tor-browser

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

highbd_quantize_neon.c (19145B)


      1 /*
      2 * Copyright (c) 2022, 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 <arm_neon.h>
     13 #include <assert.h>
     14 #include <string.h>
     15 
     16 #include "config/aom_config.h"
     17 #include "config/aom_dsp_rtcd.h"
     18 
     19 #include "aom_dsp/quantize.h"
     20 
     21 static inline uint32_t sum_abs_coeff(const uint32x4_t a) {
     22 #if AOM_ARCH_AARCH64
     23  return vaddvq_u32(a);
     24 #else
     25  const uint64x2_t b = vpaddlq_u32(a);
     26  const uint64x1_t c = vadd_u64(vget_low_u64(b), vget_high_u64(b));
     27  return (uint32_t)vget_lane_u64(c, 0);
     28 #endif
     29 }
     30 
     31 static inline uint16x4_t quantize_4(
     32    const tran_low_t *coeff_ptr, tran_low_t *qcoeff_ptr,
     33    tran_low_t *dqcoeff_ptr, int32x4_t v_quant_s32, int32x4_t v_dequant_s32,
     34    int32x4_t v_round_s32, int32x4_t v_zbin_s32, int32x4_t v_quant_shift_s32,
     35    int log_scale) {
     36  const int32x4_t v_coeff = vld1q_s32(coeff_ptr);
     37  const int32x4_t v_coeff_sign =
     38      vreinterpretq_s32_u32(vcltq_s32(v_coeff, vdupq_n_s32(0)));
     39  const int32x4_t v_abs_coeff = vabsq_s32(v_coeff);
     40  // if (abs_coeff < zbins[rc != 0]),
     41  const uint32x4_t v_zbin_mask = vcgeq_s32(v_abs_coeff, v_zbin_s32);
     42  const int32x4_t v_log_scale = vdupq_n_s32(log_scale);
     43  // const int64_t tmp = (int64_t)abs_coeff + log_scaled_round;
     44  const int32x4_t v_tmp = vaddq_s32(v_abs_coeff, v_round_s32);
     45  //  const int32_t tmpw32 = tmp * wt;
     46  const int32x4_t v_tmpw32 = vmulq_s32(v_tmp, vdupq_n_s32((1 << AOM_QM_BITS)));
     47  //  const int32_t tmp2 = (int32_t)((tmpw32 * quant64) >> 16);
     48  const int32x4_t v_tmp2 = vqdmulhq_s32(v_tmpw32, v_quant_s32);
     49  // const int32_t tmp3 =
     50  //    ((((tmp2 + tmpw32)<< log_scale) * (int64_t)(quant_shift << 15)) >> 32);
     51  const int32x4_t v_tmp3 = vqdmulhq_s32(
     52      vshlq_s32(vaddq_s32(v_tmp2, v_tmpw32), v_log_scale), v_quant_shift_s32);
     53  // const int abs_qcoeff = vmask ? (int)tmp3 >> AOM_QM_BITS : 0;
     54  const int32x4_t v_abs_qcoeff = vandq_s32(vreinterpretq_s32_u32(v_zbin_mask),
     55                                           vshrq_n_s32(v_tmp3, AOM_QM_BITS));
     56  // const tran_low_t abs_dqcoeff = (abs_qcoeff * dequant_iwt) >> log_scale;
     57  // vshlq_s32 will shift right if shift value is negative.
     58  const int32x4_t v_abs_dqcoeff =
     59      vshlq_s32(vmulq_s32(v_abs_qcoeff, v_dequant_s32), vnegq_s32(v_log_scale));
     60  //  qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
     61  const int32x4_t v_qcoeff =
     62      vsubq_s32(veorq_s32(v_abs_qcoeff, v_coeff_sign), v_coeff_sign);
     63  //  dqcoeff_ptr[rc] = (tran_low_t)((abs_dqcoeff ^ coeff_sign) - coeff_sign);
     64  const int32x4_t v_dqcoeff =
     65      vsubq_s32(veorq_s32(v_abs_dqcoeff, v_coeff_sign), v_coeff_sign);
     66 
     67  vst1q_s32(qcoeff_ptr, v_qcoeff);
     68  vst1q_s32(dqcoeff_ptr, v_dqcoeff);
     69 
     70  // Used to find eob.
