tor-browser

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

enc.c (27006B)


      1 // Copyright 2011 Google Inc. All Rights Reserved.
      2 //
      3 // Use of this source code is governed by a BSD-style license
      4 // that can be found in the COPYING file in the root of the source
      5 // tree. An additional intellectual property rights grant can be found
      6 // in the file PATENTS. All contributing project authors may
      7 // be found in the AUTHORS file in the root of the source tree.
      8 // -----------------------------------------------------------------------------
      9 //
     10 // Speed-critical encoding functions.
     11 //
     12 // Author: Skal (pascal.massimino@gmail.com)
     13 
     14 #include <assert.h>
     15 #include <stdlib.h>  // for abs()
     16 #include <string.h>
     17 
     18 #include "src/dsp/cpu.h"
     19 #include "src/dsp/dsp.h"
     20 #include "src/enc/vp8i_enc.h"
     21 #include "src/utils/utils.h"
     22 #include "src/webp/types.h"
     23 
     24 static WEBP_INLINE uint8_t clip_8b(int v) {
     25  return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
     26 }
     27 
     28 #if !WEBP_NEON_OMIT_C_CODE
     29 static WEBP_INLINE int clip_max(int v, int max) {
     30  return (v > max) ? max : v;
     31 }
     32 #endif  // !WEBP_NEON_OMIT_C_CODE
     33 
     34 //------------------------------------------------------------------------------
     35 // Compute susceptibility based on DCT-coeff histograms:
     36 // the higher, the "easier" the macroblock is to compress.
     37 
     38 const int VP8DspScan[16 + 4 + 4] = {
     39  // Luma
     40  0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
     41  0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
     42  0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
     43  0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS,
     44 
     45  0 + 0 * BPS,   4 + 0 * BPS, 0 + 4 * BPS,  4 + 4 * BPS,    // U
     46  8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
     47 };
     48 
     49 // general-purpose util function
     50 void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1],
     51                         VP8Histogram* const histo) {
     52  int max_value = 0, last_non_zero = 1;
     53  int k;
     54  for (k = 0; k <= MAX_COEFF_THRESH; ++k) {
     55    const int value = distribution[k];
     56    if (value > 0) {
     57      if (value > max_value) max_value = value;
     58      last_non_zero = k;
     59    }
     60  }
     61  histo->max_value = max_value;
     62  histo->last_non_zero = last_non_zero;
     63 }
     64 
     65 #if !WEBP_NEON_OMIT_C_CODE
     66 static void CollectHistogram_C(const uint8_t* WEBP_RESTRICT ref,
     67                               const uint8_t* WEBP_RESTRICT pred,
     68                               int start_block, int end_block,
     69                               VP8Histogram* WEBP_RESTRICT const histo) {
     70  int j;
     71  int distribution[MAX_COEFF_THRESH + 1] = { 0 };
     72  for (j = start_block; j < end_block; ++j) {
     73    int k;
     74    int16_t out[16];
     75 
     76    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
     77 
     78    // Convert coefficients to bin.
     79    for (k = 0; k < 16; ++k) {
     80      const int v = abs(out[k]) >> 3;
     81      const int clipped_value = clip_max(v, MAX_COEFF_THRESH);
     82      ++distribution[clipped_value];
     83    }
     84  }
     85  VP8SetHistogramData(distribution, histo);
     86 }
     87 #endif  // !WEBP_NEON_OMIT_C_CODE
     88 
     89 //------------------------------------------------------------------------------
     90 // run-time tables (~4k)
     91 
     92 static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
     93 
     94 // We declare this variable 'volatile' to prevent instruction reordering
     95 // and make sure it's set to true _last_ (so as to be thread-safe)
     96 static volatile int tables_ok = 0;
     97 
     98 static WEBP_TSAN_IGNORE_FUNCTION void InitTables(void) {
     99  if (!tables_ok) {
    100    int i;
    101    for (i = -255; i <= 255 + 255; ++i) {
    102      clip1[255 + i] = clip_8b(i);
    103    }
    104    tables_ok = 1;
    105  }
