// Copyright 2012 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.
// -----------------------------------------------------------------------------
//
// Author: Jyrki Alakuijala (jyrki@google.com)
//

#include <assert.h>
#include <math.h>

#include "./backward_references.h"
#include "./histogram.h"
#include "../dsp/lossless.h"
#include "../utils/color_cache.h"
#include "../utils/utils.h"

#define VALUES_IN_BYTE 256

#define HASH_MULTIPLIER (0xc6a4a7935bd1e995ULL)

#define MIN_BLOCK_SIZE 256  // minimum block size for backward references

#define MAX_ENTROPY    (1e30f)

// 1M window (4M bytes) minus 120 special codes for short distances.
#define WINDOW_SIZE ((1 << 20) - 120)

// Bounds for the match length.
#define MIN_LENGTH 2
#define MAX_LENGTH 4096

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

static const uint8_t plane_to_code_lut[128] = {
 96,   73,  55,  39,  23,  13,   5,  1,  255, 255, 255, 255, 255, 255, 255, 255,
 101,  78,  58,  42,  26,  16,   8,  2,    0,   3,  9,   17,  27,  43,  59,  79,
 102,  86,  62,  46,  32,  20,  10,  6,    4,   7,  11,  21,  33,  47,  63,  87,
 105,  90,  70,  52,  37,  28,  18,  14,  12,  15,  19,  29,  38,  53,  71,  91,
 110,  99,  82,  66,  48,  35,  30,  24,  22,  25,  31,  36,  49,  67,  83, 100,
 115, 108,  94,  76,  64,  50,  44,  40,  34,  41,  45,  51,  65,  77,  95, 109,
 118, 113, 103,  92,  80,  68,  60,  56,  54,  57,  61,  69,  81,  93, 104, 114,
 119, 116, 111, 106,  97,  88,  84,  74,  72,  75,  85,  89,  98, 107, 112, 117
};

static int DistanceToPlaneCode(int xsize, int dist) {
  const int yoffset = dist / xsize;
  const int xoffset = dist - yoffset * xsize;
  if (xoffset <= 8 && yoffset < 8) {
    return plane_to_code_lut[yoffset * 16 + 8 - xoffset] + 1;
  } else if (xoffset > xsize - 8 && yoffset < 7) {
    return plane_to_code_lut[(yoffset + 1) * 16 + 8 + (xsize - xoffset)] + 1;
  }
  return dist + 120;
}

static WEBP_INLINE int FindMatchLength(const uint32_t* const array1,
                                       const uint32_t* const array2,
                                       const int max_limit) {
  int match_len = 0;
  while (match_len < max_limit && array1[match_len] == array2[match_len]) {
    ++match_len;
  }
  return match_len;
}

// -----------------------------------------------------------------------------
//  VP8LBackwardRefs

struct PixOrCopyBlock {
  PixOrCopyBlock* next_;   // next block (or NULL)
  PixOrCopy* start_;       // data start
  int size_;               // currently used size
};

static void ClearBackwardRefs(VP8LBackwardRefs* const refs) {
  assert(refs != NULL);
  if (refs->tail_ != NULL) {
    *refs->tail_ = refs->free_blocks_;  // recycle all blocks at once
  }
  refs->free_blocks_ = refs->refs_;
  refs->tail_ = &refs->refs_;
  refs->last_block_ = NULL;
  refs->refs_ = NULL;
}

void VP8LBackwardRefsClear(VP8LBackwardRefs* const refs) {
  assert(refs != NULL);
  ClearBackwardRefs(refs);
  while (refs->free_blocks_ != NULL) {
    PixOrCopyBlock* const next = refs->free_blocks_->next_;
    WebPSafeFree(refs->free_blocks_);
    refs->free_blocks_ = next;
  }
}

void VP8LBackwardRefsInit(VP8LBackwardRefs* const refs, int block_size) {
  assert(refs != NULL);
  memset(refs, 0, sizeof(*refs));
  refs->tail_ = &refs->refs_;
  refs->block_size_ =
      (block_size < MIN_BLOCK_SIZE) ? MIN_BLOCK_SIZE : block_size;
}

VP8LRefsCursor VP8LRefsCursorInit(const VP8LBackwardRefs* const refs) {
  VP8LRefsCursor c;
  c.cur_block_ = refs->refs_;
  if (refs->refs_ != NULL) {
    c.cur_pos = c.cur_block_->start_;
    c.last_pos_ = c.cur_pos + c.cur_block_->size_;
  } else {
    c.cur_pos = NULL;
    c.last_pos_ = NULL;
  }
  return c;
}

void VP8LRefsCursorNextBlock(VP8LRefsCursor* const c) {
  PixOrCopyBlock* const b = c->cur_block_->next_;
  c->cur_pos = (b == NULL) ? NULL : b->start_;
  c->last_pos_ = (b == NULL) ? NULL : b->start_ + b->size_;
  c->cur_block_ = b;
}

