Adding upstream version 110.0.1.upstream/110.0.1upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

29 files changed, 15359 insertions, 0 deletions

diff --git a/media/libwebp/src/enc/alpha_enc.c b/media/libwebp/src/enc/alpha_enc.c
new file mode 100644
index 0000000000..f7c02690e3
--- /dev/null
+++ b/media/libwebp/src/enc/alpha_enc.c
@@ -0,0 +1,443 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Alpha-plane compression.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include "src/enc/vp8i_enc.h"
+#include "src/dsp/dsp.h"
+#include "src/utils/filters_utils.h"
+#include "src/utils/quant_levels_utils.h"
+#include "src/utils/utils.h"
+#include "src/webp/format_constants.h"
+
+// -----------------------------------------------------------------------------
+// Encodes the given alpha data via specified compression method 'method'.
+// The pre-processing (quantization) is performed if 'quality' is less than 100.
+// For such cases, the encoding is lossy. The valid range is [0, 100] for
+// 'quality' and [0, 1] for 'method':
+//   'method = 0' - No compression;
+//   'method = 1' - Use lossless coder on the alpha plane only
+// 'filter' values [0, 4] correspond to prediction modes none, horizontal,
+// vertical & gradient filters. The prediction mode 4 will try all the
+// prediction modes 0 to 3 and pick the best one.
+// 'effort_level': specifies how much effort must be spent to try and reduce
+//  the compressed output size. In range 0 (quick) to 6 (slow).
+//
+// 'output' corresponds to the buffer containing compressed alpha data.
+//          This buffer is allocated by this method and caller should call
+//          WebPSafeFree(*output) when done.
+// 'output_size' corresponds to size of this compressed alpha buffer.
+//
+// Returns 1 on successfully encoding the alpha and
+//         0 if either:
+//           invalid quality or method, or
+//           memory allocation for the compressed data fails.
+
+#include "src/enc/vp8li_enc.h"
+
+static int EncodeLossless(const uint8_t* const data, int width, int height,
+                          int effort_level,  // in [0..6] range
+                          int use_quality_100, VP8LBitWriter* const bw,
+                          WebPAuxStats* const stats) {
+  int ok = 0;
+  WebPConfig config;
+  WebPPicture picture;
+
+  WebPPictureInit(&picture);
+  picture.width = width;
+  picture.height = height;
+  picture.use_argb = 1;
+  picture.stats = stats;
+  if (!WebPPictureAlloc(&picture)) return 0;
+
+  // Transfer the alpha values to the green channel.
+  WebPDispatchAlphaToGreen(data, width, picture.width, picture.height,
+                           picture.argb, picture.argb_stride);
+
+  WebPConfigInit(&config);
+  config.lossless = 1;
+  // Enable exact, or it would alter RGB values of transparent alpha, which is
+  // normally OK but not here since we are not encoding the input image but  an
+  // internal encoding-related image containing necessary exact information in
+  // RGB channels.
+  config.exact = 1;
+  config.method = effort_level;  // impact is very small
+  // Set a low default quality for encoding alpha. Ensure that Alpha quality at
+  // lower methods (3 and below) is less than the threshold for triggering
+  // costly 'BackwardReferencesTraceBackwards'.
+  // If the alpha quality is set to 100 and the method to 6, allow for a high
+  // lossless quality to trigger the cruncher.
+  config.quality =
+      (use_quality_100 && effort_level == 6) ? 100 : 8.f * effort_level;
+  assert(config.quality >= 0 && config.quality <= 100.f);
+
+  // TODO(urvang): Temporary fix to avoid generating images that trigger
+  // a decoder bug related to alpha with color cache.
+  // See: https://code.google.com/p/webp/issues/detail?id=239
+  // Need to re-enable this later.
+  ok = VP8LEncodeStream(&config, &picture, bw, /*use_cache=*/0);
+  WebPPictureFree(&picture);
+  ok = ok && !bw->error_;
+  if (!ok) {
+    VP8LBitWriterWipeOut(bw);
+    return 0;
+  }
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+
+// Small struct to hold the result of a filter mode compression attempt.
+typedef struct {
+  size_t score;
+  VP8BitWriter bw;
+  WebPAuxStats stats;
+} FilterTrial;
+
+// This function always returns an initialized 'bw' object, even upon error.
+static int EncodeAlphaInternal(const uint8_t* const data, int width, int height,
+                               int method, int filter, int reduce_levels,
+                               int effort_level,  // in [0..6] range
+                               uint8_t* const tmp_alpha,
+                               FilterTrial* result) {
+  int ok = 0;
+  const uint8_t* alpha_src;
+  WebPFilterFunc filter_func;
+  uint8_t header;
+  const size_t data_size = width * height;
+  const uint8_t* output = NULL;
+  size_t output_size = 0;
+  VP8LBitWriter tmp_bw;
+
+  assert((uint64_t)data_size == (uint64_t)width * height);  // as per spec
+  assert(filter >= 0 && filter < WEBP_FILTER_LAST);
+  assert(method >= ALPHA_NO_COMPRESSION);
+  assert(method <= ALPHA_LOSSLESS_COMPRESSION);
+  assert(sizeof(header) == ALPHA_HEADER_LEN);
+
+  filter_func = WebPFilters[filter];
+  if (filter_func != NULL) {
+    filter_func(data, width, height, width, tmp_alpha);
+    alpha_src = tmp_alpha;
+  }  else {
+    alpha_src = data;
+  }
+
+  if (method != ALPHA_NO_COMPRESSION) {
+    ok = VP8LBitWriterInit(&tmp_bw, data_size >> 3);
+    ok = ok && EncodeLossless(alpha_src, width, height, effort_level,
+                              !reduce_levels, &tmp_bw, &result->stats);
+    if (ok) {
+      output = VP8LBitWriterFinish(&tmp_bw);
+      output_size = VP8LBitWriterNumBytes(&tmp_bw);
+      if (output_size > data_size) {
+        // compressed size is larger than source! Revert to uncompressed mode.
