tor-browser

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

selfguided_sse4.c (26475B)


      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 <smmintrin.h>
     13 
     14 #include "config/aom_config.h"
     15 #include "config/av1_rtcd.h"
     16 
     17 #include "av1/common/restoration.h"
     18 #include "aom_dsp/x86/synonyms.h"
     19 
     20 // Load 4 bytes from the possibly-misaligned pointer p, extend each byte to
     21 // 32-bit precision and return them in an SSE register.
     22 static __m128i xx_load_extend_8_32(const void *p) {
     23  return _mm_cvtepu8_epi32(xx_loadl_32(p));
     24 }
     25 
     26 // Load 4 halfwords from the possibly-misaligned pointer p, extend each
     27 // halfword to 32-bit precision and return them in an SSE register.
     28 static __m128i xx_load_extend_16_32(const void *p) {
     29  return _mm_cvtepu16_epi32(xx_loadl_64(p));
     30 }
     31 
     32 // Compute the scan of an SSE register holding 4 32-bit integers. If the
     33 // register holds x0..x3 then the scan will hold x0, x0+x1, x0+x1+x2,
     34 // x0+x1+x2+x3
     35 static __m128i scan_32(__m128i x) {
     36  const __m128i x01 = _mm_add_epi32(x, _mm_slli_si128(x, 4));
     37  return _mm_add_epi32(x01, _mm_slli_si128(x01, 8));
     38 }
     39 
     40 // Compute two integral images from src. B sums elements; A sums their
     41 // squares. The images are offset by one pixel, so will have width and height
     42 // equal to width + 1, height + 1 and the first row and column will be zero.
     43 //
     44 // A+1 and B+1 should be aligned to 16 bytes. buf_stride should be a multiple
     45 // of 4.
     46 static void integral_images(const uint8_t *src, int src_stride, int width,
     47                            int height, int32_t *A, int32_t *B,
     48                            int buf_stride) {
     49  // Write out the zero top row
     50  memset(A, 0, sizeof(*A) * (width + 1));
     51  memset(B, 0, sizeof(*B) * (width + 1));
     52 
     53  const __m128i zero = _mm_setzero_si128();
     54  for (int i = 0; i < height; ++i) {
     55    // Zero the left column.
     56    A[(i + 1) * buf_stride] = B[(i + 1) * buf_stride] = 0;
     57 
     58    // ldiff is the difference H - D where H is the output sample immediately
     59    // to the left and D is the output sample above it. These are scalars,
     60    // replicated across the four lanes.
     61    __m128i ldiff1 = zero, ldiff2 = zero;
     62    for (int j = 0; j < width; j += 4) {
     63      const int ABj = 1 + j;
     64 
     65      const __m128i above1 = xx_load_128(B + ABj + i * buf_stride);
     66      const __m128i above2 = xx_load_128(A + ABj + i * buf_stride);
     67 
     68      const __m128i x1 = xx_load_extend_8_32(src + j + i * src_stride);
     69      const __m128i x2 = _mm_madd_epi16(x1, x1);
     70 
     71      const __m128i sc1 = scan_32(x1);
     72      const __m128i sc2 = scan_32(x2);
     73 
     74      const __m128i row1 = _mm_add_epi32(_mm_add_epi32(sc1, above1), ldiff1);
     75      const __m128i row2 = _mm_add_epi32(_mm_add_epi32(sc2, above2), ldiff2);
     76 
     77      xx_store_128(B + ABj + (i + 1) * buf_stride, row1);
     78      xx_store_128(A + ABj + (i + 1) * buf_stride, row2);
     79 
     80      // Calculate the new H - D.
     81      ldiff1 = _mm_shuffle_epi32(_mm_sub_epi32(row1, above1), 0xff);
     82      ldiff2 = _mm_shuffle_epi32(_mm_sub_epi32(row2, above2), 0xff);
     83    }
     84  }
     85 }
     86 
     87 // Compute two integral images from src. B sums elements; A sums their squares
     88 //
     89 // A and B should be aligned to 16 bytes. buf_stride should be a multiple of 4.
     90 static void integral_images_highbd(const uint16_t *src, int src_stride,
     91                                   int width, int height, int32_t *A,
     92                                   int32_t *B, int buf_stride) {
     93  // Write out the zero top row
     94  memset(A, 0, sizeof(*A) * (width + 1));
     95  memset(B, 0, sizeof(*B) * (width + 1));
     96 
     97  const __m128i zero = _mm_setzero_si128();
     98  for (int i = 0; i < height; ++i) {
     99    // Zero the left column.
