tor-browser

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

highbd_sadxd_neon.c (25781B)


      1 /*
      2 * Copyright (c) 2023 The WebM project authors. All rights reserved.
      3 * Copyright (c) 2023, Alliance for Open Media. All rights reserved.
      4 *
      5 * This source code is subject to the terms of the BSD 2 Clause License and
      6 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
      7 * was not distributed with this source code in the LICENSE file, you can
      8 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
      9 * Media Patent License 1.0 was not distributed with this source code in the
     10 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
     11 */
     12 
     13 #include <arm_neon.h>
     14 
     15 #include "config/aom_config.h"
     16 #include "config/aom_dsp_rtcd.h"
     17 
     18 #include "aom/aom_integer.h"
     19 #include "aom_dsp/arm/mem_neon.h"
     20 #include "aom_dsp/arm/sum_neon.h"
     21 
     22 static inline void highbd_sad4xhx4d_small_neon(const uint8_t *src_ptr,
     23                                               int src_stride,
     24                                               const uint8_t *const ref_ptr[4],
     25                                               int ref_stride, uint32_t res[4],
     26                                               int h) {
     27  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
     28  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
     29  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
     30  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
     31  const uint16_t *ref16_ptr3 = CONVERT_TO_SHORTPTR(ref_ptr[3]);
     32 
     33  uint32x4_t sum[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
     34                        vdupq_n_u32(0) };
     35 
     36  int i = 0;
     37  do {
     38    uint16x4_t s = vld1_u16(src16_ptr + i * src_stride);
     39    uint16x4_t r0 = vld1_u16(ref16_ptr0 + i * ref_stride);
     40    uint16x4_t r1 = vld1_u16(ref16_ptr1 + i * ref_stride);
     41    uint16x4_t r2 = vld1_u16(ref16_ptr2 + i * ref_stride);
     42    uint16x4_t r3 = vld1_u16(ref16_ptr3 + i * ref_stride);
     43 
     44    sum[0] = vabal_u16(sum[0], s, r0);
     45    sum[1] = vabal_u16(sum[1], s, r1);
     46    sum[2] = vabal_u16(sum[2], s, r2);
     47    sum[3] = vabal_u16(sum[3], s, r3);
     48 
     49  } while (++i < h);
     50 
     51  vst1q_u32(res, horizontal_add_4d_u32x4(sum));
     52 }
     53 
     54 static inline void highbd_sad8xhx4d_small_neon(const uint8_t *src_ptr,
     55                                               int src_stride,
     56                                               const uint8_t *const ref_ptr[4],
     57                                               int ref_stride, uint32_t res[4],
     58                                               int h) {
     59  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
     60  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
     61  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
     62  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
     63  const uint16_t *ref16_ptr3 = CONVERT_TO_SHORTPTR(ref_ptr[3]);
     64 
     65  uint16x8_t sum[4] = { vdupq_n_u16(0), vdupq_n_u16(0), vdupq_n_u16(0),
     66                        vdupq_n_u16(0) };
     67  uint32x4_t sum_u32[4];
     68 
     69  int i = 0;
     70  do {
     71    uint16x8_t s = vld1q_u16(src16_ptr + i * src_stride);
     72 
     73    sum[0] = vabaq_u16(sum[0], s, vld1q_u16(ref16_ptr0 + i * ref_stride));
     74    sum[1] = vabaq_u16(sum[1], s, vld1q_u16(ref16_ptr1 + i * ref_stride));
     75    sum[2] = vabaq_u16(sum[2], s, vld1q_u16(ref16_ptr2 + i * ref_stride));
