tmp, 12);
     76  partial6 = tmp;
     77 
     78  /* Partial sums for lines 2 and 3. */
     79  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[2], 10));
     80  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[2], 6));
     81  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[3], 8));
     82  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[3], 8));
     83  tmp = v128_add_16(lines[2], lines[3]);
     84  partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 8));
     85  partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 8));
     86  partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 6));
     87  partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 10));
     88  partial6 = v128_add_16(partial6, tmp);
     89 
     90  /* Partial sums for lines 4 and 5. */
     91  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[4], 6));
     92  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[4], 10));
     93  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[5], 4));
     94  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[5], 12));
     95  tmp = v128_add_16(lines[4], lines[5]);
     96  partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 6));
     97  partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 10));
     98  partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 8));
     99  partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 8));
    100  partial6 = v128_add_16(partial6, tmp);
    101 
    102  /* Partial sums for lines 6 and 7. */
    103  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[6], 2));
    104  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[6], 14));
    105  partial4a = v128_add_16(partial4a, lines[7]);
    106  tmp = v128_add_16(lines[6], lines[7]);
    107  partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 4));
    108  partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 12));
    109  partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 10));
    110  partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 6));
    111  partial6 = v128_add_16(partial6, tmp);
    112 
    113  /* Compute costs in terms of partial sums. */
    114  partial4a =
    115      fold_mul_and_sum(partial4a, partial4b, v128_from_32(210, 280, 420, 840),
    116                       v128_from_32(105, 120, 140, 168));
    117  partial7a =
    118      fold_mul_and_sum(partial7a, partial7b, v128_from_32(210, 420, 0, 0),
    119                       v128_from_32(105, 105, 105, 140));
    120  partial5a =
    121      fold_mul_and_sum(partial5a, partial5b, v128_from_32(210, 420, 0, 0),
    122                       v128_from_32(105, 105, 105, 140));
    123  partial6 = v128_madd_s16(partial6, partial6);
    124  partial6 = v128_mullo_s32(partial6, v128_dup_32(105));
    125 
    126  partial4a = hsum4(partial4a, partial5a, partial6, partial7a);
    127  v128_store_unaligned(tmp_cost1, partial4a);
    128  return partial4a;
    129 }
    130 
    131 /* transpose and reverse the order of the lines -- equivalent to a 90-degree
    132   counter-clockwise rotation of the pixels. */
    133 static inline void array_reverse_transpose_8x8(v128 *in, v128 *res) {
    134  const v128 tr0_0 = v128_ziplo_16(in[1], in[0]);
    135  const v128 tr0_1 = v128_ziplo_16(in[3], in[2]);
    136  const v128 tr0_2 = v128_ziphi_16(in[1], in[0]);
    137  const v128 tr0_3 = v128_ziphi_16(in[3], in[2]);
    138  const v128 tr0_4 = v128_ziplo_16(in[5], in[4]);
    139  const v128 tr0_5 = v128_ziplo_16(in[7], in[6]);
    140  const v128 tr0_6 = v128_ziphi_16(in[5], in[4]);
    141  const v128 tr0_7 = v128_ziphi_16(in[7], in[6]);
    142 
    143  const v128 tr1_0 = v128_ziplo_32(tr0_1, tr0_0);
    144  const v128 tr1_1 = v128_ziplo_32(tr0_5, tr0_4);
    145  const v128 tr1_2 = v128_ziphi_32(tr0_1, tr0_0);
    146  const v128 tr1_3 = v128_ziphi_32(tr0_5, tr0_4);
    147  const v128 tr1_4 = v128_ziplo_32(tr0_3, tr0_2);
    148  const v128 tr1_5 = v128_ziplo_32(tr0_7, tr0_6);
    149  const v128 tr1_6 = v128_ziphi_32(tr0_3, tr0_2);
    150  const v128 tr1_7 = v128_ziphi_32(tr0_7, tr0_6);
    151 
    152  res[7] = v128_ziplo_64(tr1_1, tr1_0);
    153  res[6] = v128_ziphi_64(tr1_1, tr1_0);
    154  res[5] = v128_ziplo_64(tr1_3, tr1_2);
    155  res[4] = v128_ziphi_64(tr1_3, tr1_2);
    156  res[3] = v128_ziplo_64(tr1_5, tr1_4);
    157  res[2] = v128_ziphi_64(tr1_5, tr1_4);
    158  res[1] = v128_ziplo_64(tr1_7, tr1_6);
    159  res[0] = v128_ziphi_64(tr1_7, tr1_6);
    160 }
    161 
    162 int SIMD_FUNC(cdef_find_dir)(const uint16_t *img, int stride, int32_t *var,
    163                             int coeff_shift) {
    164  int i;
    165  int32_t cost[8];
    166  int32_t best_cost = 0;
    167  int best_dir = 0;
    168  v128 lines[8];
    169  for (i = 0; i < 8; i++) {
    170    lines[i] = v128_load_unaligned(&img[i * stride]);
    171    lines[i] =
    172        v128_sub_16(v128_shr_s16(lines[i], coeff_shift), v128_dup_16(128));
