tor-browser

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

cfl_neon.c (25127B)


      1 /*
      2 * Copyright (c) 2017, 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 #include <arm_neon.h>
     12 
     13 #include "config/aom_config.h"
     14 #include "config/av1_rtcd.h"
     15 
     16 #include "aom_dsp/arm/mem_neon.h"
     17 #include "av1/common/cfl.h"
     18 
     19 static inline void vldsubstq_s16(int16_t *dst, const uint16_t *src, int offset,
     20                                 int16x8_t sub) {
     21  vst1q_s16(dst + offset,
     22            vsubq_s16(vreinterpretq_s16_u16(vld1q_u16(src + offset)), sub));
     23 }
     24 
     25 static inline uint16x8_t vldaddq_u16(const uint16_t *buf, size_t offset) {
     26  return vaddq_u16(vld1q_u16(buf), vld1q_u16(buf + offset));
     27 }
     28 
     29 // Load half of a vector and duplicated in other half
     30 static inline uint8x8_t vldh_dup_u8(const uint8_t *ptr) {
     31  return vreinterpret_u8_u32(vld1_dup_u32((const uint32_t *)ptr));
     32 }
     33 
     34 // Store half of a vector.
     35 static inline void vsth_u16(uint16_t *ptr, uint16x4_t val) {
     36  vst1_lane_u32((uint32_t *)ptr, vreinterpret_u32_u16(val), 0);
     37 }
     38 
     39 // Store half of a vector.
     40 static inline void vsth_u8(uint8_t *ptr, uint8x8_t val) {
     41  vst1_lane_u32((uint32_t *)ptr, vreinterpret_u32_u8(val), 0);
     42 }
     43 
     44 static void cfl_luma_subsampling_420_lbd_neon(const uint8_t *input,
     45                                              int input_stride,
     46                                              uint16_t *pred_buf_q3, int width,
     47                                              int height) {
     48  const uint16_t *end = pred_buf_q3 + (height >> 1) * CFL_BUF_LINE;
     49  const int luma_stride = input_stride << 1;
     50  do {
     51    if (width == 4) {
     52      const uint16x4_t top = vpaddl_u8(vldh_dup_u8(input));
     53      const uint16x4_t sum = vpadal_u8(top, vldh_dup_u8(input + input_stride));
     54      vsth_u16(pred_buf_q3, vshl_n_u16(sum, 1));
     55    } else if (width == 8) {
     56      const uint16x4_t top = vpaddl_u8(vld1_u8(input));
     57      const uint16x4_t sum = vpadal_u8(top, vld1_u8(input + input_stride));
     58      vst1_u16(pred_buf_q3, vshl_n_u16(sum, 1));
     59    } else if (width == 16) {
     60      const uint16x8_t top = vpaddlq_u8(vld1q_u8(input));
     61      const uint16x8_t sum = vpadalq_u8(top, vld1q_u8(input + input_stride));
     62      vst1q_u16(pred_buf_q3, vshlq_n_u16(sum, 1));
     63    } else {
     64      const uint8x8x4_t top = vld4_u8(input);
     65      const uint8x8x4_t bot = vld4_u8(input + input_stride);
     66      // equivalent to a vpaddlq_u8 (because vld4q interleaves)
     67      const uint16x8_t top_0 = vaddl_u8(top.val[0], top.val[1]);
     68      // equivalent to a vpaddlq_u8 (because vld4q interleaves)
     69      const uint16x8_t bot_0 = vaddl_u8(bot.val[0], bot.val[1]);
     70      // equivalent to a vpaddlq_u8 (because vld4q interleaves)
     71      const uint16x8_t top_1 = vaddl_u8(top.val[2], top.val[3]);
     72      // equivalent to a vpaddlq_u8 (because vld4q interleaves)
     73      const uint16x8_t bot_1 = vaddl_u8(bot.val[2], bot.val[3]);
     74      uint16x8x2_t sum;
     75      sum.val[0] = vshlq_n_u16(vaddq_u16(top_0, bot_0), 1);
     76      sum.val[1] = vshlq_n_u16(vaddq_u16(top_1, bot_1), 1);
     77      vst2q_u16(pred_buf_q3, sum);
     78    }
     79    input += luma_stride;
     80  } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
     81 }
     82 
     83 static void cfl_luma_subsampling_422_lbd_neon(const uint8_t *input,
     84                                              int input_stride,
     85                                              uint16_t *pred_buf_q3, int width,