     76    sum[3] = vabaq_u16(sum[3], s, vld1q_u16(ref16_ptr3 + i * ref_stride));
     77 
     78  } while (++i < h);
     79 
     80  sum_u32[0] = vpaddlq_u16(sum[0]);
     81  sum_u32[1] = vpaddlq_u16(sum[1]);
     82  sum_u32[2] = vpaddlq_u16(sum[2]);
     83  sum_u32[3] = vpaddlq_u16(sum[3]);
     84  vst1q_u32(res, horizontal_add_4d_u32x4(sum_u32));
     85 }
     86 
     87 static inline void sad8_neon(uint16x8_t src, uint16x8_t ref,
     88                             uint32x4_t *const sad_sum) {
     89  uint16x8_t abs_diff = vabdq_u16(src, ref);
     90  *sad_sum = vpadalq_u16(*sad_sum, abs_diff);
     91 }
     92 
     93 #if !CONFIG_REALTIME_ONLY
     94 static inline void highbd_sad8xhx4d_large_neon(const uint8_t *src_ptr,
     95                                               int src_stride,
     96                                               const uint8_t *const ref_ptr[4],
     97                                               int ref_stride, uint32_t res[4],
     98                                               int h) {
     99  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
    100  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
    101  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
    102  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
    103  const uint16_t *ref16_ptr3 = CONVERT_TO_SHORTPTR(ref_ptr[3]);
    104 
    105  uint32x4_t sum[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
    106                        vdupq_n_u32(0) };
    107 
    108  int i = 0;
    109  do {
    110    uint16x8_t s = vld1q_u16(src16_ptr + i * src_stride);
    111    sad8_neon(s, vld1q_u16(ref16_ptr0 + i * ref_stride), &sum[0]);
    112    sad8_neon(s, vld1q_u16(ref16_ptr1 + i * ref_stride), &sum[1]);
    113    sad8_neon(s, vld1q_u16(ref16_ptr2 + i * ref_stride), &sum[2]);
    114    sad8_neon(s, vld1q_u16(ref16_ptr3 + i * ref_stride), &sum[3]);
    115 
    116  } while (++i < h);
    117 
    118  vst1q_u32(res, horizontal_add_4d_u32x4(sum));
    119 }
    120 #endif  // !CONFIG_REALTIME_ONLY
    121 
    122 static inline void highbd_sad16xhx4d_large_neon(const uint8_t *src_ptr,
    123                                                int src_stride,
    124                                                const uint8_t *const ref_ptr[4],
    125                                                int ref_stride, uint32_t res[4],
    126                                                int h) {
    127  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
    128  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
    129  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
    130  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
    131  const uint16_t *ref16_ptr3 = CONVERT_TO_SHORTPTR(ref_ptr[3]);
    132 
    133  uint32x4_t sum_lo[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
    134                           vdupq_n_u32(0) };
    135  uint32x4_t sum_hi[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
    136                           vdupq_n_u32(0) };
    137  uint32x4_t sum[4];
    138 
    139  int i = 0;
    140  do {
    141    uint16x8_t s0 = vld1q_u16(src16_ptr + i * src_stride);
    142    sad8_neon(s0, vld1q_u16(ref16_ptr0 + i * ref_stride), &sum_lo[0]);
    143    sad8_neon(s0, vld1q_u16(ref16_ptr1 + i * ref_stride), &sum_lo[1]);
    144    sad8_neon(s0, vld1q_u16(ref16_ptr2 + i * ref_stride), &sum_lo[2]);
    145    sad8_neon(s0, vld1q_u16(ref16_ptr3 + i * ref_stride), &sum_lo[3]);
    146 
    147    uint16x8_t s1 = vld1q_u16(src16_ptr + i * src_stride + 8);
    148    sad8_neon(s1, vld1q_u16(ref16_ptr0 + i * ref_stride + 8), &sum_hi[0]);
    149    sad8_neon(s1, vld1q_u16(ref16_ptr1 + i * ref_stride + 8), &sum_hi[1]);
    150    sad8_neon(s1, vld1q_u16(ref16_ptr2 + i * ref_stride + 8), &sum_hi[2]);