    173  }
    174 
    175  /* Compute "mostly vertical" directions. */
    176  v128 dir47 = compute_directions(lines, cost + 4);
    177 
    178  array_reverse_transpose_8x8(lines, lines);
    179 
    180  /* Compute "mostly horizontal" directions. */
    181  v128 dir03 = compute_directions(lines, cost);
    182 
    183  v128 max = v128_max_s32(dir03, dir47);
    184  max = v128_max_s32(max, v128_align(max, max, 8));
    185  max = v128_max_s32(max, v128_align(max, max, 4));
    186  best_cost = v128_low_u32(max);
    187  v128 t =
    188      v128_pack_s32_s16(v128_cmpeq_32(max, dir47), v128_cmpeq_32(max, dir03));
    189  best_dir = v128_movemask_8(v128_pack_s16_s8(t, t));
    190  best_dir = get_msb(best_dir ^ (best_dir - 1));  // Count trailing zeros
    191 
    192  /* Difference between the optimal variance and the variance along the
    193     orthogonal direction. Again, the sum(x^2) terms cancel out. */
    194  *var = best_cost - cost[(best_dir + 4) & 7];
    195  /* We'd normally divide by 840, but dividing by 1024 is close enough
    196     for what we're going to do with this. */
    197  *var >>= 10;
    198  return best_dir;
    199 }
    200 
    201 // Work around compiler out of memory issues with Win32 builds. This issue has
    202 // been observed with Visual Studio 2017, 2019, and 2022 (version 17.10.3).
    203 #if defined(_MSC_VER) && defined(_M_IX86)
    204 #define CDEF_INLINE static inline
    205 #else
    206 #define CDEF_INLINE SIMD_INLINE
    207 #endif
    208 
    209 // sign(a-b) * min(abs(a-b), max(0, threshold - (abs(a-b) >> adjdamp)))
    210 CDEF_INLINE v256 constrain16(v256 a, v256 b, unsigned int threshold,
    211                             unsigned int adjdamp) {
    212  v256 diff = v256_sub_16(a, b);
    213  const v256 sign = v256_shr_n_s16(diff, 15);
    214  diff = v256_abs_s16(diff);
    215  const v256 s =
    216      v256_ssub_u16(v256_dup_16(threshold), v256_shr_u16(diff, adjdamp));
    217  return v256_xor(v256_add_16(sign, v256_min_s16(diff, s)), sign);
    218 }
    219 
    220 SIMD_INLINE v256 get_max_primary(const int is_lowbd, v256 *tap, v256 max,
    221                                 v256 cdef_large_value_mask) {
    222  if (is_lowbd) {
    223    v256 max_u8;
    224    max_u8 = tap[0];
    225    max_u8 = v256_max_u8(max_u8, tap[1]);
    226    max_u8 = v256_max_u8(max_u8, tap[2]);
    227    max_u8 = v256_max_u8(max_u8, tap[3]);
    228    /* The source is 16 bits, however, we only really care about the lower
    229    8 bits.  The upper 8 bits contain the "large" flag.  After the final
    230    primary max has been calculated, zero out the upper 8 bits.  Use this
    231    to find the "16 bit" max. */
    232    max = v256_max_s16(max, v256_and(max_u8, cdef_large_value_mask));
    233  } else {
    234    /* Convert CDEF_VERY_LARGE to 0 before calculating max. */
    235    max = v256_max_s16(max, v256_and(tap[0], cdef_large_value_mask));
    236    max = v256_max_s16(max, v256_and(tap[1], cdef_large_value_mask));
    237    max = v256_max_s16(max, v256_and(tap[2], cdef_large_value_mask));
    238    max = v256_max_s16(max, v256_and(tap[3], cdef_large_value_mask));
    239  }
    240  return max;
    241 }
    242 
    243 SIMD_INLINE v256 get_max_secondary(const int is_lowbd, v256 *tap, v256 max,
    244                                   v256 cdef_large_value_mask) {
    245  if (is_lowbd) {
    246    v256 max_u8;
    247    max_u8 = tap[0];
    248    max_u8 = v256_max_u8(max_u8, tap[1]);
    249    max_u8 = v256_max_u8(max_u8, tap[2]);
    250    max_u8 = v256_max_u8(max_u8, tap[3]);
    251    max_u8 = v256_max_u8(max_u8, tap[4]);
    252    max_u8 = v256_max_u8(max_u8, tap[5]);
    253    max_u8 = v256_max_u8(max_u8, tap[6]);
    254    max_u8 = v256_max_u8(max_u8, tap[7]);
    255    /* The source is 16 bits, however, we only really care about the lower
    256    8 bits.  The upper 8 bits contain the "large" flag.  After the final
    257    primary max has been calculated, zero out the upper 8 bits.  Use this
    258    to find the "16 bit" max. */
    259    max = v256_max_s16(max, v256_and(max_u8, cdef_large_value_mask));
    260  } else {
    261    /* Convert CDEF_VERY_LARGE to 0 before calculating max. */
    262    max = v256_max_s16(max, v256_and(tap[0], cdef_large_value_mask));
    263    max = v256_max_s16(max, v256_and(tap[1], cdef_large_value_mask));
    264    max = v256_max_s16(max, v256_and(tap[2], cdef_large_value_mask));
    265    max = v256_max_s16(max, v256_and(tap[3], cdef_large_value_mask));
    266    max = v256_max_s16(max, v256_and(tap[4], cdef_large_value_mask));
    267    max = v256_max_s16(max, v256_and(tap[5], cdef_large_value_mask));
    268    max = v256_max_s16(max, v256_and(tap[6], cdef_large_value_mask));
    269    max = v256_max_s16(max, v256_and(tap[7], cdef_large_value_mask));
    270  }
    271  return max;
    272 }
    273 
    274 // MSVC takes far too much time optimizing these.
