tor-browser

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

pickrst_sse4.c (63278B)


      1 /*
      2 * Copyright (c) 2018, 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 <assert.h>
     13 #include <smmintrin.h>
     14 #include "aom_dsp/x86/mem_sse2.h"
     15 #include "aom_dsp/x86/synonyms.h"
     16 
     17 #include "config/av1_rtcd.h"
     18 #include "av1/common/restoration.h"
     19 #include "av1/encoder/pickrst.h"
     20 
     21 static inline void acc_stat_sse41(int32_t *dst, const uint8_t *src,
     22                                  const __m128i *shuffle, const __m128i *kl) {
     23  const __m128i s = _mm_shuffle_epi8(xx_loadu_128(src), *shuffle);
     24  const __m128i d0 = _mm_madd_epi16(*kl, _mm_cvtepu8_epi16(s));
     25  const __m128i d1 =
     26      _mm_madd_epi16(*kl, _mm_cvtepu8_epi16(_mm_srli_si128(s, 8)));
     27  const __m128i dst0 = xx_loadu_128(dst);
     28  const __m128i dst1 = xx_loadu_128(dst + 4);
     29  const __m128i r0 = _mm_add_epi32(dst0, d0);
     30  const __m128i r1 = _mm_add_epi32(dst1, d1);
     31  xx_storeu_128(dst, r0);
     32  xx_storeu_128(dst + 4, r1);
     33 }
     34 
     35 static inline void acc_stat_win7_one_line_sse4_1(
     36    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end,
     37    int dgd_stride, const __m128i *shuffle, int32_t *sumX,
     38    int32_t sumY[WIENER_WIN][WIENER_WIN], int32_t M_int[WIENER_WIN][WIENER_WIN],
     39    int32_t H_int[WIENER_WIN2][WIENER_WIN * 8]) {
     40  const int wiener_win = 7;
     41  int j, k, l;
     42  // Main loop handles two pixels at a time
     43  // We can assume that h_start is even, since it will always be aligned to
     44  // a tile edge + some number of restoration units, and both of those will
     45  // be 64-pixel aligned.
     46  // However, at the edge of the image, h_end may be odd, so we need to handle
     47  // that case correctly.
     48  assert(h_start % 2 == 0);
     49  const int h_end_even = h_end & ~1;
     50  const int has_odd_pixel = h_end & 1;
     51  for (j = h_start; j < h_end_even; j += 2) {
     52    const uint8_t *dgd_ij = dgd + j;
     53    const uint8_t X1 = src[j];
     54    const uint8_t X2 = src[j + 1];
     55    *sumX += X1 + X2;
     56    for (k = 0; k < wiener_win; k++) {
     57      const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride;
     58      for (l = 0; l < wiener_win; l++) {
     59        int32_t *H_ = &H_int[(l * wiener_win + k)][0];
     60        const uint8_t D1 = dgd_ijk[l];
     61        const uint8_t D2 = dgd_ijk[l + 1];
     62        sumY[k][l] += D1 + D2;
     63        M_int[k][l] += D1 * X1 + D2 * X2;
     64 
     65        const __m128i kl =
     66            _mm_cvtepu8_epi16(_mm_set1_epi16(loadu_int16(dgd_ijk + l)));
     67        acc_stat_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl);
     68        acc_stat_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl);
     69        acc_stat_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl);
     70        acc_stat_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl);
     71        acc_stat_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl);
     72        acc_stat_sse41(H_ + 5 * 8, dgd_ij + 5 * dgd_stride, shuffle, &kl);
     73        acc_stat_sse41(H_ + 6 * 8, dgd_ij + 6 * dgd_stride, shuffle, &kl);
     74      }
     75    }
     76  }
     77  // If the width is odd, add in the final pixel
     78  if (has_odd_pixel) {
     79    const uint8_t *dgd_ij = dgd + j;
     80    const uint8_t X1 = src[j];
     81    *sumX += X1;
     82    for (k = 0; k < wiener_win; k++) {
     83      const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride;
     84      for (l = 0; l < wiener_win; l++) {
     85        int32_t *H_ = &H_int[(l * wiener_win + k)][0];
     86        const uint8_t D1 = dgd_ijk[l];
     87        sumY[k][l] += D1;
     88        M_int[k][l] += D1 * X1;
     89 
     90        // The `acc_stat_sse41` function wants its input to have interleaved
     91        // copies of two pixels, but we only have one. However, the pixels
     92        // are (effectively) used as inputs to a multiply-accumulate.
     93        // So if we set the extra pixel slot to 0, then it is effectively
     94        // ignored.
     95        const __m128i kl = _mm_cvtepu8_epi16(_mm_set1_epi16((int16_t)D1));
     96        acc_stat_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl);
     97        acc_stat_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl);
     98        acc_stat_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl);
     99        acc_stat_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl);
    100        acc_stat_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl);
    101        acc_stat_sse41(H_ + 5 * 8, dgd_ij + 5 * dgd_stride, shuffle, &kl);
    102        acc_stat_sse41(H_ + 6 * 8, dgd_ij + 6 * dgd_stride, shuffle, &kl);
    103      }
    104    }
    105  }
    106 }
    107 
    108 static inline void compute_stats_win7_opt_sse4_1(
    109    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start,
    110    int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H,
    111    int use_downsampled_wiener_stats) {
    112  int i, j, k, l, m, n;
    113  const int wiener_win = WIENER_WIN;
    114  const int pixel_count = (h_end - h_start) * (v_end - v_start);
    115  const int wiener_win2 = wiener_win * wiener_win;
    116  const int wiener_halfwin = (wiener_win >> 1);
    117  const uint8_t avg =
    118      find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
    119 
    120  int32_t M_int32[WIENER_WIN][WIENER_WIN] = { { 0 } };
    121  int32_t M_int32_row[WIENER_WIN][WIENER_WIN] = { { 0 } };
    122  int64_t M_int64[WIENER_WIN][WIENER_WIN] = { { 0 } };
    123  int32_t H_int32[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } };
    124  int32_t H_int32_row[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } };
    125  int64_t H_int64[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } };
    126  int32_t sumY[WIENER_WIN][WIENER_WIN] = { { 0 } };
    127  int32_t sumX = 0;
    128  const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin;
    129  int downsample_factor =
    130      use_downsampled_wiener_stats ? WIENER_STATS_DOWNSAMPLE_FACTOR : 1;
    131  int32_t sumX_row = 0;
    132  int32_t sumY_row[WIENER_WIN][WIENER_WIN] = { { 0 } };
    133 
    134  const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data);
    135  for (j = v_start; j < v_end; j += 64) {
    136    const int vert_end = AOMMIN(64, v_end - j) + j;
    137    for (i = j; i < vert_end; i = i + downsample_factor) {
    138      if (use_downsampled_wiener_stats &&
    139          (vert_end - i < WIENER_STATS_DOWNSAMPLE_FACTOR)) {
    140        downsample_factor = vert_end - i;
    141      }
    142      sumX_row = 0;
    143      memset(sumY_row, 0, sizeof(int32_t) * WIENER_WIN * WIENER_WIN);
    144      memset(M_int32_row, 0, sizeof(int32_t) * WIENER_WIN * WIENER_WIN);
    145      memset(H_int32_row, 0, sizeof(int32_t) * WIENER_WIN2 * (WIENER_WIN * 8));
    146      acc_stat_win7_one_line_sse4_1(
    147          dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end,
    148          dgd_stride, &shuffle, &sumX_row, sumY_row, M_int32_row, H_int32_row);
    149      sumX += sumX_row * downsample_factor;
    150      // Scale M matrix based on the downsampling factor
    151      for (k = 0; k < wiener_win; ++k) {
    152        for (l = 0; l < wiener_win; ++l) {
    153          sumY[k][l] += (sumY_row[k][l] * downsample_factor);
    154          M_int32[k][l] += (M_int32_row[k][l] * downsample_factor);
    155        }
    156      }
    157      // Scale H matrix based on the downsampling factor
    158      for (k = 0; k < WIENER_WIN2; ++k) {
    159        for (l = 0; l < WIENER_WIN * 8; ++l) {
    160          H_int32[k][l] += (H_int32_row[k][l] * downsample_factor);
    161        }
    162      }
    163    }
    164    for (k = 0; k < wiener_win; ++k) {
    165      for (l = 0; l < wiener_win; ++l) {
    166        M_int64[k][l] += M_int32[k][l];
    167        M_int32[k][l] = 0;
    168      }
    169    }
    170    for (k = 0; k < WIENER_WIN2; ++k) {
    171      for (l = 0; l < WIENER_WIN * 8; ++l) {
    172        H_int64[k][l] += H_int32[k][l];
    173        H_int32[k][l] = 0;
    174      }
    175    }
    176  }
    177 
    178  const int64_t avg_square_sum = (int64_t)avg * (int64_t)avg * pixel_count;
    179  for (k = 0; k < wiener_win; k++) {
    180    for (l = 0; l < wiener_win; l++) {
    181      const int32_t idx0 = l * wiener_win + k;
    182      M[idx0] =
    183          M_int64[k][l] + (avg_square_sum - (int64_t)avg * (sumX + sumY[k][l]));
    184      int64_t *H_ = H + idx0 * wiener_win2;
    185      int64_t *H_int_ = &H_int64[idx0][0];
    186      for (m = 0; m < wiener_win; m++) {
    187        for (n = 0; n < wiener_win; n++) {
    188          H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum -
    189                                   (int64_t)avg * (sumY[k][l] + sumY[n][m]);
    190        }
    191      }
    192    }
    193  }
    194 }
    195 
    196 #if CONFIG_AV1_HIGHBITDEPTH
    197 static inline void acc_stat_highbd_sse41(int64_t *dst, const uint16_t *dgd,
    198                                         const __m128i *shuffle,
    199                                         const __m128i *dgd_ijkl) {
    200  // Load 256 bits from dgd in two chunks
    201  const __m128i s0l = xx_loadu_128(dgd);
    202  const __m128i s0h = xx_loadu_128(dgd + 4);
    203  // s0l = [7 6 5 4 3 2 1 0] as u16 values (dgd indices)
    204  // s0h = [11 10 9 8 7 6 5 4] as u16 values (dgd indices)
    205  // (Slightly strange order so we can apply the same shuffle to both halves)
    206 
    207  // Shuffle the u16 values in each half (actually using 8-bit shuffle mask)
    208  const __m128i s1l = _mm_shuffle_epi8(s0l, *shuffle);
    209  const __m128i s1h = _mm_shuffle_epi8(s0h, *shuffle);
    210  // s1l = [4 3 3 2 2 1 1 0] as u16 values (dgd indices)
    211  // s1h = [8 7 7 6 6 5 5 4] as u16 values (dgd indices)
    212 
    213  // Multiply s1 by dgd_ijkl resulting in 8x u32 values
    214  // Horizontally add pairs of u32 resulting in 4x u32
    215  const __m128i dl = _mm_madd_epi16(*dgd_ijkl, s1l);
    216  const __m128i dh = _mm_madd_epi16(*dgd_ijkl, s1h);
    217  // dl = [d c b a] as u32 values
    218  // dh = [h g f e] as u32 values
    219 
    220  // Add these 8x u32 results on to dst in four parts
    221  const __m128i dll = _mm_cvtepu32_epi64(dl);
    222  const __m128i dlh = _mm_cvtepu32_epi64(_mm_srli_si128(dl, 8));
    223  const __m128i dhl = _mm_cvtepu32_epi64(dh);
    224  const __m128i dhh = _mm_cvtepu32_epi64(_mm_srli_si128(dh, 8));
    225  // dll = [b a] as u64 values, etc.