    151    sad8_neon(s1, vld1q_u16(ref16_ptr3 + i * ref_stride + 8), &sum_hi[3]);
    152 
    153  } while (++i < h);
    154 
    155  sum[0] = vaddq_u32(sum_lo[0], sum_hi[0]);
    156  sum[1] = vaddq_u32(sum_lo[1], sum_hi[1]);
    157  sum[2] = vaddq_u32(sum_lo[2], sum_hi[2]);
    158  sum[3] = vaddq_u32(sum_lo[3], sum_hi[3]);
    159 
    160  vst1q_u32(res, horizontal_add_4d_u32x4(sum));
    161 }
    162 
    163 static inline void highbd_sadwxhx4d_large_neon(const uint8_t *src_ptr,
    164                                               int src_stride,
    165                                               const uint8_t *const ref_ptr[4],
    166                                               int ref_stride, uint32_t res[4],
    167                                               int w, int h) {
    168  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
    169  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
    170  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
    171  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
    172  const uint16_t *ref16_ptr3 = CONVERT_TO_SHORTPTR(ref_ptr[3]);
    173 
    174  uint32x4_t sum_lo[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
    175                           vdupq_n_u32(0) };
    176  uint32x4_t sum_hi[4] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0),
    177                           vdupq_n_u32(0) };
    178  uint32x4_t sum[4];
    179 
    180  int i = 0;
    181  do {
    182    int j = 0;
    183    do {
    184      uint16x8_t s0 = vld1q_u16(src16_ptr + i * src_stride + j);
    185      sad8_neon(s0, vld1q_u16(ref16_ptr0 + i * ref_stride + j), &sum_lo[0]);
    186      sad8_neon(s0, vld1q_u16(ref16_ptr1 + i * ref_stride + j), &sum_lo[1]);
    187      sad8_neon(s0, vld1q_u16(ref16_ptr2 + i * ref_stride + j), &sum_lo[2]);
    188      sad8_neon(s0, vld1q_u16(ref16_ptr3 + i * ref_stride + j), &sum_lo[3]);
    189 
    190      uint16x8_t s1 = vld1q_u16(src16_ptr + i * src_stride + j + 8);
    191      sad8_neon(s1, vld1q_u16(ref16_ptr0 + i * ref_stride + j + 8), &sum_hi[0]);
    192      sad8_neon(s1, vld1q_u16(ref16_ptr1 + i * ref_stride + j + 8), &sum_hi[1]);
    193      sad8_neon(s1, vld1q_u16(ref16_ptr2 + i * ref_stride + j + 8), &sum_hi[2]);
    194      sad8_neon(s1, vld1q_u16(ref16_ptr3 + i * ref_stride + j + 8), &sum_hi[3]);
    195 
    196      uint16x8_t s2 = vld1q_u16(src16_ptr + i * src_stride + j + 16);
    197      sad8_neon(s2, vld1q_u16(ref16_ptr0 + i * ref_stride + j + 16),
    198                &sum_lo[0]);
    199      sad8_neon(s2, vld1q_u16(ref16_ptr1 + i * ref_stride + j + 16),
    200                &sum_lo[1]);
    201      sad8_neon(s2, vld1q_u16(ref16_ptr2 + i * ref_stride + j + 16),
    202                &sum_lo[2]);
    203      sad8_neon(s2, vld1q_u16(ref16_ptr3 + i * ref_stride + j + 16),
    204                &sum_lo[3]);
    205 
    206      uint16x8_t s3 = vld1q_u16(src16_ptr + i * src_stride + j + 24);
    207      sad8_neon(s3, vld1q_u16(ref16_ptr0 + i * ref_stride + j + 24),
    208                &sum_hi[0]);
    209      sad8_neon(s3, vld1q_u16(ref16_ptr1 + i * ref_stride + j + 24),
    210                &sum_hi[1]);
    211      sad8_neon(s3, vld1q_u16(ref16_ptr2 + i * ref_stride + j + 24),
    212                &sum_hi[2]);
    213      sad8_neon(s3, vld1q_u16(ref16_ptr3 + i * ref_stride + j + 24),
    214                &sum_hi[3]);
    215 
    216      j += 32;
    217    } while (j < w);
    218 
    219  } while (++i < h);
    220 
    221  sum[0] = vaddq_u32(sum_lo[0], sum_hi[0]);
    222  sum[1] = vaddq_u32(sum_lo[1], sum_hi[1]);
    223  sum[2] = vaddq_u32(sum_lo[2], sum_hi[2]);
    224  sum[3] = vaddq_u32(sum_lo[3], sum_hi[3]);