    275 // https://bugs.chromium.org/p/aomedia/issues/detail?id=3395
    276 #if defined(_MSC_VER) && !defined(__clang__)
    277 #pragma optimize("", off)
    278 #endif
    279 
    280 CDEF_INLINE void filter_block_4x4(const int is_lowbd, void *dest, int dstride,
    281                                  const uint16_t *in, int pri_strength,
    282                                  int sec_strength, int dir, int pri_damping,
    283                                  int sec_damping, int coeff_shift, int height,
    284                                  int enable_primary, int enable_secondary) {
    285  uint8_t *dst8 = (uint8_t *)dest;
    286  uint16_t *dst16 = (uint16_t *)dest;
    287  const int clipping_required = enable_primary && enable_secondary;
    288  v256 p0, p1, p2, p3;
    289  v256 sum, row, res;
    290  v256 max, min;
    291  const v256 cdef_large_value_mask = v256_dup_16((uint16_t)~CDEF_VERY_LARGE);
    292  const int po1 = cdef_directions[dir][0];
    293  const int po2 = cdef_directions[dir][1];
    294  const int s1o1 = cdef_directions[dir + 2][0];
    295  const int s1o2 = cdef_directions[dir + 2][1];
    296  const int s2o1 = cdef_directions[dir - 2][0];
    297  const int s2o2 = cdef_directions[dir - 2][1];
    298  const int *pri_taps = cdef_pri_taps[(pri_strength >> coeff_shift) & 1];
    299  const int *sec_taps = cdef_sec_taps;
    300  int i;
    301 
    302  if (enable_primary && pri_strength)
    303    pri_damping = AOMMAX(0, pri_damping - get_msb(pri_strength));
    304  if (enable_secondary && sec_strength)
    305    sec_damping = AOMMAX(0, sec_damping - get_msb(sec_strength));
    306 
    307  for (i = 0; i < height; i += 4) {
    308    sum = v256_zero();
    309    row = v256_from_v64(v64_load_aligned(&in[(i + 0) * CDEF_BSTRIDE]),
    310                        v64_load_aligned(&in[(i + 1) * CDEF_BSTRIDE]),
    311                        v64_load_aligned(&in[(i + 2) * CDEF_BSTRIDE]),
    312                        v64_load_aligned(&in[(i + 3) * CDEF_BSTRIDE]));
    313    max = min = row;
    314 
    315    if (enable_primary) {
    316      v256 tap[4];
    317      // Primary near taps
    318      tap[0] =
    319          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + po1]),
    320                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + po1]),
    321                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + po1]),
    322                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + po1]));
    323      p0 = constrain16(tap[0], row, pri_strength, pri_damping);
    324      tap[1] =
    325          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - po1]),
    326                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - po1]),
    327                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - po1]),
    328                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - po1]));
    329      p1 = constrain16(tap[1], row, pri_strength, pri_damping);
    330 
    331      // sum += pri_taps[0] * (p0 + p1)
    332      sum = v256_add_16(
    333          sum, v256_mullo_s16(v256_dup_16(pri_taps[0]), v256_add_16(p0, p1)));
    334 
    335      // Primary far taps
    336      tap[2] =
    337          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + po2]),
    338                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + po2]),
    339                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + po2]),
    340                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + po2]));
    341      p0 = constrain16(tap[2], row, pri_strength, pri_damping);
    342      tap[3] =
    343          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - po2]),
    344                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - po2]),
    345                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - po2]),
    346                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - po2]));
    347      p1 = constrain16(tap[3], row, pri_strength, pri_damping);
    348 
    349      // sum += pri_taps[1] * (p0 + p1)
    350      sum = v256_add_16(
    351          sum, v256_mullo_s16(v256_dup_16(pri_taps[1]), v256_add_16(p0, p1)));
    352      if (clipping_required) {
    353        max = get_max_primary(is_lowbd, tap, max, cdef_large_value_mask);
    354 
    355        min = v256_min_s16(min, tap[0]);
    356        min = v256_min_s16(min, tap[1]);
    357        min = v256_min_s16(min, tap[2]);
    358        min = v256_min_s16(min, tap[3]);
    359      }
    360    }
    361 
    362    if (enable_secondary) {
    363      v256 tap[8];
    364      // Secondary near taps
    365      tap[0] =
    366          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + s1o1]),
    367                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s1o1]),
    368                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + s1o1]),
