quant_enc.c 44.6 KB
Newer Older
wester committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283
// Copyright 2011 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING file in the root of the source
// tree. An additional intellectual property rights grant can be found
// in the file PATENTS. All contributing project authors may
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
//   Quantization
//
// Author: Skal (pascal.massimino@gmail.com)

#include <assert.h>
#include <math.h>
#include <stdlib.h>  // for abs()

#include "./vp8i_enc.h"
#include "./cost_enc.h"

#define DO_TRELLIS_I4  1
#define DO_TRELLIS_I16 1   // not a huge gain, but ok at low bitrate.
#define DO_TRELLIS_UV  0   // disable trellis for UV. Risky. Not worth.
#define USE_TDISTO 1

#define MID_ALPHA 64      // neutral value for susceptibility
#define MIN_ALPHA 30      // lowest usable value for susceptibility
#define MAX_ALPHA 100     // higher meaningful value for susceptibility

#define SNS_TO_DQ 0.9     // Scaling constant between the sns value and the QP
                          // power-law modulation. Must be strictly less than 1.

// number of non-zero coeffs below which we consider the block very flat
// (and apply a penalty to complex predictions)
#define FLATNESS_LIMIT_I16 10      // I16 mode
#define FLATNESS_LIMIT_I4  3       // I4 mode
#define FLATNESS_LIMIT_UV  2       // UV mode
#define FLATNESS_PENALTY   140     // roughly ~1bit per block

#define MULT_8B(a, b) (((a) * (b) + 128) >> 8)

#define RD_DISTO_MULT      256  // distortion multiplier (equivalent of lambda)

// #define DEBUG_BLOCK

//------------------------------------------------------------------------------

#if defined(DEBUG_BLOCK)

#include <stdio.h>
#include <stdlib.h>

static void PrintBlockInfo(const VP8EncIterator* const it,
                           const VP8ModeScore* const rd) {
  int i, j;
  const int is_i16 = (it->mb_->type_ == 1);
  const uint8_t* const y_in = it->yuv_in_ + Y_OFF_ENC;
  const uint8_t* const y_out = it->yuv_out_ + Y_OFF_ENC;
  const uint8_t* const uv_in = it->yuv_in_ + U_OFF_ENC;
  const uint8_t* const uv_out = it->yuv_out_ + U_OFF_ENC;
  printf("SOURCE / OUTPUT / ABS DELTA\n");
  for (j = 0; j < 16; ++j) {
    for (i = 0; i < 16; ++i) printf("%3d ", y_in[i + j * BPS]);
    printf("     ");
    for (i = 0; i < 16; ++i) printf("%3d ", y_out[i + j * BPS]);
    printf("     ");
    for (i = 0; i < 16; ++i) {
      printf("%1d ", abs(y_in[i + j * BPS] - y_out[i + j * BPS]));
    }
    printf("\n");
  }
  printf("\n");   // newline before the U/V block
  for (j = 0; j < 8; ++j) {
    for (i = 0; i < 8; ++i) printf("%3d ", uv_in[i + j * BPS]);
    printf(" ");
    for (i = 8; i < 16; ++i) printf("%3d ", uv_in[i + j * BPS]);
    printf("    ");
    for (i = 0; i < 8; ++i) printf("%3d ", uv_out[i + j * BPS]);
    printf(" ");
    for (i = 8; i < 16; ++i) printf("%3d ", uv_out[i + j * BPS]);
    printf("   ");
    for (i = 0; i < 8; ++i) {
      printf("%1d ", abs(uv_out[i + j * BPS] - uv_in[i + j * BPS]));
    }
    printf(" ");
    for (i = 8; i < 16; ++i) {
      printf("%1d ", abs(uv_out[i + j * BPS] - uv_in[i + j * BPS]));
    }
    printf("\n");
  }
  printf("\nD:%d SD:%d R:%d H:%d nz:0x%x score:%d\n",
    (int)rd->D, (int)rd->SD, (int)rd->R, (int)rd->H, (int)rd->nz,
    (int)rd->score);
  if (is_i16) {
    printf("Mode: %d\n", rd->mode_i16);
    printf("y_dc_levels:");
    for (i = 0; i < 16; ++i) printf("%3d ", rd->y_dc_levels[i]);
    printf("\n");
  } else {
    printf("Modes[16]: ");
    for (i = 0; i < 16; ++i) printf("%d ", rd->modes_i4[i]);
    printf("\n");
  }
  printf("y_ac_levels:\n");
  for (j = 0; j < 16; ++j) {
    for (i = is_i16 ? 1 : 0; i < 16; ++i) {
      printf("%4d ", rd->y_ac_levels[j][i]);
    }
    printf("\n");
  }
  printf("\n");
  printf("uv_levels (mode=%d):\n", rd->mode_uv);
  for (j = 0; j < 8; ++j) {
    for (i = 0; i < 16; ++i) {
      printf("%4d ", rd->uv_levels[j][i]);
    }
    printf("\n");
  }
}

#endif   // DEBUG_BLOCK

//------------------------------------------------------------------------------

static WEBP_INLINE int clip(int v, int m, int M) {
  return v < m ? m : v > M ? M : v;
}

static const uint8_t kZigzag[16] = {
  0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
};

static const uint8_t kDcTable[128] = {
  4,     5,   6,   7,   8,   9,  10,  10,
  11,   12,  13,  14,  15,  16,  17,  17,
  18,   19,  20,  20,  21,  21,  22,  22,
  23,   23,  24,  25,  25,  26,  27,  28,
  29,   30,  31,  32,  33,  34,  35,  36,
  37,   37,  38,  39,  40,  41,  42,  43,
  44,   45,  46,  46,  47,  48,  49,  50,
  51,   52,  53,  54,  55,  56,  57,  58,
  59,   60,  61,  62,  63,  64,  65,  66,
  67,   68,  69,  70,  71,  72,  73,  74,
  75,   76,  76,  77,  78,  79,  80,  81,
  82,   83,  84,  85,  86,  87,  88,  89,
  91,   93,  95,  96,  98, 100, 101, 102,
  104, 106, 108, 110, 112, 114, 116, 118,
  122, 124, 126, 128, 130, 132, 134, 136,
  138, 140, 143, 145, 148, 151, 154, 157
};

static const uint16_t kAcTable[128] = {
  4,     5,   6,   7,   8,   9,  10,  11,
  12,   13,  14,  15,  16,  17,  18,  19,
  20,   21,  22,  23,  24,  25,  26,  27,
  28,   29,  30,  31,  32,  33,  34,  35,
  36,   37,  38,  39,  40,  41,  42,  43,
  44,   45,  46,  47,  48,  49,  50,  51,
  52,   53,  54,  55,  56,  57,  58,  60,
  62,   64,  66,  68,  70,  72,  74,  76,
  78,   80,  82,  84,  86,  88,  90,  92,
  94,   96,  98, 100, 102, 104, 106, 108,
  110, 112, 114, 116, 119, 122, 125, 128,
  131, 134, 137, 140, 143, 146, 149, 152,
  155, 158, 161, 164, 167, 170, 173, 177,
  181, 185, 189, 193, 197, 201, 205, 209,
  213, 217, 221, 225, 229, 234, 239, 245,
  249, 254, 259, 264, 269, 274, 279, 284
};