    106 }
    107 
    108 
    109 //------------------------------------------------------------------------------
    110 // Transforms (Paragraph 14.4)
    111 
    112 #if !WEBP_NEON_OMIT_C_CODE
    113 
    114 #define STORE(x, y, v) \
    115  dst[(x) + (y) * BPS] = clip_8b(ref[(x) + (y) * BPS] + ((v) >> 3))
    116 
    117 static WEBP_INLINE void ITransformOne(const uint8_t* WEBP_RESTRICT ref,
    118                                      const int16_t* WEBP_RESTRICT in,
    119                                      uint8_t* WEBP_RESTRICT dst) {
    120  int C[4 * 4], *tmp;
    121  int i;
    122  tmp = C;
    123  for (i = 0; i < 4; ++i) {    // vertical pass
    124    const int a = in[0] + in[8];
    125    const int b = in[0] - in[8];
    126    const int c =
    127        WEBP_TRANSFORM_AC3_MUL2(in[4]) - WEBP_TRANSFORM_AC3_MUL1(in[12]);
    128    const int d =
    129        WEBP_TRANSFORM_AC3_MUL1(in[4]) + WEBP_TRANSFORM_AC3_MUL2(in[12]);
    130    tmp[0] = a + d;
    131    tmp[1] = b + c;
    132    tmp[2] = b - c;
    133    tmp[3] = a - d;
    134    tmp += 4;
    135    in++;
    136  }
    137 
    138  tmp = C;
    139  for (i = 0; i < 4; ++i) {    // horizontal pass
    140    const int dc = tmp[0] + 4;
    141    const int a = dc + tmp[8];
    142    const int b = dc - tmp[8];
    143    const int c =
    144        WEBP_TRANSFORM_AC3_MUL2(tmp[4]) - WEBP_TRANSFORM_AC3_MUL1(tmp[12]);
    145    const int d =
    146        WEBP_TRANSFORM_AC3_MUL1(tmp[4]) + WEBP_TRANSFORM_AC3_MUL2(tmp[12]);
    147    STORE(0, i, a + d);
    148    STORE(1, i, b + c);
    149    STORE(2, i, b - c);
    150    STORE(3, i, a - d);
    151    tmp++;
    152  }
    153 }
    154 
    155 static void ITransform_C(const uint8_t* WEBP_RESTRICT ref,
    156                         const int16_t* WEBP_RESTRICT in,
    157                         uint8_t* WEBP_RESTRICT dst,
    158                         int do_two) {
    159  ITransformOne(ref, in, dst);
    160  if (do_two) {
    161    ITransformOne(ref + 4, in + 16, dst + 4);
    162  }
    163 }
    164 
    165 static void FTransform_C(const uint8_t* WEBP_RESTRICT src,
    166                         const uint8_t* WEBP_RESTRICT ref,
    167                         int16_t* WEBP_RESTRICT out) {
    168  int i;
    169  int tmp[16];
    170  for (i = 0; i < 4; ++i, src += BPS, ref += BPS) {
    171    const int d0 = src[0] - ref[0];   // 9bit dynamic range ([-255,255])
    172    const int d1 = src[1] - ref[1];
    173    const int d2 = src[2] - ref[2];
    174    const int d3 = src[3] - ref[3];
    175    const int a0 = (d0 + d3);         // 10b                      [-510,510]
    176    const int a1 = (d1 + d2);
    177    const int a2 = (d1 - d2);
    178    const int a3 = (d0 - d3);
    179    tmp[0 + i * 4] = (a0 + a1) * 8;   // 14b                      [-8160,8160]
    180    tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9;      // [-7536,7542]
    181    tmp[2 + i * 4] = (a0 - a1) * 8;
    182    tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 +  937) >> 9;
    183  }
    184  for (i = 0; i < 4; ++i) {
    185    const int a0 = (tmp[0 + i] + tmp[12 + i]);  // 15b
    186    const int a1 = (tmp[4 + i] + tmp[ 8 + i]);
    187    const int a2 = (tmp[4 + i] - tmp[ 8 + i]);
    188    const int a3 = (tmp[0 + i] - tmp[12 + i]);
    189    out[0 + i] = (a0 + a1 + 7) >> 4;            // 12b
    190    out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0);
    191    out[8 + i] = (a0 - a1 + 7) >> 4;
    192    out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16);
    193  }
    194 }
    195 #endif  // !WEBP_NEON_OMIT_C_CODE
    196 
    197 static void FTransform2_C(const uint8_t* WEBP_RESTRICT src,
    198                          const uint8_t* WEBP_RESTRICT ref,
    199                          int16_t* WEBP_RESTRICT out) {
    200  VP8FTransform(src, ref, out);
    201  VP8FTransform(src + 4, ref + 4, out + 16);
    202 }
    203 
    204 #if !WEBP_NEON_OMIT_C_CODE
    205 static void FTransformWHT_C(const int16_t* WEBP_RESTRICT in,
    206                            int16_t* WEBP_RESTRICT out) {
    207  // input is 12b signed
    208  int32_t tmp[16];
    209  int i;
    210  for (i = 0; i < 4; ++i, in += 64) {