    369                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + s1o1]));
    370      p0 = constrain16(tap[0], row, sec_strength, sec_damping);
    371      tap[1] =
    372          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - s1o1]),
    373                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s1o1]),
    374                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - s1o1]),
    375                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - s1o1]));
    376      p1 = constrain16(tap[1], row, sec_strength, sec_damping);
    377      tap[2] =
    378          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + s2o1]),
    379                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s2o1]),
    380                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + s2o1]),
    381                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + s2o1]));
    382      p2 = constrain16(tap[2], row, sec_strength, sec_damping);
    383      tap[3] =
    384          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - s2o1]),
    385                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s2o1]),
    386                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - s2o1]),
    387                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - s2o1]));
    388      p3 = constrain16(tap[3], row, sec_strength, sec_damping);
    389 
    390      // sum += sec_taps[0] * (p0 + p1 + p2 + p3)
    391      sum = v256_add_16(sum, v256_mullo_s16(v256_dup_16(sec_taps[0]),
    392                                            v256_add_16(v256_add_16(p0, p1),
    393                                                        v256_add_16(p2, p3))));
    394 
    395      // Secondary far taps
    396      tap[4] =
    397          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + s1o2]),
    398                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s1o2]),
    399                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + s1o2]),
    400                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + s1o2]));
    401      p0 = constrain16(tap[4], row, sec_strength, sec_damping);
    402      tap[5] =
    403          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - s1o2]),
    404                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s1o2]),
    405                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - s1o2]),
    406                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - s1o2]));
    407      p1 = constrain16(tap[5], row, sec_strength, sec_damping);
    408      tap[6] =
    409          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + s2o2]),
    410                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s2o2]),
    411                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + s2o2]),
    412                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + s2o2]));
    413      p2 = constrain16(tap[6], row, sec_strength, sec_damping);
    414      tap[7] =
    415          v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - s2o2]),
    416                        v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s2o2]),
    417                        v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - s2o2]),
    418                        v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - s2o2]));
    419      p3 = constrain16(tap[7], row, sec_strength, sec_damping);
    420 
    421      // sum += sec_taps[1] * (p0 + p1 + p2 + p3)
    422      sum = v256_add_16(sum, v256_mullo_s16(v256_dup_16(sec_taps[1]),
    423                                            v256_add_16(v256_add_16(p0, p1),
    424                                                        v256_add_16(p2, p3))));
    425 
    426      if (clipping_required) {
    427        max = get_max_secondary(is_lowbd, tap, max, cdef_large_value_mask);
    428 
    429        min = v256_min_s16(min, tap[0]);
    430        min = v256_min_s16(min, tap[1]);
    431        min = v256_min_s16(min, tap[2]);
    432        min = v256_min_s16(min, tap[3]);
    433        min = v256_min_s16(min, tap[4]);
    434        min = v256_min_s16(min, tap[5]);
    435        min = v256_min_s16(min, tap[6]);
    436        min = v256_min_s16(min, tap[7]);
    437      }
    438    }
    439 
    440    // res = row + ((sum - (sum < 0) + 8) >> 4)
    441    sum = v256_add_16(sum, v256_cmplt_s16(sum, v256_zero()));
    442    res = v256_add_16(sum, v256_dup_16(8));
    443    res = v256_shr_n_s16(res, 4);
    444    res = v256_add_16(row, res);
    445    if (clipping_required) {
    446      res = v256_min_s16(v256_max_s16(res, min), max);
    447    }
    448 
    449    if (is_lowbd) {
    450      const v128 res_128 = v256_low_v128(v256_pack_s16_u8(res, res));
    451      u32_store_aligned(&dst8[(i + 0) * dstride],
    452                        v64_high_u32(v128_high_v64(res_128)));