static const uint16_t kAcTable2[128] = {
  8,     8,   9,  10,  12,  13,  15,  17,
  18,   20,  21,  23,  24,  26,  27,  29,
  31,   32,  34,  35,  37,  38,  40,  41,
  43,   44,  46,  48,  49,  51,  52,  54,
  55,   57,  58,  60,  62,  63,  65,  66,
  68,   69,  71,  72,  74,  75,  77,  79,
  80,   82,  83,  85,  86,  88,  89,  93,
  96,   99, 102, 105, 108, 111, 114, 117,
  120, 124, 127, 130, 133, 136, 139, 142,
  145, 148, 151, 155, 158, 161, 164, 167,
  170, 173, 176, 179, 184, 189, 193, 198,
  203, 207, 212, 217, 221, 226, 230, 235,
  240, 244, 249, 254, 258, 263, 268, 274,
  280, 286, 292, 299, 305, 311, 317, 323,
  330, 336, 342, 348, 354, 362, 370, 379,
  385, 393, 401, 409, 416, 424, 432, 440
};

static const uint8_t kBiasMatrices[3][2] = {  // [luma-ac,luma-dc,chroma][dc,ac]
  { 96, 110 }, { 96, 108 }, { 110, 115 }
};

// Sharpening by (slightly) raising the hi-frequency coeffs.
// Hack-ish but helpful for mid-bitrate range. Use with care.
#define SHARPEN_BITS 11  // number of descaling bits for sharpening bias
static const uint8_t kFreqSharpening[16] = {
  0,  30, 60, 90,
  30, 60, 90, 90,
  60, 90, 90, 90,
  90, 90, 90, 90
};

//------------------------------------------------------------------------------
// Initialize quantization parameters in VP8Matrix

// Returns the average quantizer
static int ExpandMatrix(VP8Matrix* const m, int type) {
  int i, sum;
  for (i = 0; i < 2; ++i) {
    const int is_ac_coeff = (i > 0);
    const int bias = kBiasMatrices[type][is_ac_coeff];
    m->iq_[i] = (1 << QFIX) / m->q_[i];
    m->bias_[i] = BIAS(bias);
    // zthresh_ is the exact value such that QUANTDIV(coeff, iQ, B) is:
    //   * zero if coeff <= zthresh
    //   * non-zero if coeff > zthresh
    m->zthresh_[i] = ((1 << QFIX) - 1 - m->bias_[i]) / m->iq_[i];
  }
  for (i = 2; i < 16; ++i) {
    m->q_[i] = m->q_[1];
    m->iq_[i] = m->iq_[1];
    m->bias_[i] = m->bias_[1];
    m->zthresh_[i] = m->zthresh_[1];
  }
  for (sum = 0, i = 0; i < 16; ++i) {
    if (type == 0) {  // we only use sharpening for AC luma coeffs
      m->sharpen_[i] = (kFreqSharpening[i] * m->q_[i]) >> SHARPEN_BITS;
    } else {
      m->sharpen_[i] = 0;
    }
    sum += m->q_[i];
  }
  return (sum + 8) >> 4;
}

static void CheckLambdaValue(int* const v) { if (*v < 1) *v = 1; }

static void SetupMatrices(VP8Encoder* enc) {
  int i;
  const int tlambda_scale =
    (enc->method_ >= 4) ? enc->config_->sns_strength
                        : 0;
  const int num_segments = enc->segment_hdr_.num_segments_;
  for (i = 0; i < num_segments; ++i) {
    VP8SegmentInfo* const m = &enc->dqm_[i];
    const int q = m->quant_;
    int q_i4, q_i16, q_uv;
    m->y1_.q_[0] = kDcTable[clip(q + enc->dq_y1_dc_, 0, 127)];
    m->y1_.q_[1] = kAcTable[clip(q,                  0, 127)];

    m->y2_.q_[0] = kDcTable[ clip(q + enc->dq_y2_dc_, 0, 127)] * 2;
    m->y2_.q_[1] = kAcTable2[clip(q + enc->dq_y2_ac_, 0, 127)];

    m->uv_.q_[0] = kDcTable[clip(q + enc->dq_uv_dc_, 0, 117)];
    m->uv_.q_[1] = kAcTable[clip(q + enc->dq_uv_ac_, 0, 127)];

    q_i4  = ExpandMatrix(&m->y1_, 0);
    q_i16 = ExpandMatrix(&m->y2_, 1);
    q_uv  = ExpandMatrix(&m->uv_, 2);

    m->lambda_i4_          = (3 * q_i4 * q_i4) >> 7;
    m->lambda_i16_         = (3 * q_i16 * q_i16);
    m->lambda_uv_          = (3 * q_uv * q_uv) >> 6;
    m->lambda_mode_        = (1 * q_i4 * q_i4) >> 7;
    m->lambda_trellis_i4_  = (7 * q_i4 * q_i4) >> 3;
    m->lambda_trellis_i16_ = (q_i16 * q_i16) >> 2;
    m->lambda_trellis_uv_  = (q_uv * q_uv) << 1;
    m->tlambda_            = (tlambda_scale * q_i4) >> 5;

    // none of these constants should be < 1
    CheckLambdaValue(&m->lambda_i4_);
    CheckLambdaValue(&m->lambda_i16_);
    CheckLambdaValue(&m->lambda_uv_);
    CheckLambdaValue(&m->lambda_mode_);
    CheckLambdaValue(&m->lambda_trellis_i4_);
    CheckLambdaValue(&m->lambda_trellis_i16_);
    CheckLambdaValue(&m->lambda_trellis_uv_);
    CheckLambdaValue(&m->tlambda_);

    m->min_disto_ = 20 * m->y1_.q_[0];   // quantization-aware min disto
    m->max_edge_  = 0;

    m->i4_penalty_ = 1000 * q_i4 * q_i4;
  }
}

//------------------------------------------------------------------------------
// Initialize filtering parameters