    100    A[(i + 1) * buf_stride] = B[(i + 1) * buf_stride] = 0;
    101 
    102    // ldiff is the difference H - D where H is the output sample immediately
    103    // to the left and D is the output sample above it. These are scalars,
    104    // replicated across the four lanes.
    105    __m128i ldiff1 = zero, ldiff2 = zero;
    106    for (int j = 0; j < width; j += 4) {
    107      const int ABj = 1 + j;
    108 
    109      const __m128i above1 = xx_load_128(B + ABj + i * buf_stride);
    110      const __m128i above2 = xx_load_128(A + ABj + i * buf_stride);
    111 
    112      const __m128i x1 = xx_load_extend_16_32(src + j + i * src_stride);
    113      const __m128i x2 = _mm_madd_epi16(x1, x1);
    114 
    115      const __m128i sc1 = scan_32(x1);
    116      const __m128i sc2 = scan_32(x2);
    117 
    118      const __m128i row1 = _mm_add_epi32(_mm_add_epi32(sc1, above1), ldiff1);
    119      const __m128i row2 = _mm_add_epi32(_mm_add_epi32(sc2, above2), ldiff2);
    120 
    121      xx_store_128(B + ABj + (i + 1) * buf_stride, row1);
    122      xx_store_128(A + ABj + (i + 1) * buf_stride, row2);
    123 
    124      // Calculate the new H - D.
    125      ldiff1 = _mm_shuffle_epi32(_mm_sub_epi32(row1, above1), 0xff);
    126      ldiff2 = _mm_shuffle_epi32(_mm_sub_epi32(row2, above2), 0xff);
    127    }
    128  }
    129 }
    130 
    131 // Compute 4 values of boxsum from the given integral image. ii should point
    132 // at the middle of the box (for the first value). r is the box radius.
    133 static inline __m128i boxsum_from_ii(const int32_t *ii, int stride, int r) {
    134  const __m128i tl = xx_loadu_128(ii - (r + 1) - (r + 1) * stride);
    135  const __m128i tr = xx_loadu_128(ii + (r + 0) - (r + 1) * stride);
    136  const __m128i bl = xx_loadu_128(ii - (r + 1) + r * stride);
    137  const __m128i br = xx_loadu_128(ii + (r + 0) + r * stride);
    138  const __m128i u = _mm_sub_epi32(tr, tl);
    139  const __m128i v = _mm_sub_epi32(br, bl);
    140  return _mm_sub_epi32(v, u);
    141 }
    142 
    143 static __m128i round_for_shift(unsigned shift) {
    144  return _mm_set1_epi32((1 << shift) >> 1);
    145 }
    146 
    147 static __m128i compute_p(__m128i sum1, __m128i sum2, int bit_depth, int n) {
    148  __m128i an, bb;
    149  if (bit_depth > 8) {
    150    const __m128i rounding_a = round_for_shift(2 * (bit_depth - 8));
    151    const __m128i rounding_b = round_for_shift(bit_depth - 8);
    152    const __m128i shift_a = _mm_cvtsi32_si128(2 * (bit_depth - 8));
    153    const __m128i shift_b = _mm_cvtsi32_si128(bit_depth - 8);
    154    const __m128i a = _mm_srl_epi32(_mm_add_epi32(sum2, rounding_a), shift_a);
    155    const __m128i b = _mm_srl_epi32(_mm_add_epi32(sum1, rounding_b), shift_b);
    156    // b < 2^14, so we can use a 16-bit madd rather than a 32-bit
    157    // mullo to square it
    158    bb = _mm_madd_epi16(b, b);
    159    an = _mm_max_epi32(_mm_mullo_epi32(a, _mm_set1_epi32(n)), bb);
    160  } else {
    161    bb = _mm_madd_epi16(sum1, sum1);
    162    an = _mm_mullo_epi32(sum2, _mm_set1_epi32(n));
    163  }
    164  return _mm_sub_epi32(an, bb);
    165 }
    166 
    167 // Assumes that C, D are integral images for the original buffer which has been
    168 // extended to have a padding of SGRPROJ_BORDER_VERT/SGRPROJ_BORDER_HORZ pixels
    169 // on the sides. A, B, C, D point at logical position (0, 0).