    453      u32_store_aligned(&dst8[(i + 1) * dstride],
    454                        v64_low_u32(v128_high_v64(res_128)));
    455      u32_store_aligned(&dst8[(i + 2) * dstride],
    456                        v64_high_u32(v128_low_v64(res_128)));
    457      u32_store_aligned(&dst8[(i + 3) * dstride],
    458                        v64_low_u32(v128_low_v64(res_128)));
    459    } else {
    460      v64_store_aligned(&dst16[(i + 0) * dstride],
    461                        v128_high_v64(v256_high_v128(res)));
    462      v64_store_aligned(&dst16[(i + 1) * dstride],
    463                        v128_low_v64(v256_high_v128(res)));
    464      v64_store_aligned(&dst16[(i + 2) * dstride],
    465                        v128_high_v64(v256_low_v128(res)));
    466      v64_store_aligned(&dst16[(i + 3) * dstride],
    467                        v128_low_v64(v256_low_v128(res)));
    468    }
    469  }
    470 }
    471 
    472 CDEF_INLINE void filter_block_8x8(const int is_lowbd, void *dest, int dstride,
    473                                  const uint16_t *in, int pri_strength,
    474                                  int sec_strength, int dir, int pri_damping,
    475                                  int sec_damping, int coeff_shift, int height,
    476                                  int enable_primary, int enable_secondary) {
    477  uint8_t *dst8 = (uint8_t *)dest;
    478  uint16_t *dst16 = (uint16_t *)dest;
    479  const int clipping_required = enable_primary && enable_secondary;
    480  int i;
    481  v256 sum, p0, p1, p2, p3, row, res;
    482  const v256 cdef_large_value_mask = v256_dup_16((uint16_t)~CDEF_VERY_LARGE);
    483  v256 max, min;
    484  const int po1 = cdef_directions[dir][0];
    485  const int po2 = cdef_directions[dir][1];
    486  const int s1o1 = cdef_directions[dir + 2][0];
    487  const int s1o2 = cdef_directions[dir + 2][1];
    488  const int s2o1 = cdef_directions[dir - 2][0];
    489  const int s2o2 = cdef_directions[dir - 2][1];
    490  const int *pri_taps = cdef_pri_taps[(pri_strength >> coeff_shift) & 1];
    491  const int *sec_taps = cdef_sec_taps;
    492 
    493  if (enable_primary && pri_strength)
    494    pri_damping = AOMMAX(0, pri_damping - get_msb(pri_strength));
    495  if (enable_secondary && sec_strength)
    496    sec_damping = AOMMAX(0, sec_damping - get_msb(sec_strength));
    497 
    498  for (i = 0; i < height; i += 2) {
    499    v256 tap[8];
    500    sum = v256_zero();
    501    row = v256_from_v128(v128_load_aligned(&in[i * CDEF_BSTRIDE]),
    502                         v128_load_aligned(&in[(i + 1) * CDEF_BSTRIDE]));
    503 
    504    min = max = row;
    505    if (enable_primary) {
    506      // Primary near taps
    507      tap[0] = v256_from_v128(
    508          v128_load_unaligned(&in[i * CDEF_BSTRIDE + po1]),
    509          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + po1]));
    510      tap[1] = v256_from_v128(
    511          v128_load_unaligned(&in[i * CDEF_BSTRIDE - po1]),
    512          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - po1]));
    513      p0 = constrain16(tap[0], row, pri_strength, pri_damping);
    514      p1 = constrain16(tap[1], row, pri_strength, pri_damping);
    515 
    516      // sum += pri_taps[0] * (p0 + p1)
    517      sum = v256_add_16(
    518          sum, v256_mullo_s16(v256_dup_16(pri_taps[0]), v256_add_16(p0, p1)));
    519 
    520      // Primary far taps
    521      tap[2] = v256_from_v128(
    522          v128_load_unaligned(&in[i * CDEF_BSTRIDE + po2]),
    523          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + po2]));
    524      tap[3] = v256_from_v128(
    525          v128_load_unaligned(&in[i * CDEF_BSTRIDE - po2]),
    526          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - po2]));
    527      p0 = constrain16(tap[2], row, pri_strength, pri_damping);
    528      p1 = constrain16(tap[3], row, pri_strength, pri_damping);
    529 
    530      // sum += pri_taps[1] * (p0 + p1)
    531      sum = v256_add_16(
    532          sum, v256_mullo_s16(v256_dup_16(pri_taps[1]), v256_add_16(p0, p1)));
    533 
    534      if (clipping_required) {
    535        max = get_max_primary(is_lowbd, tap, max, cdef_large_value_mask);
    536 
    537        min = v256_min_s16(min, tap[0]);
    538        min = v256_min_s16(min, tap[1]);
    539        min = v256_min_s16(min, tap[2]);
    540        min = v256_min_s16(min, tap[3]);
    541      }
    542      // End primary
    543    }
    544 
    545    if (enable_secondary) {
    546      // Secondary near taps
    547      tap[0] = v256_from_v128(
    548          v128_load_unaligned(&in[i * CDEF_BSTRIDE + s1o1]),
    549          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s1o1]));
    550      tap[1] = v256_from_v128(
    551          v128_load_unaligned(&in[i * CDEF_BSTRIDE - s1o1]),
    552          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s1o1]));
    553      tap[2] = v256_from_v128(