// Very small filter-strength values have close to no visual effect. So we can
// save a little decoding-CPU by turning filtering off for these.
#define FSTRENGTH_CUTOFF 2

static void SetupFilterStrength(VP8Encoder* const enc) {
  int i;
  // level0 is in [0..500]. Using '-f 50' as filter_strength is mid-filtering.
  const int level0 = 5 * enc->config_->filter_strength;
  for (i = 0; i < NUM_MB_SEGMENTS; ++i) {
    VP8SegmentInfo* const m = &enc->dqm_[i];
    // We focus on the quantization of AC coeffs.
    const int qstep = kAcTable[clip(m->quant_, 0, 127)] >> 2;
    const int base_strength =
        VP8FilterStrengthFromDelta(enc->filter_hdr_.sharpness_, qstep);
    // Segments with lower complexity ('beta') will be less filtered.
    const int f = base_strength * level0 / (256 + m->beta_);
    m->fstrength_ = (f < FSTRENGTH_CUTOFF) ? 0 : (f > 63) ? 63 : f;
  }
  // We record the initial strength (mainly for the case of 1-segment only).
  enc->filter_hdr_.level_ = enc->dqm_[0].fstrength_;
  enc->filter_hdr_.simple_ = (enc->config_->filter_type == 0);
  enc->filter_hdr_.sharpness_ = enc->config_->filter_sharpness;
}

//------------------------------------------------------------------------------

// Note: if you change the values below, remember that the max range
// allowed by the syntax for DQ_UV is [-16,16].
#define MAX_DQ_UV (6)
#define MIN_DQ_UV (-4)

// We want to emulate jpeg-like behaviour where the expected "good" quality
// is around q=75. Internally, our "good" middle is around c=50. So we
// map accordingly using linear piece-wise function
static double QualityToCompression(double c) {
  const double linear_c = (c < 0.75) ? c * (2. / 3.) : 2. * c - 1.;
  // The file size roughly scales as pow(quantizer, 3.). Actually, the
  // exponent is somewhere between 2.8 and 3.2, but we're mostly interested
  // in the mid-quant range. So we scale the compressibility inversely to
  // this power-law: quant ~= compression ^ 1/3. This law holds well for
  // low quant. Finer modeling for high-quant would make use of kAcTable[]
  // more explicitly.
  const double v = pow(linear_c, 1 / 3.);
  return v;
}

static double QualityToJPEGCompression(double c, double alpha) {
  // We map the complexity 'alpha' and quality setting 'c' to a compression
  // exponent empirically matched to the compression curve of libjpeg6b.
  // On average, the WebP output size will be roughly similar to that of a
  // JPEG file compressed with same quality factor.
  const double amin = 0.30;
  const double amax = 0.85;
  const double exp_min = 0.4;
  const double exp_max = 0.9;
  const double slope = (exp_min - exp_max) / (amax - amin);
  // Linearly interpolate 'expn' from exp_min to exp_max
  // in the [amin, amax] range.
  const double expn = (alpha > amax) ? exp_min
                    : (alpha < amin) ? exp_max
                    : exp_max + slope * (alpha - amin);
  const double v = pow(c, expn);
  return v;
}

static int SegmentsAreEquivalent(const VP8SegmentInfo* const S1,
                                 const VP8SegmentInfo* const S2) {
  return (S1->quant_ == S2->quant_) && (S1->fstrength_ == S2->fstrength_);
}

static void SimplifySegments(VP8Encoder* const enc) {
  int map[NUM_MB_SEGMENTS] = { 0, 1, 2, 3 };
  // 'num_segments_' is previously validated and <= NUM_MB_SEGMENTS, but an
  // explicit check is needed to avoid a spurious warning about 'i' exceeding
  // array bounds of 'dqm_' with some compilers (noticed with gcc-4.9).
  const int num_segments = (enc->segment_hdr_.num_segments_ < NUM_MB_SEGMENTS)
                               ? enc->segment_hdr_.num_segments_
                               : NUM_MB_SEGMENTS;
  int num_final_segments = 1;
  int s1, s2;
  for (s1 = 1; s1 < num_segments; ++s1) {    // find similar segments
    const VP8SegmentInfo* const S1 = &enc->dqm_[s1];
    int found = 0;
    // check if we already have similar segment
    for (s2 = 0; s2 < num_final_segments; ++s2) {
      const VP8SegmentInfo* const S2 = &enc->dqm_[s2];
      if (SegmentsAreEquivalent(S1, S2)) {
        found = 1;
        break;
      }
    }
    map[s1] = s2;
    if (!found) {
      if (num_final_segments != s1) {
        enc->dqm_[num_final_segments] = enc->dqm_[s1];
      }
      ++num_final_segments;
    }
  }
  if (num_final_segments < num_segments) {  // Remap
    int i = enc->mb_w_ * enc->mb_h_;
    while (i-- > 0) enc->mb_info_[i].segment_ = map[enc->mb_info_[i].segment_];
    enc->segment_hdr_.num_segments_ = num_final_segments;
    // Replicate the trailing segment infos (it's mostly cosmetics)
    for (i = num_final_segments; i < num_segments; ++i) {
      enc->dqm_[i] = enc->dqm_[num_final_segments - 1];
    }
  }
}

void VP8SetSegmentParams(VP8Encoder* const enc, float quality) {
  int i;
  int dq_uv_ac, dq_uv_dc;
  const int num_segments = enc->segment_hdr_.num_segments_;
  const double amp = SNS_TO_DQ * enc->config_->sns_strength / 100. / 128.;
  const double Q = quality / 100.;
  const double c_base = enc->config_->emulate_jpeg_size ?
      QualityToJPEGCompression(Q, enc->alpha_ / 255.) :
      QualityToCompression(Q);
  for (i = 0; i < num_segments; ++i) {
    // We modulate the base coefficient to accommodate for the quantization
    // susceptibility and allow denser segments to be quantized more.
    const double expn = 1. - amp * enc->dqm_[i].alpha_;
    const double c = pow(c_base, expn);
    const int q = (int)(127. * (1. - c));
    assert(expn > 0.);
    enc->dqm_[i].quant_ = clip(q, 0, 127);
  }

  // purely indicative in the bitstream (except for the 1-segment case)
  enc->base_quant_ = enc->dqm_[0].quant_;

  // fill-in values for the unused segments (required by the syntax)
  for (i = num_segments; i < NUM_MB_SEGMENTS; ++i) {
    enc->dqm_[i].quant_ = enc->base_quant_;
  }

  // uv_alpha_ is normally spread around ~60. The useful range is
  // typically ~30 (quite bad) to ~100 (ok to decimate UV more).
  // We map it to the safe maximal range of MAX/MIN_DQ_UV for dq_uv.
  dq_uv_ac = (enc->uv_alpha_ - MID_ALPHA) * (MAX_DQ_UV - MIN_DQ_UV)
                                          / (MAX_ALPHA - MIN_ALPHA);
  // we rescale by the user-defined strength of adaptation
  dq_uv_ac = dq_uv_ac * enc->config_->sns_strength / 100;
  // and make it safe.
  dq_uv_ac = clip(dq_uv_ac, MIN_DQ_UV, MAX_DQ_UV);
  // We also boost the dc-uv-quant a little, based on sns-strength, since
  // U/V channels are quite more reactive to high quants (flat DC-blocks
  // tend to appear, and are unpleasant).
  dq_uv_dc = -4 * enc->config_->sns_strength / 100;
  dq_uv_dc = clip(dq_uv_dc, -15, 15);   // 4bit-signed max allowed

  enc->dq_y1_dc_ = 0;       // TODO(skal): dq-lum
  enc->dq_y2_dc_ = 0;
  enc->dq_y2_ac_ = 0;
  enc->dq_uv_dc_ = dq_uv_dc;
  enc->dq_uv_ac_ = dq_uv_ac;