     71  const uint32x4_t nz_qcoeff_mask = vcgtq_s32(v_abs_qcoeff, vdupq_n_s32(0));
     72  return vmovn_u32(nz_qcoeff_mask);
     73 }
     74 
     75 static inline int16x8_t get_max_lane_eob(const int16_t *iscan,
     76                                         int16x8_t v_eobmax,
     77                                         uint16x8_t v_mask) {
     78  const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
     79  const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, vdupq_n_s16(1));
     80  const int16x8_t v_nz_iscan = vbslq_s16(v_mask, v_iscan_plus1, vdupq_n_s16(0));
     81  return vmaxq_s16(v_eobmax, v_nz_iscan);
     82 }
     83 
     84 #if !CONFIG_REALTIME_ONLY
     85 static inline void get_min_max_lane_eob(const int16_t *iscan,
     86                                        int16x8_t *v_eobmin,
     87                                        int16x8_t *v_eobmax, uint16x8_t v_mask,
     88                                        intptr_t n_coeffs) {
     89  const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
     90  const int16x8_t v_nz_iscan_max = vbslq_s16(v_mask, v_iscan, vdupq_n_s16(-1));
     91 #if SKIP_EOB_FACTOR_ADJUST
     92  const int16x8_t v_nz_iscan_min =
     93      vbslq_s16(v_mask, v_iscan, vdupq_n_s16((int16_t)n_coeffs));
     94  *v_eobmin = vminq_s16(*v_eobmin, v_nz_iscan_min);
     95 #else
     96  (void)v_eobmin;
     97 #endif
     98  *v_eobmax = vmaxq_s16(*v_eobmax, v_nz_iscan_max);
     99 }
    100 #endif  // !CONFIG_REALTIME_ONLY
    101 
    102 static inline uint16_t get_max_eob(int16x8_t v_eobmax) {
    103 #if AOM_ARCH_AARCH64
    104  return (uint16_t)vmaxvq_s16(v_eobmax);
    105 #else
    106  const int16x4_t v_eobmax_3210 =
    107      vmax_s16(vget_low_s16(v_eobmax), vget_high_s16(v_eobmax));
    108  const int64x1_t v_eobmax_xx32 =
    109      vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
    110  const int16x4_t v_eobmax_tmp =
    111      vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
    112  const int64x1_t v_eobmax_xxx3 =
    113      vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
    114  const int16x4_t v_eobmax_final =
    115      vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
    116  return (uint16_t)vget_lane_s16(v_eobmax_final, 0);
    117 #endif
    118 }
    119 
    120 #if SKIP_EOB_FACTOR_ADJUST && !CONFIG_REALTIME_ONLY
    121 static inline uint16_t get_min_eob(int16x8_t v_eobmin) {
    122 #if AOM_ARCH_AARCH64
    123  return (uint16_t)vminvq_s16(v_eobmin);
    124 #else
    125  const int16x4_t v_eobmin_3210 =
    126      vmin_s16(vget_low_s16(v_eobmin), vget_high_s16(v_eobmin));
    127  const int64x1_t v_eobmin_xx32 =
    128      vshr_n_s64(vreinterpret_s64_s16(v_eobmin_3210), 32);
    129  const int16x4_t v_eobmin_tmp =
    130      vmin_s16(v_eobmin_3210, vreinterpret_s16_s64(v_eobmin_xx32));