+        method = ALPHA_NO_COMPRESSION;
+        VP8LBitWriterWipeOut(&tmp_bw);
+      }
+    } else {
+      VP8LBitWriterWipeOut(&tmp_bw);
+      return 0;
+    }
+  }
+
+  if (method == ALPHA_NO_COMPRESSION) {
+    output = alpha_src;
+    output_size = data_size;
+    ok = 1;
+  }
+
+  // Emit final result.
+  header = method | (filter << 2);
+  if (reduce_levels) header |= ALPHA_PREPROCESSED_LEVELS << 4;
+
+  VP8BitWriterInit(&result->bw, ALPHA_HEADER_LEN + output_size);
+  ok = ok && VP8BitWriterAppend(&result->bw, &header, ALPHA_HEADER_LEN);
+  ok = ok && VP8BitWriterAppend(&result->bw, output, output_size);
+
+  if (method != ALPHA_NO_COMPRESSION) {
+    VP8LBitWriterWipeOut(&tmp_bw);
+  }
+  ok = ok && !result->bw.error_;
+  result->score = VP8BitWriterSize(&result->bw);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+
+static int GetNumColors(const uint8_t* data, int width, int height,
+                        int stride) {
+  int j;
+  int colors = 0;
+  uint8_t color[256] = { 0 };
+
+  for (j = 0; j < height; ++j) {
+    int i;
+    const uint8_t* const p = data + j * stride;
+    for (i = 0; i < width; ++i) {
+      color[p[i]] = 1;
+    }
+  }
+  for (j = 0; j < 256; ++j) {
+    if (color[j] > 0) ++colors;
+  }
+  return colors;
+}
+
+#define FILTER_TRY_NONE (1 << WEBP_FILTER_NONE)
+#define FILTER_TRY_ALL ((1 << WEBP_FILTER_LAST) - 1)
+
+// Given the input 'filter' option, return an OR'd bit-set of filters to try.
+static uint32_t GetFilterMap(const uint8_t* alpha, int width, int height,
+                             int filter, int effort_level) {
+  uint32_t bit_map = 0U;
+  if (filter == WEBP_FILTER_FAST) {
+    // Quick estimate of the best candidate.
+    int try_filter_none = (effort_level > 3);
+    const int kMinColorsForFilterNone = 16;
+    const int kMaxColorsForFilterNone = 192;
+    const int num_colors = GetNumColors(alpha, width, height, width);
+    // For low number of colors, NONE yields better compression.
+    filter = (num_colors <= kMinColorsForFilterNone)
+        ? WEBP_FILTER_NONE
+        : WebPEstimateBestFilter(alpha, width, height, width);
+    bit_map |= 1 << filter;
+    // For large number of colors, try FILTER_NONE in addition to the best
+    // filter as well.
+    if (try_filter_none || num_colors > kMaxColorsForFilterNone) {
+      bit_map |= FILTER_TRY_NONE;
+    }
+  } else if (filter == WEBP_FILTER_NONE) {
+    bit_map = FILTER_TRY_NONE;
+  } else {  // WEBP_FILTER_BEST -> try all
+    bit_map = FILTER_TRY_ALL;
+  }
+  return bit_map;
+}
+
+static void InitFilterTrial(FilterTrial* const score) {
+  score->score = (size_t)~0U;
+  VP8BitWriterInit(&score->bw, 0);
+}
+
+static int ApplyFiltersAndEncode(const uint8_t* alpha, int width, int height,
+                                 size_t data_size, int method, int filter,
+                                 int reduce_levels, int effort_level,
+                                 uint8_t** const output,
+                                 size_t* const output_size,
+                                 WebPAuxStats* const stats) {
+  int ok = 1;
+  FilterTrial best;
+  uint32_t try_map =
+      GetFilterMap(alpha, width, height, filter, effort_level);
+  InitFilterTrial(&best);
+
+  if (try_map != FILTER_TRY_NONE) {
+    uint8_t* filtered_alpha =  (uint8_t*)WebPSafeMalloc(1ULL, data_size);
+    if (filtered_alpha == NULL) return 0;
+
+    for (filter = WEBP_FILTER_NONE; ok && try_map; ++filter, try_map >>= 1) {
+      if (try_map & 1) {
+        FilterTrial trial;
+        ok = EncodeAlphaInternal(alpha, width, height, method, filter,
+                                 reduce_levels, effort_level, filtered_alpha,
+                                 &trial);
+        if (ok && trial.score < best.score) {
+          VP8BitWriterWipeOut(&best.bw);
+          best = trial;
+        } else {
+          VP8BitWriterWipeOut(&trial.bw);
+        }
+      }
+    }
+    WebPSafeFree(filtered_alpha);
+  } else {
+    ok = EncodeAlphaInternal(alpha, width, height, method, WEBP_FILTER_NONE,
+                             reduce_levels, effort_level, NULL, &best);
+  }
+  if (ok) {
+#if !defined(WEBP_DISABLE_STATS)
+    if (stats != NULL) {
+      stats->lossless_features = best.stats.lossless_features;
+      stats->histogram_bits = best.stats.histogram_bits;
+      stats->transform_bits = best.stats.transform_bits;
+      stats->cache_bits = best.stats.cache_bits;
+      stats->palette_size = best.stats.palette_size;
+      stats->lossless_size = best.stats.lossless_size;
+      stats->lossless_hdr_size = best.stats.lossless_hdr_size;
+      stats->lossless_data_size = best.stats.lossless_data_size;
+    }
+#else
+    (void)stats;
+#endif
+    *output_size = VP8BitWriterSize(&best.bw);
+    *output = VP8BitWriterBuf(&best.bw);
+  } else {
+    VP8BitWriterWipeOut(&best.bw);
+  }
+  return ok;
+}
+
+static int EncodeAlpha(VP8Encoder* const enc,
+                       int quality, int method, int filter,
+                       int effort_level,
+                       uint8_t** const output, size_t* const output_size) {
+  const WebPPicture* const pic = enc->pic_;
+  const int width = pic->width;
+  const int height = pic->height;
+
+  uint8_t* quant_alpha = NULL;
+  const size_t data_size = width * height;
+  uint64_t sse = 0;
+  int ok = 1;
+  const int reduce_levels = (quality < 100);
+
+  // quick correctness checks
+  assert((uint64_t)data_size == (uint64_t)width * height);  // as per spec
+  assert(enc != NULL && pic != NULL && pic->a != NULL);
+  assert(output != NULL && output_size != NULL);
+  assert(width > 0 && height > 0);
+  assert(pic->a_stride >= width);
+  assert(filter >= WEBP_FILTER_NONE && filter <= WEBP_FILTER_FAST);
+
+  if (quality < 0 || quality > 100) {
+    return 0;
+  }
+
+  if (method < ALPHA_NO_COMPRESSION || method > ALPHA_LOSSLESS_COMPRESSION) {
+    return 0;
+  }
+
+  if (method == ALPHA_NO_COMPRESSION) {
+    // Don't filter, as filtering will make no impact on compressed size.