  SetupFilterStrength(enc);   // initialize segments' filtering, eventually

  if (num_segments > 1) SimplifySegments(enc);

  SetupMatrices(enc);         // finalize quantization matrices
}

//------------------------------------------------------------------------------
// Form the predictions in cache

// Must be ordered using {DC_PRED, TM_PRED, V_PRED, H_PRED} as index
const int VP8I16ModeOffsets[4] = { I16DC16, I16TM16, I16VE16, I16HE16 };
const int VP8UVModeOffsets[4] = { C8DC8, C8TM8, C8VE8, C8HE8 };

// Must be indexed using {B_DC_PRED -> B_HU_PRED} as index
const int VP8I4ModeOffsets[NUM_BMODES] = {
  I4DC4, I4TM4, I4VE4, I4HE4, I4RD4, I4VR4, I4LD4, I4VL4, I4HD4, I4HU4
};

void VP8MakeLuma16Preds(const VP8EncIterator* const it) {
  const uint8_t* const left = it->x_ ? it->y_left_ : NULL;
  const uint8_t* const top = it->y_ ? it->y_top_ : NULL;
  VP8EncPredLuma16(it->yuv_p_, left, top);
}

void VP8MakeChroma8Preds(const VP8EncIterator* const it) {
  const uint8_t* const left = it->x_ ? it->u_left_ : NULL;
  const uint8_t* const top = it->y_ ? it->uv_top_ : NULL;
  VP8EncPredChroma8(it->yuv_p_, left, top);
}

void VP8MakeIntra4Preds(const VP8EncIterator* const it) {
  VP8EncPredLuma4(it->yuv_p_, it->i4_top_);
}

//------------------------------------------------------------------------------
// Quantize

// Layout:
// +----+----+
// |YYYY|UUVV| 0
// |YYYY|UUVV| 4
// |YYYY|....| 8
// |YYYY|....| 12
// +----+----+

const int VP8Scan[16] = {  // Luma
  0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
  0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
  0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
  0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS,
};

static const int VP8ScanUV[4 + 4] = {
  0 + 0 * BPS,   4 + 0 * BPS, 0 + 4 * BPS,  4 + 4 * BPS,    // U
  8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
};

//------------------------------------------------------------------------------
// Distortion measurement

static const uint16_t kWeightY[16] = {
  38, 32, 20, 9, 32, 28, 17, 7, 20, 17, 10, 4, 9, 7, 4, 2
};

static const uint16_t kWeightTrellis[16] = {
#if USE_TDISTO == 0
  16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16
#else
  30, 27, 19, 11,
  27, 24, 17, 10,
  19, 17, 12,  8,
  11, 10,  8,  6
#endif
};

// Init/Copy the common fields in score.
static void InitScore(VP8ModeScore* const rd) {
  rd->D  = 0;
  rd->SD = 0;
  rd->R  = 0;
  rd->H  = 0;
  rd->nz = 0;
  rd->score = MAX_COST;
}

static void CopyScore(VP8ModeScore* const dst, const VP8ModeScore* const src) {
  dst->D  = src->D;
  dst->SD = src->SD;
  dst->R  = src->R;
  dst->H  = src->H;
  dst->nz = src->nz;      // note that nz is not accumulated, but just copied.
  dst->score = src->score;
}

static void AddScore(VP8ModeScore* const dst, const VP8ModeScore* const src) {
  dst->D  += src->D;
  dst->SD += src->SD;
  dst->R  += src->R;
  dst->H  += src->H;
  dst->nz |= src->nz;     // here, new nz bits are accumulated.
  dst->score += src->score;
}

//------------------------------------------------------------------------------
// Performs trellis-optimized quantization.

// Trellis node
typedef struct {
  int8_t prev;            // best previous node
  int8_t sign;            // sign of coeff_i
  int16_t level;          // level
} Node;

// Score state
typedef struct {
  score_t score;          // partial RD score
  const uint16_t* costs;  // shortcut to cost tables
} ScoreState;

// If a coefficient was quantized to a value Q (using a neutral bias),
// we test all alternate possibilities between [Q-MIN_DELTA, Q+MAX_DELTA]
// We don't test negative values though.
#define MIN_DELTA 0   // how much lower level to try
#define MAX_DELTA 1   // how much higher
#define NUM_NODES (MIN_DELTA + 1 + MAX_DELTA)
#define NODE(n, l) (nodes[(n)][(l) + MIN_DELTA])
#define SCORE_STATE(n, l) (score_states[n][(l) + MIN_DELTA])

static WEBP_INLINE void SetRDScore(int lambda, VP8ModeScore* const rd) {
  rd->score = (rd->R + rd->H) * lambda + RD_DISTO_MULT * (rd->D + rd->SD);
}

static WEBP_INLINE score_t RDScoreTrellis(int lambda, score_t rate,
                                          score_t distortion) {
  return rate * lambda + RD_DISTO_MULT * distortion;
}

static int TrellisQuantizeBlock(const VP8Encoder* const enc,
                                int16_t in[16], int16_t out[16],
                                int ctx0, int coeff_type,
                                const VP8Matrix* const mtx,
                                int lambda) {
  const ProbaArray* const probas = enc->proba_.coeffs_[coeff_type];
  CostArrayPtr const costs =
      (CostArrayPtr)enc->proba_.remapped_costs_[coeff_type];
  const int first = (coeff_type == 0) ? 1 : 0;
  Node nodes[16][NUM_NODES];
  ScoreState score_states[2][NUM_NODES];
  ScoreState* ss_cur = &SCORE_STATE(0, MIN_DELTA);
  ScoreState* ss_prev = &SCORE_STATE(1, MIN_DELTA);
  int best_path[3] = {-1, -1, -1};   // store best-last/best-level/best-previous
  score_t best_score;
  int n, m, p, last;

  {
    score_t cost;
    const int thresh = mtx->q_[1] * mtx->q_[1] / 4;
    const int last_proba = probas[VP8EncBands[first]][ctx0][0];

    // compute the position of the last interesting coefficient
    last = first - 1;
    for (n = 15; n >= first; --n) {
      const int j = kZigzag[n];
      const int err = in[j] * in[j];
      if (err > thresh) {
        last = n;
        break;
      }
    }
    // we don't need to go inspect up to n = 16 coeffs. We can just go up
    // to last + 1 (inclusive) without losing much.
    if (last < 15) ++last;