// Create a new block, either from the free list or allocated
static PixOrCopyBlock* BackwardRefsNewBlock(VP8LBackwardRefs* const refs) {
  PixOrCopyBlock* b = refs->free_blocks_;
  if (b == NULL) {   // allocate new memory chunk
    const size_t total_size =
        sizeof(*b) + refs->block_size_ * sizeof(*b->start_);
    b = (PixOrCopyBlock*)WebPSafeMalloc(1ULL, total_size);
    if (b == NULL) {
      refs->error_ |= 1;
      return NULL;
    }
    b->start_ = (PixOrCopy*)((uint8_t*)b + sizeof(*b));  // not always aligned
  } else {  // recycle from free-list
    refs->free_blocks_ = b->next_;
  }
  *refs->tail_ = b;
  refs->tail_ = &b->next_;
  refs->last_block_ = b;
  b->next_ = NULL;
  b->size_ = 0;
  return b;
}

static WEBP_INLINE void BackwardRefsCursorAdd(VP8LBackwardRefs* const refs,
                                              const PixOrCopy v) {
  PixOrCopyBlock* b = refs->last_block_;
  if (b == NULL || b->size_ == refs->block_size_) {
    b = BackwardRefsNewBlock(refs);
    if (b == NULL) return;   // refs->error_ is set
  }
  b->start_[b->size_++] = v;
}

int VP8LBackwardRefsCopy(const VP8LBackwardRefs* const src,
                         VP8LBackwardRefs* const dst) {
  const PixOrCopyBlock* b = src->refs_;
  ClearBackwardRefs(dst);
  assert(src->block_size_ == dst->block_size_);
  while (b != NULL) {
    PixOrCopyBlock* const new_b = BackwardRefsNewBlock(dst);
    if (new_b == NULL) return 0;   // dst->error_ is set
    memcpy(new_b->start_, b->start_, b->size_ * sizeof(*b->start_));
    new_b->size_ = b->size_;
    b = b->next_;
  }
  return 1;
}

// -----------------------------------------------------------------------------
// Hash chains

// initialize as empty
static void HashChainInit(VP8LHashChain* const p) {
  int i;
  assert(p != NULL);
  for (i = 0; i < p->size_; ++i) {
    p->chain_[i] = -1;
  }
  for (i = 0; i < HASH_SIZE; ++i) {
    p->hash_to_first_index_[i] = -1;
  }
}

int VP8LHashChainInit(VP8LHashChain* const p, int size) {
  assert(p->size_ == 0);
  assert(p->chain_ == NULL);
  assert(size > 0);
  p->chain_ = (int*)WebPSafeMalloc(size, sizeof(*p->chain_));
  if (p->chain_ == NULL) return 0;
  p->size_ = size;
  HashChainInit(p);
  return 1;
}

void VP8LHashChainClear(VP8LHashChain* const p) {
  assert(p != NULL);
  WebPSafeFree(p->chain_);
  p->size_ = 0;
  p->chain_ = NULL;
}

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

static WEBP_INLINE uint64_t GetPixPairHash64(const uint32_t* const argb) {
  uint64_t key = ((uint64_t)argb[1] << 32) | argb[0];
  key = (key * HASH_MULTIPLIER) >> (64 - HASH_BITS);
  return key;
}

// Insertion of two pixels at a time.
static void HashChainInsert(VP8LHashChain* const p,
                            const uint32_t* const argb, int pos) {
  const uint64_t hash_code = GetPixPairHash64(argb);
  p->chain_[pos] = p->hash_to_first_index_[hash_code];
  p->hash_to_first_index_[hash_code] = pos;
}

static void GetParamsForHashChainFindCopy(int quality, int xsize,
                                          int cache_bits, int* window_size,
                                          int* iter_pos, int* iter_limit) {
  const int iter_mult = (quality < 27) ? 1 : 1 + ((quality - 27) >> 4);
  const int iter_neg = -iter_mult * (quality >> 1);
  // Limit the backward-ref window size for lower qualities.
  const int max_window_size = (quality > 50) ? WINDOW_SIZE
                            : (quality > 25) ? (xsize << 8)
                            : (xsize << 4);
  assert(xsize > 0);
  *window_size = (max_window_size > WINDOW_SIZE) ? WINDOW_SIZE
               : max_window_size;
  *iter_pos = 8 + (quality >> 3);
  // For lower entropy images, the rigorous search loop in HashChainFindCopy
  // can be relaxed.
  *iter_limit = (cache_bits > 0) ? iter_neg : iter_neg / 2;
}