    226 
    227  const __m128i rll = _mm_add_epi64(xx_loadu_128(dst), dll);
    228  xx_storeu_128(dst, rll);
    229  const __m128i rlh = _mm_add_epi64(xx_loadu_128(dst + 2), dlh);
    230  xx_storeu_128(dst + 2, rlh);
    231  const __m128i rhl = _mm_add_epi64(xx_loadu_128(dst + 4), dhl);
    232  xx_storeu_128(dst + 4, rhl);
    233  const __m128i rhh = _mm_add_epi64(xx_loadu_128(dst + 6), dhh);
    234  xx_storeu_128(dst + 6, rhh);
    235 }
    236 
    237 static inline void acc_stat_highbd_win7_one_line_sse4_1(
    238    const uint16_t *dgd, const uint16_t *src, int h_start, int h_end,
    239    int dgd_stride, const __m128i *shuffle, int32_t *sumX,
    240    int32_t sumY[WIENER_WIN][WIENER_WIN], int64_t M_int[WIENER_WIN][WIENER_WIN],
    241    int64_t H_int[WIENER_WIN2][WIENER_WIN * 8]) {
    242  int j, k, l;
    243  const int wiener_win = WIENER_WIN;
    244  // Main loop handles two pixels at a time
    245  // We can assume that h_start is even, since it will always be aligned to
    246  // a tile edge + some number of restoration units, and both of those will
    247  // be 64-pixel aligned.
    248  // However, at the edge of the image, h_end may be odd, so we need to handle
    249  // that case correctly.
    250  assert(h_start % 2 == 0);
    251  const int h_end_even = h_end & ~1;
    252  const int has_odd_pixel = h_end & 1;
    253  for (j = h_start; j < h_end_even; j += 2) {
    254    const uint16_t X1 = src[j];
    255    const uint16_t X2 = src[j + 1];
    256    *sumX += X1 + X2;
    257    const uint16_t *dgd_ij = dgd + j;
    258    for (k = 0; k < wiener_win; k++) {
    259      const uint16_t *dgd_ijk = dgd_ij + k * dgd_stride;
    260      for (l = 0; l < wiener_win; l++) {
    261        int64_t *H_ = &H_int[(l * wiener_win + k)][0];
    262        const uint16_t D1 = dgd_ijk[l];
    263        const uint16_t D2 = dgd_ijk[l + 1];
    264        sumY[k][l] += D1 + D2;
    265        M_int[k][l] += D1 * X1 + D2 * X2;
    266 
    267        // Load two u16 values from dgd as a single u32
    268        // Then broadcast to 4x u32 slots of a 128
    269        const __m128i dgd_ijkl = _mm_set1_epi32(loadu_int32(dgd_ijk + l));
    270        // dgd_ijkl = [y x y x y x y x] as u16
    271 
    272        acc_stat_highbd_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle,
    273                              &dgd_ijkl);
    274        acc_stat_highbd_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle,
    275                              &dgd_ijkl);
    276        acc_stat_highbd_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle,
    277                              &dgd_ijkl);
    278        acc_stat_highbd_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle,
    279                              &dgd_ijkl);
    280        acc_stat_highbd_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle,
    281                              &dgd_ijkl);
    282        acc_stat_highbd_sse41(H_ + 5 * 8, dgd_ij + 5 * dgd_stride, shuffle,
    283                              &dgd_ijkl);
    284        acc_stat_highbd_sse41(H_ + 6 * 8, dgd_ij + 6 * dgd_stride, shuffle,
    285                              &dgd_ijkl);
    286      }
    287    }
    288  }
    289  // If the width is odd, add in the final pixel
    290  if (has_odd_pixel) {
    291    const uint16_t X1 = src[j];
    292    *sumX += X1;
    293    const uint16_t *dgd_ij = dgd + j;
    294    for (k = 0; k < wiener_win; k++) {
    295      const uint16_t *dgd_ijk = dgd_ij + k * dgd_stride;
    296      for (l = 0; l < wiener_win; l++) {
    297        int64_t *H_ = &H_int[(l * wiener_win + k)][0];
    298        const uint16_t D1 = dgd_ijk[l];
    299        sumY[k][l] += D1;
    300        M_int[k][l] += D1 * X1;
    301 
    302        // The `acc_stat_highbd_sse41` function wants its input to have
    303        // interleaved copies of two pixels, but we only have one. However, the
    304        // pixels are (effectively) used as inputs to a multiply-accumulate. So
    305        // if we set the extra pixel slot to 0, then it is effectively ignored.
    306        const __m128i dgd_ijkl = _mm_set1_epi32((int)D1);
    307 
    308        acc_stat_highbd_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle,
    309                              &dgd_ijkl);
    310        acc_stat_highbd_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle,
    311                              &dgd_ijkl);
    312        acc_stat_highbd_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle,
    313                              &dgd_ijkl);
    314        acc_stat_highbd_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle,
    315                              &dgd_ijkl);
    316        acc_stat_highbd_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle,
    317                              &dgd_ijkl);
    318        acc_stat_highbd_sse41(H_ + 5 * 8, dgd_ij + 5 * dgd_stride, shuffle,
    319                              &dgd_ijkl);
    320        acc_stat_highbd_sse41(H_ + 6 * 8, dgd_ij + 6 * dgd_stride, shuffle,
    321                              &dgd_ijkl);
    322      }
    323    }
    324  }
    325 }
    326 
    327 static inline void compute_stats_highbd_win7_opt_sse4_1(
    328    const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end,
    329    int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M,
    330    int64_t *H, aom_bit_depth_t bit_depth) {
    331  int i, j, k, l, m, n;
    332  const int wiener_win = WIENER_WIN;
    333  const int pixel_count = (h_end - h_start) * (v_end - v_start);
    334  const int wiener_win2 = wiener_win * wiener_win;
    335  const int wiener_halfwin = (wiener_win >> 1);
    336  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
    337  const uint16_t *dgd = CONVERT_TO_SHORTPTR(dgd8);
    338  const uint16_t avg =
    339      find_average_highbd(dgd, h_start, h_end, v_start, v_end, dgd_stride);
    340 
    341  int64_t M_int[WIENER_WIN][WIENER_WIN] = { { 0 } };
    342  int64_t H_int[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } };
    343  int32_t sumY[WIENER_WIN][WIENER_WIN] = { { 0 } };
    344  int32_t sumX = 0;
    345  const uint16_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin;
    346 
    347  // Load just half of the 256-bit shuffle control used for the AVX2 version
    348  const __m128i shuffle = xx_loadu_128(g_shuffle_stats_highbd_data);
    349  for (j = v_start; j < v_end; j += 64) {
    350    const int vert_end = AOMMIN(64, v_end - j) + j;
    351    for (i = j; i < vert_end; i++) {
    352      acc_stat_highbd_win7_one_line_sse4_1(
    353          dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end,
    354          dgd_stride, &shuffle, &sumX, sumY, M_int, H_int);
    355    }
    356  }
    357 
    358  uint8_t bit_depth_divider = 1;
    359  if (bit_depth == AOM_BITS_12)
    360    bit_depth_divider = 16;
    361  else if (bit_depth == AOM_BITS_10)
    362    bit_depth_divider = 4;
    363 
    364  const int64_t avg_square_sum = (int64_t)avg * (int64_t)avg * pixel_count;
    365  for (k = 0; k < wiener_win; k++) {
    366    for (l = 0; l < wiener_win; l++) {
    367      const int32_t idx0 = l * wiener_win + k;
    368      M[idx0] = (M_int[k][l] +
    369                 (avg_square_sum - (int64_t)avg * (sumX + sumY[k][l]))) /
    370                bit_depth_divider;
    371      int64_t *H_ = H + idx0 * wiener_win2;
    372      int64_t *H_int_ = &H_int[idx0][0];
    373      for (m = 0; m < wiener_win; m++) {
    374        for (n = 0; n < wiener_win; n++) {
    375          H_[m * wiener_win + n] =
    376              (H_int_[n * 8 + m] +
    377               (avg_square_sum - (int64_t)avg * (sumY[k][l] + sumY[n][m]))) /
    378              bit_depth_divider;
    379        }
    380      }
    381    }
    382  }
    383 }
    384 
    385 static inline void acc_stat_highbd_win5_one_line_sse4_1(
    386    const uint16_t *dgd, const uint16_t *src, int h_start, int h_end,
    387    int dgd_stride, const __m128i *shuffle, int32_t *sumX,
    388    int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA],
    389    int64_t M_int[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA],
    390    int64_t H_int[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8]) {
    391  int j, k, l;
    392  const int wiener_win = WIENER_WIN_CHROMA;
    393  // Main loop handles two pixels at a time
    394  // We can assume that h_start is even, since it will always be aligned to
    395  // a tile edge + some number of restoration units, and both of those will
    396  // be 64-pixel aligned.