     86                                              int height) {
     87  const uint16_t *end = pred_buf_q3 + height * CFL_BUF_LINE;
     88  do {
     89    if (width == 4) {
     90      const uint16x4_t top = vpaddl_u8(vldh_dup_u8(input));
     91      vsth_u16(pred_buf_q3, vshl_n_u16(top, 2));
     92    } else if (width == 8) {
     93      const uint16x4_t top = vpaddl_u8(vld1_u8(input));
     94      vst1_u16(pred_buf_q3, vshl_n_u16(top, 2));
     95    } else if (width == 16) {
     96      const uint16x8_t top = vpaddlq_u8(vld1q_u8(input));
     97      vst1q_u16(pred_buf_q3, vshlq_n_u16(top, 2));
     98    } else {
     99      const uint8x8x4_t top = vld4_u8(input);
    100      uint16x8x2_t sum;
    101      // vaddl_u8 is equivalent to a vpaddlq_u8 (because vld4q interleaves)
    102      sum.val[0] = vshlq_n_u16(vaddl_u8(top.val[0], top.val[1]), 2);
    103      sum.val[1] = vshlq_n_u16(vaddl_u8(top.val[2], top.val[3]), 2);
    104      vst2q_u16(pred_buf_q3, sum);
    105    }
    106    input += input_stride;
    107  } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    108 }
    109 
    110 static void cfl_luma_subsampling_444_lbd_neon(const uint8_t *input,
    111                                              int input_stride,
    112                                              uint16_t *pred_buf_q3, int width,
    113                                              int height) {
    114  const uint16_t *end = pred_buf_q3 + height * CFL_BUF_LINE;
    115  do {
    116    if (width == 4) {
    117      const uint16x8_t top = vshll_n_u8(vldh_dup_u8(input), 3);
    118      vst1_u16(pred_buf_q3, vget_low_u16(top));
    119    } else if (width == 8) {
    120      const uint16x8_t top = vshll_n_u8(vld1_u8(input), 3);
    121      vst1q_u16(pred_buf_q3, top);
    122    } else {
    123      const uint8x16_t top = vld1q_u8(input);
    124      vst1q_u16(pred_buf_q3, vshll_n_u8(vget_low_u8(top), 3));
    125      vst1q_u16(pred_buf_q3 + 8, vshll_n_u8(vget_high_u8(top), 3));
    126      if (width == 32) {
    127        const uint8x16_t next_top = vld1q_u8(input + 16);
    128        vst1q_u16(pred_buf_q3 + 16, vshll_n_u8(vget_low_u8(next_top), 3));
    129        vst1q_u16(pred_buf_q3 + 24, vshll_n_u8(vget_high_u8(next_top), 3));
    130      }
    131    }
    132    input += input_stride;
    133  } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    134 }
    135 
    136 #if CONFIG_AV1_HIGHBITDEPTH
    137 #if !AOM_ARCH_AARCH64
    138 static uint16x8_t vpaddq_u16(uint16x8_t a, uint16x8_t b) {
    139  return vcombine_u16(vpadd_u16(vget_low_u16(a), vget_high_u16(a)),
    140                      vpadd_u16(vget_low_u16(b), vget_high_u16(b)));
    141 }
    142 #endif
    143 
    144 static void cfl_luma_subsampling_420_hbd_neon(const uint16_t *input,
    145                                              int input_stride,
    146                                              uint16_t *pred_buf_q3, int width,
    147                                              int height) {
    148  const uint16_t *end = pred_buf_q3 + (height >> 1) * CFL_BUF_LINE;
    149  const int luma_stride = input_stride << 1;
    150  do {
    151    if (width == 4) {
    152      const uint16x4_t top = vld1_u16(input);
    153      const uint16x4_t bot = vld1_u16(input + input_stride);
    154      const uint16x4_t sum = vadd_u16(top, bot);
    155      const uint16x4_t hsum = vpadd_u16(sum, sum);
    156      vsth_u16(pred_buf_q3, vshl_n_u16(hsum, 1));
    157    } else if (width < 32) {
    158      const uint16x8_t top = vld1q_u16(input);
    159      const uint16x8_t bot = vld1q_u16(input + input_stride);
    160      const uint16x8_t sum = vaddq_u16(top, bot);
    161      if (width == 8) {
    162        const uint16x4_t hsum = vget_low_u16(vpaddq_u16(sum, sum));
    163        vst1_u16(pred_buf_q3, vshl_n_u16(hsum, 1));
    164      } else {
    165        const uint16x8_t top_1 = vld1q_u16(input + 8);