    // compute 'skip' score. This is the max score one can do.
    cost = VP8BitCost(0, last_proba);
    best_score = RDScoreTrellis(lambda, cost, 0);

    // initialize source node.
    for (m = -MIN_DELTA; m <= MAX_DELTA; ++m) {
      const score_t rate = (ctx0 == 0) ? VP8BitCost(1, last_proba) : 0;
      ss_cur[m].score = RDScoreTrellis(lambda, rate, 0);
      ss_cur[m].costs = costs[first][ctx0];
    }
  }

  // traverse trellis.
  for (n = first; n <= last; ++n) {
    const int j = kZigzag[n];
    const uint32_t Q  = mtx->q_[j];
    const uint32_t iQ = mtx->iq_[j];
    const uint32_t B = BIAS(0x00);     // neutral bias
    // note: it's important to take sign of the _original_ coeff,
    // so we don't have to consider level < 0 afterward.
    const int sign = (in[j] < 0);
    const uint32_t coeff0 = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
    int level0 = QUANTDIV(coeff0, iQ, B);
    int thresh_level = QUANTDIV(coeff0, iQ, BIAS(0x80));
    if (thresh_level > MAX_LEVEL) thresh_level = MAX_LEVEL;
    if (level0 > MAX_LEVEL) level0 = MAX_LEVEL;

    {   // Swap current and previous score states
      ScoreState* const tmp = ss_cur;
      ss_cur = ss_prev;
      ss_prev = tmp;
    }

    // test all alternate level values around level0.
    for (m = -MIN_DELTA; m <= MAX_DELTA; ++m) {
      Node* const cur = &NODE(n, m);
      int level = level0 + m;
      const int ctx = (level > 2) ? 2 : level;
      const int band = VP8EncBands[n + 1];
      score_t base_score;
      score_t best_cur_score = MAX_COST;
      int best_prev = 0;   // default, in case

      ss_cur[m].score = MAX_COST;
      ss_cur[m].costs = costs[n + 1][ctx];
      if (level < 0 || level > thresh_level) {
        // Node is dead.
        continue;
      }

      {
        // Compute delta_error = how much coding this level will
        // subtract to max_error as distortion.
        // Here, distortion = sum of (|coeff_i| - level_i * Q_i)^2
        const int new_error = coeff0 - level * Q;
        const int delta_error =
            kWeightTrellis[j] * (new_error * new_error - coeff0 * coeff0);
        base_score = RDScoreTrellis(lambda, 0, delta_error);
      }

      // Inspect all possible non-dead predecessors. Retain only the best one.
      for (p = -MIN_DELTA; p <= MAX_DELTA; ++p) {
        // Dead nodes (with ss_prev[p].score >= MAX_COST) are automatically
        // eliminated since their score can't be better than the current best.
        const score_t cost = VP8LevelCost(ss_prev[p].costs, level);
        // Examine node assuming it's a non-terminal one.
        const score_t score =
            base_score + ss_prev[p].score + RDScoreTrellis(lambda, cost, 0);
        if (score < best_cur_score) {
          best_cur_score = score;
          best_prev = p;
        }
      }
      // Store best finding in current node.
      cur->sign = sign;
      cur->level = level;
      cur->prev = best_prev;
      ss_cur[m].score = best_cur_score;

      // Now, record best terminal node (and thus best entry in the graph).
      if (level != 0) {
        const score_t last_pos_cost =
            (n < 15) ? VP8BitCost(0, probas[band][ctx][0]) : 0;
        const score_t last_pos_score = RDScoreTrellis(lambda, last_pos_cost, 0);
        const score_t score = best_cur_score + last_pos_score;
        if (score < best_score) {
          best_score = score;
          best_path[0] = n;                     // best eob position
          best_path[1] = m;                     // best node index
          best_path[2] = best_prev;             // best predecessor
        }
      }
    }
  }

  // Fresh start
  memset(in + first, 0, (16 - first) * sizeof(*in));
  memset(out + first, 0, (16 - first) * sizeof(*out));
  if (best_path[0] == -1) {
    return 0;   // skip!
  }

  {
    // Unwind the best path.
    // Note: best-prev on terminal node is not necessarily equal to the
    // best_prev for non-terminal. So we patch best_path[2] in.
    int nz = 0;
    int best_node = best_path[1];
    n = best_path[0];
    NODE(n, best_node).prev = best_path[2];   // force best-prev for terminal

    for (; n >= first; --n) {
      const Node* const node = &NODE(n, best_node);
      const int j = kZigzag[n];
      out[n] = node->sign ? -node->level : node->level;
      nz |= node->level;
      in[j] = out[n] * mtx->q_[j];
      best_node = node->prev;
    }
    return (nz != 0);
  }
}

#undef NODE

//------------------------------------------------------------------------------
// Performs: difference, transform, quantize, back-transform, add
// all at once. Output is the reconstructed block in *yuv_out, and the
// quantized levels in *levels.

static int ReconstructIntra16(VP8EncIterator* const it,
                              VP8ModeScore* const rd,
                              uint8_t* const yuv_out,
                              int mode) {
  const VP8Encoder* const enc = it->enc_;
  const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
  const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC;
  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
  int nz = 0;
  int n;
  int16_t tmp[16][16], dc_tmp[16];

  for (n = 0; n < 16; n += 2) {
    VP8FTransform2(src + VP8Scan[n], ref + VP8Scan[n], tmp[n]);
  }
  VP8FTransformWHT(tmp[0], dc_tmp);
  nz |= VP8EncQuantizeBlockWHT(dc_tmp, rd->y_dc_levels, &dqm->y2_) << 24;

  if (DO_TRELLIS_I16 && it->do_trellis_) {
    int x, y;
    VP8IteratorNzToBytes(it);
    for (y = 0, n = 0; y < 4; ++y) {
      for (x = 0; x < 4; ++x, ++n) {
        const int ctx = it->top_nz_[x] + it->left_nz_[y];
        const int non_zero =
            TrellisQuantizeBlock(enc, tmp[n], rd->y_ac_levels[n], ctx, 0,
                                 &dqm->y1_, dqm->lambda_trellis_i16_);
        it->top_nz_[x] = it->left_nz_[y] = non_zero;
        rd->y_ac_levels[n][0] = 0;
        nz |= non_zero << n;
      }
    }
  } else {
    for (n = 0; n < 16; n += 2) {
      // Zero-out the first coeff, so that: a) nz is correct below, and
      // b) finding 'last' non-zero coeffs in SetResidualCoeffs() is simplified.
      tmp[n][0] = tmp[n + 1][0] = 0;
      nz |= VP8EncQuantize2Blocks(tmp[n], rd->y_ac_levels[n], &dqm->y1_) << n;
      assert(rd->y_ac_levels[n + 0][0] == 0);
      assert(rd->y_ac_levels[n + 1][0] == 0);
    }
  }