    397  // However, at the edge of the image, h_end may be odd, so we need to handle
    398  // that case correctly.
    399  assert(h_start % 2 == 0);
    400  const int h_end_even = h_end & ~1;
    401  const int has_odd_pixel = h_end & 1;
    402  for (j = h_start; j < h_end_even; j += 2) {
    403    const uint16_t X1 = src[j];
    404    const uint16_t X2 = src[j + 1];
    405    *sumX += X1 + X2;
    406    const uint16_t *dgd_ij = dgd + j;
    407    for (k = 0; k < wiener_win; k++) {
    408      const uint16_t *dgd_ijk = dgd_ij + k * dgd_stride;
    409      for (l = 0; l < wiener_win; l++) {
    410        int64_t *H_ = &H_int[(l * wiener_win + k)][0];
    411        const uint16_t D1 = dgd_ijk[l];
    412        const uint16_t D2 = dgd_ijk[l + 1];
    413        sumY[k][l] += D1 + D2;
    414        M_int[k][l] += D1 * X1 + D2 * X2;
    415 
    416        // Load two u16 values from dgd as a single u32
    417        // then broadcast to 4x u32 slots of a 128
    418        const __m128i dgd_ijkl = _mm_set1_epi32(loadu_int32(dgd_ijk + l));
    419        // dgd_ijkl = [y x y x y x y x] as u16
    420 
    421        acc_stat_highbd_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle,
    422                              &dgd_ijkl);
    423        acc_stat_highbd_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle,
    424                              &dgd_ijkl);
    425        acc_stat_highbd_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle,
    426                              &dgd_ijkl);
    427        acc_stat_highbd_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle,
    428                              &dgd_ijkl);
    429        acc_stat_highbd_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle,
    430                              &dgd_ijkl);
    431      }
    432    }
    433  }
    434  // If the width is odd, add in the final pixel
    435  if (has_odd_pixel) {
    436    const uint16_t X1 = src[j];
    437    *sumX += X1;
    438    const uint16_t *dgd_ij = dgd + j;
    439    for (k = 0; k < wiener_win; k++) {
    440      const uint16_t *dgd_ijk = dgd_ij + k * dgd_stride;
    441      for (l = 0; l < wiener_win; l++) {
    442        int64_t *H_ = &H_int[(l * wiener_win + k)][0];
    443        const uint16_t D1 = dgd_ijk[l];
    444        sumY[k][l] += D1;
    445        M_int[k][l] += D1 * X1;
    446 
    447        // The `acc_stat_highbd_sse41` function wants its input to have
    448        // interleaved copies of two pixels, but we only have one. However, the
    449        // pixels are (effectively) used as inputs to a multiply-accumulate. So
    450        // if we set the extra pixel slot to 0, then it is effectively ignored.
    451        const __m128i dgd_ijkl = _mm_set1_epi32((int)D1);
    452 
    453        acc_stat_highbd_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle,
    454                              &dgd_ijkl);
    455        acc_stat_highbd_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle,
    456                              &dgd_ijkl);
    457        acc_stat_highbd_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle,
    458                              &dgd_ijkl);
    459        acc_stat_highbd_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle,
    460                              &dgd_ijkl);
    461        acc_stat_highbd_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle,
    462                              &dgd_ijkl);
    463      }
    464    }
    465  }
    466 }
    467 
    468 static inline void compute_stats_highbd_win5_opt_sse4_1(
    469    const uint8_t *dgd8, const uint8_t *src8, int h_start, int h_end,
    470    int v_start, int v_end, int dgd_stride, int src_stride, int64_t *M,
    471    int64_t *H, aom_bit_depth_t bit_depth) {
    472  int i, j, k, l, m, n;
    473  const int wiener_win = WIENER_WIN_CHROMA;
    474  const int pixel_count = (h_end - h_start) * (v_end - v_start);
    475  const int wiener_win2 = wiener_win * wiener_win;
    476  const int wiener_halfwin = (wiener_win >> 1);
    477  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
    478  const uint16_t *dgd = CONVERT_TO_SHORTPTR(dgd8);
    479  const uint16_t avg =
    480      find_average_highbd(dgd, h_start, h_end, v_start, v_end, dgd_stride);
    481 
    482  int64_t M_int[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
    483  int64_t H_int[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } };
    484  int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
    485  int32_t sumX = 0;
    486  const uint16_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin;
    487 
    488  // Load just half of the 256-bit shuffle control used for the AVX2 version
    489  const __m128i shuffle = xx_loadu_128(g_shuffle_stats_highbd_data);
    490  for (j = v_start; j < v_end; j += 64) {
    491    const int vert_end = AOMMIN(64, v_end - j) + j;
    492    for (i = j; i < vert_end; i++) {
    493      acc_stat_highbd_win5_one_line_sse4_1(
    494          dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end,
    495          dgd_stride, &shuffle, &sumX, sumY, M_int, H_int);
    496    }
    497  }
    498 
    499  uint8_t bit_depth_divider = 1;
    500  if (bit_depth == AOM_BITS_12)
    501    bit_depth_divider = 16;
    502  else if (bit_depth == AOM_BITS_10)
    503    bit_depth_divider = 4;
    504 
    505  const int64_t avg_square_sum = (int64_t)avg * (int64_t)avg * pixel_count;
    506  for (k = 0; k < wiener_win; k++) {
    507    for (l = 0; l < wiener_win; l++) {
    508      const int32_t idx0 = l * wiener_win + k;
    509      M[idx0] = (M_int[k][l] +
    510                 (avg_square_sum - (int64_t)avg * (sumX + sumY[k][l]))) /
    511                bit_depth_divider;
    512      int64_t *H_ = H + idx0 * wiener_win2;
    513      int64_t *H_int_ = &H_int[idx0][0];
    514      for (m = 0; m < wiener_win; m++) {
    515        for (n = 0; n < wiener_win; n++) {
    516          H_[m * wiener_win + n] =
    517              (H_int_[n * 8 + m] +
    518               (avg_square_sum - (int64_t)avg * (sumY[k][l] + sumY[n][m]))) /
    519              bit_depth_divider;
    520        }
    521      }
    522    }
    523  }
    524 }
    525 
    526 void av1_compute_stats_highbd_sse4_1(int wiener_win, const uint8_t *dgd8,
    527                                     const uint8_t *src8, int16_t *dgd_avg,
    528                                     int16_t *src_avg, int h_start, int h_end,
    529                                     int v_start, int v_end, int dgd_stride,
    530                                     int src_stride, int64_t *M, int64_t *H,
    531                                     aom_bit_depth_t bit_depth) {
    532  if (wiener_win == WIENER_WIN) {
    533    (void)dgd_avg;
    534    (void)src_avg;
    535    compute_stats_highbd_win7_opt_sse4_1(dgd8, src8, h_start, h_end, v_start,
    536                                         v_end, dgd_stride, src_stride, M, H,
    537                                         bit_depth);
    538  } else if (wiener_win == WIENER_WIN_CHROMA) {
    539    (void)dgd_avg;
    540    (void)src_avg;
    541    compute_stats_highbd_win5_opt_sse4_1(dgd8, src8, h_start, h_end, v_start,
    542                                         v_end, dgd_stride, src_stride, M, H,
    543                                         bit_depth);
    544  } else {
    545    av1_compute_stats_highbd_c(wiener_win, dgd8, src8, dgd_avg, src_avg,
    546                               h_start, h_end, v_start, v_end, dgd_stride,
    547                               src_stride, M, H, bit_depth);
    548  }
    549 }
    550 #endif  // CONFIG_AV1_HIGHBITDEPTH
    551 
    552 static inline void acc_stat_win5_one_line_sse4_1(
    553    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end,
    554    int dgd_stride, const __m128i *shuffle, int32_t *sumX,
    555    int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA],
    556    int32_t M_int[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA],
    557    int32_t H_int[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8]) {
    558  const int wiener_win = WIENER_WIN_CHROMA;
    559  int j, k, l;
    560  // Main loop handles two pixels at a time
    561  // We can assume that h_start is even, since it will always be aligned to
    562  // a tile edge + some number of restoration units, and both of those will
    563  // be 64-pixel aligned.
    564  // However, at the edge of the image, h_end may be odd, so we need to handle
    565  // that case correctly.
    566  assert(h_start % 2 == 0);
    567  const int h_end_even = h_end & ~1;
    568  const int has_odd_pixel = h_end & 1;
    569  for (j = h_start; j < h_end_even; j += 2) {
    570    const uint8_t *dgd_ij = dgd + j;
    571    const uint8_t X1 = src[j];
    572    const uint8_t X2 = src[j + 1];
    573    *sumX += X1 + X2;
    574    for (k = 0; k < wiener_win; k++) {
    575      const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride;
    576      for (l = 0; l < wiener_win; l++) {
    577        int32_t *H_ = &H_int[(l * wiener_win + k)][0];
    578        const uint8_t D1 = dgd_ijk[l];
    579        const uint8_t D2 = dgd_ijk[l + 1];
    580        sumY[k][l] += D1 + D2;
    581        M_int[k][l] += D1 * X1 + D2 * X2;
    582 
    583        const __m128i kl =
    584            _mm_cvtepu8_epi16(_mm_set1_epi16(loadu_int16(dgd_ijk + l)));
    585        acc_stat_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl);
    586        acc_stat_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl);
    587        acc_stat_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl);
    588        acc_stat_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl);
    589        acc_stat_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl);
    590      }
    591    }
    592  }
    593  // If the width is odd, add in the final pixel
    594  if (has_odd_pixel) {
    595    const uint8_t *dgd_ij = dgd + j;
    596    const uint8_t X1 = src[j];
    597    *sumX += X1;
    598    for (k = 0; k < wiener_win; k++) {
    599      const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride;
    600      for (l = 0; l < wiener_win; l++) {
    601        int32_t *H_ = &H_int[(l * wiener_win + k)][0];
    602        const uint8_t D1 = dgd_ijk[l];
    603        sumY[k][l] += D1;
    604        M_int[k][l] += D1 * X1;
    605 
    606        // The `acc_stat_sse41` function wants its input to have interleaved
    607        // copies of two pixels, but we only have one. However, the pixels
    608        // are (effectively) used as inputs to a multiply-accumulate.