    225 
    226  vst1q_u32(res, horizontal_add_4d_u32x4(sum));
    227 }
    228 
    229 static inline void highbd_sad128xhx4d_large_neon(
    230    const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_ptr[4],
    231    int ref_stride, uint32_t res[4], int h) {
    232  highbd_sadwxhx4d_large_neon(src_ptr, src_stride, ref_ptr, ref_stride, res,
    233                              128, h);
    234 }
    235 
    236 static inline void highbd_sad64xhx4d_large_neon(const uint8_t *src_ptr,
    237                                                int src_stride,
    238                                                const uint8_t *const ref_ptr[4],
    239                                                int ref_stride, uint32_t res[4],
    240                                                int h) {
    241  highbd_sadwxhx4d_large_neon(src_ptr, src_stride, ref_ptr, ref_stride, res, 64,
    242                              h);
    243 }
    244 
    245 static inline void highbd_sad32xhx4d_large_neon(const uint8_t *src_ptr,
    246                                                int src_stride,
    247                                                const uint8_t *const ref_ptr[4],
    248                                                int ref_stride, uint32_t res[4],
    249                                                int h) {
    250  highbd_sadwxhx4d_large_neon(src_ptr, src_stride, ref_ptr, ref_stride, res, 32,
    251                              h);
    252 }
    253 
    254 #define HBD_SAD_WXH_4D_SMALL_NEON(w, h)                                      \
    255  void aom_highbd_sad##w##x##h##x4d_neon(                                    \
    256      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4], \
    257      int ref_stride, uint32_t sad_array[4]) {                               \
    258    highbd_sad##w##xhx4d_small_neon(src, src_stride, ref_array, ref_stride,  \
    259                                    sad_array, (h));                         \
    260  }
    261 
    262 #define HBD_SAD_WXH_4D_LARGE_NEON(w, h)                                      \
    263  void aom_highbd_sad##w##x##h##x4d_neon(                                    \
    264      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4], \
    265      int ref_stride, uint32_t sad_array[4]) {                               \
    266    highbd_sad##w##xhx4d_large_neon(src, src_stride, ref_array, ref_stride,  \
    267                                    sad_array, (h));                         \
    268  }
    269 
    270 HBD_SAD_WXH_4D_SMALL_NEON(4, 4)
    271 HBD_SAD_WXH_4D_SMALL_NEON(4, 8)
    272 
    273 HBD_SAD_WXH_4D_SMALL_NEON(8, 4)
    274 HBD_SAD_WXH_4D_SMALL_NEON(8, 8)
    275 HBD_SAD_WXH_4D_SMALL_NEON(8, 16)
    276 
    277 HBD_SAD_WXH_4D_LARGE_NEON(16, 8)
    278 HBD_SAD_WXH_4D_LARGE_NEON(16, 16)
    279 HBD_SAD_WXH_4D_LARGE_NEON(16, 32)
    280 
    281 HBD_SAD_WXH_4D_LARGE_NEON(32, 16)
    282 HBD_SAD_WXH_4D_LARGE_NEON(32, 32)
    283 HBD_SAD_WXH_4D_LARGE_NEON(32, 64)
    284 
    285 HBD_SAD_WXH_4D_LARGE_NEON(64, 32)
    286 HBD_SAD_WXH_4D_LARGE_NEON(64, 64)
    287 HBD_SAD_WXH_4D_LARGE_NEON(64, 128)
    288 
    289 HBD_SAD_WXH_4D_LARGE_NEON(128, 64)
    290 HBD_SAD_WXH_4D_LARGE_NEON(128, 128)
    291 
    292 #if !CONFIG_REALTIME_ONLY
    293 HBD_SAD_WXH_4D_SMALL_NEON(4, 16)
    294 
    295 HBD_SAD_WXH_4D_LARGE_NEON(8, 32)
    296 
    297 HBD_SAD_WXH_4D_LARGE_NEON(16, 4)
    298 HBD_SAD_WXH_4D_LARGE_NEON(16, 64)
    299 
    300 HBD_SAD_WXH_4D_LARGE_NEON(32, 8)
    301 
    302 HBD_SAD_WXH_4D_LARGE_NEON(64, 16)
    303 #endif  // !CONFIG_REALTIME_ONLY
    304 
    305 #define HBD_SAD_SKIP_WXH_4D_SMALL_NEON(w, h)                                 \
    306  void aom_highbd_sad_skip_##w##x##h##x4d_neon(                              \