static int HashChainFindCopy(const VP8LHashChain* const p,
                             int base_position, int xsize_signed,
                             const uint32_t* const argb, int max_len,
                             int window_size, int iter_pos, int iter_limit,
                             int* const distance_ptr,
                             int* const length_ptr) {
  const uint32_t* const argb_start = argb + base_position;
  uint64_t best_val = 0;
  uint32_t best_length = 1;
  uint32_t best_distance = 0;
  const uint32_t xsize = (uint32_t)xsize_signed;
  const int min_pos =
      (base_position > window_size) ? base_position - window_size : 0;
  int pos;
  assert(xsize > 0);
  if (max_len > MAX_LENGTH) {
    max_len = MAX_LENGTH;
  }
  for (pos = p->hash_to_first_index_[GetPixPairHash64(argb_start)];
       pos >= min_pos;
       pos = p->chain_[pos]) {
    uint64_t val;
    uint32_t curr_length;
    uint32_t distance;
    const uint32_t* const ptr1 = (argb + pos + best_length - 1);
    const uint32_t* const ptr2 = (argb_start + best_length - 1);

    if (iter_pos < 0) {
      if (iter_pos < iter_limit || best_val >= 0xff0000) {
        break;
      }
    }
    --iter_pos;

    // Before 'expensive' linear match, check if the two arrays match at the
    // current best length index and also for the succeeding elements.
    if (ptr1[0] != ptr2[0] || ptr1[1] != ptr2[1]) continue;

    curr_length = FindMatchLength(argb + pos, argb_start, max_len);
    if (curr_length < best_length) continue;

    distance = (uint32_t)(base_position - pos);
    val = curr_length << 16;
    // Favoring 2d locality here gives savings for certain images.
    if (distance < 9 * xsize) {
      const uint32_t y = distance / xsize;
      uint32_t x = distance % xsize;
      if (x > (xsize >> 1)) {
        x = xsize - x;
      }
      if (x <= 7) {
        val += 9 * 9 + 9 * 9;
        val -= y * y + x * x;
      }
    }
    if (best_val < val) {
      best_val = val;
      best_length = curr_length;
      best_distance = distance;
      if (curr_length >= (uint32_t)max_len) {
        break;
      }
      if ((best_distance == 1 || distance == xsize) &&
          best_length >= 128) {
        break;
      }
    }
  }
  *distance_ptr = (int)best_distance;
  *length_ptr = best_length;
  return (best_length >= MIN_LENGTH);
}

static WEBP_INLINE void PushBackCopy(VP8LBackwardRefs* const refs, int length) {
  while (length >= MAX_LENGTH) {
    BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(1, MAX_LENGTH));
    length -= MAX_LENGTH;
  }
  if (length > 0) {
    BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(1, length));
  }
}

static int BackwardReferencesRle(int xsize, int ysize,
                                 const uint32_t* const argb,
                                 VP8LBackwardRefs* const refs) {
  const int pix_count = xsize * ysize;
  int match_len = 0;
  int i;
  ClearBackwardRefs(refs);
  PushBackCopy(refs, match_len);    // i=0 case
  BackwardRefsCursorAdd(refs, PixOrCopyCreateLiteral(argb[0]));
  for (i = 1; i < pix_count; ++i) {
    if (argb[i] == argb[i - 1]) {
      ++match_len;
    } else {
      PushBackCopy(refs, match_len);
      match_len = 0;
      BackwardRefsCursorAdd(refs, PixOrCopyCreateLiteral(argb[i]));
    }
  }
  PushBackCopy(refs, match_len);
  return !refs->error_;
}

static int BackwardReferencesHashChain(int xsize, int ysize,
                                       const uint32_t* const argb,
                                       int cache_bits, int quality,
                                       VP8LHashChain* const hash_chain,
                                       VP8LBackwardRefs* const refs) {
  int i;
  int ok = 0;
  int cc_init = 0;
  const int use_color_cache = (cache_bits > 0);
  const int pix_count = xsize * ysize;
  VP8LColorCache hashers;
  int window_size = WINDOW_SIZE;
  int iter_pos = 1;
  int iter_limit = -1;

  if (use_color_cache) {
    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
    if (!cc_init) goto Error;
  }