    609        // So if we set the extra pixel slot to 0, then it is effectively
    610        // ignored.
    611        const __m128i kl = _mm_cvtepu8_epi16(_mm_set1_epi16((int16_t)D1));
    612        acc_stat_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl);
    613        acc_stat_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl);
    614        acc_stat_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl);
    615        acc_stat_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl);
    616        acc_stat_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl);
    617      }
    618    }
    619  }
    620 }
    621 
    622 static inline void compute_stats_win5_opt_sse4_1(
    623    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start,
    624    int v_end, int dgd_stride, int src_stride, int64_t *M, int64_t *H,
    625    int use_downsampled_wiener_stats) {
    626  int i, j, k, l, m, n;
    627  const int wiener_win = WIENER_WIN_CHROMA;
    628  const int pixel_count = (h_end - h_start) * (v_end - v_start);
    629  const int wiener_win2 = wiener_win * wiener_win;
    630  const int wiener_halfwin = (wiener_win >> 1);
    631  const uint8_t avg =
    632      find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
    633 
    634  int32_t M_int32[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
    635  int32_t M_int32_row[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
    636  int64_t M_int64[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
    637  int32_t H_int32[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } };
    638  int32_t H_int32_row[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } };
    639  int64_t H_int64[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } };
    640  int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
    641  int32_t sumX = 0;
    642  const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin;
    643  int downsample_factor =
    644      use_downsampled_wiener_stats ? WIENER_STATS_DOWNSAMPLE_FACTOR : 1;
    645  int32_t sumX_row = 0;
    646  int32_t sumY_row[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
    647 
    648  const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data);
    649  for (j = v_start; j < v_end; j += 64) {
    650    const int vert_end = AOMMIN(64, v_end - j) + j;
    651    for (i = j; i < vert_end; i = i + downsample_factor) {
    652      if (use_downsampled_wiener_stats &&
    653          (vert_end - i < WIENER_STATS_DOWNSAMPLE_FACTOR)) {
    654        downsample_factor = vert_end - i;
    655      }
    656      sumX_row = 0;
    657      memset(sumY_row, 0,
    658             sizeof(int32_t) * WIENER_WIN_CHROMA * WIENER_WIN_CHROMA);
    659      memset(M_int32_row, 0,
    660             sizeof(int32_t) * WIENER_WIN_CHROMA * WIENER_WIN_CHROMA);
    661      memset(H_int32_row, 0,
    662             sizeof(int32_t) * WIENER_WIN2_CHROMA * (WIENER_WIN_CHROMA * 8));
    663      acc_stat_win5_one_line_sse4_1(
    664          dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end,
    665          dgd_stride, &shuffle, &sumX_row, sumY_row, M_int32_row, H_int32_row);
    666      sumX += sumX_row * downsample_factor;
    667      // Scale M matrix based on the downsampling factor
    668      for (k = 0; k < wiener_win; ++k) {
    669        for (l = 0; l < wiener_win; ++l) {
    670          sumY[k][l] += (sumY_row[k][l] * downsample_factor);
    671          M_int32[k][l] += (M_int32_row[k][l] * downsample_factor);
    672        }
    673      }
    674      // Scale H matrix based on the downsampling factor
    675      for (k = 0; k < WIENER_WIN_CHROMA * WIENER_WIN_CHROMA; ++k) {
    676        for (l = 0; l < WIENER_WIN_CHROMA * 8; ++l) {
    677          H_int32[k][l] += (H_int32_row[k][l] * downsample_factor);
    678        }
    679      }
    680    }
    681    for (k = 0; k < wiener_win; ++k) {
    682      for (l = 0; l < wiener_win; ++l) {
    683        M_int64[k][l] += M_int32[k][l];
    684        M_int32[k][l] = 0;
    685      }
    686    }
    687    for (k = 0; k < WIENER_WIN_CHROMA * WIENER_WIN_CHROMA; ++k) {
    688      for (l = 0; l < WIENER_WIN_CHROMA * 8; ++l) {
    689        H_int64[k][l] += H_int32[k][l];
    690        H_int32[k][l] = 0;
    691      }
    692    }
    693  }
    694 
    695  const int64_t avg_square_sum = (int64_t)avg * (int64_t)avg * pixel_count;
    696  for (k = 0; k < wiener_win; k++) {
    697    for (l = 0; l < wiener_win; l++) {
    698      const int32_t idx0 = l * wiener_win + k;
    699      M[idx0] =
    700          M_int64[k][l] + (avg_square_sum - (int64_t)avg * (sumX + sumY[k][l]));
    701      int64_t *H_ = H + idx0 * wiener_win2;
    702      int64_t *H_int_ = &H_int64[idx0][0];
    703      for (m = 0; m < wiener_win; m++) {
    704        for (n = 0; n < wiener_win; n++) {
    705          H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum -
    706                                   (int64_t)avg * (sumY[k][l] + sumY[n][m]);
    707        }
    708      }
    709    }
    710  }
    711 }
    712 void av1_compute_stats_sse4_1(int wiener_win, const uint8_t *dgd,
    713                              const uint8_t *src, int16_t *dgd_avg,
    714                              int16_t *src_avg, int h_start, int h_end,
    715                              int v_start, int v_end, int dgd_stride,
    716                              int src_stride, int64_t *M, int64_t *H,
    717                              int use_downsampled_wiener_stats) {
    718  if (wiener_win == WIENER_WIN) {
    719    compute_stats_win7_opt_sse4_1(dgd, src, h_start, h_end, v_start, v_end,
    720                                  dgd_stride, src_stride, M, H,
    721                                  use_downsampled_wiener_stats);
    722  } else if (wiener_win == WIENER_WIN_CHROMA) {
    723    compute_stats_win5_opt_sse4_1(dgd, src, h_start, h_end, v_start, v_end,
    724                                  dgd_stride, src_stride, M, H,
    725                                  use_downsampled_wiener_stats);
    726  } else {
    727    av1_compute_stats_c(wiener_win, dgd, src, dgd_avg, src_avg, h_start, h_end,
    728                        v_start, v_end, dgd_stride, src_stride, M, H,
    729                        use_downsampled_wiener_stats);
    730  }
    731 }
    732 
    733 static inline __m128i pair_set_epi16(int a, int b) {
    734  return _mm_set1_epi32(
    735      (int32_t)(((uint16_t)(a)) | (((uint32_t)(uint16_t)(b)) << 16)));
    736 }
    737 
    738 int64_t av1_lowbd_pixel_proj_error_sse4_1(
    739    const uint8_t *src8, int width, int height, int src_stride,
    740    const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride,
    741    int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params) {
    742  int i, j, k;
    743  const int32_t shift = SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS;
    744  const __m128i rounding = _mm_set1_epi32(1 << (shift - 1));
    745  __m128i sum64 = _mm_setzero_si128();
    746  const uint8_t *src = src8;
    747  const uint8_t *dat = dat8;
    748  int64_t err = 0;
    749  if (params->r[0] > 0 && params->r[1] > 0) {
    750    __m128i xq_coeff = pair_set_epi16(xq[0], xq[1]);
    751    for (i = 0; i < height; ++i) {
    752      __m128i sum32 = _mm_setzero_si128();
    753      for (j = 0; j <= width - 8; j += 8) {
    754        const __m128i d0 = _mm_cvtepu8_epi16(xx_loadl_64(dat + j));
    755        const __m128i s0 = _mm_cvtepu8_epi16(xx_loadl_64(src + j));
    756        const __m128i flt0_16b =
    757            _mm_packs_epi32(xx_loadu_128(flt0 + j), xx_loadu_128(flt0 + j + 4));
    758        const __m128i flt1_16b =
    759            _mm_packs_epi32(xx_loadu_128(flt1 + j), xx_loadu_128(flt1 + j + 4));
    760        const __m128i u0 = _mm_slli_epi16(d0, SGRPROJ_RST_BITS);
    761        const __m128i flt0_0_sub_u = _mm_sub_epi16(flt0_16b, u0);
    762        const __m128i flt1_0_sub_u = _mm_sub_epi16(flt1_16b, u0);
    763        const __m128i v0 = _mm_madd_epi16(
    764            xq_coeff, _mm_unpacklo_epi16(flt0_0_sub_u, flt1_0_sub_u));
    765        const __m128i v1 = _mm_madd_epi16(
    766            xq_coeff, _mm_unpackhi_epi16(flt0_0_sub_u, flt1_0_sub_u));
    767        const __m128i vr0 = _mm_srai_epi32(_mm_add_epi32(v0, rounding), shift);
    768        const __m128i vr1 = _mm_srai_epi32(_mm_add_epi32(v1, rounding), shift);