    131  const int64x1_t v_eobmin_xxx3 =
    132      vshr_n_s64(vreinterpret_s64_s16(v_eobmin_tmp), 16);
    133  const int16x4_t v_eobmin_final =
    134      vmin_s16(v_eobmin_tmp, vreinterpret_s16_s64(v_eobmin_xxx3));
    135  return (uint16_t)vget_lane_s16(v_eobmin_final, 0);
    136 #endif
    137 }
    138 #endif  // SKIP_EOB_FACTOR_ADJUST && !CONFIG_REALTIME_ONLY
    139 
    140 static void highbd_quantize_b_neon(
    141    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
    142    const int16_t *round_ptr, const int16_t *quant_ptr,
    143    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
    144    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
    145    const int16_t *scan, const int16_t *iscan, const int log_scale) {
    146  (void)scan;
    147  const int16x4_t v_quant = vld1_s16(quant_ptr);
    148  const int16x4_t v_dequant = vld1_s16(dequant_ptr);
    149  const int16x4_t v_zero = vdup_n_s16(0);
    150  const uint16x4_t v_round_select = vcgt_s16(vdup_n_s16(log_scale), v_zero);
    151  const int16x4_t v_round_no_scale = vld1_s16(round_ptr);
    152  const int16x4_t v_round_log_scale =
    153      vqrdmulh_n_s16(v_round_no_scale, (int16_t)(1 << (15 - log_scale)));
    154  const int16x4_t v_round =
    155      vbsl_s16(v_round_select, v_round_log_scale, v_round_no_scale);
    156  const int16x4_t v_quant_shift = vld1_s16(quant_shift_ptr);
    157  const int16x4_t v_zbin_no_scale = vld1_s16(zbin_ptr);
    158  const int16x4_t v_zbin_log_scale =
    159      vqrdmulh_n_s16(v_zbin_no_scale, (int16_t)(1 << (15 - log_scale)));
    160  const int16x4_t v_zbin =
    161      vbsl_s16(v_round_select, v_zbin_log_scale, v_zbin_no_scale);
    162  int32x4_t v_round_s32 = vmovl_s16(v_round);
    163  int32x4_t v_quant_s32 = vshlq_n_s32(vmovl_s16(v_quant), 15);
    164  int32x4_t v_dequant_s32 = vmovl_s16(v_dequant);
    165  int32x4_t v_quant_shift_s32 = vshlq_n_s32(vmovl_s16(v_quant_shift), 15);
    166  int32x4_t v_zbin_s32 = vmovl_s16(v_zbin);
    167  uint16x4_t v_mask_lo, v_mask_hi;
    168  int16x8_t v_eobmax = vdupq_n_s16(-1);
    169 
    170  intptr_t non_zero_count = n_coeffs;
    171 
    172  assert(n_coeffs > 8);
    173  // Pre-scan pass
    174  const int32x4_t v_zbin_s32x = vdupq_lane_s32(vget_low_s32(v_zbin_s32), 1);
    175  intptr_t i = n_coeffs;
    176  do {
    177    const int32x4_t v_coeff_a = vld1q_s32(coeff_ptr + i - 4);
    178    const int32x4_t v_coeff_b = vld1q_s32(coeff_ptr + i - 8);
    179    const int32x4_t v_abs_coeff_a = vabsq_s32(v_coeff_a);
    180    const int32x4_t v_abs_coeff_b = vabsq_s32(v_coeff_b);
    181    const uint32x4_t v_mask_a = vcgeq_s32(v_abs_coeff_a, v_zbin_s32x);
    182    const uint32x4_t v_mask_b = vcgeq_s32(v_abs_coeff_b, v_zbin_s32x);
    183    // If the coefficient is in the base ZBIN range, then discard.