  // Transform back
  VP8TransformWHT(dc_tmp, tmp[0]);
  for (n = 0; n < 16; n += 2) {
    VP8ITransform(ref + VP8Scan[n], tmp[n], yuv_out + VP8Scan[n], 1);
  }

  return nz;
}

static int ReconstructIntra4(VP8EncIterator* const it,
                             int16_t levels[16],
                             const uint8_t* const src,
                             uint8_t* const yuv_out,
                             int mode) {
  const VP8Encoder* const enc = it->enc_;
  const uint8_t* const ref = it->yuv_p_ + VP8I4ModeOffsets[mode];
  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
  int nz = 0;
  int16_t tmp[16];

  VP8FTransform(src, ref, tmp);
  if (DO_TRELLIS_I4 && it->do_trellis_) {
    const int x = it->i4_ & 3, y = it->i4_ >> 2;
    const int ctx = it->top_nz_[x] + it->left_nz_[y];
    nz = TrellisQuantizeBlock(enc, tmp, levels, ctx, 3, &dqm->y1_,
                              dqm->lambda_trellis_i4_);
  } else {
    nz = VP8EncQuantizeBlock(tmp, levels, &dqm->y1_);
  }
  VP8ITransform(ref, tmp, yuv_out, 0);
  return nz;
}

static int ReconstructUV(VP8EncIterator* const it, VP8ModeScore* const rd,
                         uint8_t* const yuv_out, int mode) {
  const VP8Encoder* const enc = it->enc_;
  const uint8_t* const ref = it->yuv_p_ + VP8UVModeOffsets[mode];
  const uint8_t* const src = it->yuv_in_ + U_OFF_ENC;
  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
  int nz = 0;
  int n;
  int16_t tmp[8][16];

  for (n = 0; n < 8; n += 2) {
    VP8FTransform2(src + VP8ScanUV[n], ref + VP8ScanUV[n], tmp[n]);
  }
  if (DO_TRELLIS_UV && it->do_trellis_) {
    int ch, x, y;
    for (ch = 0, n = 0; ch <= 2; ch += 2) {
      for (y = 0; y < 2; ++y) {
        for (x = 0; x < 2; ++x, ++n) {
          const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
          const int non_zero =
              TrellisQuantizeBlock(enc, tmp[n], rd->uv_levels[n], ctx, 2,
                                   &dqm->uv_, dqm->lambda_trellis_uv_);
          it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = non_zero;
          nz |= non_zero << n;
        }
      }
    }
  } else {
    for (n = 0; n < 8; n += 2) {
      nz |= VP8EncQuantize2Blocks(tmp[n], rd->uv_levels[n], &dqm->uv_) << n;
    }
  }

  for (n = 0; n < 8; n += 2) {
    VP8ITransform(ref + VP8ScanUV[n], tmp[n], yuv_out + VP8ScanUV[n], 1);
  }
  return (nz << 16);
}

//------------------------------------------------------------------------------
// RD-opt decision. Reconstruct each modes, evalue distortion and bit-cost.
// Pick the mode is lower RD-cost = Rate + lambda * Distortion.

static void StoreMaxDelta(VP8SegmentInfo* const dqm, const int16_t DCs[16]) {
  // We look at the first three AC coefficients to determine what is the average
  // delta between each sub-4x4 block.
  const int v0 = abs(DCs[1]);
  const int v1 = abs(DCs[2]);
  const int v2 = abs(DCs[4]);
  int max_v = (v1 > v0) ? v1 : v0;
  max_v = (v2 > max_v) ? v2 : max_v;
  if (max_v > dqm->max_edge_) dqm->max_edge_ = max_v;
}

static void SwapModeScore(VP8ModeScore** a, VP8ModeScore** b) {
  VP8ModeScore* const tmp = *a;
  *a = *b;
  *b = tmp;
}

static void SwapPtr(uint8_t** a, uint8_t** b) {
  uint8_t* const tmp = *a;
  *a = *b;
  *b = tmp;
}

static void SwapOut(VP8EncIterator* const it) {
  SwapPtr(&it->yuv_out_, &it->yuv_out2_);
}

static score_t IsFlat(const int16_t* levels, int num_blocks, score_t thresh) {
  score_t score = 0;
  while (num_blocks-- > 0) {      // TODO(skal): refine positional scoring?
    int i;
    for (i = 1; i < 16; ++i) {    // omit DC, we're only interested in AC
      score += (levels[i] != 0);
      if (score > thresh) return 0;
    }
    levels += 16;
  }
  return 1;
}

static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* rd) {
  const int kNumBlocks = 16;
  VP8SegmentInfo* const dqm = &it->enc_->dqm_[it->mb_->segment_];
  const int lambda = dqm->lambda_i16_;
  const int tlambda = dqm->tlambda_;
  const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC;
  VP8ModeScore rd_tmp;
  VP8ModeScore* rd_cur = &rd_tmp;
  VP8ModeScore* rd_best = rd;
  int mode;

  rd->mode_i16 = -1;
  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
    uint8_t* const tmp_dst = it->yuv_out2_ + Y_OFF_ENC;  // scratch buffer
    rd_cur->mode_i16 = mode;

    // Reconstruct
    rd_cur->nz = ReconstructIntra16(it, rd_cur, tmp_dst, mode);

    // Measure RD-score
    rd_cur->D = VP8SSE16x16(src, tmp_dst);
    rd_cur->SD =
        tlambda ? MULT_8B(tlambda, VP8TDisto16x16(src, tmp_dst, kWeightY)) : 0;
    rd_cur->H = VP8FixedCostsI16[mode];
    rd_cur->R = VP8GetCostLuma16(it, rd_cur);
    if (mode > 0 &&
        IsFlat(rd_cur->y_ac_levels[0], kNumBlocks, FLATNESS_LIMIT_I16)) {
      // penalty to avoid flat area to be mispredicted by complex mode
      rd_cur->R += FLATNESS_PENALTY * kNumBlocks;
    }

    // Since we always examine Intra16 first, we can overwrite *rd directly.
    SetRDScore(lambda, rd_cur);
    if (mode == 0 || rd_cur->score < rd_best->score) {
      SwapModeScore(&rd_cur, &rd_best);
      SwapOut(it);
    }
  }
  if (rd_best != rd) {
    memcpy(rd, rd_best, sizeof(*rd));
  }
  SetRDScore(dqm->lambda_mode_, rd);   // finalize score for mode decision.
  VP8SetIntra16Mode(it, rd->mode_i16);

  // we have a blocky macroblock (only DCs are non-zero) with fairly high
  // distortion, record max delta so we can later adjust the minimal filtering
  // strength needed to smooth these blocks out.
  if ((rd->nz & 0x100ffff) == 0x1000000 && rd->D > dqm->min_disto_) {
    StoreMaxDelta(dqm, rd->y_dc_levels);
  }
}