+    filter = WEBP_FILTER_NONE;
+  }
+
+  quant_alpha = (uint8_t*)WebPSafeMalloc(1ULL, data_size);
+  if (quant_alpha == NULL) {
+    return 0;
+  }
+
+  // Extract alpha data (width x height) from raw_data (stride x height).
+  WebPCopyPlane(pic->a, pic->a_stride, quant_alpha, width, width, height);
+
+  if (reduce_levels) {  // No Quantization required for 'quality = 100'.
+    // 16 alpha levels gives quite a low MSE w.r.t original alpha plane hence
+    // mapped to moderate quality 70. Hence Quality:[0, 70] -> Levels:[2, 16]
+    // and Quality:]70, 100] -> Levels:]16, 256].
+    const int alpha_levels = (quality <= 70) ? (2 + quality / 5)
+                                             : (16 + (quality - 70) * 8);
+    ok = QuantizeLevels(quant_alpha, width, height, alpha_levels, &sse);
+  }
+
+  if (ok) {
+    VP8FiltersInit();
+    ok = ApplyFiltersAndEncode(quant_alpha, width, height, data_size, method,
+                               filter, reduce_levels, effort_level, output,
+                               output_size, pic->stats);
+#if !defined(WEBP_DISABLE_STATS)
+    if (pic->stats != NULL) {  // need stats?
+      pic->stats->coded_size += (int)(*output_size);
+      enc->sse_[3] = sse;
+    }
+#endif
+  }
+
+  WebPSafeFree(quant_alpha);
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+// Main calls
+
+static int CompressAlphaJob(void* arg1, void* unused) {
+  VP8Encoder* const enc = (VP8Encoder*)arg1;
+  const WebPConfig* config = enc->config_;
+  uint8_t* alpha_data = NULL;
+  size_t alpha_size = 0;
+  const int effort_level = config->method;  // maps to [0..6]
+  const WEBP_FILTER_TYPE filter =
+      (config->alpha_filtering == 0) ? WEBP_FILTER_NONE :
+      (config->alpha_filtering == 1) ? WEBP_FILTER_FAST :
+                                       WEBP_FILTER_BEST;
+  if (!EncodeAlpha(enc, config->alpha_quality, config->alpha_compression,
+                   filter, effort_level, &alpha_data, &alpha_size)) {
+    return 0;
+  }
+  if (alpha_size != (uint32_t)alpha_size) {  // Soundness check.
+    WebPSafeFree(alpha_data);
+    return 0;
+  }
+  enc->alpha_data_size_ = (uint32_t)alpha_size;
+  enc->alpha_data_ = alpha_data;
+  (void)unused;
+  return 1;
+}
+
+void VP8EncInitAlpha(VP8Encoder* const enc) {
+  WebPInitAlphaProcessing();
+  enc->has_alpha_ = WebPPictureHasTransparency(enc->pic_);
+  enc->alpha_data_ = NULL;
+  enc->alpha_data_size_ = 0;
+  if (enc->thread_level_ > 0) {
+    WebPWorker* const worker = &enc->alpha_worker_;
+    WebPGetWorkerInterface()->Init(worker);
+    worker->data1 = enc;
+    worker->data2 = NULL;
+    worker->hook = CompressAlphaJob;
+  }
+}
+
+int VP8EncStartAlpha(VP8Encoder* const enc) {
+  if (enc->has_alpha_) {
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      // Makes sure worker is good to go.
+      if (!WebPGetWorkerInterface()->Reset(worker)) {
+        return 0;
+      }
+      WebPGetWorkerInterface()->Launch(worker);
+      return 1;
+    } else {
+      return CompressAlphaJob(enc, NULL);   // just do the job right away
+    }
+  }
+  return 1;
+}
+
+int VP8EncFinishAlpha(VP8Encoder* const enc) {
+  if (enc->has_alpha_) {
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      if (!WebPGetWorkerInterface()->Sync(worker)) return 0;  // error
+    }
+  }
+  return WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
+}
+
+int VP8EncDeleteAlpha(VP8Encoder* const enc) {
+  int ok = 1;
+  if (enc->thread_level_ > 0) {
+    WebPWorker* const worker = &enc->alpha_worker_;
+    // finish anything left in flight
+    ok = WebPGetWorkerInterface()->Sync(worker);
+    // still need to end the worker, even if !ok
+    WebPGetWorkerInterface()->End(worker);
+  }
+  WebPSafeFree(enc->alpha_data_);
+  enc->alpha_data_ = NULL;
+  enc->alpha_data_size_ = 0;
+  enc->has_alpha_ = 0;
+  return ok;
+}
diff --git a/media/libwebp/src/enc/analysis_enc.c b/media/libwebp/src/enc/analysis_enc.c
new file mode 100644
index 0000000000..ebb784261c
--- /dev/null
+++ b/media/libwebp/src/enc/analysis_enc.c
@@ -0,0 +1,475 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Macroblock analysis
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+#include <string.h>
+#include <assert.h>
+
+#include "src/enc/vp8i_enc.h"
+#include "src/enc/cost_enc.h"
+#include "src/utils/utils.h"
+
+#define MAX_ITERS_K_MEANS  6
+
+//------------------------------------------------------------------------------
+// Smooth the segment map by replacing isolated block by the majority of its
+// neighbours.