    211    const int a0 = (in[0 * 16] + in[2 * 16]);  // 13b
    212    const int a1 = (in[1 * 16] + in[3 * 16]);
    213    const int a2 = (in[1 * 16] - in[3 * 16]);
    214    const int a3 = (in[0 * 16] - in[2 * 16]);
    215    tmp[0 + i * 4] = a0 + a1;   // 14b
    216    tmp[1 + i * 4] = a3 + a2;
    217    tmp[2 + i * 4] = a3 - a2;
    218    tmp[3 + i * 4] = a0 - a1;
    219  }
    220  for (i = 0; i < 4; ++i) {
    221    const int a0 = (tmp[0 + i] + tmp[8 + i]);  // 15b
    222    const int a1 = (tmp[4 + i] + tmp[12+ i]);
    223    const int a2 = (tmp[4 + i] - tmp[12+ i]);
    224    const int a3 = (tmp[0 + i] - tmp[8 + i]);
    225    const int b0 = a0 + a1;    // 16b
    226    const int b1 = a3 + a2;
    227    const int b2 = a3 - a2;
    228    const int b3 = a0 - a1;
    229    out[ 0 + i] = b0 >> 1;     // 15b
    230    out[ 4 + i] = b1 >> 1;
    231    out[ 8 + i] = b2 >> 1;
    232    out[12 + i] = b3 >> 1;
    233  }
    234 }
    235 #endif  // !WEBP_NEON_OMIT_C_CODE
    236 
    237 #undef STORE
    238 
    239 //------------------------------------------------------------------------------
    240 // Intra predictions
    241 
    242 static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
    243  int j;
    244  for (j = 0; j < size; ++j) {
    245    memset(dst + j * BPS, value, size);
    246  }
    247 }
    248 
    249 static WEBP_INLINE void VerticalPred(uint8_t* WEBP_RESTRICT dst,
    250                                     const uint8_t* WEBP_RESTRICT top,
    251                                     int size) {
    252  int j;
    253  if (top != NULL) {
    254    for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
    255  } else {
    256    Fill(dst, 127, size);
    257  }
    258 }
    259 
    260 static WEBP_INLINE void HorizontalPred(uint8_t* WEBP_RESTRICT dst,
    261                                       const uint8_t* WEBP_RESTRICT left,
    262                                       int size) {
    263  if (left != NULL) {
    264    int j;
    265    for (j = 0; j < size; ++j) {
    266      memset(dst + j * BPS, left[j], size);
    267    }
    268  } else {
    269    Fill(dst, 129, size);
    270  }
    271 }
    272 
    273 static WEBP_INLINE void TrueMotion(uint8_t* WEBP_RESTRICT dst,
    274                                   const uint8_t* WEBP_RESTRICT left,
    275                                   const uint8_t* WEBP_RESTRICT top, int size) {
    276  int y;
    277  if (left != NULL) {
    278    if (top != NULL) {
    279      const uint8_t* const clip = clip1 + 255 - left[-1];
    280      for (y = 0; y < size; ++y) {
    281        const uint8_t* const clip_table = clip + left[y];
    282        int x;
    283        for (x = 0; x < size; ++x) {
    284          dst[x] = clip_table[top[x]];
    285        }
    286        dst += BPS;
    287      }
    288    } else {
    289      HorizontalPred(dst, left, size);
    290    }
    291  } else {
    292    // true motion without left samples (hence: with default 129 value)
    293    // is equivalent to VE prediction where you just copy the top samples.
    294    // Note that if top samples are not available, the default value is
    295    // then 129, and not 127 as in the VerticalPred case.
    296    if (top != NULL) {
    297      VerticalPred(dst, top, size);
    298    } else {
    299      Fill(dst, 129, size);
    300    }
    301  }
    302 }
    303 
    304 static WEBP_INLINE void DCMode(uint8_t* WEBP_RESTRICT dst,
    305                               const uint8_t* WEBP_RESTRICT left,
    306                               const uint8_t* WEBP_RESTRICT top,
    307                               int size, int round, int shift) {
    308  int DC = 0;
    309  int j;
    310  if (top != NULL) {
    311    for (j = 0; j < size; ++j) DC += top[j];
    312    if (left != NULL) {   // top and left present
    313      for (j = 0; j < size; ++j) DC += left[j];
    314    } else {      // top, but no left
    315      DC += DC;
    316    }
    317    DC = (DC + round) >> shift;
    318  } else if (left != NULL) {   // left but no top
    319    for (j = 0; j < size; ++j) DC += left[j];
    320    DC += DC;
    321    DC = (DC + round) >> shift;
    322  } else {   // no top, no left, nothing.