  ClearBackwardRefs(refs);
  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                &window_size, &iter_pos, &iter_limit);
  HashChainInit(hash_chain);
  for (i = 0; i < pix_count; ) {
    // Alternative#1: Code the pixels starting at 'i' using backward reference.
    int offset = 0;
    int len = 0;
    if (i < pix_count - 1) {  // FindCopy(i,..) reads pixels at [i] and [i + 1].
      int max_len = pix_count - i;
      HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
                        window_size, iter_pos, iter_limit,
                        &offset, &len);
    }
    if (len >= MIN_LENGTH) {
      // Alternative#2: Insert the pixel at 'i' as literal, and code the
      // pixels starting at 'i + 1' using backward reference.
      int offset2 = 0;
      int len2 = 0;
      int k;
      HashChainInsert(hash_chain, &argb[i], i);
      if (i < pix_count - 2) {  // FindCopy(i+1,..) reads [i + 1] and [i + 2].
        int max_len = pix_count - (i + 1);
        HashChainFindCopy(hash_chain, i + 1, xsize, argb, max_len,
                          window_size, iter_pos, iter_limit,
                          &offset2, &len2);
        if (len2 > len + 1) {
          const uint32_t pixel = argb[i];
          // Alternative#2 is a better match. So push pixel at 'i' as literal.
          PixOrCopy v;
          if (use_color_cache && VP8LColorCacheContains(&hashers, pixel)) {
            const int ix = VP8LColorCacheGetIndex(&hashers, pixel);
            v = PixOrCopyCreateCacheIdx(ix);
          } else {
            if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
            v = PixOrCopyCreateLiteral(pixel);
          }
          BackwardRefsCursorAdd(refs, v);
          i++;  // Backward reference to be done for next pixel.
          len = len2;
          offset = offset2;
        }
      }
      if (len >= MAX_LENGTH) {
        len = MAX_LENGTH - 1;
      }
      BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(offset, len));
      if (use_color_cache) {
        for (k = 0; k < len; ++k) {
          VP8LColorCacheInsert(&hashers, argb[i + k]);
        }
      }
      // Add to the hash_chain (but cannot add the last pixel).
      {
        const int last = (len < pix_count - 1 - i) ? len : pix_count - 1 - i;
        for (k = 1; k < last; ++k) {
          HashChainInsert(hash_chain, &argb[i + k], i + k);
        }
      }
      i += len;
    } else {
      const uint32_t pixel = argb[i];
      PixOrCopy v;
      if (use_color_cache && VP8LColorCacheContains(&hashers, pixel)) {
        // push pixel as a PixOrCopyCreateCacheIdx pixel
        const int ix = VP8LColorCacheGetIndex(&hashers, pixel);
        v = PixOrCopyCreateCacheIdx(ix);
      } else {
        if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
        v = PixOrCopyCreateLiteral(pixel);
      }
      BackwardRefsCursorAdd(refs, v);
      if (i + 1 < pix_count) {
        HashChainInsert(hash_chain, &argb[i], i);
      }
      ++i;
    }
  }
  ok = !refs->error_;
Error:
  if (cc_init) VP8LColorCacheClear(&hashers);
  return ok;
}

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

typedef struct {
  double alpha_[VALUES_IN_BYTE];
  double red_[VALUES_IN_BYTE];
  double literal_[PIX_OR_COPY_CODES_MAX];
  double blue_[VALUES_IN_BYTE];
  double distance_[NUM_DISTANCE_CODES];
} CostModel;

static int BackwardReferencesTraceBackwards(
    int xsize, int ysize, int recursive_cost_model,
    const uint32_t* const argb, int quality, int cache_bits,
    VP8LHashChain* const hash_chain,
    VP8LBackwardRefs* const refs);

static void ConvertPopulationCountTableToBitEstimates(
    int num_symbols, const uint32_t population_counts[], double output[]) {
  uint32_t sum = 0;
  int nonzeros = 0;
  int i;
  for (i = 0; i < num_symbols; ++i) {
    sum += population_counts[i];
    if (population_counts[i] > 0) {
      ++nonzeros;
    }
  }
  if (nonzeros <= 1) {
    memset(output, 0, num_symbols * sizeof(*output));
  } else {
    const double logsum = VP8LFastLog2(sum);
    for (i = 0; i < num_symbols; ++i) {
      output[i] = logsum - VP8LFastLog2(population_counts[i]);
    }
  }
}

static int CostModelBuild(CostModel* const m, int xsize, int ysize,
                          int recursion_level, const uint32_t* const argb,
                          int quality, int cache_bits,
                          VP8LHashChain* const hash_chain,
                          VP8LBackwardRefs* const refs) {
  int ok = 0;
  VP8LHistogram* histo = NULL;