+
+static void SmoothSegmentMap(VP8Encoder* const enc) {
+  int n, x, y;
+  const int w = enc->mb_w_;
+  const int h = enc->mb_h_;
+  const int majority_cnt_3_x_3_grid = 5;
+  uint8_t* const tmp = (uint8_t*)WebPSafeMalloc(w * h, sizeof(*tmp));
+  assert((uint64_t)(w * h) == (uint64_t)w * h);   // no overflow, as per spec
+
+  if (tmp == NULL) return;
+  for (y = 1; y < h - 1; ++y) {
+    for (x = 1; x < w - 1; ++x) {
+      int cnt[NUM_MB_SEGMENTS] = { 0 };
+      const VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
+      int majority_seg = mb->segment_;
+      // Check the 8 neighbouring segment values.
+      cnt[mb[-w - 1].segment_]++;  // top-left
+      cnt[mb[-w + 0].segment_]++;  // top
+      cnt[mb[-w + 1].segment_]++;  // top-right
+      cnt[mb[   - 1].segment_]++;  // left
+      cnt[mb[   + 1].segment_]++;  // right
+      cnt[mb[ w - 1].segment_]++;  // bottom-left
+      cnt[mb[ w + 0].segment_]++;  // bottom
+      cnt[mb[ w + 1].segment_]++;  // bottom-right
+      for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
+        if (cnt[n] >= majority_cnt_3_x_3_grid) {
+          majority_seg = n;
+          break;
+        }
+      }
+      tmp[x + y * w] = majority_seg;
+    }
+  }
+  for (y = 1; y < h - 1; ++y) {
+    for (x = 1; x < w - 1; ++x) {
+      VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
+      mb->segment_ = tmp[x + y * w];
+    }
+  }
+  WebPSafeFree(tmp);
+}
+
+//------------------------------------------------------------------------------
+// set segment susceptibility alpha_ / beta_
+
+static WEBP_INLINE int clip(int v, int m, int M) {
+  return (v < m) ? m : (v > M) ? M : v;
+}
+
+static void SetSegmentAlphas(VP8Encoder* const enc,
+                             const int centers[NUM_MB_SEGMENTS],
+                             int mid) {
+  const int nb = enc->segment_hdr_.num_segments_;
+  int min = centers[0], max = centers[0];
+  int n;
+
+  if (nb > 1) {
+    for (n = 0; n < nb; ++n) {
+      if (min > centers[n]) min = centers[n];
+      if (max < centers[n]) max = centers[n];
+    }
+  }
+  if (max == min) max = min + 1;
+  assert(mid <= max && mid >= min);
+  for (n = 0; n < nb; ++n) {
+    const int alpha = 255 * (centers[n] - mid) / (max - min);
+    const int beta = 255 * (centers[n] - min) / (max - min);
+    enc->dqm_[n].alpha_ = clip(alpha, -127, 127);
+    enc->dqm_[n].beta_ = clip(beta, 0, 255);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Compute susceptibility based on DCT-coeff histograms:
+// the higher, the "easier" the macroblock is to compress.
+
+#define MAX_ALPHA 255                // 8b of precision for susceptibilities.
+#define ALPHA_SCALE (2 * MAX_ALPHA)  // scaling factor for alpha.
+#define DEFAULT_ALPHA (-1)
+#define IS_BETTER_ALPHA(alpha, best_alpha) ((alpha) > (best_alpha))
+
+static int FinalAlphaValue(int alpha) {
+  alpha = MAX_ALPHA - alpha;
+  return clip(alpha, 0, MAX_ALPHA);
+}
+
+static int GetAlpha(const VP8Histogram* const histo) {
+  // 'alpha' will later be clipped to [0..MAX_ALPHA] range, clamping outer
+  // values which happen to be mostly noise. This leaves the maximum precision
+  // for handling the useful small values which contribute most.
+  const int max_value = histo->max_value;
+  const int last_non_zero = histo->last_non_zero;
+  const int alpha =
+      (max_value > 1) ? ALPHA_SCALE * last_non_zero / max_value : 0;
+  return alpha;
+}
+
+static void InitHistogram(VP8Histogram* const histo) {
+  histo->max_value = 0;
+  histo->last_non_zero = 1;
+}
+
+//------------------------------------------------------------------------------
+// Simplified k-Means, to assign Nb segments based on alpha-histogram
+
+static void AssignSegments(VP8Encoder* const enc,
+                           const int alphas[MAX_ALPHA + 1]) {
+  // 'num_segments_' is previously validated and <= NUM_MB_SEGMENTS, but an
+  // explicit check is needed to avoid spurious warning about 'n + 1' exceeding
+  // array bounds of 'centers' with some compilers (noticed with gcc-4.9).
+  const int nb = (enc->segment_hdr_.num_segments_ < NUM_MB_SEGMENTS) ?