    769        const __m128i e0 =
    770            _mm_sub_epi16(_mm_add_epi16(_mm_packs_epi32(vr0, vr1), d0), s0);
    771        const __m128i err0 = _mm_madd_epi16(e0, e0);
    772        sum32 = _mm_add_epi32(sum32, err0);
    773      }
    774      for (k = j; k < width; ++k) {
    775        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
    776        int32_t v = xq[0] * (flt0[k] - u) + xq[1] * (flt1[k] - u);
    777        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
    778        err += ((int64_t)e * e);
    779      }
    780      dat += dat_stride;
    781      src += src_stride;
    782      flt0 += flt0_stride;
    783      flt1 += flt1_stride;
    784      const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32);
    785      const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8));
    786      sum64 = _mm_add_epi64(sum64, sum64_0);
    787      sum64 = _mm_add_epi64(sum64, sum64_1);
    788    }
    789  } else if (params->r[0] > 0 || params->r[1] > 0) {
    790    const int xq_active = (params->r[0] > 0) ? xq[0] : xq[1];
    791    const __m128i xq_coeff =
    792        pair_set_epi16(xq_active, -xq_active * (1 << SGRPROJ_RST_BITS));
    793    const int32_t *flt = (params->r[0] > 0) ? flt0 : flt1;
    794    const int flt_stride = (params->r[0] > 0) ? flt0_stride : flt1_stride;
    795    for (i = 0; i < height; ++i) {
    796      __m128i sum32 = _mm_setzero_si128();
    797      for (j = 0; j <= width - 8; j += 8) {
    798        const __m128i d0 = _mm_cvtepu8_epi16(xx_loadl_64(dat + j));
    799        const __m128i s0 = _mm_cvtepu8_epi16(xx_loadl_64(src + j));
    800        const __m128i flt_16b =
    801            _mm_packs_epi32(xx_loadu_128(flt + j), xx_loadu_128(flt + j + 4));
    802        const __m128i v0 =
    803            _mm_madd_epi16(xq_coeff, _mm_unpacklo_epi16(flt_16b, d0));
    804        const __m128i v1 =
    805            _mm_madd_epi16(xq_coeff, _mm_unpackhi_epi16(flt_16b, d0));
    806        const __m128i vr0 = _mm_srai_epi32(_mm_add_epi32(v0, rounding), shift);
    807        const __m128i vr1 = _mm_srai_epi32(_mm_add_epi32(v1, rounding), shift);
    808        const __m128i e0 =
    809            _mm_sub_epi16(_mm_add_epi16(_mm_packs_epi32(vr0, vr1), d0), s0);
    810        const __m128i err0 = _mm_madd_epi16(e0, e0);
    811        sum32 = _mm_add_epi32(sum32, err0);
    812      }
    813      for (k = j; k < width; ++k) {
    814        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
    815        int32_t v = xq_active * (flt[k] - u);
    816        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
    817        err += ((int64_t)e * e);
    818      }
    819      dat += dat_stride;
    820      src += src_stride;
    821      flt += flt_stride;
    822      const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32);
    823      const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8));
    824      sum64 = _mm_add_epi64(sum64, sum64_0);
    825      sum64 = _mm_add_epi64(sum64, sum64_1);
    826    }
    827  } else {
    828    __m128i sum32 = _mm_setzero_si128();
    829    for (i = 0; i < height; ++i) {
    830      for (j = 0; j <= width - 16; j += 16) {
    831        const __m128i d = xx_loadu_128(dat + j);
    832        const __m128i s = xx_loadu_128(src + j);
    833        const __m128i d0 = _mm_cvtepu8_epi16(d);
    834        const __m128i d1 = _mm_cvtepu8_epi16(_mm_srli_si128(d, 8));
    835        const __m128i s0 = _mm_cvtepu8_epi16(s);
    836        const __m128i s1 = _mm_cvtepu8_epi16(_mm_srli_si128(s, 8));
    837        const __m128i diff0 = _mm_sub_epi16(d0, s0);
    838        const __m128i diff1 = _mm_sub_epi16(d1, s1);
    839        const __m128i err0 = _mm_madd_epi16(diff0, diff0);
    840        const __m128i err1 = _mm_madd_epi16(diff1, diff1);
    841        sum32 = _mm_add_epi32(sum32, err0);
    842        sum32 = _mm_add_epi32(sum32, err1);
    843      }
    844      for (k = j; k < width; ++k) {
    845        const int32_t e = (int32_t)(dat[k]) - src[k];
    846        err += ((int64_t)e * e);
    847      }
    848      dat += dat_stride;
    849      src += src_stride;
    850    }
    851    const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32);
    852    const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8));
    853    sum64 = _mm_add_epi64(sum64_0, sum64_1);
    854  }
    855  int64_t sum[2];
    856  xx_storeu_128(sum, sum64);
    857  err += sum[0] + sum[1];
    858  return err;
    859 }
    860 
    861 // When params->r[0] > 0 and params->r[1] > 0. In this case all elements of
    862 // C and H need to be computed.
    863 static inline void calc_proj_params_r0_r1_sse4_1(
    864    const uint8_t *src8, int width, int height, int src_stride,
    865    const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride,
    866    int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2]) {
    867  const int size = width * height;
    868  const uint8_t *src = src8;
    869  const uint8_t *dat = dat8;
    870  __m128i h00, h01, h11, c0, c1;
    871  const __m128i zero = _mm_setzero_si128();
    872  h01 = h11 = c0 = c1 = h00 = zero;
    873 
    874  for (int i = 0; i < height; ++i) {
    875    for (int j = 0; j < width; j += 4) {
    876      const __m128i u_load = _mm_cvtepu8_epi32(
    877          _mm_cvtsi32_si128(*((int *)(dat + i * dat_stride + j))));
    878      const __m128i s_load = _mm_cvtepu8_epi32(
    879          _mm_cvtsi32_si128(*((int *)(src + i * src_stride + j))));
    880      __m128i f1 = _mm_loadu_si128((__m128i *)(flt0 + i * flt0_stride + j));
    881      __m128i f2 = _mm_loadu_si128((__m128i *)(flt1 + i * flt1_stride + j));
    882      __m128i d = _mm_slli_epi32(u_load, SGRPROJ_RST_BITS);
    883      __m128i s = _mm_slli_epi32(s_load, SGRPROJ_RST_BITS);
    884      s = _mm_sub_epi32(s, d);
    885      f1 = _mm_sub_epi32(f1, d);
    886      f2 = _mm_sub_epi32(f2, d);
    887 
    888      const __m128i h00_even = _mm_mul_epi32(f1, f1);
    889      const __m128i h00_odd =
    890          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(f1, 32));
    891      h00 = _mm_add_epi64(h00, h00_even);
    892      h00 = _mm_add_epi64(h00, h00_odd);
    893 
    894      const __m128i h01_even = _mm_mul_epi32(f1, f2);
    895      const __m128i h01_odd =
    896          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(f2, 32));
    897      h01 = _mm_add_epi64(h01, h01_even);
    898      h01 = _mm_add_epi64(h01, h01_odd);
    899 
    900      const __m128i h11_even = _mm_mul_epi32(f2, f2);
    901      const __m128i h11_odd =
    902          _mm_mul_epi32(_mm_srli_epi64(f2, 32), _mm_srli_epi64(f2, 32));
    903      h11 = _mm_add_epi64(h11, h11_even);
    904      h11 = _mm_add_epi64(h11, h11_odd);
    905 
    906      const __m128i c0_even = _mm_mul_epi32(f1, s);
    907      const __m128i c0_odd =
    908          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(s, 32));
    909      c0 = _mm_add_epi64(c0, c0_even);
    910      c0 = _mm_add_epi64(c0, c0_odd);
    911 
    912      const __m128i c1_even = _mm_mul_epi32(f2, s);
    913      const __m128i c1_odd =
    914          _mm_mul_epi32(_mm_srli_epi64(f2, 32), _mm_srli_epi64(s, 32));
    915      c1 = _mm_add_epi64(c1, c1_even);
    916      c1 = _mm_add_epi64(c1, c1_odd);
    917    }
    918  }
    919 
    920  __m128i c_low = _mm_unpacklo_epi64(c0, c1);
    921  const __m128i c_high = _mm_unpackhi_epi64(c0, c1);
    922  c_low = _mm_add_epi64(c_low, c_high);
    923 
    924  __m128i h0x_low = _mm_unpacklo_epi64(h00, h01);
    925  const __m128i h0x_high = _mm_unpackhi_epi64(h00, h01);
    926  h0x_low = _mm_add_epi64(h0x_low, h0x_high);
    927 
    928  // Using the symmetric properties of H,  calculations of H[1][0] are not
    929  // needed.
    930  __m128i h1x_low = _mm_unpacklo_epi64(zero, h11);
    931  const __m128i h1x_high = _mm_unpackhi_epi64(zero, h11);
    932  h1x_low = _mm_add_epi64(h1x_low, h1x_high);
    933 
    934  xx_storeu_128(C, c_low);
    935  xx_storeu_128(H[0], h0x_low);
    936  xx_storeu_128(H[1], h1x_low);
    937 
    938  H[0][0] /= size;
    939  H[0][1] /= size;
    940  H[1][1] /= size;
    941 
    942  // Since H is a symmetric matrix
    943  H[1][0] = H[0][1];
    944  C[0] /= size;
    945  C[1] /= size;
    946 }
    947 
    948 // When only params->r[0] > 0. In this case only H[0][0] and C[0] are
    949 // non-zero and need to be computed.