    170 static void calc_ab(int32_t *A, int32_t *B, const int32_t *C, const int32_t *D,
    171                    int width, int height, int buf_stride, int bit_depth,
    172                    int sgr_params_idx, int radius_idx) {
    173  const sgr_params_type *const params = &av1_sgr_params[sgr_params_idx];
    174  const int r = params->r[radius_idx];
    175  const int n = (2 * r + 1) * (2 * r + 1);
    176  const __m128i s = _mm_set1_epi32(params->s[radius_idx]);
    177  // one_over_n[n-1] is 2^12/n, so easily fits in an int16
    178  const __m128i one_over_n = _mm_set1_epi32(av1_one_by_x[n - 1]);
    179 
    180  const __m128i rnd_z = round_for_shift(SGRPROJ_MTABLE_BITS);
    181  const __m128i rnd_res = round_for_shift(SGRPROJ_RECIP_BITS);
    182 
    183  // Set up masks
    184  const __m128i ones32 = _mm_set_epi32(0, 0, ~0, ~0);
    185  __m128i mask[4];
    186  for (int idx = 0; idx < 4; idx++) {
    187    const __m128i shift = _mm_cvtsi32_si128(8 * (4 - idx));
    188    mask[idx] = _mm_cvtepi8_epi32(_mm_srl_epi64(ones32, shift));
    189  }
    190 
    191  for (int i = -1; i < height + 1; ++i) {
    192    for (int j = -1; j < width + 1; j += 4) {
    193      const int32_t *Cij = C + i * buf_stride + j;
    194      const int32_t *Dij = D + i * buf_stride + j;
    195 
    196      __m128i sum1 = boxsum_from_ii(Dij, buf_stride, r);
    197      __m128i sum2 = boxsum_from_ii(Cij, buf_stride, r);
    198 
    199      // When width + 2 isn't a multiple of 4, sum1 and sum2 will contain
    200      // some uninitialised data in their upper words. We use a mask to
    201      // ensure that these bits are set to 0.
    202      int idx = AOMMIN(4, width + 1 - j);
    203      assert(idx >= 1);
    204 
    205      if (idx < 4) {
    206        sum1 = _mm_and_si128(mask[idx], sum1);
    207        sum2 = _mm_and_si128(mask[idx], sum2);
    208      }
    209 
    210      const __m128i p = compute_p(sum1, sum2, bit_depth, n);
    211 
    212      const __m128i z = _mm_min_epi32(
    213          _mm_srli_epi32(_mm_add_epi32(_mm_mullo_epi32(p, s), rnd_z),
    214                         SGRPROJ_MTABLE_BITS),
    215          _mm_set1_epi32(255));
    216 
    217      // 'Gather' type instructions are not available pre-AVX2, so synthesize a
    218      // gather using scalar loads.
    219      const __m128i a_res =
    220          _mm_set_epi32(av1_x_by_xplus1[_mm_extract_epi32(z, 3)],
    221                        av1_x_by_xplus1[_mm_extract_epi32(z, 2)],
    222                        av1_x_by_xplus1[_mm_extract_epi32(z, 1)],
    223                        av1_x_by_xplus1[_mm_extract_epi32(z, 0)]);
    224 
    225      xx_storeu_128(A + i * buf_stride + j, a_res);
    226 
    227      const __m128i a_complement =
    228          _mm_sub_epi32(_mm_set1_epi32(SGRPROJ_SGR), a_res);
    229 
    230      // sum1 might have lanes greater than 2^15, so we can't use madd to do
    231      // multiplication involving sum1. However, a_complement and one_over_n
    232      // are both less than 256, so we can multiply them first.
    233      const __m128i a_comp_over_n = _mm_madd_epi16(a_complement, one_over_n);
    234      const __m128i b_int = _mm_mullo_epi32(a_comp_over_n, sum1);
    235      const __m128i b_res =
    236          _mm_srli_epi32(_mm_add_epi32(b_int, rnd_res), SGRPROJ_RECIP_BITS);
    237 
    238      xx_storeu_128(B + i * buf_stride + j, b_res);
    239    }
    240  }
    241 }
    242 
    243 // Calculate 4 values of the "cross sum" starting at buf. This is a 3x3 filter
    244 // where the outer four corners have weight 3 and all other pixels have weight
    245 // 4.