+                 enc->segment_hdr_.num_segments_ : NUM_MB_SEGMENTS;
+  int centers[NUM_MB_SEGMENTS];
+  int weighted_average = 0;
+  int map[MAX_ALPHA + 1];
+  int a, n, k;
+  int min_a = 0, max_a = MAX_ALPHA, range_a;
+  // 'int' type is ok for histo, and won't overflow
+  int accum[NUM_MB_SEGMENTS], dist_accum[NUM_MB_SEGMENTS];
+
+  assert(nb >= 1);
+  assert(nb <= NUM_MB_SEGMENTS);
+
+  // bracket the input
+  for (n = 0; n <= MAX_ALPHA && alphas[n] == 0; ++n) {}
+  min_a = n;
+  for (n = MAX_ALPHA; n > min_a && alphas[n] == 0; --n) {}
+  max_a = n;
+  range_a = max_a - min_a;
+
+  // Spread initial centers evenly
+  for (k = 0, n = 1; k < nb; ++k, n += 2) {
+    assert(n < 2 * nb);
+    centers[k] = min_a + (n * range_a) / (2 * nb);
+  }
+
+  for (k = 0; k < MAX_ITERS_K_MEANS; ++k) {     // few iters are enough
+    int total_weight;
+    int displaced;
+    // Reset stats
+    for (n = 0; n < nb; ++n) {
+      accum[n] = 0;
+      dist_accum[n] = 0;
+    }
+    // Assign nearest center for each 'a'
+    n = 0;    // track the nearest center for current 'a'
+    for (a = min_a; a <= max_a; ++a) {
+      if (alphas[a]) {
+        while (n + 1 < nb && abs(a - centers[n + 1]) < abs(a - centers[n])) {
+          n++;
+        }
+        map[a] = n;
+        // accumulate contribution into best centroid
+        dist_accum[n] += a * alphas[a];
+        accum[n] += alphas[a];
+      }
+    }
+    // All point are classified. Move the centroids to the
+    // center of their respective cloud.
+    displaced = 0;
+    weighted_average = 0;
+    total_weight = 0;
+    for (n = 0; n < nb; ++n) {
+      if (accum[n]) {
+        const int new_center = (dist_accum[n] + accum[n] / 2) / accum[n];
+        displaced += abs(centers[n] - new_center);
+        centers[n] = new_center;
+        weighted_average += new_center * accum[n];
+        total_weight += accum[n];
+      }
+    }
+    weighted_average = (weighted_average + total_weight / 2) / total_weight;
+    if (displaced < 5) break;   // no need to keep on looping...
+  }
+
+  // Map each original value to the closest centroid
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    VP8MBInfo* const mb = &enc->mb_info_[n];
+    const int alpha = mb->alpha_;
+    mb->segment_ = map[alpha];
+    mb->alpha_ = centers[map[alpha]];  // for the record.
+  }
+
+  if (nb > 1) {
+    const int smooth = (enc->config_->preprocessing & 1);
+    if (smooth) SmoothSegmentMap(enc);
+  }
+
+  SetSegmentAlphas(enc, centers, weighted_average);  // pick some alphas.
+}
+
+//------------------------------------------------------------------------------
+// Macroblock analysis: collect histogram for each mode, deduce the maximal
+// susceptibility and set best modes for this macroblock.
+// Segment assignment is done later.
+
+// Number of modes to inspect for alpha_ evaluation. We don't need to test all
+// the possible modes during the analysis phase: we risk falling into a local
+// optimum, or be subject to boundary effect
+#define MAX_INTRA16_MODE 2
+#define MAX_INTRA4_MODE  2
+#define MAX_UV_MODE      2
+
+static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
+  const int max_mode = MAX_INTRA16_MODE;
+  int mode;
+  int best_alpha = DEFAULT_ALPHA;
+  int best_mode = 0;
+
+  VP8MakeLuma16Preds(it);
+  for (mode = 0; mode < max_mode; ++mode) {
+    VP8Histogram histo;
+    int alpha;
+
+    InitHistogram(&histo);
+    VP8CollectHistogram(it->yuv_in_ + Y_OFF_ENC,
+                        it->yuv_p_ + VP8I16ModeOffsets[mode],
+                        0, 16, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
+      best_alpha = alpha;
+      best_mode = mode;
+    }
+  }
+  VP8SetIntra16Mode(it, best_mode);
+  return best_alpha;
+}
+
+static int FastMBAnalyze(VP8EncIterator* const it) {
+  // Empirical cut-off value, should be around 16 (~=block size). We use the
+  // [8-17] range and favor intra4 at high quality, intra16 for low quality.
+  const int q = (int)it->enc_->config_->quality;
+  const uint32_t kThreshold = 8 + (17 - 8) * q / 100;
+  int k;
+  uint32_t dc[16], m, m2;
+  for (k = 0; k < 16; k += 4) {
+    VP8Mean16x4(it->yuv_in_ + Y_OFF_ENC + k * BPS, &dc[k]);
+  }
+  for (m = 0, m2 = 0, k = 0; k < 16; ++k) {
+    m += dc[k];
+    m2 += dc[k] * dc[k];
+  }
+  if (kThreshold * m2 < m * m) {
+    VP8SetIntra16Mode(it, 0);   // DC16
+  } else {
+    const uint8_t modes[16] = { 0 };  // DC4
+    VP8SetIntra4Mode(it, modes);
+  }
+  return 0;
+}
+
+static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
+  int best_alpha = DEFAULT_ALPHA;
+  int smallest_alpha = 0;
+  int best_mode = 0;
+  const int max_mode = MAX_UV_MODE;
+  int mode;
+
+  VP8MakeChroma8Preds(it);
+  for (mode = 0; mode < max_mode; ++mode) {
+    VP8Histogram histo;
+    int alpha;
+    InitHistogram(&histo);
+    VP8CollectHistogram(it->yuv_in_ + U_OFF_ENC,
+                        it->yuv_p_ + VP8UVModeOffsets[mode],
+                        16, 16 + 4 + 4, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
+      best_alpha = alpha;
+    }
+    // The best prediction mode tends to be the one with the smallest alpha.