//------------------------------------------------------------------------------

// return the cost array corresponding to the surrounding prediction modes.
static const uint16_t* GetCostModeI4(VP8EncIterator* const it,
                                     const uint8_t modes[16]) {
  const int preds_w = it->enc_->preds_w_;
  const int x = (it->i4_ & 3), y = it->i4_ >> 2;
  const int left = (x == 0) ? it->preds_[y * preds_w - 1] : modes[it->i4_ - 1];
  const int top = (y == 0) ? it->preds_[-preds_w + x] : modes[it->i4_ - 4];
  return VP8FixedCostsI4[top][left];
}

static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
  const VP8Encoder* const enc = it->enc_;
  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
  const int lambda = dqm->lambda_i4_;
  const int tlambda = dqm->tlambda_;
  const uint8_t* const src0 = it->yuv_in_ + Y_OFF_ENC;
  uint8_t* const best_blocks = it->yuv_out2_ + Y_OFF_ENC;
  int total_header_bits = 0;
  VP8ModeScore rd_best;

  if (enc->max_i4_header_bits_ == 0) {
    return 0;
  }

  InitScore(&rd_best);
  rd_best.H = 211;  // '211' is the value of VP8BitCost(0, 145)
  SetRDScore(dqm->lambda_mode_, &rd_best);
  VP8IteratorStartI4(it);
  do {
    const int kNumBlocks = 1;
    VP8ModeScore rd_i4;
    int mode;
    int best_mode = -1;
    const uint8_t* const src = src0 + VP8Scan[it->i4_];
    const uint16_t* const mode_costs = GetCostModeI4(it, rd->modes_i4);
    uint8_t* best_block = best_blocks + VP8Scan[it->i4_];
    uint8_t* tmp_dst = it->yuv_p_ + I4TMP;    // scratch buffer.

    InitScore(&rd_i4);
    VP8MakeIntra4Preds(it);
    for (mode = 0; mode < NUM_BMODES; ++mode) {
      VP8ModeScore rd_tmp;
      int16_t tmp_levels[16];

      // Reconstruct
      rd_tmp.nz =
          ReconstructIntra4(it, tmp_levels, src, tmp_dst, mode) << it->i4_;

      // Compute RD-score
      rd_tmp.D = VP8SSE4x4(src, tmp_dst);
      rd_tmp.SD =
          tlambda ? MULT_8B(tlambda, VP8TDisto4x4(src, tmp_dst, kWeightY))
                  : 0;
      rd_tmp.H = mode_costs[mode];

      // Add flatness penalty
      if (mode > 0 && IsFlat(tmp_levels, kNumBlocks, FLATNESS_LIMIT_I4)) {
        rd_tmp.R = FLATNESS_PENALTY * kNumBlocks;
      } else {
        rd_tmp.R = 0;
      }

      // early-out check
      SetRDScore(lambda, &rd_tmp);
      if (best_mode >= 0 && rd_tmp.score >= rd_i4.score) continue;

      // finish computing score
      rd_tmp.R += VP8GetCostLuma4(it, tmp_levels);
      SetRDScore(lambda, &rd_tmp);

      if (best_mode < 0 || rd_tmp.score < rd_i4.score) {
        CopyScore(&rd_i4, &rd_tmp);
        best_mode = mode;
        SwapPtr(&tmp_dst, &best_block);
        memcpy(rd_best.y_ac_levels[it->i4_], tmp_levels,
               sizeof(rd_best.y_ac_levels[it->i4_]));
      }
    }
    SetRDScore(dqm->lambda_mode_, &rd_i4);
    AddScore(&rd_best, &rd_i4);
    if (rd_best.score >= rd->score) {
      return 0;
    }
    total_header_bits += (int)rd_i4.H;   // <- equal to mode_costs[best_mode];
    if (total_header_bits > enc->max_i4_header_bits_) {
      return 0;
    }
    // Copy selected samples if not in the right place already.
    if (best_block != best_blocks + VP8Scan[it->i4_]) {
      VP8Copy4x4(best_block, best_blocks + VP8Scan[it->i4_]);
    }
    rd->modes_i4[it->i4_] = best_mode;
    it->top_nz_[it->i4_ & 3] = it->left_nz_[it->i4_ >> 2] = (rd_i4.nz ? 1 : 0);
  } while (VP8IteratorRotateI4(it, best_blocks));

  // finalize state
  CopyScore(rd, &rd_best);
  VP8SetIntra4Mode(it, rd->modes_i4);
  SwapOut(it);
  memcpy(rd->y_ac_levels, rd_best.y_ac_levels, sizeof(rd->y_ac_levels));
  return 1;   // select intra4x4 over intra16x16
}

//------------------------------------------------------------------------------

static void PickBestUV(VP8EncIterator* const it, VP8ModeScore* const rd) {
  const int kNumBlocks = 8;
  const VP8SegmentInfo* const dqm = &it->enc_->dqm_[it->mb_->segment_];
  const int lambda = dqm->lambda_uv_;
  const uint8_t* const src = it->yuv_in_ + U_OFF_ENC;
  uint8_t* tmp_dst = it->yuv_out2_ + U_OFF_ENC;  // scratch buffer
  uint8_t* dst0 = it->yuv_out_ + U_OFF_ENC;
  uint8_t* dst = dst0;
  VP8ModeScore rd_best;
  int mode;

  rd->mode_uv = -1;
  InitScore(&rd_best);
  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
    VP8ModeScore rd_uv;

    // Reconstruct
    rd_uv.nz = ReconstructUV(it, &rd_uv, tmp_dst, mode);

    // Compute RD-score
    rd_uv.D  = VP8SSE16x8(src, tmp_dst);
    rd_uv.SD = 0;    // not calling TDisto here: it tends to flatten areas.
    rd_uv.H  = VP8FixedCostsUV[mode];
    rd_uv.R  = VP8GetCostUV(it, &rd_uv);
    if (mode > 0 && IsFlat(rd_uv.uv_levels[0], kNumBlocks, FLATNESS_LIMIT_UV)) {
      rd_uv.R += FLATNESS_PENALTY * kNumBlocks;
    }

    SetRDScore(lambda, &rd_uv);
    if (mode == 0 || rd_uv.score < rd_best.score) {
      CopyScore(&rd_best, &rd_uv);
      rd->mode_uv = mode;
      memcpy(rd->uv_levels, rd_uv.uv_levels, sizeof(rd->uv_levels));
      SwapPtr(&dst, &tmp_dst);
    }
  }
  VP8SetIntraUVMode(it, rd->mode_uv);
  AddScore(rd, &rd_best);
  if (dst != dst0) {   // copy 16x8 block if needed
    VP8Copy16x8(dst, dst0);
  }
}

//------------------------------------------------------------------------------
// Final reconstruction and quantization.