    166        const uint16x8_t bot_1 = vld1q_u16(input + 8 + input_stride);
    167        const uint16x8_t sum_1 = vaddq_u16(top_1, bot_1);
    168        const uint16x8_t hsum = vpaddq_u16(sum, sum_1);
    169        vst1q_u16(pred_buf_q3, vshlq_n_u16(hsum, 1));
    170      }
    171    } else {
    172      const uint16x8x4_t top = vld4q_u16(input);
    173      const uint16x8x4_t bot = vld4q_u16(input + input_stride);
    174      // equivalent to a vpaddq_u16 (because vld4q interleaves)
    175      const uint16x8_t top_0 = vaddq_u16(top.val[0], top.val[1]);
    176      // equivalent to a vpaddq_u16 (because vld4q interleaves)
    177      const uint16x8_t bot_0 = vaddq_u16(bot.val[0], bot.val[1]);
    178      // equivalent to a vpaddq_u16 (because vld4q interleaves)
    179      const uint16x8_t top_1 = vaddq_u16(top.val[2], top.val[3]);
    180      // equivalent to a vpaddq_u16 (because vld4q interleaves)
    181      const uint16x8_t bot_1 = vaddq_u16(bot.val[2], bot.val[3]);
    182      uint16x8x2_t sum;
    183      sum.val[0] = vshlq_n_u16(vaddq_u16(top_0, bot_0), 1);
    184      sum.val[1] = vshlq_n_u16(vaddq_u16(top_1, bot_1), 1);
    185      vst2q_u16(pred_buf_q3, sum);
    186    }
    187    input += luma_stride;
    188  } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    189 }
    190 
    191 static void cfl_luma_subsampling_422_hbd_neon(const uint16_t *input,
    192                                              int input_stride,
    193                                              uint16_t *pred_buf_q3, int width,
    194                                              int height) {
    195  const uint16_t *end = pred_buf_q3 + height * CFL_BUF_LINE;
    196  do {
    197    if (width == 4) {
    198      const uint16x4_t top = vld1_u16(input);
    199      const uint16x4_t hsum = vpadd_u16(top, top);
    200      vsth_u16(pred_buf_q3, vshl_n_u16(hsum, 2));
    201    } else if (width == 8) {
    202      const uint16x4x2_t top = vld2_u16(input);
    203      // equivalent to a vpadd_u16 (because vld2 interleaves)
    204      const uint16x4_t hsum = vadd_u16(top.val[0], top.val[1]);
    205      vst1_u16(pred_buf_q3, vshl_n_u16(hsum, 2));
    206    } else if (width == 16) {
    207      const uint16x8x2_t top = vld2q_u16(input);
    208      // equivalent to a vpaddq_u16 (because vld2q interleaves)
    209      const uint16x8_t hsum = vaddq_u16(top.val[0], top.val[1]);
    210      vst1q_u16(pred_buf_q3, vshlq_n_u16(hsum, 2));
    211    } else {
    212      const uint16x8x4_t top = vld4q_u16(input);
    213      // equivalent to a vpaddq_u16 (because vld4q interleaves)
    214      const uint16x8_t hsum_0 = vaddq_u16(top.val[0], top.val[1]);
    215      // equivalent to a vpaddq_u16 (because vld4q interleaves)
    216      const uint16x8_t hsum_1 = vaddq_u16(top.val[2], top.val[3]);
    217      uint16x8x2_t result = { { vshlq_n_u16(hsum_0, 2),
    218                                vshlq_n_u16(hsum_1, 2) } };
    219      vst2q_u16(pred_buf_q3, result);
    220    }
    221    input += input_stride;
    222  } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    223 }
    224 
    225 static void cfl_luma_subsampling_444_hbd_neon(const uint16_t *input,
    226                                              int input_stride,
    227                                              uint16_t *pred_buf_q3, int width,
    228                                              int height) {
    229  const uint16_t *end = pred_buf_q3 + height * CFL_BUF_LINE;
    230  do {
    231    if (width == 4) {
    232      const uint16x4_t top = vld1_u16(input);
    233      vst1_u16(pred_buf_q3, vshl_n_u16(top, 3));
    234    } else if (width == 8) {
    235      const uint16x8_t top = vld1q_u16(input);
    236      vst1q_u16(pred_buf_q3, vshlq_n_u16(top, 3));
    237    } else if (width == 16) {