    246 //
    247 // Pixels are indexed like this:
    248 // xtl  xt   xtr
    249 // xl    x   xr
    250 // xbl  xb   xbr
    251 //
    252 // buf points to x
    253 //
    254 // fours = xl + xt + xr + xb + x
    255 // threes = xtl + xtr + xbr + xbl
    256 // cross_sum = 4 * fours + 3 * threes
    257 //           = 4 * (fours + threes) - threes
    258 //           = (fours + threes) << 2 - threes
    259 static inline __m128i cross_sum(const int32_t *buf, int stride) {
    260  const __m128i xtl = xx_loadu_128(buf - 1 - stride);
    261  const __m128i xt = xx_loadu_128(buf - stride);
    262  const __m128i xtr = xx_loadu_128(buf + 1 - stride);
    263  const __m128i xl = xx_loadu_128(buf - 1);
    264  const __m128i x = xx_loadu_128(buf);
    265  const __m128i xr = xx_loadu_128(buf + 1);
    266  const __m128i xbl = xx_loadu_128(buf - 1 + stride);
    267  const __m128i xb = xx_loadu_128(buf + stride);
    268  const __m128i xbr = xx_loadu_128(buf + 1 + stride);
    269 
    270  const __m128i fours = _mm_add_epi32(
    271      xl, _mm_add_epi32(xt, _mm_add_epi32(xr, _mm_add_epi32(xb, x))));
    272  const __m128i threes =
    273      _mm_add_epi32(xtl, _mm_add_epi32(xtr, _mm_add_epi32(xbr, xbl)));
    274 
    275  return _mm_sub_epi32(_mm_slli_epi32(_mm_add_epi32(fours, threes), 2), threes);
    276 }
    277 
    278 // The final filter for self-guided restoration. Computes a weighted average
    279 // across A, B with "cross sums" (see cross_sum implementation above).
    280 static void final_filter(int32_t *dst, int dst_stride, const int32_t *A,
    281                         const int32_t *B, int buf_stride, const void *dgd8,
    282                         int dgd_stride, int width, int height, int highbd) {
    283  const int nb = 5;
    284  const __m128i rounding =
    285      round_for_shift(SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
    286  const uint8_t *dgd_real =
    287      highbd ? (const uint8_t *)CONVERT_TO_SHORTPTR(dgd8) : dgd8;
    288 
    289  for (int i = 0; i < height; ++i) {
    290    for (int j = 0; j < width; j += 4) {
    291      const __m128i a = cross_sum(A + i * buf_stride + j, buf_stride);
    292      const __m128i b = cross_sum(B + i * buf_stride + j, buf_stride);
    293      const __m128i raw =
    294          xx_loadl_64(dgd_real + ((i * dgd_stride + j) << highbd));
    295      const __m128i src =
    296          highbd ? _mm_cvtepu16_epi32(raw) : _mm_cvtepu8_epi32(raw);
    297 
    298      __m128i v = _mm_add_epi32(_mm_madd_epi16(a, src), b);
    299      __m128i w = _mm_srai_epi32(_mm_add_epi32(v, rounding),
    300                                 SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
    301 
    302      xx_storeu_128(dst + i * dst_stride + j, w);
    303    }
    304  }
    305 }
    306 
    307 // Assumes that C, D are integral images for the original buffer which has been
    308 // extended to have a padding of SGRPROJ_BORDER_VERT/SGRPROJ_BORDER_HORZ pixels
    309 // on the sides. A, B, C, D point at logical position (0, 0).
    310 static void calc_ab_fast(int32_t *A, int32_t *B, const int32_t *C,
    311                         const int32_t *D, int width, int height,
    312                         int buf_stride, int bit_depth, int sgr_params_idx,
    313                         int radius_idx) {
    314  const sgr_params_type *const params = &av1_sgr_params[sgr_params_idx];
    315  const int r = params->r[radius_idx];
    316  const int n = (2 * r + 1) * (2 * r + 1);
    317  const __m128i s = _mm_set1_epi32(params->s[radius_idx]);
    318  // one_over_n[n-1] is 2^12/n, so easily fits in an int16
    319  const __m128i one_over_n = _mm_set1_epi32(av1_one_by_x[n - 1]);
    320 
    321  const __m128i rnd_z = round_for_shift(SGRPROJ_MTABLE_BITS);
    322  const __m128i rnd_res = round_for_shift(SGRPROJ_RECIP_BITS);
    323 
    324  // Set up masks
    325  const __m128i ones32 = _mm_set_epi32(0, 0, ~0, ~0);
    326  __m128i mask[4];
    327  for (int idx = 0; idx < 4; idx++) {
    328    const __m128i shift = _mm_cvtsi32_si128(8 * (4 - idx));
    329    mask[idx] = _mm_cvtepi8_epi32(_mm_srl_epi64(ones32, shift));
    330  }
    331 
    332  for (int i = -1; i < height + 1; i += 2) {
    333    for (int j = -1; j < width + 1; j += 4) {
    334      const int32_t *Cij = C + i * buf_stride + j;
    335      const int32_t *Dij = D + i * buf_stride + j;
    336 
    337      __m128i sum1 = boxsum_from_ii(Dij, buf_stride, r);
    338      __m128i sum2 = boxsum_from_ii(Cij, buf_stride, r);
    339 
    340      // When width + 2 isn't a multiple of 4, sum1 and sum2 will contain
    341      // some uninitialised data in their upper words. We use a mask to
    342      // ensure that these bits are set to 0.