    323    DC = 0x80;
    324  }
    325  Fill(dst, DC, size);
    326 }
    327 
    328 //------------------------------------------------------------------------------
    329 // Chroma 8x8 prediction (paragraph 12.2)
    330 
    331 static void IntraChromaPreds_C(uint8_t* WEBP_RESTRICT dst,
    332                               const uint8_t* WEBP_RESTRICT left,
    333                               const uint8_t* WEBP_RESTRICT top) {
    334  // U block
    335  DCMode(C8DC8 + dst, left, top, 8, 8, 4);
    336  VerticalPred(C8VE8 + dst, top, 8);
    337  HorizontalPred(C8HE8 + dst, left, 8);
    338  TrueMotion(C8TM8 + dst, left, top, 8);
    339  // V block
    340  dst += 8;
    341  if (top != NULL) top += 8;
    342  if (left != NULL) left += 16;
    343  DCMode(C8DC8 + dst, left, top, 8, 8, 4);
    344  VerticalPred(C8VE8 + dst, top, 8);
    345  HorizontalPred(C8HE8 + dst, left, 8);
    346  TrueMotion(C8TM8 + dst, left, top, 8);
    347 }
    348 
    349 //------------------------------------------------------------------------------
    350 // luma 16x16 prediction (paragraph 12.3)
    351 
    352 #if !WEBP_NEON_OMIT_C_CODE || !WEBP_AARCH64
    353 static void Intra16Preds_C(uint8_t* WEBP_RESTRICT dst,
    354                           const uint8_t* WEBP_RESTRICT left,
    355                           const uint8_t* WEBP_RESTRICT top) {
    356  DCMode(I16DC16 + dst, left, top, 16, 16, 5);
    357  VerticalPred(I16VE16 + dst, top, 16);
    358  HorizontalPred(I16HE16 + dst, left, 16);
    359  TrueMotion(I16TM16 + dst, left, top, 16);
    360 }
    361 #endif  // !WEBP_NEON_OMIT_C_CODE || !WEBP_AARCH64
    362 
    363 //------------------------------------------------------------------------------
    364 // luma 4x4 prediction
    365 
    366 #if !WEBP_NEON_OMIT_C_CODE || !WEBP_AARCH64 || BPS != 32
    367 
    368 #define DST(x, y) dst[(x) + (y) * BPS]
    369 #define AVG3(a, b, c) ((uint8_t)(((a) + 2 * (b) + (c) + 2) >> 2))
    370 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
    371 
    372 // vertical
    373 static void VE4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    374  const uint8_t vals[4] = {
    375    AVG3(top[-1], top[0], top[1]),
    376    AVG3(top[ 0], top[1], top[2]),
    377    AVG3(top[ 1], top[2], top[3]),
    378    AVG3(top[ 2], top[3], top[4])
    379  };
    380  int i;
    381  for (i = 0; i < 4; ++i) {
    382    memcpy(dst + i * BPS, vals, 4);
    383  }
    384 }
    385 
    386 // horizontal
    387 static void HE4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    388  const int X = top[-1];
    389  const int I = top[-2];
    390  const int J = top[-3];
    391  const int K = top[-4];
    392  const int L = top[-5];
    393  WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J));
    394  WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K));
    395  WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L));
    396  WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));
    397 }
    398 
    399 static void DC4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    400  uint32_t dc = 4;
    401  int i;
    402  for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
    403  Fill(dst, dc >> 3, 4);
    404 }
    405 
    406 static void RD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    407  const int X = top[-1];
    408  const int I = top[-2];
    409  const int J = top[-3];
    410  const int K = top[-4];
    411  const int L = top[-5];
    412  const int A = top[0];
    413  const int B = top[1];
    414  const int C = top[2];
    415  const int D = top[3];
    416  DST(0, 3)                                     = AVG3(J, K, L);
    417  DST(0, 2) = DST(1, 3)                         = AVG3(I, J, K);
    418  DST(0, 1) = DST(1, 2) = DST(2, 3)             = AVG3(X, I, J);
    419  DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I);
    420  DST(1, 0) = DST(2, 1) = DST(3, 2)             = AVG3(B, A, X);
    421  DST(2, 0) = DST(3, 1)                         = AVG3(C, B, A);
    422  DST(3, 0)                                     = AVG3(D, C, B);