    238      uint16x8x2_t top = vld2q_u16(input);
    239      top.val[0] = vshlq_n_u16(top.val[0], 3);
    240      top.val[1] = vshlq_n_u16(top.val[1], 3);
    241      vst2q_u16(pred_buf_q3, top);
    242    } else {
    243      uint16x8x4_t top = vld4q_u16(input);
    244      top.val[0] = vshlq_n_u16(top.val[0], 3);
    245      top.val[1] = vshlq_n_u16(top.val[1], 3);
    246      top.val[2] = vshlq_n_u16(top.val[2], 3);
    247      top.val[3] = vshlq_n_u16(top.val[3], 3);
    248      vst4q_u16(pred_buf_q3, top);
    249    }
    250    input += input_stride;
    251  } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    252 }
    253 #endif  // CONFIG_AV1_HIGHBITDEPTH
    254 
    255 CFL_GET_SUBSAMPLE_FUNCTION(neon)
    256 
    257 static inline void subtract_average_neon(const uint16_t *src, int16_t *dst,
    258                                         int width, int height,
    259                                         int round_offset,
    260                                         const int num_pel_log2) {
    261  const uint16_t *const end = src + height * CFL_BUF_LINE;
    262 
    263  // Round offset is not needed, because NEON will handle the rounding.
    264  (void)round_offset;
    265 
    266  // To optimize the use of the CPU pipeline, we process 4 rows per iteration
    267  const int step = 4 * CFL_BUF_LINE;
    268 
    269  // At this stage, the prediction buffer contains scaled reconstructed luma
    270  // pixels, which are positive integer and only require 15 bits. By using
    271  // unsigned integer for the sum, we can do one addition operation inside 16
    272  // bits (8 lanes) before having to convert to 32 bits (4 lanes).
    273  const uint16_t *sum_buf = src;
    274  uint32x4_t sum_32x4 = vdupq_n_u32(0);
    275  do {
    276    // For all widths, we load, add and combine the data so it fits in 4 lanes.
    277    if (width == 4) {
    278      const uint16x4_t a0 =
    279          vadd_u16(vld1_u16(sum_buf), vld1_u16(sum_buf + CFL_BUF_LINE));
    280      const uint16x4_t a1 = vadd_u16(vld1_u16(sum_buf + 2 * CFL_BUF_LINE),
    281                                     vld1_u16(sum_buf + 3 * CFL_BUF_LINE));
    282      sum_32x4 = vaddq_u32(sum_32x4, vaddl_u16(a0, a1));
    283    } else if (width == 8) {
    284      const uint16x8_t a0 = vldaddq_u16(sum_buf, CFL_BUF_LINE);
    285      const uint16x8_t a1 =
    286          vldaddq_u16(sum_buf + 2 * CFL_BUF_LINE, CFL_BUF_LINE);
    287      sum_32x4 = vpadalq_u16(sum_32x4, a0);
    288      sum_32x4 = vpadalq_u16(sum_32x4, a1);
    289    } else {
    290      const uint16x8_t row0 = vldaddq_u16(sum_buf, 8);
    291      const uint16x8_t row1 = vldaddq_u16(sum_buf + CFL_BUF_LINE, 8);
    292      const uint16x8_t row2 = vldaddq_u16(sum_buf + 2 * CFL_BUF_LINE, 8);
    293      const uint16x8_t row3 = vldaddq_u16(sum_buf + 3 * CFL_BUF_LINE, 8);
    294      sum_32x4 = vpadalq_u16(sum_32x4, row0);
    295      sum_32x4 = vpadalq_u16(sum_32x4, row1);
    296      sum_32x4 = vpadalq_u16(sum_32x4, row2);
    297      sum_32x4 = vpadalq_u16(sum_32x4, row3);
    298 
    299      if (width == 32) {
    300        const uint16x8_t row0_1 = vldaddq_u16(sum_buf + 16, 8);
    301        const uint16x8_t row1_1 = vldaddq_u16(sum_buf + CFL_BUF_LINE + 16, 8);
    302        const uint16x8_t row2_1 =
    303            vldaddq_u16(sum_buf + 2 * CFL_BUF_LINE + 16, 8);
    304        const uint16x8_t row3_1 =
    305            vldaddq_u16(sum_buf + 3 * CFL_BUF_LINE + 16, 8);
    306 
    307        sum_32x4 = vpadalq_u16(sum_32x4, row0_1);
    308        sum_32x4 = vpadalq_u16(sum_32x4, row1_1);
    309        sum_32x4 = vpadalq_u16(sum_32x4, row2_1);
    310        sum_32x4 = vpadalq_u16(sum_32x4, row3_1);
    311      }
    312    }
    313    sum_buf += step;
    314  } while (sum_buf < end);
    315 
    316  // Permute and add in such a way that each lane contains the block sum.