+    if (mode == 0 || alpha < smallest_alpha) {
+      smallest_alpha = alpha;
+      best_mode = mode;
+    }
+  }
+  VP8SetIntraUVMode(it, best_mode);
+  return best_alpha;
+}
+
+static void MBAnalyze(VP8EncIterator* const it,
+                      int alphas[MAX_ALPHA + 1],
+                      int* const alpha, int* const uv_alpha) {
+  const VP8Encoder* const enc = it->enc_;
+  int best_alpha, best_uv_alpha;
+
+  VP8SetIntra16Mode(it, 0);  // default: Intra16, DC_PRED
+  VP8SetSkip(it, 0);         // not skipped
+  VP8SetSegment(it, 0);      // default segment, spec-wise.
+
+  if (enc->method_ <= 1) {
+    best_alpha = FastMBAnalyze(it);
+  } else {
+    best_alpha = MBAnalyzeBestIntra16Mode(it);
+  }
+  best_uv_alpha = MBAnalyzeBestUVMode(it);
+
+  // Final susceptibility mix
+  best_alpha = (3 * best_alpha + best_uv_alpha + 2) >> 2;
+  best_alpha = FinalAlphaValue(best_alpha);
+  alphas[best_alpha]++;
+  it->mb_->alpha_ = best_alpha;   // for later remapping.
+
+  // Accumulate for later complexity analysis.
+  *alpha += best_alpha;   // mixed susceptibility (not just luma)
+  *uv_alpha += best_uv_alpha;
+}
+
+static void DefaultMBInfo(VP8MBInfo* const mb) {
+  mb->type_ = 1;     // I16x16
+  mb->uv_mode_ = 0;
+  mb->skip_ = 0;     // not skipped
+  mb->segment_ = 0;  // default segment
+  mb->alpha_ = 0;
+}
+
+//------------------------------------------------------------------------------
+// Main analysis loop:
+// Collect all susceptibilities for each macroblock and record their
+// distribution in alphas[]. Segments is assigned a-posteriori, based on
+// this histogram.
+// We also pick an intra16 prediction mode, which shouldn't be considered
+// final except for fast-encode settings. We can also pick some intra4 modes
+// and decide intra4/intra16, but that's usually almost always a bad choice at
+// this stage.
+
+static void ResetAllMBInfo(VP8Encoder* const enc) {
+  int n;
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    DefaultMBInfo(&enc->mb_info_[n]);
+  }
+  // Default susceptibilities.
+  enc->dqm_[0].alpha_ = 0;
+  enc->dqm_[0].beta_ = 0;
+  // Note: we can't compute this alpha_ / uv_alpha_ -> set to default value.
+  enc->alpha_ = 0;
+  enc->uv_alpha_ = 0;
+  WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
+}
+
+// struct used to collect job result
+typedef struct {
+  WebPWorker worker;
+  int alphas[MAX_ALPHA + 1];
+  int alpha, uv_alpha;
+  VP8EncIterator it;
+  int delta_progress;
+} SegmentJob;
+
+// main work call
+static int DoSegmentsJob(void* arg1, void* arg2) {
+  SegmentJob* const job = (SegmentJob*)arg1;
+  VP8EncIterator* const it = (VP8EncIterator*)arg2;
+  int ok = 1;
+  if (!VP8IteratorIsDone(it)) {
+    uint8_t tmp[32 + WEBP_ALIGN_CST];
+    uint8_t* const scratch = (uint8_t*)WEBP_ALIGN(tmp);
+    do {
+      // Let's pretend we have perfect lossless reconstruction.
+      VP8IteratorImport(it, scratch);
+      MBAnalyze(it, job->alphas, &job->alpha, &job->uv_alpha);
+      ok = VP8IteratorProgress(it, job->delta_progress);
+    } while (ok && VP8IteratorNext(it));
+  }
+  return ok;
+}
+
+static void MergeJobs(const SegmentJob* const src, SegmentJob* const dst) {
+  int i;
+  for (i = 0; i <= MAX_ALPHA; ++i) dst->alphas[i] += src->alphas[i];
+  dst->alpha += src->alpha;
+  dst->uv_alpha += src->uv_alpha;
+}
+
+// initialize the job struct with some tasks to perform
+static void InitSegmentJob(VP8Encoder* const enc, SegmentJob* const job,
+                           int start_row, int end_row) {
+  WebPGetWorkerInterface()->Init(&job->worker);
+  job->worker.data1 = job;
+  job->worker.data2 = &job->it;
+  job->worker.hook = DoSegmentsJob;
+  VP8IteratorInit(enc, &job->it);
+  VP8IteratorSetRow(&job->it, start_row);
+  VP8IteratorSetCountDown(&job->it, (end_row - start_row) * enc->mb_w_);
+  memset(job->alphas, 0, sizeof(job->alphas));
+  job->alpha = 0;
+  job->uv_alpha = 0;
+  // only one of both jobs can record the progress, since we don't
+  // expect the user's hook to be multi-thread safe
+  job->delta_progress = (start_row == 0) ? 20 : 0;
+}
+
+// main entry point
+int VP8EncAnalyze(VP8Encoder* const enc) {
+  int ok = 1;
+  const int do_segments =
+      enc->config_->emulate_jpeg_size ||   // We need the complexity evaluation.
+      (enc->segment_hdr_.num_segments_ > 1) ||
+      (enc->method_ <= 1);  // for method 0 - 1, we need preds_[] to be filled.
+  if (do_segments) {
+    const int last_row = enc->mb_h_;
+    // We give a little more than a half work to the main thread.
+    const int split_row = (9 * last_row + 15) >> 4;
+    const int total_mb = last_row * enc->mb_w_;
+#ifdef WEBP_USE_THREAD
+    const int kMinSplitRow = 2;  // minimal rows needed for mt to be worth it
+    const int do_mt = (enc->thread_level_ > 0) && (split_row >= kMinSplitRow);
+#else
+    const int do_mt = 0;
+#endif
+    const WebPWorkerInterface* const worker_interface =
+        WebPGetWorkerInterface();
+    SegmentJob main_job;
+    if (do_mt) {
+      SegmentJob side_job;
+      // Note the use of '&' instead of '&&' because we must call the functions
+      // no matter what.