    184    if (sum_abs_coeff(v_mask_a) + sum_abs_coeff(v_mask_b) == 0) {
    185      non_zero_count -= 8;
    186    } else {
    187      break;
    188    }
    189    i -= 8;
    190  } while (i > 0);
    191 
    192  const intptr_t remaining_zcoeffs = n_coeffs - non_zero_count;
    193  memset(qcoeff_ptr + non_zero_count, 0,
    194         remaining_zcoeffs * sizeof(*qcoeff_ptr));
    195  memset(dqcoeff_ptr + non_zero_count, 0,
    196         remaining_zcoeffs * sizeof(*dqcoeff_ptr));
    197 
    198  // DC and first 3 AC
    199  v_mask_lo =
    200      quantize_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, v_quant_s32, v_dequant_s32,
    201                 v_round_s32, v_zbin_s32, v_quant_shift_s32, log_scale);
    202 
    203  // overwrite the DC constants with AC constants
    204  v_round_s32 = vdupq_lane_s32(vget_low_s32(v_round_s32), 1);
    205  v_quant_s32 = vdupq_lane_s32(vget_low_s32(v_quant_s32), 1);
    206  v_dequant_s32 = vdupq_lane_s32(vget_low_s32(v_dequant_s32), 1);
    207  v_quant_shift_s32 = vdupq_lane_s32(vget_low_s32(v_quant_shift_s32), 1);
    208  v_zbin_s32 = vdupq_lane_s32(vget_low_s32(v_zbin_s32), 1);
    209 
    210  // 4 more AC
    211  v_mask_hi = quantize_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
    212                         v_quant_s32, v_dequant_s32, v_round_s32, v_zbin_s32,
    213                         v_quant_shift_s32, log_scale);
    214 
    215  v_eobmax =
    216      get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
    217 
    218  intptr_t count = non_zero_count - 8;
    219  for (; count > 0; count -= 8) {
    220    coeff_ptr += 8;
    221    qcoeff_ptr += 8;
    222    dqcoeff_ptr += 8;
    223    iscan += 8;
    224    v_mask_lo = quantize_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, v_quant_s32,
    225                           v_dequant_s32, v_round_s32, v_zbin_s32,
    226                           v_quant_shift_s32, log_scale);
    227    v_mask_hi = quantize_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
    228                           v_quant_s32, v_dequant_s32, v_round_s32, v_zbin_s32,
    229                           v_quant_shift_s32, log_scale);
    230    // Find the max lane eob for 8 coeffs.
    231    v_eobmax =
    232        get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
    233  }
    234 
    235  *eob_ptr = get_max_eob(v_eobmax);
    236 }
    237 
    238 void aom_highbd_quantize_b_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
    239                                const int16_t *zbin_ptr,
    240                                const int16_t *round_ptr,
    241                                const int16_t *quant_ptr,
    242                                const int16_t *quant_shift_ptr,
    243                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
    244                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
    245                                const int16_t *scan, const int16_t *iscan) {
    246  highbd_quantize_b_neon(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr,
    247                         quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr,
    248                         eob_ptr, scan, iscan, 0);
    249 }
    250 
    251 void aom_highbd_quantize_b_32x32_neon(
    252    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
    253    const int16_t *round_ptr, const int16_t *quant_ptr,
    254    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
    255    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
    256    const int16_t *scan, const int16_t *iscan) {
    257  highbd_quantize_b_neon(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr,
    258                         quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr,
    259                         eob_ptr, scan, iscan, 1);
    260 }
    261 
    262 void aom_highbd_quantize_b_64x64_neon(
    263    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
    264    const int16_t *round_ptr, const int16_t *quant_ptr,
    265    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
    266    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
    267    const int16_t *scan, const int16_t *iscan) {