    950 static inline void calc_proj_params_r0_sse4_1(const uint8_t *src8, int width,
    951                                              int height, int src_stride,
    952                                              const uint8_t *dat8,
    953                                              int dat_stride, int32_t *flt0,
    954                                              int flt0_stride, int64_t H[2][2],
    955                                              int64_t C[2]) {
    956  const int size = width * height;
    957  const uint8_t *src = src8;
    958  const uint8_t *dat = dat8;
    959  __m128i h00, c0;
    960  const __m128i zero = _mm_setzero_si128();
    961  c0 = h00 = zero;
    962 
    963  for (int i = 0; i < height; ++i) {
    964    for (int j = 0; j < width; j += 4) {
    965      const __m128i u_load = _mm_cvtepu8_epi32(
    966          _mm_cvtsi32_si128(*((int *)(dat + i * dat_stride + j))));
    967      const __m128i s_load = _mm_cvtepu8_epi32(
    968          _mm_cvtsi32_si128(*((int *)(src + i * src_stride + j))));
    969      __m128i f1 = _mm_loadu_si128((__m128i *)(flt0 + i * flt0_stride + j));
    970      __m128i d = _mm_slli_epi32(u_load, SGRPROJ_RST_BITS);
    971      __m128i s = _mm_slli_epi32(s_load, SGRPROJ_RST_BITS);
    972      s = _mm_sub_epi32(s, d);
    973      f1 = _mm_sub_epi32(f1, d);
    974 
    975      const __m128i h00_even = _mm_mul_epi32(f1, f1);
    976      const __m128i h00_odd =
    977          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(f1, 32));
    978      h00 = _mm_add_epi64(h00, h00_even);
    979      h00 = _mm_add_epi64(h00, h00_odd);
    980 
    981      const __m128i c0_even = _mm_mul_epi32(f1, s);
    982      const __m128i c0_odd =
    983          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(s, 32));
    984      c0 = _mm_add_epi64(c0, c0_even);
    985      c0 = _mm_add_epi64(c0, c0_odd);
    986    }
    987  }
    988  const __m128i h00_val = _mm_add_epi64(h00, _mm_srli_si128(h00, 8));
    989 
    990  const __m128i c0_val = _mm_add_epi64(c0, _mm_srli_si128(c0, 8));
    991 
    992  const __m128i c = _mm_unpacklo_epi64(c0_val, zero);
    993  const __m128i h0x = _mm_unpacklo_epi64(h00_val, zero);
    994 
    995  xx_storeu_128(C, c);
    996  xx_storeu_128(H[0], h0x);
    997 
    998  H[0][0] /= size;
    999  C[0] /= size;
   1000 }
   1001 
   1002 // When only params->r[1] > 0. In this case only H[1][1] and C[1] are
   1003 // non-zero and need to be computed.
   1004 static inline void calc_proj_params_r1_sse4_1(const uint8_t *src8, int width,
   1005                                              int height, int src_stride,
   1006                                              const uint8_t *dat8,
   1007                                              int dat_stride, int32_t *flt1,
   1008                                              int flt1_stride, int64_t H[2][2],
   1009                                              int64_t C[2]) {
   1010  const int size = width * height;
   1011  const uint8_t *src = src8;
   1012  const uint8_t *dat = dat8;
   1013  __m128i h11, c1;
   1014  const __m128i zero = _mm_setzero_si128();
   1015  c1 = h11 = zero;
   1016 
   1017  for (int i = 0; i < height; ++i) {
   1018    for (int j = 0; j < width; j += 4) {
   1019      const __m128i u_load = _mm_cvtepu8_epi32(
   1020          _mm_cvtsi32_si128(*((int *)(dat + i * dat_stride + j))));
   1021      const __m128i s_load = _mm_cvtepu8_epi32(
   1022          _mm_cvtsi32_si128(*((int *)(src + i * src_stride + j))));
   1023      __m128i f2 = _mm_loadu_si128((__m128i *)(flt1 + i * flt1_stride + j));
   1024      __m128i d = _mm_slli_epi32(u_load, SGRPROJ_RST_BITS);
   1025      __m128i s = _mm_slli_epi32(s_load, SGRPROJ_RST_BITS);
   1026      s = _mm_sub_epi32(s, d);
   1027      f2 = _mm_sub_epi32(f2, d);
   1028 
   1029      const __m128i h11_even = _mm_mul_epi32(f2, f2);
   1030      const __m128i h11_odd =
   1031          _mm_mul_epi32(_mm_srli_epi64(f2, 32), _mm_srli_epi64(f2, 32));
   1032      h11 = _mm_add_epi64(h11, h11_even);
   1033      h11 = _mm_add_epi64(h11, h11_odd);
   1034 
   1035      const __m128i c1_even = _mm_mul_epi32(f2, s);
   1036      const __m128i c1_odd =
   1037          _mm_mul_epi32(_mm_srli_epi64(f2, 32), _mm_srli_epi64(s, 32));
   1038      c1 = _mm_add_epi64(c1, c1_even);
   1039      c1 = _mm_add_epi64(c1, c1_odd);
   1040    }
   1041  }
   1042 
   1043  const __m128i h11_val = _mm_add_epi64(h11, _mm_srli_si128(h11, 8));
   1044 
   1045  const __m128i c1_val = _mm_add_epi64(c1, _mm_srli_si128(c1, 8));
   1046 
   1047  const __m128i c = _mm_unpacklo_epi64(zero, c1_val);
   1048  const __m128i h1x = _mm_unpacklo_epi64(zero, h11_val);
   1049 
   1050  xx_storeu_128(C, c);
   1051  xx_storeu_128(H[1], h1x);
   1052 
   1053  H[1][1] /= size;
   1054  C[1] /= size;
   1055 }
   1056 
   1057 // SSE4.1 variant of av1_calc_proj_params_c.
   1058 void av1_calc_proj_params_sse4_1(const uint8_t *src8, int width, int height,
   1059                                 int src_stride, const uint8_t *dat8,
   1060                                 int dat_stride, int32_t *flt0, int flt0_stride,
   1061                                 int32_t *flt1, int flt1_stride,
   1062                                 int64_t H[2][2], int64_t C[2],
   1063                                 const sgr_params_type *params) {
   1064  if ((params->r[0] > 0) && (params->r[1] > 0)) {
   1065    calc_proj_params_r0_r1_sse4_1(src8, width, height, src_stride, dat8,
   1066                                  dat_stride, flt0, flt0_stride, flt1,
   1067                                  flt1_stride, H, C);
   1068  } else if (params->r[0] > 0) {
   1069    calc_proj_params_r0_sse4_1(src8, width, height, src_stride, dat8,
   1070                               dat_stride, flt0, flt0_stride, H, C);
   1071  } else if (params->r[1] > 0) {
   1072    calc_proj_params_r1_sse4_1(src8, width, height, src_stride, dat8,
   1073                               dat_stride, flt1, flt1_stride, H, C);
   1074  }
   1075 }
   1076 
   1077 #if CONFIG_AV1_HIGHBITDEPTH
   1078 static inline void calc_proj_params_r0_r1_high_bd_sse4_1(
   1079    const uint8_t *src8, int width, int height, int src_stride,
   1080    const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride,
   1081    int32_t *flt1, int flt1_stride, int64_t H[2][2], int64_t C[2]) {
   1082  const int size = width * height;
   1083  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
   1084  const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
   1085  __m128i h00, h01, h11, c0, c1;
   1086  const __m128i zero = _mm_setzero_si128();
   1087  h01 = h11 = c0 = c1 = h00 = zero;
   1088 
   1089  for (int i = 0; i < height; ++i) {
   1090    for (int j = 0; j < width; j += 4) {
   1091      const __m128i u_load = _mm_cvtepu16_epi32(
   1092          _mm_loadl_epi64((__m128i *)(dat + i * dat_stride + j)));
   1093      const __m128i s_load = _mm_cvtepu16_epi32(
   1094          _mm_loadl_epi64((__m128i *)(src + i * src_stride + j)));
   1095      __m128i f1 = _mm_loadu_si128((__m128i *)(flt0 + i * flt0_stride + j));
   1096      __m128i f2 = _mm_loadu_si128((__m128i *)(flt1 + i * flt1_stride + j));
   1097      __m128i d = _mm_slli_epi32(u_load, SGRPROJ_RST_BITS);
   1098      __m128i s = _mm_slli_epi32(s_load, SGRPROJ_RST_BITS);
   1099      s = _mm_sub_epi32(s, d);
   1100      f1 = _mm_sub_epi32(f1, d);
   1101      f2 = _mm_sub_epi32(f2, d);
   1102 
   1103      const __m128i h00_even = _mm_mul_epi32(f1, f1);
   1104      const __m128i h00_odd =
   1105          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(f1, 32));
   1106      h00 = _mm_add_epi64(h00, h00_even);
   1107      h00 = _mm_add_epi64(h00, h00_odd);
   1108 
   1109      const __m128i h01_even = _mm_mul_epi32(f1, f2);
   1110      const __m128i h01_odd =
   1111          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(f2, 32));
   1112      h01 = _mm_add_epi64(h01, h01_even);
   1113      h01 = _mm_add_epi64(h01, h01_odd);
   1114 
   1115      const __m128i h11_even = _mm_mul_epi32(f2, f2);
   1116      const __m128i h11_odd =
   1117          _mm_mul_epi32(_mm_srli_epi64(f2, 32), _mm_srli_epi64(f2, 32));
   1118      h11 = _mm_add_epi64(h11, h11_even);
   1119      h11 = _mm_add_epi64(h11, h11_odd);
   1120 
   1121      const __m128i c0_even = _mm_mul_epi32(f1, s);
   1122      const __m128i c0_odd =
   1123          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(s, 32));
   1124      c0 = _mm_add_epi64(c0, c0_even);
   1125      c0 = _mm_add_epi64(c0, c0_odd);
   1126 
   1127      const __m128i c1_even = _mm_mul_epi32(f2, s);
   1128      const __m128i c1_odd =
   1129          _mm_mul_epi32(_mm_srli_epi64(f2, 32), _mm_srli_epi64(s, 32));
   1130      c1 = _mm_add_epi64(c1, c1_even);
   1131      c1 = _mm_add_epi64(c1, c1_odd);
   1132    }
   1133  }
   1134 
   1135  __m128i c_low = _mm_unpacklo_epi64(c0, c1);
   1136  const __m128i c_high = _mm_unpackhi_epi64(c0, c1);
   1137  c_low = _mm_add_epi64(c_low, c_high);
   1138 
   1139  __m128i h0x_low = _mm_unpacklo_epi64(h00, h01);
   1140  const __m128i h0x_high = _mm_unpackhi_epi64(h00, h01);
   1141  h0x_low = _mm_add_epi64(h0x_low, h0x_high);
   1142 
   1143  // Using the symmetric properties of H,  calculations of H[1][0] are not
   1144  // needed.