    423 }
    424 
    425 static void LD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    426  const int A = top[0];
    427  const int B = top[1];
    428  const int C = top[2];
    429  const int D = top[3];
    430  const int E = top[4];
    431  const int F = top[5];
    432  const int G = top[6];
    433  const int H = top[7];
    434  DST(0, 0)                                     = AVG3(A, B, C);
    435  DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
    436  DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
    437  DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
    438  DST(3, 1) = DST(2, 2) = DST(1, 3)             = AVG3(E, F, G);
    439  DST(3, 2) = DST(2, 3)                         = AVG3(F, G, H);
    440  DST(3, 3)                                     = AVG3(G, H, H);
    441 }
    442 
    443 static void VR4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    444  const int X = top[-1];
    445  const int I = top[-2];
    446  const int J = top[-3];
    447  const int K = top[-4];
    448  const int A = top[0];
    449  const int B = top[1];
    450  const int C = top[2];
    451  const int D = top[3];
    452  DST(0, 0) = DST(1, 2) = AVG2(X, A);
    453  DST(1, 0) = DST(2, 2) = AVG2(A, B);
    454  DST(2, 0) = DST(3, 2) = AVG2(B, C);
    455  DST(3, 0)             = AVG2(C, D);
    456 
    457  DST(0, 3) =             AVG3(K, J, I);
    458  DST(0, 2) =             AVG3(J, I, X);
    459  DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
    460  DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
    461  DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
    462  DST(3, 1) =             AVG3(B, C, D);
    463 }
    464 
    465 static void VL4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    466  const int A = top[0];
    467  const int B = top[1];
    468  const int C = top[2];
    469  const int D = top[3];
    470  const int E = top[4];
    471  const int F = top[5];
    472  const int G = top[6];
    473  const int H = top[7];
    474  DST(0, 0) =             AVG2(A, B);
    475  DST(1, 0) = DST(0, 2) = AVG2(B, C);
    476  DST(2, 0) = DST(1, 2) = AVG2(C, D);
    477  DST(3, 0) = DST(2, 2) = AVG2(D, E);
    478 
    479  DST(0, 1) =             AVG3(A, B, C);
    480  DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
    481  DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
    482  DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
    483              DST(3, 2) = AVG3(E, F, G);
    484              DST(3, 3) = AVG3(F, G, H);
    485 }
    486 
    487 static void HU4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    488  const int I = top[-2];
    489  const int J = top[-3];
    490  const int K = top[-4];
    491  const int L = top[-5];
    492  DST(0, 0) =             AVG2(I, J);
    493  DST(2, 0) = DST(0, 1) = AVG2(J, K);
    494  DST(2, 1) = DST(0, 2) = AVG2(K, L);
    495  DST(1, 0) =             AVG3(I, J, K);
    496  DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
    497  DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
    498  DST(3, 2) = DST(2, 2) =
    499  DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
    500 }
    501 
    502 static void HD4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    503  const int X = top[-1];
    504  const int I = top[-2];
    505  const int J = top[-3];
    506  const int K = top[-4];
    507  const int L = top[-5];
    508  const int A = top[0];
    509  const int B = top[1];
    510  const int C = top[2];
    511 
    512  DST(0, 0) = DST(2, 1) = AVG2(I, X);
    513  DST(0, 1) = DST(2, 2) = AVG2(J, I);
    514  DST(0, 2) = DST(2, 3) = AVG2(K, J);
    515  DST(0, 3)             = AVG2(L, K);
    516 
    517  DST(3, 0)             = AVG3(A, B, C);
    518  DST(2, 0)             = AVG3(X, A, B);
    519  DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
    520  DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
    521  DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
    522  DST(1, 3)             = AVG3(L, K, J);
    523 }
    524 