    317  // [A+C+B+D, B+D+A+C, C+A+D+B, D+B+C+A]
    318 #if AOM_ARCH_AARCH64
    319  sum_32x4 = vpaddq_u32(sum_32x4, sum_32x4);
    320  sum_32x4 = vpaddq_u32(sum_32x4, sum_32x4);
    321 #else
    322  uint32x4_t flip =
    323      vcombine_u32(vget_high_u32(sum_32x4), vget_low_u32(sum_32x4));
    324  sum_32x4 = vaddq_u32(sum_32x4, flip);
    325  sum_32x4 = vaddq_u32(sum_32x4, vrev64q_u32(sum_32x4));
    326 #endif
    327 
    328  // Computing the average could be done using scalars, but getting off the NEON
    329  // engine introduces latency, so we use vqrshrn.
    330  int16x4_t avg_16x4;
    331  // Constant propagation makes for some ugly code.
    332  switch (num_pel_log2) {
    333    case 4: avg_16x4 = vreinterpret_s16_u16(vqrshrn_n_u32(sum_32x4, 4)); break;
    334    case 5: avg_16x4 = vreinterpret_s16_u16(vqrshrn_n_u32(sum_32x4, 5)); break;
    335    case 6: avg_16x4 = vreinterpret_s16_u16(vqrshrn_n_u32(sum_32x4, 6)); break;
    336    case 7: avg_16x4 = vreinterpret_s16_u16(vqrshrn_n_u32(sum_32x4, 7)); break;
    337    case 8: avg_16x4 = vreinterpret_s16_u16(vqrshrn_n_u32(sum_32x4, 8)); break;
    338    case 9: avg_16x4 = vreinterpret_s16_u16(vqrshrn_n_u32(sum_32x4, 9)); break;
    339    case 10:
    340      avg_16x4 = vreinterpret_s16_u16(vqrshrn_n_u32(sum_32x4, 10));
    341      break;
    342    default: assert(0);
    343  }
    344 
    345  if (width == 4) {
    346    do {
    347      vst1_s16(dst, vsub_s16(vreinterpret_s16_u16(vld1_u16(src)), avg_16x4));
    348      src += CFL_BUF_LINE;
    349      dst += CFL_BUF_LINE;
    350    } while (src < end);
    351  } else {
    352    const int16x8_t avg_16x8 = vcombine_s16(avg_16x4, avg_16x4);
    353    do {
    354      vldsubstq_s16(dst, src, 0, avg_16x8);
    355      vldsubstq_s16(dst, src, CFL_BUF_LINE, avg_16x8);
    356      vldsubstq_s16(dst, src, 2 * CFL_BUF_LINE, avg_16x8);
    357      vldsubstq_s16(dst, src, 3 * CFL_BUF_LINE, avg_16x8);
    358 
    359      if (width > 8) {
    360        vldsubstq_s16(dst, src, 8, avg_16x8);
    361        vldsubstq_s16(dst, src, 8 + CFL_BUF_LINE, avg_16x8);
    362        vldsubstq_s16(dst, src, 8 + 2 * CFL_BUF_LINE, avg_16x8);
    363        vldsubstq_s16(dst, src, 8 + 3 * CFL_BUF_LINE, avg_16x8);
    364      }
    365      if (width == 32) {
    366        vldsubstq_s16(dst, src, 16, avg_16x8);
    367        vldsubstq_s16(dst, src, 16 + CFL_BUF_LINE, avg_16x8);
    368        vldsubstq_s16(dst, src, 16 + 2 * CFL_BUF_LINE, avg_16x8);
    369        vldsubstq_s16(dst, src, 16 + 3 * CFL_BUF_LINE, avg_16x8);
    370        vldsubstq_s16(dst, src, 24, avg_16x8);
    371        vldsubstq_s16(dst, src, 24 + CFL_BUF_LINE, avg_16x8);
    372        vldsubstq_s16(dst, src, 24 + 2 * CFL_BUF_LINE, avg_16x8);
    373        vldsubstq_s16(dst, src, 24 + 3 * CFL_BUF_LINE, avg_16x8);
    374      }
    375      src += step;
    376      dst += step;
    377    } while (src < end);
    378  }
    379 }
    380 
    381 CFL_SUB_AVG_FN(neon)
    382 
    383 // Saturating negate 16-bit integers in a when the corresponding signed 16-bit
    384 // integer in b is negative.