+      InitSegmentJob(enc, &main_job, 0, split_row);
+      InitSegmentJob(enc, &side_job, split_row, last_row);
+      // we don't need to call Reset() on main_job.worker, since we're calling
+      // WebPWorkerExecute() on it
+      ok &= worker_interface->Reset(&side_job.worker);
+      // launch the two jobs in parallel
+      if (ok) {
+        worker_interface->Launch(&side_job.worker);
+        worker_interface->Execute(&main_job.worker);
+        ok &= worker_interface->Sync(&side_job.worker);
+        ok &= worker_interface->Sync(&main_job.worker);
+      }
+      worker_interface->End(&side_job.worker);
+      if (ok) MergeJobs(&side_job, &main_job);  // merge results together
+    } else {
+      // Even for single-thread case, we use the generic Worker tools.
+      InitSegmentJob(enc, &main_job, 0, last_row);
+      worker_interface->Execute(&main_job.worker);
+      ok &= worker_interface->Sync(&main_job.worker);
+    }
+    worker_interface->End(&main_job.worker);
+    if (ok) {
+      enc->alpha_ = main_job.alpha / total_mb;
+      enc->uv_alpha_ = main_job.uv_alpha / total_mb;
+      AssignSegments(enc, main_job.alphas);
+    }
+  } else {   // Use only one default segment.
+    ResetAllMBInfo(enc);
+  }
+  return ok;
+}
+
diff --git a/media/libwebp/src/enc/backward_references_cost_enc.c b/media/libwebp/src/enc/backward_references_cost_enc.c
new file mode 100644
index 0000000000..6968ef3c9f
--- /dev/null
+++ b/media/libwebp/src/enc/backward_references_cost_enc.c
@@ -0,0 +1,795 @@
+// Copyright 2017 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.
+// -----------------------------------------------------------------------------
+//
+// Improves a given set of backward references by analyzing its bit cost.
+// The algorithm is similar to the Zopfli compression algorithm but tailored to
+// images.
+//
+// Author: Vincent Rabaud (vrabaud@google.com)
+//
+
+#include <assert.h>
+#include <float.h>
+
+#include "src/dsp/lossless_common.h"
+#include "src/enc/backward_references_enc.h"
+#include "src/enc/histogram_enc.h"
+#include "src/utils/color_cache_utils.h"
+#include "src/utils/utils.h"
+
+#define VALUES_IN_BYTE 256
+
+extern void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs);
+extern int VP8LDistanceToPlaneCode(int xsize, int dist);
+extern void VP8LBackwardRefsCursorAdd(VP8LBackwardRefs* const refs,
+                                      const PixOrCopy v);
+
+typedef struct {
+  float alpha_[VALUES_IN_BYTE];
+  float red_[VALUES_IN_BYTE];
+  float blue_[VALUES_IN_BYTE];
+  float distance_[NUM_DISTANCE_CODES];
+  float* literal_;
+} CostModel;
+
+static void ConvertPopulationCountTableToBitEstimates(
+    int num_symbols, const uint32_t population_counts[], float 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 float 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 cache_bits,
+                          const VP8LBackwardRefs* const refs) {
+  int ok = 0;
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  VP8LHistogram* const histo = VP8LAllocateHistogram(cache_bits);
+  if (histo == NULL) goto Error;
+
+  // The following code is similar to VP8LHistogramCreate but converts the
+  // distance to plane code.
+  VP8LHistogramInit(histo, cache_bits, /*init_arrays=*/ 1);
+  while (VP8LRefsCursorOk(&c)) {
+    VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos, VP8LDistanceToPlaneCode,
+                                    xsize);
+    VP8LRefsCursorNext(&c);
+  }
+
+  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 float 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 float 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 float 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 float GetDistanceCost(const CostModel* const m,
+                                         uint32_t distance) {
+  int code, extra_bits;
+  VP8LPrefixEncodeBits(distance, &code, &extra_bits);
+  return m->distance_[code] + extra_bits;
+}
+
+static WEBP_INLINE void AddSingleLiteralWithCostModel(
+    const uint32_t* const argb, VP8LColorCache* const hashers,
+    const CostModel* const cost_model, int idx, int use_color_cache,
+    float prev_cost, float* const cost, uint16_t* const dist_array) {
+  float cost_val = prev_cost;
+  const uint32_t color = argb[idx];
+  const int ix = use_color_cache ? VP8LColorCacheContains(hashers, color) : -1;
+  if (ix >= 0) {
+    // use_color_cache is true and hashers contains color
+    const float mul0 = 0.68f;
+    cost_val += GetCacheCost(cost_model, ix) * mul0;
+  } else {
+    const float mul1 = 0.82f;
+    if (use_color_cache) VP8LColorCacheInsert(hashers, color);
+    cost_val += GetLiteralCost(cost_model, color) * mul1;
+  }
+  if (cost[idx] > cost_val) {
+    cost[idx] = cost_val;
+    dist_array[idx] = 1;  // only one is inserted.
+  }
+}
+
+// -----------------------------------------------------------------------------
+// CostManager and interval handling
+
+// Empirical value to avoid high memory consumption but good for performance.
+#define COST_CACHE_INTERVAL_SIZE_MAX 500
+
+// To perform backward reference every pixel at index index_ is considered and
+// the cost for the MAX_LENGTH following pixels computed. Those following pixels
+// at index index_ + k (k from 0 to MAX_LENGTH) have a cost of:
+//     cost_ = distance cost at index + GetLengthCost(cost_model, k)
+// and the minimum value is kept. GetLengthCost(cost_model, k) is cached in an
+// array of size MAX_LENGTH.
+// Instead of performing MAX_LENGTH comparisons per pixel, we keep track of the
+// minimal values using intervals of constant cost.
+// An interval is defined by the index_ of the pixel that generated it and
+// is only useful in a range of indices from start_ to end_ (exclusive), i.e.