    343      int idx = AOMMIN(4, width + 1 - j);
    344      assert(idx >= 1);
    345 
    346      if (idx < 4) {
    347        sum1 = _mm_and_si128(mask[idx], sum1);
    348        sum2 = _mm_and_si128(mask[idx], sum2);
    349      }
    350 
    351      const __m128i p = compute_p(sum1, sum2, bit_depth, n);
    352 
    353      const __m128i z = _mm_min_epi32(
    354          _mm_srli_epi32(_mm_add_epi32(_mm_mullo_epi32(p, s), rnd_z),
    355                         SGRPROJ_MTABLE_BITS),
    356          _mm_set1_epi32(255));
    357 
    358      // 'Gather' type instructions are not available pre-AVX2, so synthesize a
    359      // gather using scalar loads.
    360      const __m128i a_res =
    361          _mm_set_epi32(av1_x_by_xplus1[_mm_extract_epi32(z, 3)],
    362                        av1_x_by_xplus1[_mm_extract_epi32(z, 2)],
    363                        av1_x_by_xplus1[_mm_extract_epi32(z, 1)],
    364                        av1_x_by_xplus1[_mm_extract_epi32(z, 0)]);
    365 
    366      xx_storeu_128(A + i * buf_stride + j, a_res);
    367 
    368      const __m128i a_complement =
    369          _mm_sub_epi32(_mm_set1_epi32(SGRPROJ_SGR), a_res);
    370 
    371      // sum1 might have lanes greater than 2^15, so we can't use madd to do
    372      // multiplication involving sum1. However, a_complement and one_over_n
    373      // are both less than 256, so we can multiply them first.
    374      const __m128i a_comp_over_n = _mm_madd_epi16(a_complement, one_over_n);
    375      const __m128i b_int = _mm_mullo_epi32(a_comp_over_n, sum1);
    376      const __m128i b_res =
    377          _mm_srli_epi32(_mm_add_epi32(b_int, rnd_res), SGRPROJ_RECIP_BITS);
    378 
    379      xx_storeu_128(B + i * buf_stride + j, b_res);
    380    }
    381  }
    382 }
    383 
    384 // Calculate 4 values of the "cross sum" starting at buf.
    385 //
    386 // Pixels are indexed like this:
    387 // xtl  xt   xtr
    388 //  -   buf   -
    389 // xbl  xb   xbr
    390 //
    391 // Pixels are weighted like this:
    392 //  5    6    5
    393 //  0    0    0
    394 //  5    6    5
    395 //
    396 // fives = xtl + xtr + xbl + xbr
    397 // sixes = xt + xb
    398 // cross_sum = 6 * sixes + 5 * fives
    399 //           = 5 * (fives + sixes) - sixes
    400 //           = (fives + sixes) << 2 + (fives + sixes) + sixes
    401 static inline __m128i cross_sum_fast_even_row(const int32_t *buf, int stride) {
    402  const __m128i xtl = xx_loadu_128(buf - 1 - stride);
    403  const __m128i xt = xx_loadu_128(buf - stride);
    404  const __m128i xtr = xx_loadu_128(buf + 1 - stride);
    405  const __m128i xbl = xx_loadu_128(buf - 1 + stride);
    406  const __m128i xb = xx_loadu_128(buf + stride);
    407  const __m128i xbr = xx_loadu_128(buf + 1 + stride);
    408 
    409  const __m128i fives =
    410      _mm_add_epi32(xtl, _mm_add_epi32(xtr, _mm_add_epi32(xbr, xbl)));
    411  const __m128i sixes = _mm_add_epi32(xt, xb);
    412  const __m128i fives_plus_sixes = _mm_add_epi32(fives, sixes);
    413 
    414  return _mm_add_epi32(
    415      _mm_add_epi32(_mm_slli_epi32(fives_plus_sixes, 2), fives_plus_sixes),
    416      sixes);
    417 }
    418 
    419 // Calculate 4 values of the "cross sum" starting at buf.