    385 // Notes:
    386 //   * Negating INT16_MIN results in INT16_MIN. However, this cannot occur in
    387 //   practice, as scaled_luma is the multiplication of two absolute values.
    388 //   * In the Intel equivalent, elements in a are zeroed out when the
    389 //   corresponding elements in b are zero. Because vsign is used twice in a
    390 //   row, with b in the first call becoming a in the second call, there's no
    391 //   impact from not zeroing out.
    392 static int16x4_t vsign_s16(int16x4_t a, int16x4_t b) {
    393  const int16x4_t mask = vshr_n_s16(b, 15);
    394  return veor_s16(vadd_s16(a, mask), mask);
    395 }
    396 
    397 // Saturating negate 16-bit integers in a when the corresponding signed 16-bit
    398 // integer in b is negative.
    399 // Notes:
    400 //   * Negating INT16_MIN results in INT16_MIN. However, this cannot occur in
    401 //   practice, as scaled_luma is the multiplication of two absolute values.
    402 //   * In the Intel equivalent, elements in a are zeroed out when the
    403 //   corresponding elements in b are zero. Because vsignq is used twice in a
    404 //   row, with b in the first call becoming a in the second call, there's no
    405 //   impact from not zeroing out.
    406 static int16x8_t vsignq_s16(int16x8_t a, int16x8_t b) {
    407  const int16x8_t mask = vshrq_n_s16(b, 15);
    408  return veorq_s16(vaddq_s16(a, mask), mask);
    409 }
    410 
    411 static inline int16x4_t predict_w4(const int16_t *pred_buf_q3,
    412                                   int16x4_t alpha_sign, int abs_alpha_q12,
    413                                   int16x4_t dc) {
    414  const int16x4_t ac_q3 = vld1_s16(pred_buf_q3);
    415  const int16x4_t ac_sign = veor_s16(alpha_sign, ac_q3);
    416  int16x4_t scaled_luma = vqrdmulh_n_s16(vabs_s16(ac_q3), abs_alpha_q12);
    417  return vadd_s16(vsign_s16(scaled_luma, ac_sign), dc);
    418 }
    419 
    420 static inline int16x8_t predict_w8(const int16_t *pred_buf_q3,
    421                                   int16x8_t alpha_sign, int abs_alpha_q12,
    422                                   int16x8_t dc) {
    423  const int16x8_t ac_q3 = vld1q_s16(pred_buf_q3);
    424  const int16x8_t ac_sign = veorq_s16(alpha_sign, ac_q3);
    425  int16x8_t scaled_luma = vqrdmulhq_n_s16(vabsq_s16(ac_q3), abs_alpha_q12);
    426  return vaddq_s16(vsignq_s16(scaled_luma, ac_sign), dc);
    427 }
    428 
    429 static inline int16x8x2_t predict_w16(const int16_t *pred_buf_q3,
    430                                      int16x8_t alpha_sign, int abs_alpha_q12,
    431                                      int16x8_t dc) {
    432  const int16x8x2_t ac_q3 = vld1q_s16_x2(pred_buf_q3);
    433  const int16x8_t ac_sign_0 = veorq_s16(alpha_sign, ac_q3.val[0]);
    434  const int16x8_t ac_sign_1 = veorq_s16(alpha_sign, ac_q3.val[1]);
    435  const int16x8_t scaled_luma_0 =
    436      vqrdmulhq_n_s16(vabsq_s16(ac_q3.val[0]), abs_alpha_q12);
    437  const int16x8_t scaled_luma_1 =
    438      vqrdmulhq_n_s16(vabsq_s16(ac_q3.val[1]), abs_alpha_q12);
    439  int16x8x2_t result;
    440  result.val[0] = vaddq_s16(vsignq_s16(scaled_luma_0, ac_sign_0), dc);
    441  result.val[1] = vaddq_s16(vsignq_s16(scaled_luma_1, ac_sign_1), dc);
    442  return result;
    443 }
    444 
    445 static inline int16x8x4_t predict_w32(const int16_t *pred_buf_q3,