    525 static void TM4(uint8_t* WEBP_RESTRICT dst, const uint8_t* WEBP_RESTRICT top) {
    526  int x, y;
    527  const uint8_t* const clip = clip1 + 255 - top[-1];
    528  for (y = 0; y < 4; ++y) {
    529    const uint8_t* const clip_table = clip + top[-2 - y];
    530    for (x = 0; x < 4; ++x) {
    531      dst[x] = clip_table[top[x]];
    532    }
    533    dst += BPS;
    534  }
    535 }
    536 
    537 #undef DST
    538 #undef AVG3
    539 #undef AVG2
    540 
    541 // Left samples are top[-5 .. -2], top_left is top[-1], top are
    542 // located at top[0..3], and top right is top[4..7]
    543 static void Intra4Preds_C(uint8_t* WEBP_RESTRICT dst,
    544                          const uint8_t* WEBP_RESTRICT top) {
    545  DC4(I4DC4 + dst, top);
    546  TM4(I4TM4 + dst, top);
    547  VE4(I4VE4 + dst, top);
    548  HE4(I4HE4 + dst, top);
    549  RD4(I4RD4 + dst, top);
    550  VR4(I4VR4 + dst, top);
    551  LD4(I4LD4 + dst, top);
    552  VL4(I4VL4 + dst, top);
    553  HD4(I4HD4 + dst, top);
    554  HU4(I4HU4 + dst, top);
    555 }
    556 
    557 #endif  // !WEBP_NEON_OMIT_C_CODE || !WEBP_AARCH64 || BPS != 32
    558 
    559 //------------------------------------------------------------------------------
    560 // Metric
    561 
    562 #if !WEBP_NEON_OMIT_C_CODE
    563 static WEBP_INLINE int GetSSE(const uint8_t* WEBP_RESTRICT a,
    564                              const uint8_t* WEBP_RESTRICT b,
    565                              int w, int h) {
    566  int count = 0;
    567  int y, x;
    568  for (y = 0; y < h; ++y) {
    569    for (x = 0; x < w; ++x) {
    570      const int diff = (int)a[x] - b[x];
    571      count += diff * diff;
    572    }
    573    a += BPS;
    574    b += BPS;
    575  }
    576  return count;
    577 }
    578 
    579 static int SSE16x16_C(const uint8_t* WEBP_RESTRICT a,
    580                      const uint8_t* WEBP_RESTRICT b) {
    581  return GetSSE(a, b, 16, 16);
    582 }
    583 static int SSE16x8_C(const uint8_t* WEBP_RESTRICT a,
    584                     const uint8_t* WEBP_RESTRICT b) {
    585  return GetSSE(a, b, 16, 8);
    586 }
    587 static int SSE8x8_C(const uint8_t* WEBP_RESTRICT a,
    588                    const uint8_t* WEBP_RESTRICT b) {
    589  return GetSSE(a, b, 8, 8);
    590 }
    591 static int SSE4x4_C(const uint8_t* WEBP_RESTRICT a,
    592                    const uint8_t* WEBP_RESTRICT b) {
    593  return GetSSE(a, b, 4, 4);
    594 }
    595 #endif  // !WEBP_NEON_OMIT_C_CODE
    596 
    597 static void Mean16x4_C(const uint8_t* WEBP_RESTRICT ref, uint32_t dc[4]) {
    598  int k, x, y;
    599  for (k = 0; k < 4; ++k) {
    600    uint32_t avg = 0;
    601    for (y = 0; y < 4; ++y) {
    602      for (x = 0; x < 4; ++x) {
    603        avg += ref[x + y * BPS];
    604      }
    605    }
    606    dc[k] = avg;
    607    ref += 4;   // go to next 4x4 block.
    608  }
    609 }
    610 
    611 //------------------------------------------------------------------------------
    612 // Texture distortion
    613 //
    614 // We try to match the spectral content (weighted) between source and
    615 // reconstructed samples.
    616 
    617 #if !WEBP_NEON_OMIT_C_CODE
    618 // Hadamard transform
    619 // Returns the weighted sum of the absolute value of transformed coefficients.
    620 // w[] contains a row-major 4 by 4 symmetric matrix.
    621 static int TTransform(const uint8_t* WEBP_RESTRICT in,
    622                      const uint16_t* WEBP_RESTRICT w) {
    623  int sum = 0;
    624  int tmp[16];
    625  int i;
    626  // horizontal pass
    627  for (i = 0; i < 4; ++i, in += BPS) {
    628    const int a0 = in[0] + in[2];
    629    const int a1 = in[1] + in[3];
    630    const int a2 = in[1] - in[3];
    631    const int a3 = in[0] - in[2];
    632    tmp[0 + i * 4] = a0 + a1;
    633    tmp[1 + i * 4] = a3 + a2;
    634    tmp[2 + i * 4] = a3 - a2;
    635    tmp[3 + i * 4] = a0 - a1;
    636  }
    637  // vertical pass
    638  for (i = 0; i < 4; ++i, ++w) {
    639    const int a0 = tmp[0 + i] + tmp[8 + i];