    554          v128_load_unaligned(&in[i * CDEF_BSTRIDE + s2o1]),
    555          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s2o1]));
    556      tap[3] = v256_from_v128(
    557          v128_load_unaligned(&in[i * CDEF_BSTRIDE - s2o1]),
    558          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s2o1]));
    559      p0 = constrain16(tap[0], row, sec_strength, sec_damping);
    560      p1 = constrain16(tap[1], row, sec_strength, sec_damping);
    561      p2 = constrain16(tap[2], row, sec_strength, sec_damping);
    562      p3 = constrain16(tap[3], row, sec_strength, sec_damping);
    563 
    564      // sum += sec_taps[0] * (p0 + p1 + p2 + p3)
    565      sum = v256_add_16(sum, v256_mullo_s16(v256_dup_16(sec_taps[0]),
    566                                            v256_add_16(v256_add_16(p0, p1),
    567                                                        v256_add_16(p2, p3))));
    568 
    569      // Secondary far taps
    570      tap[4] = v256_from_v128(
    571          v128_load_unaligned(&in[i * CDEF_BSTRIDE + s1o2]),
    572          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s1o2]));
    573      tap[5] = v256_from_v128(
    574          v128_load_unaligned(&in[i * CDEF_BSTRIDE - s1o2]),
    575          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s1o2]));
    576      tap[6] = v256_from_v128(
    577          v128_load_unaligned(&in[i * CDEF_BSTRIDE + s2o2]),
    578          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s2o2]));
    579      tap[7] = v256_from_v128(
    580          v128_load_unaligned(&in[i * CDEF_BSTRIDE - s2o2]),
    581          v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s2o2]));
    582      p0 = constrain16(tap[4], row, sec_strength, sec_damping);
    583      p1 = constrain16(tap[5], row, sec_strength, sec_damping);
    584      p2 = constrain16(tap[6], row, sec_strength, sec_damping);
    585      p3 = constrain16(tap[7], row, sec_strength, sec_damping);
    586 
    587      // sum += sec_taps[1] * (p0 + p1 + p2 + p3)
    588      sum = v256_add_16(sum, v256_mullo_s16(v256_dup_16(sec_taps[1]),
    589                                            v256_add_16(v256_add_16(p0, p1),
    590                                                        v256_add_16(p2, p3))));
    591 
    592      if (clipping_required) {
    593        max = get_max_secondary(is_lowbd, tap, max, cdef_large_value_mask);
    594 
    595        min = v256_min_s16(min, tap[0]);
    596        min = v256_min_s16(min, tap[1]);
    597        min = v256_min_s16(min, tap[2]);
    598        min = v256_min_s16(min, tap[3]);
    599        min = v256_min_s16(min, tap[4]);
    600        min = v256_min_s16(min, tap[5]);
    601        min = v256_min_s16(min, tap[6]);
    602        min = v256_min_s16(min, tap[7]);
    603      }
    604      // End secondary
    605    }
    606 
    607    // res = row + ((sum - (sum < 0) + 8) >> 4)
    608    sum = v256_add_16(sum, v256_cmplt_s16(sum, v256_zero()));
    609    res = v256_add_16(sum, v256_dup_16(8));
    610    res = v256_shr_n_s16(res, 4);
    611    res = v256_add_16(row, res);
    612    if (clipping_required) {
    613      res = v256_min_s16(v256_max_s16(res, min), max);
    614    }
    615 
    616    if (is_lowbd) {
    617      const v128 res_128 = v256_low_v128(v256_pack_s16_u8(res, res));
    618      v64_store_aligned(&dst8[i * dstride], v128_high_v64(res_128));
    619      v64_store_aligned(&dst8[(i + 1) * dstride], v128_low_v64(res_128));
    620    } else {
    621      v128_store_unaligned(&dst16[i * dstride], v256_high_v128(res));
    622      v128_store_unaligned(&dst16[(i + 1) * dstride], v256_low_v128(res));
    623    }
    624  }
    625 }
    626 
    627 #if defined(_MSC_VER) && !defined(__clang__)
    628 #pragma optimize("", on)
    629 #endif
    630 
    631 SIMD_INLINE void copy_block_4xh(const int is_lowbd, void *dest, int dstride,
    632                                const uint16_t *in, int height) {
    633  uint8_t *dst8 = (uint8_t *)dest;
    634  uint16_t *dst16 = (uint16_t *)dest;
    635  int i;
    636  for (i = 0; i < height; i += 4) {
    637    const v128 row0 =
    638        v128_from_v64(v64_load_aligned(&in[(i + 0) * CDEF_BSTRIDE]),
    639                      v64_load_aligned(&in[(i + 1) * CDEF_BSTRIDE]));
    640    const v128 row1 =
    641        v128_from_v64(v64_load_aligned(&in[(i + 2) * CDEF_BSTRIDE]),
    642                      v64_load_aligned(&in[(i + 3) * CDEF_BSTRIDE]));
    643    if (is_lowbd) {
    644      /* Note: v128_pack_s16_u8(). The parameter order is swapped internally */
    645      const v128 res_128 = v128_pack_s16_u8(row1, row0);
    646      u32_store_aligned(&dst8[(i + 0) * dstride],
    647                        v64_high_u32(v128_low_v64(res_128)));
    648      u32_store_aligned(&dst8[(i + 1) * dstride],
    649                        v64_low_u32(v128_low_v64(res_128)));
    650      u32_store_aligned(&dst8[(i + 2) * dstride],