    307      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4], \
    308      int ref_stride, uint32_t sad_array[4]) {                               \
    309    highbd_sad##w##xhx4d_small_neon(src, 2 * src_stride, ref_array,          \
    310                                    2 * ref_stride, sad_array, ((h) >> 1));  \
    311    sad_array[0] <<= 1;                                                      \
    312    sad_array[1] <<= 1;                                                      \
    313    sad_array[2] <<= 1;                                                      \
    314    sad_array[3] <<= 1;                                                      \
    315  }
    316 
    317 #define HBD_SAD_SKIP_WXH_4D_LARGE_NEON(w, h)                                 \
    318  void aom_highbd_sad_skip_##w##x##h##x4d_neon(                              \
    319      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4], \
    320      int ref_stride, uint32_t sad_array[4]) {                               \
    321    highbd_sad##w##xhx4d_large_neon(src, 2 * src_stride, ref_array,          \
    322                                    2 * ref_stride, sad_array, ((h) >> 1));  \
    323    sad_array[0] <<= 1;                                                      \
    324    sad_array[1] <<= 1;                                                      \
    325    sad_array[2] <<= 1;                                                      \
    326    sad_array[3] <<= 1;                                                      \
    327  }
    328 
    329 HBD_SAD_SKIP_WXH_4D_SMALL_NEON(8, 16)
    330 
    331 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(16, 16)
    332 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(16, 32)
    333 
    334 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(32, 16)
    335 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(32, 32)
    336 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(32, 64)
    337 
    338 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(64, 32)
    339 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(64, 64)
    340 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(64, 128)
    341 
    342 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(128, 64)
    343 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(128, 128)
    344 
    345 #if !CONFIG_REALTIME_ONLY
    346 HBD_SAD_SKIP_WXH_4D_SMALL_NEON(4, 16)
    347 
    348 HBD_SAD_SKIP_WXH_4D_SMALL_NEON(8, 32)
    349 
    350 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(16, 64)
    351 
    352 HBD_SAD_SKIP_WXH_4D_LARGE_NEON(64, 16)
    353 #endif  // !CONFIG_REALTIME_ONLY
    354 
    355 static inline void highbd_sad4xhx3d_small_neon(const uint8_t *src_ptr,
    356                                               int src_stride,
    357                                               const uint8_t *const ref_ptr[4],
    358                                               int ref_stride, uint32_t res[4],
    359                                               int h) {
    360  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
    361  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
    362  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
    363  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
    364 
    365  uint32x4_t sum[3] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0) };
    366 
    367  int i = 0;
    368  do {
    369    uint16x4_t s = vld1_u16(src16_ptr + i * src_stride);
    370    uint16x4_t r0 = vld1_u16(ref16_ptr0 + i * ref_stride);
    371    uint16x4_t r1 = vld1_u16(ref16_ptr1 + i * ref_stride);
    372    uint16x4_t r2 = vld1_u16(ref16_ptr2 + i * ref_stride);
    373 
    374    sum[0] = vabal_u16(sum[0], s, r0);