    640    const int a1 = tmp[4 + i] + tmp[12+ i];
    641    const int a2 = tmp[4 + i] - tmp[12+ i];
    642    const int a3 = tmp[0 + i] - tmp[8 + i];
    643    const int b0 = a0 + a1;
    644    const int b1 = a3 + a2;
    645    const int b2 = a3 - a2;
    646    const int b3 = a0 - a1;
    647 
    648    sum += w[ 0] * abs(b0);
    649    sum += w[ 4] * abs(b1);
    650    sum += w[ 8] * abs(b2);
    651    sum += w[12] * abs(b3);
    652  }
    653  return sum;
    654 }
    655 
    656 static int Disto4x4_C(const uint8_t* WEBP_RESTRICT const a,
    657                      const uint8_t* WEBP_RESTRICT const b,
    658                      const uint16_t* WEBP_RESTRICT const w) {
    659  const int sum1 = TTransform(a, w);
    660  const int sum2 = TTransform(b, w);
    661  return abs(sum2 - sum1) >> 5;
    662 }
    663 
    664 static int Disto16x16_C(const uint8_t* WEBP_RESTRICT const a,
    665                        const uint8_t* WEBP_RESTRICT const b,
    666                        const uint16_t* WEBP_RESTRICT const w) {
    667  int D = 0;
    668  int x, y;
    669  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
    670    for (x = 0; x < 16; x += 4) {
    671      D += Disto4x4_C(a + x + y, b + x + y, w);
    672    }
    673  }
    674  return D;
    675 }
    676 #endif  // !WEBP_NEON_OMIT_C_CODE
    677 
    678 //------------------------------------------------------------------------------
    679 // Quantization
    680 //
    681 
    682 #if !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC
    683 static const uint8_t kZigzag[16] = {
    684  0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
    685 };
    686 
    687 // Simple quantization
    688 static int QuantizeBlock_C(int16_t in[16], int16_t out[16],
    689                           const VP8Matrix* WEBP_RESTRICT const mtx) {
    690  int last = -1;
    691  int n;
    692  for (n = 0; n < 16; ++n) {
    693    const int j = kZigzag[n];
    694    const int sign = (in[j] < 0);
    695    const uint32_t coeff = (sign ? -in[j] : in[j]) + mtx->sharpen[j];
    696    if (coeff > mtx->zthresh[j]) {
    697      const uint32_t Q = mtx->q[j];
    698      const uint32_t iQ = mtx->iq[j];
    699      const uint32_t B = mtx->bias[j];
    700      int level = QUANTDIV(coeff, iQ, B);
    701      if (level > MAX_LEVEL) level = MAX_LEVEL;
    702      if (sign) level = -level;
    703      in[j] = level * (int)Q;
    704      out[n] = level;
    705      if (level) last = n;
    706    } else {
    707      out[n] = 0;
    708      in[j] = 0;
    709    }
    710  }
    711  return (last >= 0);
    712 }
    713 
    714 static int Quantize2Blocks_C(int16_t in[32], int16_t out[32],
    715                             const VP8Matrix* WEBP_RESTRICT const mtx) {
    716  int nz;
    717  nz  = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0;
    718  nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1;
    719  return nz;
    720 }
    721 #endif  // !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC
    722 
    723 //------------------------------------------------------------------------------
    724 // Block copy
    725 
    726 static WEBP_INLINE void Copy(const uint8_t* WEBP_RESTRICT src,
    727                             uint8_t* WEBP_RESTRICT dst, int w, int h) {
    728  int y;
    729  for (y = 0; y < h; ++y) {
    730    memcpy(dst, src, w);
    731    src += BPS;
    732    dst += BPS;
    733  }
    734 }
    735 
    736 static void Copy4x4_C(const uint8_t* WEBP_RESTRICT src,
    737                      uint8_t* WEBP_RESTRICT dst) {
    738  Copy(src, dst, 4, 4);
    739 }
    740 
    741 static void Copy16x8_C(const uint8_t* WEBP_RESTRICT src,
    742                       uint8_t* WEBP_RESTRICT dst) {
    743  Copy(src, dst, 16, 8);
    744 }
    745 
    746 //------------------------------------------------------------------------------
    747 // Initialization
    748 
    749 // Speed-critical function pointers. We have to initialize them to the default
    750 // implementations within VP8EncDspInit().