  ClearBackwardRefs(refs);
  if (recursion_level > 0) {
    if (!BackwardReferencesTraceBackwards(xsize, ysize, recursion_level - 1,
                                          argb, quality, cache_bits, hash_chain,
                                          refs)) {
      goto Error;
    }
  } else {
    if (!BackwardReferencesHashChain(xsize, ysize, argb, cache_bits, quality,
                                     hash_chain, refs)) {
      goto Error;
    }
  }
  histo = VP8LAllocateHistogram(cache_bits);
  if (histo == NULL) goto Error;

  VP8LHistogramCreate(histo, refs, cache_bits);

  ConvertPopulationCountTableToBitEstimates(
      VP8LHistogramNumCodes(histo->palette_code_bits_),
      histo->literal_, m->literal_);
  ConvertPopulationCountTableToBitEstimates(
      VALUES_IN_BYTE, histo->red_, m->red_);
  ConvertPopulationCountTableToBitEstimates(
      VALUES_IN_BYTE, histo->blue_, m->blue_);
  ConvertPopulationCountTableToBitEstimates(
      VALUES_IN_BYTE, histo->alpha_, m->alpha_);
  ConvertPopulationCountTableToBitEstimates(
      NUM_DISTANCE_CODES, histo->distance_, m->distance_);
  ok = 1;

 Error:
  VP8LFreeHistogram(histo);
  return ok;
}

static WEBP_INLINE double GetLiteralCost(const CostModel* const m, uint32_t v) {
  return m->alpha_[v >> 24] +
         m->red_[(v >> 16) & 0xff] +
         m->literal_[(v >> 8) & 0xff] +
         m->blue_[v & 0xff];
}

static WEBP_INLINE double GetCacheCost(const CostModel* const m, uint32_t idx) {
  const int literal_idx = VALUES_IN_BYTE + NUM_LENGTH_CODES + idx;
  return m->literal_[literal_idx];
}

static WEBP_INLINE double GetLengthCost(const CostModel* const m,
                                        uint32_t length) {
  int code, extra_bits;
  VP8LPrefixEncodeBits(length, &code, &extra_bits);
  return m->literal_[VALUES_IN_BYTE + code] + extra_bits;
}

static WEBP_INLINE double GetDistanceCost(const CostModel* const m,
                                          uint32_t distance) {
  int code, extra_bits;
  VP8LPrefixEncodeBits(distance, &code, &extra_bits);
  return m->distance_[code] + extra_bits;
}

static int BackwardReferencesHashChainDistanceOnly(
    int xsize, int ysize, int recursive_cost_model, const uint32_t* const argb,
    int quality, int cache_bits, VP8LHashChain* const hash_chain,
    VP8LBackwardRefs* const refs, uint32_t* const dist_array) {
  int i;
  int ok = 0;
  int cc_init = 0;
  const int pix_count = xsize * ysize;
  const int use_color_cache = (cache_bits > 0);
  float* const cost =
      (float*)WebPSafeMalloc(pix_count, sizeof(*cost));
  CostModel* cost_model = (CostModel*)WebPSafeMalloc(1ULL, sizeof(*cost_model));
  VP8LColorCache hashers;
  const double mul0 = (recursive_cost_model != 0) ? 1.0 : 0.68;
  const double mul1 = (recursive_cost_model != 0) ? 1.0 : 0.82;
  const int min_distance_code = 2;  // TODO(vikasa): tune as function of quality
  int window_size = WINDOW_SIZE;
  int iter_pos = 1;
  int iter_limit = -1;

  if (cost == NULL || cost_model == NULL) goto Error;

  if (use_color_cache) {
    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
    if (!cc_init) goto Error;
  }

  if (!CostModelBuild(cost_model, xsize, ysize, recursive_cost_model, argb,
                      quality, cache_bits, hash_chain, refs)) {
    goto Error;
  }

  for (i = 0; i < pix_count; ++i) cost[i] = 1e38f;