    375    sum[1] = vabal_u16(sum[1], s, r1);
    376    sum[2] = vabal_u16(sum[2], s, r2);
    377 
    378  } while (++i < h);
    379 
    380  res[0] = horizontal_add_u32x4(sum[0]);
    381  res[1] = horizontal_add_u32x4(sum[1]);
    382  res[2] = horizontal_add_u32x4(sum[2]);
    383 }
    384 
    385 static inline void highbd_sad8xhx3d_small_neon(const uint8_t *src_ptr,
    386                                               int src_stride,
    387                                               const uint8_t *const ref_ptr[4],
    388                                               int ref_stride, uint32_t res[4],
    389                                               int h) {
    390  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
    391  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
    392  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
    393  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
    394 
    395  uint16x8_t sum[3] = { vdupq_n_u16(0), vdupq_n_u16(0), vdupq_n_u16(0) };
    396 
    397  int i = 0;
    398  do {
    399    uint16x8_t s = vld1q_u16(src16_ptr + i * src_stride);
    400 
    401    sum[0] = vabaq_u16(sum[0], s, vld1q_u16(ref16_ptr0 + i * ref_stride));
    402    sum[1] = vabaq_u16(sum[1], s, vld1q_u16(ref16_ptr1 + i * ref_stride));
    403    sum[2] = vabaq_u16(sum[2], s, vld1q_u16(ref16_ptr2 + i * ref_stride));
    404 
    405  } while (++i < h);
    406 
    407  res[0] = horizontal_add_u32x4(vpaddlq_u16(sum[0]));
    408  res[1] = horizontal_add_u32x4(vpaddlq_u16(sum[1]));
    409  res[2] = horizontal_add_u32x4(vpaddlq_u16(sum[2]));
    410 }
    411 
    412 #if !CONFIG_REALTIME_ONLY
    413 static inline void highbd_sad8xhx3d_large_neon(const uint8_t *src_ptr,
    414                                               int src_stride,
    415                                               const uint8_t *const ref_ptr[4],
    416                                               int ref_stride, uint32_t res[4],
    417                                               int h) {
    418  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
    419  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
    420  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
    421  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
    422 
    423  uint32x4_t sum[3] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0) };
    424 
    425  int i = 0;
    426  do {
    427    uint16x8_t s = vld1q_u16(src16_ptr + i * src_stride);
    428    uint16x8_t r0 = vld1q_u16(ref16_ptr0 + i * ref_stride);
    429    uint16x8_t r1 = vld1q_u16(ref16_ptr1 + i * ref_stride);
    430    uint16x8_t r2 = vld1q_u16(ref16_ptr2 + i * ref_stride);
    431 
    432    sad8_neon(s, r0, &sum[0]);
    433    sad8_neon(s, r1, &sum[1]);
    434    sad8_neon(s, r2, &sum[2]);
    435 
    436  } while (++i < h);
    437 
    438  res[0] = horizontal_add_u32x4(sum[0]);
    439  res[1] = horizontal_add_u32x4(sum[1]);
    440  res[2] = horizontal_add_u32x4(sum[2]);
    441 }
    442 #endif  // !CONFIG_REALTIME_ONLY
    443 
    444 static inline void highbd_sad16xhx3d_large_neon(const uint8_t *src_ptr,
    445                                                int src_stride,
    446                                                const uint8_t *const ref_ptr[4],
    447                                                int ref_stride, uint32_t res[4],
    448                                                int h) {
    449  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
    450  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
    451  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
    452  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
    453 