    651                        v64_high_u32(v128_high_v64(res_128)));
    652      u32_store_aligned(&dst8[(i + 3) * dstride],
    653                        v64_low_u32(v128_high_v64(res_128)));
    654    } else {
    655      v64_store_aligned(&dst16[(i + 0) * dstride], v128_high_v64(row0));
    656      v64_store_aligned(&dst16[(i + 1) * dstride], v128_low_v64(row0));
    657      v64_store_aligned(&dst16[(i + 2) * dstride], v128_high_v64(row1));
    658      v64_store_aligned(&dst16[(i + 3) * dstride], v128_low_v64(row1));
    659    }
    660  }
    661 }
    662 
    663 SIMD_INLINE void copy_block_8xh(const int is_lowbd, void *dest, int dstride,
    664                                const uint16_t *in, int height) {
    665  uint8_t *dst8 = (uint8_t *)dest;
    666  uint16_t *dst16 = (uint16_t *)dest;
    667  int i;
    668  for (i = 0; i < height; i += 2) {
    669    const v128 row0 = v128_load_aligned(&in[i * CDEF_BSTRIDE]);
    670    const v128 row1 = v128_load_aligned(&in[(i + 1) * CDEF_BSTRIDE]);
    671    if (is_lowbd) {
    672      /* Note: v128_pack_s16_u8(). The parameter order is swapped internally */
    673      const v128 res_128 = v128_pack_s16_u8(row1, row0);
    674      v64_store_aligned(&dst8[i * dstride], v128_low_v64(res_128));
    675      v64_store_aligned(&dst8[(i + 1) * dstride], v128_high_v64(res_128));
    676    } else {
    677      v128_store_unaligned(&dst16[i * dstride], row0);
    678      v128_store_unaligned(&dst16[(i + 1) * dstride], row1);
    679    }
    680  }
    681 }
    682 
    683 void SIMD_FUNC(cdef_filter_8_0)(void *dest, int dstride, const uint16_t *in,
    684                                int pri_strength, int sec_strength, int dir,
    685                                int pri_damping, int sec_damping,
    686                                int coeff_shift, int block_width,
    687                                int block_height) {
    688  if (block_width == 8) {
    689    filter_block_8x8(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
    690                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    691                     block_height, /*enable_primary=*/1,
    692                     /*enable_secondary=*/1);
    693  } else {
    694    filter_block_4x4(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
    695                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    696                     block_height, /*enable_primary=*/1,
    697                     /*enable_secondary=*/1);
    698  }
    699 }
    700 
    701 void SIMD_FUNC(cdef_filter_8_1)(void *dest, int dstride, const uint16_t *in,
    702                                int pri_strength, int sec_strength, int dir,
    703                                int pri_damping, int sec_damping,
    704                                int coeff_shift, int block_width,
    705                                int block_height) {
    706  if (block_width == 8) {
    707    filter_block_8x8(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
    708                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    709                     block_height, /*enable_primary=*/1,
    710                     /*enable_secondary=*/0);
    711  } else {
    712    filter_block_4x4(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
    713                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    714                     block_height, /*enable_primary=*/1,
    715                     /*enable_secondary=*/0);
    716  }
    717 }
    718 void SIMD_FUNC(cdef_filter_8_2)(void *dest, int dstride, const uint16_t *in,
    719                                int pri_strength, int sec_strength, int dir,
    720                                int pri_damping, int sec_damping,
    721                                int coeff_shift, int block_width,
    722                                int block_height) {
    723  if (block_width == 8) {
    724    filter_block_8x8(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
    725                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    726                     block_height, /*enable_primary=*/0,
    727                     /*enable_secondary=*/1);
    728  } else {
    729    filter_block_4x4(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
    730                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    731                     block_height, /*enable_primary=*/0,
    732                     /*enable_secondary=*/1);
    733  }
    734 }
    735 
    736 void SIMD_FUNC(cdef_filter_8_3)(void *dest, int dstride, const uint16_t *in,
    737                                int pri_strength, int sec_strength, int dir,
    738                                int pri_damping, int sec_damping,
    739                                int coeff_shift, int block_width,
    740                                int block_height) {
    741  (void)pri_strength;
    742  (void)sec_strength;