    420 //
    421 // Pixels are indexed like this:
    422 // xl    x   xr
    423 //
    424 // Pixels are weighted like this:
    425 //  5    6    5
    426 //
    427 // buf points to x
    428 //
    429 // fives = xl + xr
    430 // sixes = x
    431 // cross_sum = 5 * fives + 6 * sixes
    432 //           = 4 * (fives + sixes) + (fives + sixes) + sixes
    433 //           = (fives + sixes) << 2 + (fives + sixes) + sixes
    434 static inline __m128i cross_sum_fast_odd_row(const int32_t *buf) {
    435  const __m128i xl = xx_loadu_128(buf - 1);
    436  const __m128i x = xx_loadu_128(buf);
    437  const __m128i xr = xx_loadu_128(buf + 1);
    438 
    439  const __m128i fives = _mm_add_epi32(xl, xr);
    440  const __m128i sixes = x;
    441 
    442  const __m128i fives_plus_sixes = _mm_add_epi32(fives, sixes);
    443 
    444  return _mm_add_epi32(
    445      _mm_add_epi32(_mm_slli_epi32(fives_plus_sixes, 2), fives_plus_sixes),
    446      sixes);
    447 }
    448 
    449 // The final filter for the self-guided restoration. Computes a
    450 // weighted average across A, B with "cross sums" (see cross_sum_...
    451 // implementations above).
    452 static void final_filter_fast(int32_t *dst, int dst_stride, const int32_t *A,
    453                              const int32_t *B, int buf_stride,
    454                              const void *dgd8, int dgd_stride, int width,
    455                              int height, int highbd) {
    456  const int nb0 = 5;
    457  const int nb1 = 4;
    458 
    459  const __m128i rounding0 =
    460      round_for_shift(SGRPROJ_SGR_BITS + nb0 - SGRPROJ_RST_BITS);
    461  const __m128i rounding1 =
    462      round_for_shift(SGRPROJ_SGR_BITS + nb1 - SGRPROJ_RST_BITS);
    463 
    464  const uint8_t *dgd_real =
    465      highbd ? (const uint8_t *)CONVERT_TO_SHORTPTR(dgd8) : dgd8;
    466 
    467  for (int i = 0; i < height; ++i) {
    468    if (!(i & 1)) {  // even row
    469      for (int j = 0; j < width; j += 4) {
    470        const __m128i a =
    471            cross_sum_fast_even_row(A + i * buf_stride + j, buf_stride);
    472        const __m128i b =
    473            cross_sum_fast_even_row(B + i * buf_stride + j, buf_stride);
    474        const __m128i raw =
    475            xx_loadl_64(dgd_real + ((i * dgd_stride + j) << highbd));
    476        const __m128i src =
    477            highbd ? _mm_cvtepu16_epi32(raw) : _mm_cvtepu8_epi32(raw);
    478 
    479        __m128i v = _mm_add_epi32(_mm_madd_epi16(a, src), b);
    480        __m128i w = _mm_srai_epi32(_mm_add_epi32(v, rounding0),
    481                                   SGRPROJ_SGR_BITS + nb0 - SGRPROJ_RST_BITS);
    482 
    483        xx_storeu_128(dst + i * dst_stride + j, w);
    484      }
    485    } else {  // odd row
    486      for (int j = 0; j < width; j += 4) {
    487        const __m128i a = cross_sum_fast_odd_row(A + i * buf_stride + j);
    488        const __m128i b = cross_sum_fast_odd_row(B + i * buf_stride + j);
    489        const __m128i raw =
    490            xx_loadl_64(dgd_real + ((i * dgd_stride + j) << highbd));
    491        const __m128i src =
    492            highbd ? _mm_cvtepu16_epi32(raw) : _mm_cvtepu8_epi32(raw);
    493 
    494        __m128i v = _mm_add_epi32(_mm_madd_epi16(a, src), b);
    495        __m128i w = _mm_srai_epi32(_mm_add_epi32(v, rounding1),
    496                                   SGRPROJ_SGR_BITS + nb1 - SGRPROJ_RST_BITS);
    497 
    498        xx_storeu_128(dst + i * dst_stride + j, w);
    499      }
    500    }
    501  }
    502 }
    503 
    504 int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width,
    505                                      int height, int dgd_stride, int32_t *flt0,
    506                                      int32_t *flt1, int flt_stride,
    507                                      int sgr_params_idx, int bit_depth,
    508                                      int highbd) {
    509  int32_t *buf = (int32_t *)aom_memalign(
    510      16, 4 * sizeof(*buf) * RESTORATION_PROC_UNIT_PELS);
    511  if (!buf) return -1;
    512  memset(buf, 0, 4 * sizeof(*buf) * RESTORATION_PROC_UNIT_PELS);
    513 
    514  const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
    515  const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
    516 
    517  // Adjusting the stride of A and B here appears to avoid bad cache effects,
    518  // leading to a significant speed improvement.