static void SimpleQuantize(VP8EncIterator* const it, VP8ModeScore* const rd) {
  const VP8Encoder* const enc = it->enc_;
  const int is_i16 = (it->mb_->type_ == 1);
  int nz = 0;

  if (is_i16) {
    nz = ReconstructIntra16(it, rd, it->yuv_out_ + Y_OFF_ENC, it->preds_[0]);
  } else {
    VP8IteratorStartI4(it);
    do {
      const int mode =
          it->preds_[(it->i4_ & 3) + (it->i4_ >> 2) * enc->preds_w_];
      const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC + VP8Scan[it->i4_];
      uint8_t* const dst = it->yuv_out_ + Y_OFF_ENC + VP8Scan[it->i4_];
      VP8MakeIntra4Preds(it);
      nz |= ReconstructIntra4(it, rd->y_ac_levels[it->i4_],
                              src, dst, mode) << it->i4_;
    } while (VP8IteratorRotateI4(it, it->yuv_out_ + Y_OFF_ENC));
  }

  nz |= ReconstructUV(it, rd, it->yuv_out_ + U_OFF_ENC, it->mb_->uv_mode_);
  rd->nz = nz;
}

// Refine intra16/intra4 sub-modes based on distortion only (not rate).
static void RefineUsingDistortion(VP8EncIterator* const it,
                                  int try_both_modes, int refine_uv_mode,
                                  VP8ModeScore* const rd) {
  score_t best_score = MAX_COST;
  int nz = 0;
  int mode;
  int is_i16 = try_both_modes || (it->mb_->type_ == 1);

  const VP8SegmentInfo* const dqm = &it->enc_->dqm_[it->mb_->segment_];
  // Some empiric constants, of approximate order of magnitude.
  const int lambda_d_i16 = 106;
  const int lambda_d_i4 = 11;
  const int lambda_d_uv = 120;
  score_t score_i4 = dqm->i4_penalty_;
  score_t i4_bit_sum = 0;
  const score_t bit_limit = try_both_modes ? it->enc_->mb_header_limit_
                                           : MAX_COST;  // no early-out allowed

  if (is_i16) {   // First, evaluate Intra16 distortion
    int best_mode = -1;
    const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC;
    for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
      const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
      const score_t score = VP8SSE16x16(src, ref) * RD_DISTO_MULT
                          + VP8FixedCostsI16[mode] * lambda_d_i16;
      if (mode > 0 && VP8FixedCostsI16[mode] > bit_limit) {
        continue;
      }
      if (score < best_score) {
        best_mode = mode;
        best_score = score;
      }
    }
    VP8SetIntra16Mode(it, best_mode);
    // we'll reconstruct later, if i16 mode actually gets selected
  }

  // Next, evaluate Intra4
  if (try_both_modes || !is_i16) {
    // We don't evaluate the rate here, but just account for it through a
    // constant penalty (i4 mode usually needs more bits compared to i16).
    is_i16 = 0;
    VP8IteratorStartI4(it);
    do {
      int best_i4_mode = -1;
      score_t best_i4_score = MAX_COST;
      const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC + VP8Scan[it->i4_];
      const uint16_t* const mode_costs = GetCostModeI4(it, rd->modes_i4);

      VP8MakeIntra4Preds(it);
      for (mode = 0; mode < NUM_BMODES; ++mode) {
        const uint8_t* const ref = it->yuv_p_ + VP8I4ModeOffsets[mode];
        const score_t score = VP8SSE4x4(src, ref) * RD_DISTO_MULT
                            + mode_costs[mode] * lambda_d_i4;
        if (score < best_i4_score) {
          best_i4_mode = mode;
          best_i4_score = score;
        }
      }
      i4_bit_sum += mode_costs[best_i4_mode];
      rd->modes_i4[it->i4_] = best_i4_mode;
      score_i4 += best_i4_score;
      if (score_i4 >= best_score || i4_bit_sum > bit_limit) {
        // Intra4 won't be better than Intra16. Bail out and pick Intra16.
        is_i16 = 1;
        break;
      } else {  // reconstruct partial block inside yuv_out2_ buffer
        uint8_t* const tmp_dst = it->yuv_out2_ + Y_OFF_ENC + VP8Scan[it->i4_];
        nz |= ReconstructIntra4(it, rd->y_ac_levels[it->i4_],
                                src, tmp_dst, best_i4_mode) << it->i4_;
      }
    } while (VP8IteratorRotateI4(it, it->yuv_out2_ + Y_OFF_ENC));
  }

  // Final reconstruction, depending on which mode is selected.
  if (!is_i16) {
    VP8SetIntra4Mode(it, rd->modes_i4);
    SwapOut(it);
    best_score = score_i4;
  } else {
    nz = ReconstructIntra16(it, rd, it->yuv_out_ + Y_OFF_ENC, it->preds_[0]);
  }

  // ... and UV!
  if (refine_uv_mode) {
    int best_mode = -1;
    score_t best_uv_score = MAX_COST;
    const uint8_t* const src = it->yuv_in_ + U_OFF_ENC;
    for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
      const uint8_t* const ref = it->yuv_p_ + VP8UVModeOffsets[mode];
      const score_t score = VP8SSE16x8(src, ref) * RD_DISTO_MULT
                          + VP8FixedCostsUV[mode] * lambda_d_uv;
      if (score < best_uv_score) {
        best_mode = mode;
        best_uv_score = score;
      }
    }
    VP8SetIntraUVMode(it, best_mode);
  }
  nz |= ReconstructUV(it, rd, it->yuv_out_ + U_OFF_ENC, it->mb_->uv_mode_);

  rd->nz = nz;
  rd->score = best_score;
}

//------------------------------------------------------------------------------
// Entry point

int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd,
                VP8RDLevel rd_opt) {
  int is_skipped;
  const int method = it->enc_->method_;

  InitScore(rd);

  // We can perform predictions for Luma16x16 and Chroma8x8 already.
  // Luma4x4 predictions needs to be done as-we-go.
  VP8MakeLuma16Preds(it);
  VP8MakeChroma8Preds(it);

  if (rd_opt > RD_OPT_NONE) {
    it->do_trellis_ = (rd_opt >= RD_OPT_TRELLIS_ALL);
    PickBestIntra16(it, rd);
    if (method >= 2) {
      PickBestIntra4(it, rd);
    }
    PickBestUV(it, rd);
    if (rd_opt == RD_OPT_TRELLIS) {   // finish off with trellis-optim now
      it->do_trellis_ = 1;
      SimpleQuantize(it, rd);
    }
  } else {
    // At this point we have heuristically decided intra16 / intra4.
    // For method >= 2, pick the best intra4/intra16 based on SSE (~tad slower).
    // For method <= 1, we don't re-examine the decision but just go ahead with
    // quantization/reconstruction.
    RefineUsingDistortion(it, (method >= 2), (method >= 1), rd);
  }
  is_skipped = (rd->nz == 0);
  VP8SetSkip(it, is_skipped);
  return is_skipped;
}