  // We loop one pixel at a time, but store all currently best points to
  // non-processed locations from this point.
  dist_array[0] = 0;
  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                &window_size, &iter_pos, &iter_limit);
  HashChainInit(hash_chain);
  for (i = 0; i < pix_count; ++i) {
    double prev_cost = 0.0;
    int shortmax;
    if (i > 0) {
      prev_cost = cost[i - 1];
    }
    for (shortmax = 0; shortmax < 2; ++shortmax) {
      int offset = 0;
      int len = 0;
      if (i < pix_count - 1) {  // FindCopy reads pixels at [i] and [i + 1].
        int max_len = shortmax ? 2 : pix_count - i;
        HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
                          window_size, iter_pos, iter_limit,
                          &offset, &len);
      }
      if (len >= MIN_LENGTH) {
        const int code = DistanceToPlaneCode(xsize, offset);
        const double distance_cost =
            prev_cost + GetDistanceCost(cost_model, code);
        int k;
        for (k = 1; k < len; ++k) {
          const double cost_val = distance_cost + GetLengthCost(cost_model, k);
          if (cost[i + k] > cost_val) {
            cost[i + k] = (float)cost_val;
            dist_array[i + k] = k + 1;
          }
        }
        // This if is for speedup only. It roughly doubles the speed, and
        // makes compression worse by .1 %.
        if (len >= 128 && code <= min_distance_code) {
          // Long copy for short distances, let's skip the middle
          // lookups for better copies.
          // 1) insert the hashes.
          if (use_color_cache) {
            for (k = 0; k < len; ++k) {
              VP8LColorCacheInsert(&hashers, argb[i + k]);
            }
          }
          // 2) Add to the hash_chain (but cannot add the last pixel)
          {
            const int last = (len + i < pix_count - 1) ? len + i
                                                       : pix_count - 1;
            for (k = i; k < last; ++k) {
              HashChainInsert(hash_chain, &argb[k], k);
            }
          }
          // 3) jump.
          i += len - 1;  // for loop does ++i, thus -1 here.
          goto next_symbol;
        }
      }
    }
    if (i < pix_count - 1) {
      HashChainInsert(hash_chain, &argb[i], i);
    }
    {
      // inserting a literal pixel
      double cost_val = prev_cost;
      if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
        const int ix = VP8LColorCacheGetIndex(&hashers, argb[i]);
        cost_val += GetCacheCost(cost_model, ix) * mul0;
      } else {
        if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
        cost_val += GetLiteralCost(cost_model, argb[i]) * mul1;
      }
      if (cost[i] > cost_val) {
        cost[i] = (float)cost_val;
        dist_array[i] = 1;  // only one is inserted.
      }
    }
 next_symbol: ;
  }
  // Last pixel still to do, it can only be a single step if not reached
  // through cheaper means already.
  ok = !refs->error_;
Error:
  if (cc_init) VP8LColorCacheClear(&hashers);
  WebPSafeFree(cost_model);
  WebPSafeFree(cost);
  return ok;
}

// We pack the path at the end of *dist_array and return
// a pointer to this part of the array. Example:
// dist_array = [1x2xx3x2] => packed [1x2x1232], chosen_path = [1232]
static void TraceBackwards(uint32_t* const dist_array,
                           int dist_array_size,
                           uint32_t** const chosen_path,
                           int* const chosen_path_size) {
  uint32_t* path = dist_array + dist_array_size;
  uint32_t* cur = dist_array + dist_array_size - 1;
  while (cur >= dist_array) {
    const int k = *cur;
    --path;
    *path = k;
    cur -= k;
  }
  *chosen_path = path;
  *chosen_path_size = (int)(dist_array + dist_array_size - path);
}

static int BackwardReferencesHashChainFollowChosenPath(
    int xsize, int ysize, const uint32_t* const argb,
    int quality, int cache_bits,
    const uint32_t* const chosen_path, int chosen_path_size,
    VP8LHashChain* const hash_chain,
    VP8LBackwardRefs* const refs) {
  const int pix_count = xsize * ysize;
  const int use_color_cache = (cache_bits > 0);
  int size = 0;
  int i = 0;
  int k;
  int ix;
  int ok = 0;
  int cc_init = 0;
  int window_size = WINDOW_SIZE;
  int iter_pos = 1;
  int iter_limit = -1;
  VP8LColorCache hashers;

  if (use_color_cache) {
    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
    if (!cc_init) goto Error;
  }