   1145  __m128i h1x_low = _mm_unpacklo_epi64(zero, h11);
   1146  const __m128i h1x_high = _mm_unpackhi_epi64(zero, h11);
   1147  h1x_low = _mm_add_epi64(h1x_low, h1x_high);
   1148 
   1149  xx_storeu_128(C, c_low);
   1150  xx_storeu_128(H[0], h0x_low);
   1151  xx_storeu_128(H[1], h1x_low);
   1152 
   1153  H[0][0] /= size;
   1154  H[0][1] /= size;
   1155  H[1][1] /= size;
   1156 
   1157  // Since H is a symmetric matrix
   1158  H[1][0] = H[0][1];
   1159  C[0] /= size;
   1160  C[1] /= size;
   1161 }
   1162 
   1163 // When only params->r[0] > 0. In this case only H[0][0] and C[0] are
   1164 // non-zero and need to be computed.
   1165 static inline void calc_proj_params_r0_high_bd_sse4_1(
   1166    const uint8_t *src8, int width, int height, int src_stride,
   1167    const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride,
   1168    int64_t H[2][2], int64_t C[2]) {
   1169  const int size = width * height;
   1170  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
   1171  const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
   1172  __m128i h00, c0;
   1173  const __m128i zero = _mm_setzero_si128();
   1174  c0 = h00 = zero;
   1175 
   1176  for (int i = 0; i < height; ++i) {
   1177    for (int j = 0; j < width; j += 4) {
   1178      const __m128i u_load = _mm_cvtepu16_epi32(
   1179          _mm_loadl_epi64((__m128i *)(dat + i * dat_stride + j)));
   1180      const __m128i s_load = _mm_cvtepu16_epi32(
   1181          _mm_loadl_epi64((__m128i *)(src + i * src_stride + j)));
   1182      __m128i f1 = _mm_loadu_si128((__m128i *)(flt0 + i * flt0_stride + j));
   1183      __m128i d = _mm_slli_epi32(u_load, SGRPROJ_RST_BITS);
   1184      __m128i s = _mm_slli_epi32(s_load, SGRPROJ_RST_BITS);
   1185      s = _mm_sub_epi32(s, d);
   1186      f1 = _mm_sub_epi32(f1, d);
   1187 
   1188      const __m128i h00_even = _mm_mul_epi32(f1, f1);
   1189      const __m128i h00_odd =
   1190          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(f1, 32));
   1191      h00 = _mm_add_epi64(h00, h00_even);
   1192      h00 = _mm_add_epi64(h00, h00_odd);
   1193 
   1194      const __m128i c0_even = _mm_mul_epi32(f1, s);
   1195      const __m128i c0_odd =
   1196          _mm_mul_epi32(_mm_srli_epi64(f1, 32), _mm_srli_epi64(s, 32));
   1197      c0 = _mm_add_epi64(c0, c0_even);
   1198      c0 = _mm_add_epi64(c0, c0_odd);
   1199    }
   1200  }
   1201  const __m128i h00_val = _mm_add_epi64(h00, _mm_srli_si128(h00, 8));
   1202 
   1203  const __m128i c0_val = _mm_add_epi64(c0, _mm_srli_si128(c0, 8));
   1204 
   1205  const __m128i c = _mm_unpacklo_epi64(c0_val, zero);
   1206  const __m128i h0x = _mm_unpacklo_epi64(h00_val, zero);
   1207 
   1208  xx_storeu_128(C, c);
   1209  xx_storeu_128(H[0], h0x);
   1210 
   1211  H[0][0] /= size;
   1212  C[0] /= size;
   1213 }
   1214 
   1215 // When only params->r[1] > 0. In this case only H[1][1] and C[1] are
   1216 // non-zero and need to be computed.
   1217 static inline void calc_proj_params_r1_high_bd_sse4_1(
   1218    const uint8_t *src8, int width, int height, int src_stride,
   1219    const uint8_t *dat8, int dat_stride, int32_t *flt1, int flt1_stride,
   1220    int64_t H[2][2], int64_t C[2]) {
   1221  const int size = width * height;
   1222  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
   1223  const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
   1224  __m128i h11, c1;
   1225  const __m128i zero = _mm_setzero_si128();
   1226  c1 = h11 = zero;
   1227 
   1228  for (int i = 0; i < height; ++i) {
   1229    for (int j = 0; j < width; j += 4) {
   1230      const __m128i u_load = _mm_cvtepu16_epi32(
   1231          _mm_loadl_epi64((__m128i *)(dat + i * dat_stride + j)));
   1232      const __m128i s_load = _mm_cvtepu16_epi32(
   1233          _mm_loadl_epi64((__m128i *)(src + i * src_stride + j)));
   1234      __m128i f2 = _mm_loadu_si128((__m128i *)(flt1 + i * flt1_stride + j));
   1235      __m128i d = _mm_slli_epi32(u_load, SGRPROJ_RST_BITS);
   1236      __m128i s = _mm_slli_epi32(s_load, SGRPROJ_RST_BITS);
   1237      s = _mm_sub_epi32(s, d);
   1238      f2 = _mm_sub_epi32(f2, d);
   1239 
   1240      const __m128i h11_even = _mm_mul_epi32(f2, f2);
   1241      const __m128i h11_odd =
   1242          _mm_mul_epi32(_mm_srli_epi64(f2, 32), _mm_srli_epi64(f2, 32));
   1243      h11 = _mm_add_epi64(h11, h11_even);
   1244      h11 = _mm_add_epi64(h11, h11_odd);
   1245 
   1246      const __m128i c1_even = _mm_mul_epi32(f2, s);
   1247      const __m128i c1_odd =
   1248          _mm_mul_epi32(_mm_srli_epi64(f2, 32), _mm_srli_epi64(s, 32));
   1249      c1 = _mm_add_epi64(c1, c1_even);
   1250      c1 = _mm_add_epi64(c1, c1_odd);
   1251    }
   1252  }
   1253 
   1254  const __m128i h11_val = _mm_add_epi64(h11, _mm_srli_si128(h11, 8));
   1255 
   1256  const __m128i c1_val = _mm_add_epi64(c1, _mm_srli_si128(c1, 8));
   1257 
   1258  const __m128i c = _mm_unpacklo_epi64(zero, c1_val);
   1259  const __m128i h1x = _mm_unpacklo_epi64(zero, h11_val);
   1260 
   1261  xx_storeu_128(C, c);
   1262  xx_storeu_128(H[1], h1x);
   1263 
   1264  H[1][1] /= size;
   1265  C[1] /= size;
   1266 }
   1267 
   1268 // SSE4.1 variant of av1_calc_proj_params_high_bd_c.
   1269 void av1_calc_proj_params_high_bd_sse4_1(const uint8_t *src8, int width,
   1270                                         int height, int src_stride,
   1271                                         const uint8_t *dat8, int dat_stride,
   1272                                         int32_t *flt0, int flt0_stride,
   1273                                         int32_t *flt1, int flt1_stride,
   1274                                         int64_t H[2][2], int64_t C[2],
   1275                                         const sgr_params_type *params) {
   1276  if ((params->r[0] > 0) && (params->r[1] > 0)) {
   1277    calc_proj_params_r0_r1_high_bd_sse4_1(src8, width, height, src_stride, dat8,
   1278                                          dat_stride, flt0, flt0_stride, flt1,
   1279                                          flt1_stride, H, C);
   1280  } else if (params->r[0] > 0) {
   1281    calc_proj_params_r0_high_bd_sse4_1(src8, width, height, src_stride, dat8,
   1282                                       dat_stride, flt0, flt0_stride, H, C);
   1283  } else if (params->r[1] > 0) {
   1284    calc_proj_params_r1_high_bd_sse4_1(src8, width, height, src_stride, dat8,
   1285                                       dat_stride, flt1, flt1_stride, H, C);
   1286  }
   1287 }
   1288 
   1289 int64_t av1_highbd_pixel_proj_error_sse4_1(
   1290    const uint8_t *src8, int width, int height, int src_stride,
   1291    const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride,
   1292    int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params) {
   1293  int i, j, k;
   1294  const int32_t shift = SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS;
   1295  const __m128i rounding = _mm_set1_epi32(1 << (shift - 1));
   1296  __m128i sum64 = _mm_setzero_si128();
   1297  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
   1298  const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
   1299  int64_t err = 0;
   1300  if (params->r[0] > 0 && params->r[1] > 0) {  // Both filters are enabled
   1301    const __m128i xq0 = _mm_set1_epi32(xq[0]);
   1302    const __m128i xq1 = _mm_set1_epi32(xq[1]);
   1303 
   1304    for (i = 0; i < height; ++i) {
   1305      __m128i sum32 = _mm_setzero_si128();
   1306      for (j = 0; j <= width - 8; j += 8) {
   1307        // Load 8x pixels from source image
   1308        const __m128i s0 = xx_loadu_128(src + j);
   1309        // s0 = [7 6 5 4 3 2 1 0] as i16 (indices of src[])
   1310 
   1311        // Load 8x pixels from corrupted image
   1312        const __m128i d0 = xx_loadu_128(dat + j);
   1313        // d0 = [7 6 5 4 3 2 1 0] as i16 (indices of dat[])
   1314 
   1315        // Shift each pixel value up by SGRPROJ_RST_BITS
   1316        const __m128i u0 = _mm_slli_epi16(d0, SGRPROJ_RST_BITS);
   1317 
   1318        // Split u0 into two halves and pad each from u16 to i32
   1319        const __m128i u0l = _mm_cvtepu16_epi32(u0);
   1320        const __m128i u0h = _mm_cvtepu16_epi32(_mm_srli_si128(u0, 8));
   1321        // u0h = [7 6 5 4] as i32, u0l = [3 2 1 0] as i32, all dat[] indices
   1322 
   1323        // Load 8 pixels from first and second filtered images