    519  // We also align the stride to a multiple of 16 bytes for efficiency.
    520  int buf_stride = ((width_ext + 3) & ~3) + 16;
    521 
    522  // The "tl" pointers point at the top-left of the initialised data for the
    523  // array. Adding 3 here ensures that column 1 is 16-byte aligned.
    524  int32_t *Atl = buf + 0 * RESTORATION_PROC_UNIT_PELS + 3;
    525  int32_t *Btl = buf + 1 * RESTORATION_PROC_UNIT_PELS + 3;
    526  int32_t *Ctl = buf + 2 * RESTORATION_PROC_UNIT_PELS + 3;
    527  int32_t *Dtl = buf + 3 * RESTORATION_PROC_UNIT_PELS + 3;
    528 
    529  // The "0" pointers are (- SGRPROJ_BORDER_VERT, -SGRPROJ_BORDER_HORZ). Note
    530  // there's a zero row and column in A, B (integral images), so we move down
    531  // and right one for them.
    532  const int buf_diag_border =
    533      SGRPROJ_BORDER_HORZ + buf_stride * SGRPROJ_BORDER_VERT;
    534 
    535  int32_t *A0 = Atl + 1 + buf_stride;
    536  int32_t *B0 = Btl + 1 + buf_stride;
    537  int32_t *C0 = Ctl + 1 + buf_stride;
    538  int32_t *D0 = Dtl + 1 + buf_stride;
    539 
    540  // Finally, A, B, C, D point at position (0, 0).
    541  int32_t *A = A0 + buf_diag_border;
    542  int32_t *B = B0 + buf_diag_border;
    543  int32_t *C = C0 + buf_diag_border;
    544  int32_t *D = D0 + buf_diag_border;
    545 
    546  const int dgd_diag_border =
    547      SGRPROJ_BORDER_HORZ + dgd_stride * SGRPROJ_BORDER_VERT;
    548  const uint8_t *dgd0 = dgd8 - dgd_diag_border;
    549 
    550  // Generate integral images from the input. C will contain sums of squares; D
    551  // will contain just sums
    552  if (highbd)
    553    integral_images_highbd(CONVERT_TO_SHORTPTR(dgd0), dgd_stride, width_ext,
    554                           height_ext, Ctl, Dtl, buf_stride);
    555  else
    556    integral_images(dgd0, dgd_stride, width_ext, height_ext, Ctl, Dtl,
    557                    buf_stride);
    558 
    559  const sgr_params_type *const params = &av1_sgr_params[sgr_params_idx];
    560  // Write to flt0 and flt1
    561  // If params->r == 0 we skip the corresponding filter. We only allow one of
    562  // the radii to be 0, as having both equal to 0 would be equivalent to
    563  // skipping SGR entirely.