    454  uint32x4_t sum_lo[3] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0) };
    455  uint32x4_t sum_hi[3] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0) };
    456 
    457  int i = 0;
    458  do {
    459    uint16x8_t s0 = vld1q_u16(src16_ptr + i * src_stride);
    460    sad8_neon(s0, vld1q_u16(ref16_ptr0 + i * ref_stride), &sum_lo[0]);
    461    sad8_neon(s0, vld1q_u16(ref16_ptr1 + i * ref_stride), &sum_lo[1]);
    462    sad8_neon(s0, vld1q_u16(ref16_ptr2 + i * ref_stride), &sum_lo[2]);
    463 
    464    uint16x8_t s1 = vld1q_u16(src16_ptr + i * src_stride + 8);
    465    sad8_neon(s1, vld1q_u16(ref16_ptr0 + i * ref_stride + 8), &sum_hi[0]);
    466    sad8_neon(s1, vld1q_u16(ref16_ptr1 + i * ref_stride + 8), &sum_hi[1]);
    467    sad8_neon(s1, vld1q_u16(ref16_ptr2 + i * ref_stride + 8), &sum_hi[2]);
    468 
    469  } while (++i < h);
    470 
    471  res[0] = horizontal_add_u32x4(vaddq_u32(sum_lo[0], sum_hi[0]));
    472  res[1] = horizontal_add_u32x4(vaddq_u32(sum_lo[1], sum_hi[1]));
    473  res[2] = horizontal_add_u32x4(vaddq_u32(sum_lo[2], sum_hi[2]));
    474 }
    475 
    476 static inline void highbd_sadwxhx3d_large_neon(const uint8_t *src_ptr,
    477                                               int src_stride,
    478                                               const uint8_t *const ref_ptr[4],
    479                                               int ref_stride, uint32_t res[4],
    480                                               int w, int h) {
    481  const uint16_t *src16_ptr = CONVERT_TO_SHORTPTR(src_ptr);
    482  const uint16_t *ref16_ptr0 = CONVERT_TO_SHORTPTR(ref_ptr[0]);
    483  const uint16_t *ref16_ptr1 = CONVERT_TO_SHORTPTR(ref_ptr[1]);
    484  const uint16_t *ref16_ptr2 = CONVERT_TO_SHORTPTR(ref_ptr[2]);
    485 
    486  uint32x4_t sum_lo[3] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0) };
    487  uint32x4_t sum_hi[3] = { vdupq_n_u32(0), vdupq_n_u32(0), vdupq_n_u32(0) };
    488  uint32x4_t sum[3];
    489 
    490  int i = 0;
    491  do {
    492    int j = 0;
    493    do {
    494      uint16x8_t s0 = vld1q_u16(src16_ptr + i * src_stride + j);
    495      sad8_neon(s0, vld1q_u16(ref16_ptr0 + i * ref_stride + j), &sum_lo[0]);
    496      sad8_neon(s0, vld1q_u16(ref16_ptr1 + i * ref_stride + j), &sum_lo[1]);
    497      sad8_neon(s0, vld1q_u16(ref16_ptr2 + i * ref_stride + j), &sum_lo[2]);
    498 
    499      uint16x8_t s1 = vld1q_u16(src16_ptr + i * src_stride + j + 8);
    500      sad8_neon(s1, vld1q_u16(ref16_ptr0 + i * ref_stride + j + 8), &sum_hi[0]);
    501      sad8_neon(s1, vld1q_u16(ref16_ptr1 + i * ref_stride + j + 8), &sum_hi[1]);
    502      sad8_neon(s1, vld1q_u16(ref16_ptr2 + i * ref_stride + j + 8), &sum_hi[2]);
    503 
    504      uint16x8_t s2 = vld1q_u16(src16_ptr + i * src_stride + j + 16);
    505      sad8_neon(s2, vld1q_u16(ref16_ptr0 + i * ref_stride + j + 16),
    506                &sum_lo[0]);
    507      sad8_neon(s2, vld1q_u16(ref16_ptr1 + i * ref_stride + j + 16),
    508                &sum_lo[1]);
    509      sad8_neon(s2, vld1q_u16(ref16_ptr2 + i * ref_stride + j + 16),
    510                &sum_lo[2]);
    511 
    512      uint16x8_t s3 = vld1q_u16(src16_ptr + i * src_stride + j + 24);
    513      sad8_neon(s3, vld1q_u16(ref16_ptr0 + i * ref_stride + j + 24),
    514                &sum_hi[0]);
    515      sad8_neon(s3, vld1q_u16(ref16_ptr1 + i * ref_stride + j + 24),
    516                &sum_hi[1]);
    517      sad8_neon(s3, vld1q_u16(ref16_ptr2 + i * ref_stride + j + 24),
    518                &sum_hi[2]);
    519 
    520      j += 32;
    521    } while (j < w);
    522 
    523  } while (++i < h);
    524 
    525  sum[0] = vaddq_u32(sum_lo[0], sum_hi[0]);
    526  sum[1] = vaddq_u32(sum_lo[1], sum_hi[1]);