    751 VP8CHisto VP8CollectHistogram;
    752 VP8Idct VP8ITransform;
    753 VP8Fdct VP8FTransform;
    754 VP8Fdct VP8FTransform2;
    755 VP8WHT VP8FTransformWHT;
    756 VP8Intra4Preds VP8EncPredLuma4;
    757 VP8IntraPreds VP8EncPredLuma16;
    758 VP8IntraPreds VP8EncPredChroma8;
    759 VP8Metric VP8SSE16x16;
    760 VP8Metric VP8SSE8x8;
    761 VP8Metric VP8SSE16x8;
    762 VP8Metric VP8SSE4x4;
    763 VP8WMetric VP8TDisto4x4;
    764 VP8WMetric VP8TDisto16x16;
    765 VP8MeanMetric VP8Mean16x4;
    766 VP8QuantizeBlock VP8EncQuantizeBlock;
    767 VP8Quantize2Blocks VP8EncQuantize2Blocks;
    768 VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
    769 VP8BlockCopy VP8Copy4x4;
    770 VP8BlockCopy VP8Copy16x8;
    771 
    772 extern VP8CPUInfo VP8GetCPUInfo;
    773 extern void VP8EncDspInitSSE2(void);
    774 extern void VP8EncDspInitSSE41(void);
    775 extern void VP8EncDspInitNEON(void);
    776 extern void VP8EncDspInitMIPS32(void);
    777 extern void VP8EncDspInitMIPSdspR2(void);
    778 extern void VP8EncDspInitMSA(void);
    779 
    780 WEBP_DSP_INIT_FUNC(VP8EncDspInit) {
    781  VP8DspInit();  // common inverse transforms
    782  InitTables();
    783 
    784  // default C implementations
    785 #if !WEBP_NEON_OMIT_C_CODE
    786  VP8ITransform = ITransform_C;
    787  VP8FTransform = FTransform_C;
    788  VP8FTransformWHT = FTransformWHT_C;
    789  VP8TDisto4x4 = Disto4x4_C;
    790  VP8TDisto16x16 = Disto16x16_C;
    791  VP8CollectHistogram = CollectHistogram_C;
    792  VP8SSE16x16 = SSE16x16_C;
    793  VP8SSE16x8 = SSE16x8_C;
    794  VP8SSE8x8 = SSE8x8_C;
    795  VP8SSE4x4 = SSE4x4_C;
    796 #endif
    797 
    798 #if !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC
    799  VP8EncQuantizeBlock = QuantizeBlock_C;
    800  VP8EncQuantize2Blocks = Quantize2Blocks_C;
    801  VP8EncQuantizeBlockWHT = QuantizeBlock_C;
    802 #endif
    803 
    804 #if !WEBP_NEON_OMIT_C_CODE || !WEBP_AARCH64 || BPS != 32
    805  VP8EncPredLuma4 = Intra4Preds_C;
    806 #endif
    807 #if !WEBP_NEON_OMIT_C_CODE || !WEBP_AARCH64
    808  VP8EncPredLuma16 = Intra16Preds_C;
    809 #endif
    810 
    811  VP8FTransform2 = FTransform2_C;
    812  VP8EncPredChroma8 = IntraChromaPreds_C;
    813  VP8Mean16x4 = Mean16x4_C;
    814  VP8Copy4x4 = Copy4x4_C;
    815  VP8Copy16x8 = Copy16x8_C;
    816 
    817  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
    818  if (VP8GetCPUInfo != NULL) {
    819 #if defined(WEBP_HAVE_SSE2)
    820    if (VP8GetCPUInfo(kSSE2)) {
    821      VP8EncDspInitSSE2();
    822 #if defined(WEBP_HAVE_SSE41)
    823      if (VP8GetCPUInfo(kSSE4_1)) {
    824        VP8EncDspInitSSE41();
    825      }
    826 #endif
    827    }
    828 #endif
    829 #if defined(WEBP_USE_MIPS32)
    830    if (VP8GetCPUInfo(kMIPS32)) {
    831      VP8EncDspInitMIPS32();
    832    }
    833 #endif
    834 #if defined(WEBP_USE_MIPS_DSP_R2)
    835    if (VP8GetCPUInfo(kMIPSdspR2)) {
    836      VP8EncDspInitMIPSdspR2();
    837    }
    838 #endif
    839 #if defined(WEBP_USE_MSA)
    840    if (VP8GetCPUInfo(kMSA)) {
    841      VP8EncDspInitMSA();
    842    }
    843 #endif
    844  }
    845 
    846 #if defined(WEBP_HAVE_NEON)
    847  if (WEBP_NEON_OMIT_C_CODE ||
    848      (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
    849    VP8EncDspInitNEON();
    850  }
    851 #endif
    852 
    853  assert(VP8ITransform != NULL);
    854  assert(VP8FTransform != NULL);
    855  assert(VP8FTransformWHT != NULL);
    856  assert(VP8TDisto4x4 != NULL);
    857  assert(VP8TDisto16x16 != NULL);
    858  assert(VP8CollectHistogram != NULL);
    859  assert(VP8SSE16x16 != NULL);
    860  assert(VP8SSE16x8 != NULL);
    861  assert(VP8SSE8x8 != NULL);
    862  assert(VP8SSE4x4 != NULL);
    863  assert(VP8EncQuantizeBlock != NULL);
    864  assert(VP8EncQuantize2Blocks != NULL);
    865  assert(VP8FTransform2 != NULL);
    866  assert(VP8EncPredLuma4 != NULL);
    867  assert(VP8EncPredLuma16 != NULL);
    868  assert(VP8EncPredChroma8 != NULL);
    869  assert(VP8Mean16x4 != NULL);
    870  assert(VP8EncQuantizeBlockWHT != NULL);
    871  assert(VP8Copy4x4 != NULL);
    872  assert(VP8Copy16x8 != NULL);
    873 }