    268  highbd_quantize_b_neon(coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr,
    269                         quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr,
    270                         eob_ptr, scan, iscan, 2);
    271 }
    272 
    273 #if !CONFIG_REALTIME_ONLY
    274 static void highbd_quantize_b_adaptive_neon(
    275    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
    276    const int16_t *round_ptr, const int16_t *quant_ptr,
    277    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
    278    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
    279    const int16_t *scan, const int16_t *iscan, const int log_scale) {
    280  (void)scan;
    281  const int16x4_t v_quant = vld1_s16(quant_ptr);
    282  const int16x4_t v_dequant = vld1_s16(dequant_ptr);
    283  const int16x4_t v_zero = vdup_n_s16(0);
    284  const uint16x4_t v_round_select = vcgt_s16(vdup_n_s16(log_scale), v_zero);
    285  const int16x4_t v_round_no_scale = vld1_s16(round_ptr);
    286  const int16x4_t v_round_log_scale =
    287      vqrdmulh_n_s16(v_round_no_scale, (int16_t)(1 << (15 - log_scale)));
    288  const int16x4_t v_round =
    289      vbsl_s16(v_round_select, v_round_log_scale, v_round_no_scale);
    290  const int16x4_t v_quant_shift = vld1_s16(quant_shift_ptr);
    291  const int16x4_t v_zbin_no_scale = vld1_s16(zbin_ptr);
    292  const int16x4_t v_zbin_log_scale =
    293      vqrdmulh_n_s16(v_zbin_no_scale, (int16_t)(1 << (15 - log_scale)));
    294  const int16x4_t v_zbin =
    295      vbsl_s16(v_round_select, v_zbin_log_scale, v_zbin_no_scale);
    296  int32x4_t v_round_s32 = vmovl_s16(v_round);
    297  int32x4_t v_quant_s32 = vshlq_n_s32(vmovl_s16(v_quant), 15);
    298  int32x4_t v_dequant_s32 = vmovl_s16(v_dequant);
    299  int32x4_t v_quant_shift_s32 = vshlq_n_s32(vmovl_s16(v_quant_shift), 15);
    300  int32x4_t v_zbin_s32 = vmovl_s16(v_zbin);
    301  uint16x4_t v_mask_lo, v_mask_hi;
    302  int16x8_t v_eobmax = vdupq_n_s16(-1);
    303  int16x8_t v_eobmin = vdupq_n_s16((int16_t)n_coeffs);
    304 
    305  assert(n_coeffs > 8);
    306  // Pre-scan pass
    307  const int32x4_t v_zbin_s32x = vdupq_lane_s32(vget_low_s32(v_zbin_s32), 1);
    308  const int prescan_add_1 =
    309      ROUND_POWER_OF_TWO(dequant_ptr[1] * EOB_FACTOR, 7 + AOM_QM_BITS);
    310  const int32x4_t v_zbin_prescan =
    311      vaddq_s32(v_zbin_s32x, vdupq_n_s32(prescan_add_1));
    312  intptr_t non_zero_count = n_coeffs;
    313  intptr_t i = n_coeffs;
    314  do {
    315    const int32x4_t v_coeff_a = vld1q_s32(coeff_ptr + i - 4);
    316    const int32x4_t v_coeff_b = vld1q_s32(coeff_ptr + i - 8);
    317    const int32x4_t v_abs_coeff_a = vabsq_s32(v_coeff_a);
    318    const int32x4_t v_abs_coeff_b = vabsq_s32(v_coeff_b);
    319    const uint32x4_t v_mask_a = vcgeq_s32(v_abs_coeff_a, v_zbin_prescan);
    320    const uint32x4_t v_mask_b = vcgeq_s32(v_abs_coeff_b, v_zbin_prescan);
    321    // If the coefficient is in the base ZBIN range, then discard.
    322    if (sum_abs_coeff(v_mask_a) + sum_abs_coeff(v_mask_b) == 0) {
    323      non_zero_count -= 8;
    324    } else {
    325      break;
    326    }
    327    i -= 8;
    328  } while (i > 0);
    329 
    330  const intptr_t remaining_zcoeffs = n_coeffs - non_zero_count;
    331  memset(qcoeff_ptr + non_zero_count, 0,
    332         remaining_zcoeffs * sizeof(*qcoeff_ptr));
    333  memset(dqcoeff_ptr + non_zero_count, 0,
    334         remaining_zcoeffs * sizeof(*dqcoeff_ptr));
    335 
    336  // DC and first 3 AC
    337  v_mask_lo =
    338      quantize_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, v_quant_s32, v_dequant_s32,
    339                 v_round_s32, v_zbin_s32, v_quant_shift_s32, log_scale);
    340 
    341  // overwrite the DC constants with AC constants
    342  v_round_s32 = vdupq_lane_s32(vget_low_s32(v_round_s32), 1);
    343  v_quant_s32 = vdupq_lane_s32(vget_low_s32(v_quant_s32), 1);
    344  v_dequant_s32 = vdupq_lane_s32(vget_low_s32(v_dequant_s32), 1);
    345  v_quant_shift_s32 = vdupq_lane_s32(vget_low_s32(v_quant_shift_s32), 1);
    346  v_zbin_s32 = vdupq_lane_s32(vget_low_s32(v_zbin_s32), 1);
    347 
    348  // 4 more AC