    564  assert(!(params->r[0] == 0 && params->r[1] == 0));
    565  assert(params->r[0] < AOMMIN(SGRPROJ_BORDER_VERT, SGRPROJ_BORDER_HORZ));
    566  assert(params->r[1] < AOMMIN(SGRPROJ_BORDER_VERT, SGRPROJ_BORDER_HORZ));
    567 
    568  if (params->r[0] > 0) {
    569    calc_ab_fast(A, B, C, D, width, height, buf_stride, bit_depth,
    570                 sgr_params_idx, 0);
    571    final_filter_fast(flt0, flt_stride, A, B, buf_stride, dgd8, dgd_stride,
    572                      width, height, highbd);
    573  }
    574 
    575  if (params->r[1] > 0) {
    576    calc_ab(A, B, C, D, width, height, buf_stride, bit_depth, sgr_params_idx,
    577            1);
    578    final_filter(flt1, flt_stride, A, B, buf_stride, dgd8, dgd_stride, width,
    579                 height, highbd);
    580  }
    581  aom_free(buf);
    582  return 0;
    583 }
    584 
    585 int av1_apply_selfguided_restoration_sse4_1(const uint8_t *dat8, int width,
    586                                            int height, int stride, int eps,
    587                                            const int *xqd, uint8_t *dst8,
    588                                            int dst_stride, int32_t *tmpbuf,
    589                                            int bit_depth, int highbd) {
    590  int32_t *flt0 = tmpbuf;
    591  int32_t *flt1 = flt0 + RESTORATION_UNITPELS_MAX;
    592  assert(width * height <= RESTORATION_UNITPELS_MAX);
    593  const int ret = av1_selfguided_restoration_sse4_1(
    594      dat8, width, height, stride, flt0, flt1, width, eps, bit_depth, highbd);
    595  if (ret != 0) return ret;
    596  const sgr_params_type *const params = &av1_sgr_params[eps];
    597  int xq[2];
    598  av1_decode_xq(xqd, xq, params);
    599 
    600  __m128i xq0 = _mm_set1_epi32(xq[0]);
    601  __m128i xq1 = _mm_set1_epi32(xq[1]);
    602 
    603  for (int i = 0; i < height; ++i) {
    604    // Calculate output in batches of 8 pixels
    605    for (int j = 0; j < width; j += 8) {
    606      const int k = i * width + j;
    607      const int m = i * dst_stride + j;
    608 
    609      const uint8_t *dat8ij = dat8 + i * stride + j;
    610      __m128i src;
    611      if (highbd) {
    612        src = xx_loadu_128(CONVERT_TO_SHORTPTR(dat8ij));
    613      } else {
    614        src = _mm_cvtepu8_epi16(xx_loadl_64(dat8ij));
    615      }
    616 
    617      const __m128i u = _mm_slli_epi16(src, SGRPROJ_RST_BITS);
    618      const __m128i u_0 = _mm_cvtepu16_epi32(u);
    619      const __m128i u_1 = _mm_cvtepu16_epi32(_mm_srli_si128(u, 8));
    620 
    621      __m128i v_0 = _mm_slli_epi32(u_0, SGRPROJ_PRJ_BITS);
    622      __m128i v_1 = _mm_slli_epi32(u_1, SGRPROJ_PRJ_BITS);
    623 
    624      if (params->r[0] > 0) {
    625        const __m128i f1_0 = _mm_sub_epi32(xx_loadu_128(&flt0[k]), u_0);
    626        v_0 = _mm_add_epi32(v_0, _mm_mullo_epi32(xq0, f1_0));
    627 
    628        const __m128i f1_1 = _mm_sub_epi32(xx_loadu_128(&flt0[k + 4]), u_1);
    629        v_1 = _mm_add_epi32(v_1, _mm_mullo_epi32(xq0, f1_1));
    630      }
    631 
    632      if (params->r[1] > 0) {
    633        const __m128i f2_0 = _mm_sub_epi32(xx_loadu_128(&flt1[k]), u_0);
    634        v_0 = _mm_add_epi32(v_0, _mm_mullo_epi32(xq1, f2_0));
    635 
    636        const __m128i f2_1 = _mm_sub_epi32(xx_loadu_128(&flt1[k + 4]), u_1);
    637        v_1 = _mm_add_epi32(v_1, _mm_mullo_epi32(xq1, f2_1));
    638      }
    639 
    640      const __m128i rounding =
    641          round_for_shift(SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
    642      const __m128i w_0 = _mm_srai_epi32(_mm_add_epi32(v_0, rounding),
    643                                         SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
    644      const __m128i w_1 = _mm_srai_epi32(_mm_add_epi32(v_1, rounding),
    645                                         SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
    646 
    647      if (highbd) {
    648        // Pack into 16 bits and clamp to [0, 2^bit_depth)
    649        const __m128i tmp = _mm_packus_epi32(w_0, w_1);
    650        const __m128i max = _mm_set1_epi16((1 << bit_depth) - 1);
    651        const __m128i res = _mm_min_epi16(tmp, max);
    652        xx_storeu_128(CONVERT_TO_SHORTPTR(dst8 + m), res);
    653      } else {
    654        // Pack into 8 bits and clamp to [0, 256)
    655        const __m128i tmp = _mm_packs_epi32(w_0, w_1);
    656        const __m128i res = _mm_packus_epi16(tmp, tmp /* "don't care" value */);
    657        xx_storel_64(dst8 + m, res);
    658      }
    659    }
    660  }
    661  return 0;
    662 }