    743  (void)dir;
    744  (void)pri_damping;
    745  (void)sec_damping;
    746  (void)coeff_shift;
    747  (void)block_width;
    748 
    749  if (block_width == 8) {
    750    copy_block_8xh(/*is_lowbd=*/1, dest, dstride, in, block_height);
    751  } else {
    752    copy_block_4xh(/*is_lowbd=*/1, dest, dstride, in, block_height);
    753  }
    754 }
    755 
    756 void SIMD_FUNC(cdef_filter_16_0)(void *dest, int dstride, const uint16_t *in,
    757                                 int pri_strength, int sec_strength, int dir,
    758                                 int pri_damping, int sec_damping,
    759                                 int coeff_shift, int block_width,
    760                                 int block_height) {
    761  if (block_width == 8) {
    762    filter_block_8x8(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
    763                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    764                     block_height, /*enable_primary=*/1,
    765                     /*enable_secondary=*/1);
    766  } else {
    767    filter_block_4x4(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
    768                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    769                     block_height, /*enable_primary=*/1,
    770                     /*enable_secondary=*/1);
    771  }
    772 }
    773 
    774 void SIMD_FUNC(cdef_filter_16_1)(void *dest, int dstride, const uint16_t *in,
    775                                 int pri_strength, int sec_strength, int dir,
    776                                 int pri_damping, int sec_damping,
    777                                 int coeff_shift, int block_width,
    778                                 int block_height) {
    779  if (block_width == 8) {
    780    filter_block_8x8(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
    781                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    782                     block_height, /*enable_primary=*/1,
    783                     /*enable_secondary=*/0);
    784  } else {
    785    filter_block_4x4(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
    786                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    787                     block_height, /*enable_primary=*/1,
    788                     /*enable_secondary=*/0);
    789  }
    790 }
    791 void SIMD_FUNC(cdef_filter_16_2)(void *dest, int dstride, const uint16_t *in,
    792                                 int pri_strength, int sec_strength, int dir,
    793                                 int pri_damping, int sec_damping,
    794                                 int coeff_shift, int block_width,
    795                                 int block_height) {
    796  if (block_width == 8) {
    797    filter_block_8x8(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
    798                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    799                     block_height, /*enable_primary=*/0,
    800                     /*enable_secondary=*/1);
    801  } else {
    802    filter_block_4x4(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
    803                     sec_strength, dir, pri_damping, sec_damping, coeff_shift,
    804                     block_height, /*enable_primary=*/0,
    805                     /*enable_secondary=*/1);
    806  }
    807 }
    808 
    809 void SIMD_FUNC(cdef_filter_16_3)(void *dest, int dstride, const uint16_t *in,
    810                                 int pri_strength, int sec_strength, int dir,
    811                                 int pri_damping, int sec_damping,
    812                                 int coeff_shift, int block_width,
    813                                 int block_height) {
    814  (void)pri_strength;
    815  (void)sec_strength;
    816  (void)dir;
    817  (void)pri_damping;
    818  (void)sec_damping;
    819  (void)coeff_shift;
    820  (void)block_width;
    821  if (block_width == 8) {
    822    copy_block_8xh(/*is_lowbd=*/0, dest, dstride, in, block_height);
    823  } else {
    824    copy_block_4xh(/*is_lowbd=*/0, dest, dstride, in, block_height);
    825  }
    826 }
    827 
    828 #if CONFIG_AV1_HIGHBITDEPTH
    829 void SIMD_FUNC(cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride,
    830                                               const uint16_t *src, int sstride,
    831                                               int width, int height) {
    832  int i, j;
    833  for (i = 0; i < height; i++) {
    834    for (j = 0; j < (width & ~0x7); j += 8) {
    835      v128 row = v128_load_unaligned(&src[i * sstride + j]);
    836      v128_store_unaligned(&dst[i * dstride + j], row);
    837    }
    838    for (; j < width; j++) {
    839      dst[i * dstride + j] = src[i * sstride + j];
    840    }
    841  }
    842 }
    843 #endif  // CONFIG_AV1_HIGHBITDEPTH
    844 
    845 #undef CDEF_INLINE
    846 
    847 #endif  // AOM_AV1_COMMON_CDEF_BLOCK_SIMD_H_