    527  sum[2] = vaddq_u32(sum_lo[2], sum_hi[2]);
    528 
    529  res[0] = horizontal_add_u32x4(sum[0]);
    530  res[1] = horizontal_add_u32x4(sum[1]);
    531  res[2] = horizontal_add_u32x4(sum[2]);
    532 }
    533 
    534 static inline void highbd_sad128xhx3d_large_neon(
    535    const uint8_t *src_ptr, int src_stride, const uint8_t *const ref_ptr[4],
    536    int ref_stride, uint32_t res[4], int h) {
    537  highbd_sadwxhx3d_large_neon(src_ptr, src_stride, ref_ptr, ref_stride, res,
    538                              128, h);
    539 }
    540 
    541 static inline void highbd_sad64xhx3d_large_neon(const uint8_t *src_ptr,
    542                                                int src_stride,
    543                                                const uint8_t *const ref_ptr[4],
    544                                                int ref_stride, uint32_t res[4],
    545                                                int h) {
    546  highbd_sadwxhx3d_large_neon(src_ptr, src_stride, ref_ptr, ref_stride, res, 64,
    547                              h);
    548 }
    549 
    550 static inline void highbd_sad32xhx3d_large_neon(const uint8_t *src_ptr,
    551                                                int src_stride,
    552                                                const uint8_t *const ref_ptr[4],
    553                                                int ref_stride, uint32_t res[4],
    554                                                int h) {
    555  highbd_sadwxhx3d_large_neon(src_ptr, src_stride, ref_ptr, ref_stride, res, 32,
    556                              h);
    557 }
    558 
    559 #define HBD_SAD_WXH_3D_SMALL_NEON(w, h)                                      \
    560  void aom_highbd_sad##w##x##h##x3d_neon(                                    \
    561      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4], \
    562      int ref_stride, uint32_t sad_array[4]) {                               \
    563    highbd_sad##w##xhx3d_small_neon(src, src_stride, ref_array, ref_stride,  \
    564                                    sad_array, (h));                         \
    565  }
    566 
    567 #define HBD_SAD_WXH_3D_LARGE_NEON(w, h)                                      \
    568  void aom_highbd_sad##w##x##h##x3d_neon(                                    \
    569      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4], \
    570      int ref_stride, uint32_t sad_array[4]) {                               \
    571    highbd_sad##w##xhx3d_large_neon(src, src_stride, ref_array, ref_stride,  \
    572                                    sad_array, (h));                         \
    573  }
    574 
    575 HBD_SAD_WXH_3D_SMALL_NEON(4, 4)
    576 HBD_SAD_WXH_3D_SMALL_NEON(4, 8)
    577 
    578 HBD_SAD_WXH_3D_SMALL_NEON(8, 4)
    579 HBD_SAD_WXH_3D_SMALL_NEON(8, 8)
    580 HBD_SAD_WXH_3D_SMALL_NEON(8, 16)
    581 
    582 HBD_SAD_WXH_3D_LARGE_NEON(16, 8)
    583 HBD_SAD_WXH_3D_LARGE_NEON(16, 16)
    584 HBD_SAD_WXH_3D_LARGE_NEON(16, 32)
    585 
    586 HBD_SAD_WXH_3D_LARGE_NEON(32, 16)
    587 HBD_SAD_WXH_3D_LARGE_NEON(32, 32)
    588 HBD_SAD_WXH_3D_LARGE_NEON(32, 64)
    589 
    590 HBD_SAD_WXH_3D_LARGE_NEON(64, 32)
    591 HBD_SAD_WXH_3D_LARGE_NEON(64, 64)
    592 HBD_SAD_WXH_3D_LARGE_NEON(64, 128)
    593 
    594 HBD_SAD_WXH_3D_LARGE_NEON(128, 64)
    595 HBD_SAD_WXH_3D_LARGE_NEON(128, 128)
    596 
    597 #if !CONFIG_REALTIME_ONLY
    598 HBD_SAD_WXH_3D_SMALL_NEON(4, 16)
    599 
    600 HBD_SAD_WXH_3D_LARGE_NEON(8, 32)
    601 
    602 HBD_SAD_WXH_3D_LARGE_NEON(16, 4)
    603 HBD_SAD_WXH_3D_LARGE_NEON(16, 64)
    604 
    605 HBD_SAD_WXH_3D_LARGE_NEON(32, 8)
    606 
    607 HBD_SAD_WXH_3D_LARGE_NEON(64, 16)
    608 #endif  // !CONFIG_REALTIME_ONLY