    349  v_mask_hi = quantize_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
    350                         v_quant_s32, v_dequant_s32, v_round_s32, v_zbin_s32,
    351                         v_quant_shift_s32, log_scale);
    352 
    353  get_min_max_lane_eob(iscan, &v_eobmin, &v_eobmax,
    354                       vcombine_u16(v_mask_lo, v_mask_hi), n_coeffs);
    355 
    356  intptr_t count = non_zero_count - 8;
    357  for (; count > 0; count -= 8) {
    358    coeff_ptr += 8;
    359    qcoeff_ptr += 8;
    360    dqcoeff_ptr += 8;
    361    iscan += 8;
    362    v_mask_lo = quantize_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, v_quant_s32,
    363                           v_dequant_s32, v_round_s32, v_zbin_s32,
    364                           v_quant_shift_s32, log_scale);
    365    v_mask_hi = quantize_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
    366                           v_quant_s32, v_dequant_s32, v_round_s32, v_zbin_s32,
    367                           v_quant_shift_s32, log_scale);
    368 
    369    get_min_max_lane_eob(iscan, &v_eobmin, &v_eobmax,
    370                         vcombine_u16(v_mask_lo, v_mask_hi), n_coeffs);
    371  }
    372 
    373  int eob = get_max_eob(v_eobmax);
    374 
    375 #if SKIP_EOB_FACTOR_ADJUST
    376  const int first = get_min_eob(v_eobmin);
    377  if (eob >= 0 && first == eob) {
    378    const int rc = scan[eob];
    379    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
    380      const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale),
    381                             ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) };
    382      const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
    383      const qm_val_t wt = (1 << AOM_QM_BITS);
    384      const int coeff = coeff_ptr[rc] * wt;
    385      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
    386      const int prescan_add_val =
    387          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
    388      if (coeff < (zbins[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val) &&
    389          coeff > (nzbins[rc != 0] * (1 << AOM_QM_BITS) - prescan_add_val)) {
    390        qcoeff_ptr[rc] = 0;
    391        dqcoeff_ptr[rc] = 0;
    392        eob = -1;
    393      }
    394    }
    395  }
    396 #endif  // SKIP_EOB_FACTOR_ADJUST
    397  *eob_ptr = eob + 1;
    398 }
    399 
    400 void aom_highbd_quantize_b_adaptive_neon(
    401    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
    402    const int16_t *round_ptr, const int16_t *quant_ptr,
    403    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
    404    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
    405    const int16_t *scan, const int16_t *iscan) {
    406  highbd_quantize_b_adaptive_neon(
    407      coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr, quant_shift_ptr,
    408      qcoeff_ptr, dqcoeff_ptr, dequant_ptr, eob_ptr, scan, iscan, 0);
    409 }
    410 
    411 void aom_highbd_quantize_b_32x32_adaptive_neon(
    412    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
    413    const int16_t *round_ptr, const int16_t *quant_ptr,
    414    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
    415    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
    416    const int16_t *scan, const int16_t *iscan) {
    417  highbd_quantize_b_adaptive_neon(
    418      coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr, quant_shift_ptr,
    419      qcoeff_ptr, dqcoeff_ptr, dequant_ptr, eob_ptr, scan, iscan, 1);
    420 }
    421 
    422 void aom_highbd_quantize_b_64x64_adaptive_neon(
    423    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
    424    const int16_t *round_ptr, const int16_t *quant_ptr,
    425    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
    426    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
    427    const int16_t *scan, const int16_t *iscan) {
    428  highbd_quantize_b_adaptive_neon(
    429      coeff_ptr, n_coeffs, zbin_ptr, round_ptr, quant_ptr, quant_shift_ptr,
    430      qcoeff_ptr, dqcoeff_ptr, dequant_ptr, eob_ptr, scan, iscan, 2);
    431 }
    432 #endif  // !CONFIG_REALTIME_ONLY