   1324        const __m128i flt0l = xx_loadu_128(flt0 + j);
   1325        const __m128i flt0h = xx_loadu_128(flt0 + j + 4);
   1326        const __m128i flt1l = xx_loadu_128(flt1 + j);
   1327        const __m128i flt1h = xx_loadu_128(flt1 + j + 4);
   1328        // flt0 = [7 6 5 4] [3 2 1 0] as i32 (indices of flt0+j)
   1329        // flt1 = [7 6 5 4] [3 2 1 0] as i32 (indices of flt1+j)
   1330 
   1331        // Subtract shifted corrupt image from each filtered image
   1332        // This gives our two basis vectors for the projection
   1333        const __m128i flt0l_subu = _mm_sub_epi32(flt0l, u0l);
   1334        const __m128i flt0h_subu = _mm_sub_epi32(flt0h, u0h);
   1335        const __m128i flt1l_subu = _mm_sub_epi32(flt1l, u0l);
   1336        const __m128i flt1h_subu = _mm_sub_epi32(flt1h, u0h);
   1337        // flt?h_subu = [ f[7]-u[7] f[6]-u[6] f[5]-u[5] f[4]-u[4] ] as i32
   1338        // flt?l_subu = [ f[3]-u[3] f[2]-u[2] f[1]-u[1] f[0]-u[0] ] as i32
   1339 
   1340        // Multiply each basis vector by the corresponding coefficient
   1341        const __m128i v0l = _mm_mullo_epi32(flt0l_subu, xq0);
   1342        const __m128i v0h = _mm_mullo_epi32(flt0h_subu, xq0);
   1343        const __m128i v1l = _mm_mullo_epi32(flt1l_subu, xq1);
   1344        const __m128i v1h = _mm_mullo_epi32(flt1h_subu, xq1);
   1345 
   1346        // Add together the contribution from each scaled basis vector
   1347        const __m128i vl = _mm_add_epi32(v0l, v1l);
   1348        const __m128i vh = _mm_add_epi32(v0h, v1h);
   1349 
   1350        // Right-shift v with appropriate rounding
   1351        const __m128i vrl = _mm_srai_epi32(_mm_add_epi32(vl, rounding), shift);
   1352        const __m128i vrh = _mm_srai_epi32(_mm_add_epi32(vh, rounding), shift);
   1353 
   1354        // Saturate each i32 value to i16 and combine lower and upper halves
   1355        const __m128i vr = _mm_packs_epi32(vrl, vrh);
   1356 
   1357        // Add twin-subspace-sgr-filter to corrupt image then subtract source
   1358        const __m128i e0 = _mm_sub_epi16(_mm_add_epi16(vr, d0), s0);
   1359 
   1360        // Calculate squared error and add adjacent values
   1361        const __m128i err0 = _mm_madd_epi16(e0, e0);
   1362 
   1363        sum32 = _mm_add_epi32(sum32, err0);
   1364      }
   1365 
   1366      const __m128i sum32l = _mm_cvtepu32_epi64(sum32);
   1367      sum64 = _mm_add_epi64(sum64, sum32l);
   1368      const __m128i sum32h = _mm_cvtepu32_epi64(_mm_srli_si128(sum32, 8));
   1369      sum64 = _mm_add_epi64(sum64, sum32h);
   1370 
   1371      // Process remaining pixels in this row (modulo 8)
   1372      for (k = j; k < width; ++k) {
   1373        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
   1374        int32_t v = xq[0] * (flt0[k] - u) + xq[1] * (flt1[k] - u);
   1375        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
   1376        err += ((int64_t)e * e);
   1377      }
   1378      dat += dat_stride;
   1379      src += src_stride;
   1380      flt0 += flt0_stride;
   1381      flt1 += flt1_stride;
   1382    }
   1383  } else if (params->r[0] > 0 || params->r[1] > 0) {  // Only one filter enabled
   1384    const int32_t xq_on = (params->r[0] > 0) ? xq[0] : xq[1];
   1385    const __m128i xq_active = _mm_set1_epi32(xq_on);
   1386    const __m128i xq_inactive =
   1387        _mm_set1_epi32(-xq_on * (1 << SGRPROJ_RST_BITS));
   1388    const int32_t *flt = (params->r[0] > 0) ? flt0 : flt1;
   1389    const int flt_stride = (params->r[0] > 0) ? flt0_stride : flt1_stride;
   1390    for (i = 0; i < height; ++i) {
   1391      __m128i sum32 = _mm_setzero_si128();
   1392      for (j = 0; j <= width - 8; j += 8) {
   1393        // Load 8x pixels from source image
   1394        const __m128i s0 = xx_loadu_128(src + j);
   1395        // s0 = [7 6 5 4 3 2 1 0] as u16 (indices of src[])
   1396 
   1397        // Load 8x pixels from corrupted image and pad each u16 to i32
   1398        const __m128i d0 = xx_loadu_128(dat + j);
   1399        const __m128i d0h = _mm_cvtepu16_epi32(_mm_srli_si128(d0, 8));
   1400        const __m128i d0l = _mm_cvtepu16_epi32(d0);
   1401        // d0h, d0l = [7 6 5 4], [3 2 1 0] as u32 (indices of dat[])
   1402 
   1403        // Load 8 pixels from the filtered image
   1404        const __m128i flth = xx_loadu_128(flt + j + 4);
   1405        const __m128i fltl = xx_loadu_128(flt + j);
   1406        // flth, fltl = [7 6 5 4], [3 2 1 0] as i32 (indices of flt+j)
   1407 
   1408        const __m128i flth_xq = _mm_mullo_epi32(flth, xq_active);
   1409        const __m128i fltl_xq = _mm_mullo_epi32(fltl, xq_active);
   1410        const __m128i d0h_xq = _mm_mullo_epi32(d0h, xq_inactive);
   1411        const __m128i d0l_xq = _mm_mullo_epi32(d0l, xq_inactive);
   1412 
   1413        const __m128i vh = _mm_add_epi32(flth_xq, d0h_xq);
   1414        const __m128i vl = _mm_add_epi32(fltl_xq, d0l_xq);
   1415        // vh = [ xq0(f[7]-d[7]) xq0(f[6]-d[6]) xq0(f[5]-d[5]) xq0(f[4]-d[4]) ]
   1416        // vl = [ xq0(f[3]-d[3]) xq0(f[2]-d[2]) xq0(f[1]-d[1]) xq0(f[0]-d[0]) ]
   1417 
   1418        // Shift this down with appropriate rounding
   1419        const __m128i vrh = _mm_srai_epi32(_mm_add_epi32(vh, rounding), shift);
   1420        const __m128i vrl = _mm_srai_epi32(_mm_add_epi32(vl, rounding), shift);
   1421 
   1422        // Saturate vr0 and vr1 from i32 to i16 then pack together
   1423        const __m128i vr = _mm_packs_epi32(vrl, vrh);
   1424 
   1425        // Subtract twin-subspace-sgr filtered from source image to get error
   1426        const __m128i e0 = _mm_sub_epi16(_mm_add_epi16(vr, d0), s0);
   1427 
   1428        // Calculate squared error and add adjacent values
   1429        const __m128i err0 = _mm_madd_epi16(e0, e0);
   1430 
   1431        sum32 = _mm_add_epi32(sum32, err0);
   1432      }
   1433 
   1434      const __m128i sum32l = _mm_cvtepu32_epi64(sum32);
   1435      sum64 = _mm_add_epi64(sum64, sum32l);
   1436      const __m128i sum32h = _mm_cvtepu32_epi64(_mm_srli_si128(sum32, 8));
   1437      sum64 = _mm_add_epi64(sum64, sum32h);
   1438 
   1439      // Process remaining pixels in this row (modulo 8)
   1440      for (k = j; k < width; ++k) {
   1441        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
   1442        int32_t v = xq_on * (flt[k] - u);
   1443        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
   1444        err += ((int64_t)e * e);
   1445      }
   1446      dat += dat_stride;
   1447      src += src_stride;
   1448      flt += flt_stride;
   1449    }
   1450  } else {  // Neither filter is enabled
   1451    for (i = 0; i < height; ++i) {
   1452      __m128i sum32 = _mm_setzero_si128();
   1453      for (j = 0; j <= width - 16; j += 16) {
   1454        // Load 2x8 u16 from source image
   1455        const __m128i s0 = xx_loadu_128(src + j);
   1456        const __m128i s1 = xx_loadu_128(src + j + 8);
   1457        // Load 2x8 u16 from corrupted image
   1458        const __m128i d0 = xx_loadu_128(dat + j);
   1459        const __m128i d1 = xx_loadu_128(dat + j + 8);
   1460 
   1461        // Subtract corrupted image from source image
   1462        const __m128i diff0 = _mm_sub_epi16(d0, s0);
   1463        const __m128i diff1 = _mm_sub_epi16(d1, s1);
   1464 
   1465        // Square error and add adjacent values
   1466        const __m128i err0 = _mm_madd_epi16(diff0, diff0);
   1467        const __m128i err1 = _mm_madd_epi16(diff1, diff1);
   1468 
   1469        sum32 = _mm_add_epi32(sum32, err0);
   1470        sum32 = _mm_add_epi32(sum32, err1);
   1471      }
   1472 
   1473      const __m128i sum32l = _mm_cvtepu32_epi64(sum32);
   1474      sum64 = _mm_add_epi64(sum64, sum32l);
   1475      const __m128i sum32h = _mm_cvtepu32_epi64(_mm_srli_si128(sum32, 8));
   1476      sum64 = _mm_add_epi64(sum64, sum32h);
   1477 
   1478      // Process remaining pixels (modulu 8)
   1479      for (k = j; k < width; ++k) {
   1480        const int32_t e = (int32_t)(dat[k]) - src[k];
   1481        err += ((int64_t)e * e);
   1482      }
   1483      dat += dat_stride;
   1484      src += src_stride;
   1485    }
   1486  }
   1487 
   1488  // Sum 4 values from sum64l and sum64h into err
   1489  int64_t sum[2];
   1490  xx_storeu_128(sum, sum64);
   1491  err += sum[0] + sum[1];
   1492  return err;
   1493 }
   1494 #endif  // CONFIG_AV1_HIGHBITDEPTH