  ClearBackwardRefs(refs);
  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
                                &window_size, &iter_pos, &iter_limit);
  HashChainInit(hash_chain);
  for (ix = 0; ix < chosen_path_size; ++ix, ++size) {
    int offset = 0;
    int len = 0;
    int max_len = chosen_path[ix];
    if (max_len != 1) {
      HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
                        window_size, iter_pos, iter_limit,
                        &offset, &len);
      assert(len == max_len);
      BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(offset, len));
      if (use_color_cache) {
        for (k = 0; k < len; ++k) {
          VP8LColorCacheInsert(&hashers, argb[i + k]);
        }
      }
      {
        const int last = (len < pix_count - 1 - i) ? len : pix_count - 1 - i;
        for (k = 0; k < last; ++k) {
          HashChainInsert(hash_chain, &argb[i + k], i + k);
        }
      }
      i += len;
    } else {
      PixOrCopy v;
      if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
        // push pixel as a color cache index
        const int idx = VP8LColorCacheGetIndex(&hashers, argb[i]);
        v = PixOrCopyCreateCacheIdx(idx);
      } else {
        if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
        v = PixOrCopyCreateLiteral(argb[i]);
      }
      BackwardRefsCursorAdd(refs, v);
      if (i + 1 < pix_count) {
        HashChainInsert(hash_chain, &argb[i], i);
      }
      ++i;
    }
  }
  ok = !refs->error_;
Error:
  if (cc_init) VP8LColorCacheClear(&hashers);
  return ok;
}

// Returns 1 on success.
static int BackwardReferencesTraceBackwards(int xsize, int ysize,
                                            int recursive_cost_model,
                                            const uint32_t* const argb,
                                            int quality, int cache_bits,
                                            VP8LHashChain* const hash_chain,
                                            VP8LBackwardRefs* const refs) {
  int ok = 0;
  const int dist_array_size = xsize * ysize;
  uint32_t* chosen_path = NULL;
  int chosen_path_size = 0;
  uint32_t* dist_array =
      (uint32_t*)WebPSafeMalloc(dist_array_size, sizeof(*dist_array));

  if (dist_array == NULL) goto Error;

  if (!BackwardReferencesHashChainDistanceOnly(
      xsize, ysize, recursive_cost_model, argb, quality, cache_bits, hash_chain,
      refs, dist_array)) {
    goto Error;
  }
  TraceBackwards(dist_array, dist_array_size, &chosen_path, &chosen_path_size);
  if (!BackwardReferencesHashChainFollowChosenPath(
      xsize, ysize, argb, quality, cache_bits, chosen_path, chosen_path_size,
      hash_chain, refs)) {
    goto Error;
  }
  ok = 1;
 Error:
  WebPSafeFree(dist_array);
  return ok;
}

static void BackwardReferences2DLocality(int xsize,
                                         const VP8LBackwardRefs* const refs) {
  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
  while (VP8LRefsCursorOk(&c)) {
    if (PixOrCopyIsCopy(c.cur_pos)) {
      const int dist = c.cur_pos->argb_or_distance;
      const int transformed_dist = DistanceToPlaneCode(xsize, dist);
      c.cur_pos->argb_or_distance = transformed_dist;
    }
    VP8LRefsCursorNext(&c);
  }
}

VP8LBackwardRefs* VP8LGetBackwardReferences(
    int width, int height, const uint32_t* const argb, int quality,
    int cache_bits, int use_2d_locality, VP8LHashChain* const hash_chain,
    VP8LBackwardRefs refs_array[2]) {
  int lz77_is_useful;
  const int num_pix = width * height;
  VP8LBackwardRefs* best = NULL;
  VP8LBackwardRefs* const refs_lz77 = &refs_array[0];
  VP8LBackwardRefs* const refs_rle = &refs_array[1];

  if (!BackwardReferencesHashChain(width, height, argb, cache_bits, quality,
                                   hash_chain, refs_lz77)) {
    return NULL;
  }
  if (!BackwardReferencesRle(width, height, argb, refs_rle)) {
    return NULL;
  }

  {
    double bit_cost_lz77, bit_cost_rle;
    VP8LHistogram* const histo = VP8LAllocateHistogram(cache_bits);
    if (histo == NULL) return NULL;
    // Evaluate LZ77 coding.
    VP8LHistogramCreate(histo, refs_lz77, cache_bits);
    bit_cost_lz77 = VP8LHistogramEstimateBits(histo);
    // Evaluate RLE coding.
    VP8LHistogramCreate(histo, refs_rle, cache_bits);
    bit_cost_rle = VP8LHistogramEstimateBits(histo);
    // Decide if LZ77 is useful.
    lz77_is_useful = (bit_cost_lz77 < bit_cost_rle);
    VP8LFreeHistogram(histo);
  }