    446                                      int16x8_t alpha_sign, int abs_alpha_q12,
    447                                      int16x8_t dc) {
    448  const int16x8x4_t ac_q3 = vld1q_s16_x4(pred_buf_q3);
    449  const int16x8_t ac_sign_0 = veorq_s16(alpha_sign, ac_q3.val[0]);
    450  const int16x8_t ac_sign_1 = veorq_s16(alpha_sign, ac_q3.val[1]);
    451  const int16x8_t ac_sign_2 = veorq_s16(alpha_sign, ac_q3.val[2]);
    452  const int16x8_t ac_sign_3 = veorq_s16(alpha_sign, ac_q3.val[3]);
    453  const int16x8_t scaled_luma_0 =
    454      vqrdmulhq_n_s16(vabsq_s16(ac_q3.val[0]), abs_alpha_q12);
    455  const int16x8_t scaled_luma_1 =
    456      vqrdmulhq_n_s16(vabsq_s16(ac_q3.val[1]), abs_alpha_q12);
    457  const int16x8_t scaled_luma_2 =
    458      vqrdmulhq_n_s16(vabsq_s16(ac_q3.val[2]), abs_alpha_q12);
    459  const int16x8_t scaled_luma_3 =
    460      vqrdmulhq_n_s16(vabsq_s16(ac_q3.val[3]), abs_alpha_q12);
    461  int16x8x4_t result;
    462  result.val[0] = vaddq_s16(vsignq_s16(scaled_luma_0, ac_sign_0), dc);
    463  result.val[1] = vaddq_s16(vsignq_s16(scaled_luma_1, ac_sign_1), dc);
    464  result.val[2] = vaddq_s16(vsignq_s16(scaled_luma_2, ac_sign_2), dc);
    465  result.val[3] = vaddq_s16(vsignq_s16(scaled_luma_3, ac_sign_3), dc);
    466  return result;
    467 }
    468 
    469 static inline void cfl_predict_lbd_neon(const int16_t *pred_buf_q3,
    470                                        uint8_t *dst, int dst_stride,
    471                                        int alpha_q3, int width, int height) {
    472  const int16_t abs_alpha_q12 = abs(alpha_q3) << 9;
    473  const int16_t *const end = pred_buf_q3 + height * CFL_BUF_LINE;
    474  if (width == 4) {
    475    const int16x4_t alpha_sign = vdup_n_s16(alpha_q3);
    476    const int16x4_t dc = vdup_n_s16(*dst);
    477    do {
    478      const int16x4_t pred =
    479          predict_w4(pred_buf_q3, alpha_sign, abs_alpha_q12, dc);
    480      vsth_u8(dst, vqmovun_s16(vcombine_s16(pred, pred)));
    481      dst += dst_stride;
    482    } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    483  } else {
    484    const int16x8_t alpha_sign = vdupq_n_s16(alpha_q3);
    485    const int16x8_t dc = vdupq_n_s16(*dst);
    486    do {
    487      if (width == 8) {
    488        vst1_u8(dst, vqmovun_s16(predict_w8(pred_buf_q3, alpha_sign,
    489                                            abs_alpha_q12, dc)));
    490      } else if (width == 16) {
    491        const int16x8x2_t pred =
    492            predict_w16(pred_buf_q3, alpha_sign, abs_alpha_q12, dc);
    493        const uint8x8x2_t predun = { { vqmovun_s16(pred.val[0]),
    494                                       vqmovun_s16(pred.val[1]) } };
    495        vst1_u8_x2(dst, predun);
    496      } else {
    497        const int16x8x4_t pred =
    498            predict_w32(pred_buf_q3, alpha_sign, abs_alpha_q12, dc);
    499        const uint8x8x4_t predun = {
    500          { vqmovun_s16(pred.val[0]), vqmovun_s16(pred.val[1]),
    501            vqmovun_s16(pred.val[2]), vqmovun_s16(pred.val[3]) }
    502        };
    503        vst1_u8_x4(dst, predun);
    504      }
    505      dst += dst_stride;
    506    } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    507  }
    508 }
    509 
    510 CFL_PREDICT_FN(neon, lbd)