  // Choose appropriate backward reference.
  if (lz77_is_useful) {
    // TraceBackwards is costly. Don't execute it at lower quality.
    const int try_lz77_trace_backwards = (quality >= 25);
    best = refs_lz77;   // default guess: lz77 is better
    if (try_lz77_trace_backwards) {
      // Set recursion level for large images using a color cache.
      const int recursion_level =
          (num_pix < 320 * 200) && (cache_bits > 0) ? 1 : 0;
      VP8LBackwardRefs* const refs_trace = &refs_array[1];
      ClearBackwardRefs(refs_trace);
      if (BackwardReferencesTraceBackwards(width, height, recursion_level, argb,
                                           quality, cache_bits, hash_chain,
                                           refs_trace)) {
        best = refs_trace;
      }
    }
  } else {
    best = refs_rle;
  }

  if (use_2d_locality) BackwardReferences2DLocality(width, best);

  return best;
}

// Returns entropy for the given cache bits.
static double ComputeCacheEntropy(const uint32_t* const argb,
                                  int xsize, int ysize,
                                  const VP8LBackwardRefs* const refs,
                                  int cache_bits) {
  int pixel_index = 0;
  uint32_t k;
  const int use_color_cache = (cache_bits > 0);
  int cc_init = 0;
  double entropy = MAX_ENTROPY;
  const double kSmallPenaltyForLargeCache = 4.0;
  VP8LColorCache hashers;
  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
  VP8LHistogram* histo = VP8LAllocateHistogram(cache_bits);
  if (histo == NULL) goto Error;

  if (use_color_cache) {
    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
    if (!cc_init) goto Error;
  }

  while (VP8LRefsCursorOk(&c)) {
    const PixOrCopy* const v = c.cur_pos;
    if (PixOrCopyIsLiteral(v)) {
      if (use_color_cache &&
          VP8LColorCacheContains(&hashers, argb[pixel_index])) {
        // push pixel as a cache index
        const int ix = VP8LColorCacheGetIndex(&hashers, argb[pixel_index]);
        const PixOrCopy token = PixOrCopyCreateCacheIdx(ix);
        VP8LHistogramAddSinglePixOrCopy(histo, &token);
      } else {
        VP8LHistogramAddSinglePixOrCopy(histo, v);
      }
    } else {
      VP8LHistogramAddSinglePixOrCopy(histo, v);
    }
    if (use_color_cache) {
      for (k = 0; k < PixOrCopyLength(v); ++k) {
        VP8LColorCacheInsert(&hashers, argb[pixel_index + k]);
      }
    }
    pixel_index += PixOrCopyLength(v);
    VP8LRefsCursorNext(&c);
  }
  assert(pixel_index == xsize * ysize);
  (void)xsize;  // xsize is not used in non-debug compilations otherwise.
  (void)ysize;  // ysize is not used in non-debug compilations otherwise.
  entropy = VP8LHistogramEstimateBits(histo) +
      kSmallPenaltyForLargeCache * cache_bits;
 Error:
  if (cc_init) VP8LColorCacheClear(&hashers);
  VP8LFreeHistogram(histo);
  return entropy;
}

// *best_cache_bits will contain how many bits are to be used for a color cache.
// Returns 0 in case of memory error.
int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
                                      int xsize, int ysize, int quality,
                                      VP8LHashChain* const hash_chain,
                                      VP8LBackwardRefs* const refs,
                                      int* const best_cache_bits) {
  int eval_low = 1;
  int eval_high = 1;
  double entropy_low = MAX_ENTROPY;
  double entropy_high = MAX_ENTROPY;
  int cache_bits_low = 0;
  int cache_bits_high = MAX_COLOR_CACHE_BITS;

  if (!BackwardReferencesHashChain(xsize, ysize, argb, 0, quality, hash_chain,
                                   refs)) {
    return 0;
  }
  // Do a binary search to find the optimal entropy for cache_bits.
  while (cache_bits_high - cache_bits_low > 1) {
    if (eval_low) {
      entropy_low =
          ComputeCacheEntropy(argb, xsize, ysize, refs, cache_bits_low);
      eval_low = 0;
    }
    if (eval_high) {
      entropy_high =
          ComputeCacheEntropy(argb, xsize, ysize, refs, cache_bits_high);
      eval_high = 0;
    }
    if (entropy_high < entropy_low) {
      *best_cache_bits = cache_bits_high;
      cache_bits_low = (cache_bits_low + cache_bits_high) / 2;
      eval_low = 1;
    } else {
      *best_cache_bits = cache_bits_low;
      cache_bits_high = (cache_bits_low + cache_bits_high) / 2;
      eval_high = 1;
    }
  }
  return 1;
}