    511 
    512 #if CONFIG_AV1_HIGHBITDEPTH
    513 static inline uint16x4_t clamp_s16(int16x4_t a, int16x4_t max) {
    514  return vreinterpret_u16_s16(vmax_s16(vmin_s16(a, max), vdup_n_s16(0)));
    515 }
    516 
    517 static inline uint16x8_t clampq_s16(int16x8_t a, int16x8_t max) {
    518  return vreinterpretq_u16_s16(vmaxq_s16(vminq_s16(a, max), vdupq_n_s16(0)));
    519 }
    520 
    521 static inline uint16x8x2_t clamp2q_s16(int16x8x2_t a, int16x8_t max) {
    522  uint16x8x2_t result;
    523  result.val[0] = vreinterpretq_u16_s16(
    524      vmaxq_s16(vminq_s16(a.val[0], max), vdupq_n_s16(0)));
    525  result.val[1] = vreinterpretq_u16_s16(
    526      vmaxq_s16(vminq_s16(a.val[1], max), vdupq_n_s16(0)));
    527  return result;
    528 }
    529 
    530 static inline uint16x8x4_t clamp4q_s16(int16x8x4_t a, int16x8_t max) {
    531  uint16x8x4_t result;
    532  result.val[0] = vreinterpretq_u16_s16(
    533      vmaxq_s16(vminq_s16(a.val[0], max), vdupq_n_s16(0)));
    534  result.val[1] = vreinterpretq_u16_s16(
    535      vmaxq_s16(vminq_s16(a.val[1], max), vdupq_n_s16(0)));
    536  result.val[2] = vreinterpretq_u16_s16(
    537      vmaxq_s16(vminq_s16(a.val[2], max), vdupq_n_s16(0)));
    538  result.val[3] = vreinterpretq_u16_s16(
    539      vmaxq_s16(vminq_s16(a.val[3], max), vdupq_n_s16(0)));
    540  return result;
    541 }
    542 
    543 static inline void cfl_predict_hbd_neon(const int16_t *pred_buf_q3,
    544                                        uint16_t *dst, int dst_stride,
    545                                        int alpha_q3, int bd, int width,
    546                                        int height) {
    547  const int max = (1 << bd) - 1;
    548  const int16_t abs_alpha_q12 = abs(alpha_q3) << 9;
    549  const int16_t *const end = pred_buf_q3 + height * CFL_BUF_LINE;
    550  if (width == 4) {
    551    const int16x4_t alpha_sign = vdup_n_s16(alpha_q3);
    552    const int16x4_t dc = vdup_n_s16(*dst);
    553    const int16x4_t max_16x4 = vdup_n_s16(max);
    554    do {
    555      const int16x4_t scaled_luma =
    556          predict_w4(pred_buf_q3, alpha_sign, abs_alpha_q12, dc);
    557      vst1_u16(dst, clamp_s16(scaled_luma, max_16x4));
    558      dst += dst_stride;
    559    } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    560  } else {
    561    const int16x8_t alpha_sign = vdupq_n_s16(alpha_q3);
    562    const int16x8_t dc = vdupq_n_s16(*dst);
    563    const int16x8_t max_16x8 = vdupq_n_s16(max);
    564    do {
    565      if (width == 8) {
    566        const int16x8_t pred =
    567            predict_w8(pred_buf_q3, alpha_sign, abs_alpha_q12, dc);
    568        vst1q_u16(dst, clampq_s16(pred, max_16x8));
    569      } else if (width == 16) {
    570        const int16x8x2_t pred =
    571            predict_w16(pred_buf_q3, alpha_sign, abs_alpha_q12, dc);
    572        vst1q_u16_x2(dst, clamp2q_s16(pred, max_16x8));
    573      } else {
    574        const int16x8x4_t pred =
    575            predict_w32(pred_buf_q3, alpha_sign, abs_alpha_q12, dc);
    576        vst1q_u16_x4(dst, clamp4q_s16(pred, max_16x8));
    577      }
    578      dst += dst_stride;
    579    } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
    580  }
    581 }
    582 
    583 CFL_PREDICT_FN(neon, hbd)
    584 #endif  // CONFIG_AV1_HIGHBITDEPTH