+// it contains the minimum value for pixels between start_ and end_.
+// Intervals are stored in a linked list and ordered by start_. When a new
+// interval has a better value, old intervals are split or removed. There are
+// therefore no overlapping intervals.
+typedef struct CostInterval CostInterval;
+struct CostInterval {
+  float cost_;
+  int start_;
+  int end_;
+  int index_;
+  CostInterval* previous_;
+  CostInterval* next_;
+};
+
+// The GetLengthCost(cost_model, k) are cached in a CostCacheInterval.
+typedef struct {
+  float cost_;
+  int start_;
+  int end_;       // Exclusive.
+} CostCacheInterval;
+
+// This structure is in charge of managing intervals and costs.
+// It caches the different CostCacheInterval, caches the different
+// GetLengthCost(cost_model, k) in cost_cache_ and the CostInterval's (whose
+// count_ is limited by COST_CACHE_INTERVAL_SIZE_MAX).
+#define COST_MANAGER_MAX_FREE_LIST 10
+typedef struct {
+  CostInterval* head_;
+  int count_;  // The number of stored intervals.
+  CostCacheInterval* cache_intervals_;
+  size_t cache_intervals_size_;
+  float cost_cache_[MAX_LENGTH];  // Contains the GetLengthCost(cost_model, k).
+  float* costs_;
+  uint16_t* dist_array_;
+  // Most of the time, we only need few intervals -> use a free-list, to avoid
+  // fragmentation with small allocs in most common cases.
+  CostInterval intervals_[COST_MANAGER_MAX_FREE_LIST];
+  CostInterval* free_intervals_;
+  // These are regularly malloc'd remains. This list can't grow larger than than
+  // size COST_CACHE_INTERVAL_SIZE_MAX - COST_MANAGER_MAX_FREE_LIST, note.
+  CostInterval* recycled_intervals_;
+} CostManager;
+
+static void CostIntervalAddToFreeList(CostManager* const manager,
+                                      CostInterval* const interval) {
+  interval->next_ = manager->free_intervals_;
+  manager->free_intervals_ = interval;
+}
+
+static int CostIntervalIsInFreeList(const CostManager* const manager,
+                                    const CostInterval* const interval) {
+  return (interval >= &manager->intervals_[0] &&
+          interval <= &manager->intervals_[COST_MANAGER_MAX_FREE_LIST - 1]);
+}
+
+static void CostManagerInitFreeList(CostManager* const manager) {
+  int i;
+  manager->free_intervals_ = NULL;
+  for (i = 0; i < COST_MANAGER_MAX_FREE_LIST; ++i) {
+    CostIntervalAddToFreeList(manager, &manager->intervals_[i]);
+  }
+}
+
+static void DeleteIntervalList(CostManager* const manager,
+                               const CostInterval* interval) {
+  while (interval != NULL) {
+    const CostInterval* const next = interval->next_;
+    if (!CostIntervalIsInFreeList(manager, interval)) {
+      WebPSafeFree((void*)interval);
+    }  // else: do nothing
+    interval = next;
+  }
+}
+
+static void CostManagerClear(CostManager* const manager) {
+  if (manager == NULL) return;
+
+  WebPSafeFree(manager->costs_);
+  WebPSafeFree(manager->cache_intervals_);
+
+  // Clear the interval lists.
+  DeleteIntervalList(manager, manager->head_);
+  manager->head_ = NULL;
+  DeleteIntervalList(manager, manager->recycled_intervals_);
+  manager->recycled_intervals_ = NULL;
+
+  // Reset pointers, count_ and cache_intervals_size_.
+  memset(manager, 0, sizeof(*manager));
+  CostManagerInitFreeList(manager);
+}
+
+static int CostManagerInit(CostManager* const manager,
+                           uint16_t* const dist_array, int pix_count,
+                           const CostModel* const cost_model) {
+  int i;
+  const int cost_cache_size = (pix_count > MAX_LENGTH) ? MAX_LENGTH : pix_count;
+
+  manager->costs_ = NULL;
+  manager->cache_intervals_ = NULL;
+  manager->head_ = NULL;
+  manager->recycled_intervals_ = NULL;
+  manager->count_ = 0;
+  manager->dist_array_ = dist_array;
+  CostManagerInitFreeList(manager);
+
+  // Fill in the cost_cache_.
+  // Has to be done in two passes due to a GCC bug on i686
+  // related to https://gcc.gnu.org/bugzilla/show_bug.cgi?id=323
+  for (i = 0; i < cost_cache_size; ++i) {
+    manager->cost_cache_[i] = GetLengthCost(cost_model, i);
+  }
+  manager->cache_intervals_size_ = 1;
+  for (i = 1; i < cost_cache_size; ++i) {
+    // Get the number of bound intervals.
+    if (manager->cost_cache_[i] != manager->cost_cache_[i - 1]) {
+      ++manager->cache_intervals_size_;
+    }
+  }
+
+  // With the current cost model, we usually have below 20 intervals.
+  // The worst case scenario with a cost model would be if every length has a
+  // different cost, hence MAX_LENGTH but that is impossible with the current
+  // implementation that spirals around a pixel.
+  assert(manager->cache_intervals_size_ <= MAX_LENGTH);
+  manager->cache_intervals_ = (CostCacheInterval*)WebPSafeMalloc(
+      manager->cache_intervals_size_, sizeof(*manager->cache_intervals_));
+  if (manager->cache_intervals_ == NULL) {
+    CostManagerClear(manager);
+    return 0;
+  }
+
+  // Fill in the cache_intervals_.
+  {
+    CostCacheInterval* cur = manager->cache_intervals_;
+
+    // Consecutive values in cost_cache_ are compared and if a big enough
+    // difference is found, a new interval is created and bounded.
+    cur->start_ = 0;
+    cur->end_ = 1;
+    cur->cost_ = manager->cost_cache_[0];
+    for (i = 1; i < cost_cache_size; ++i) {
+      const float cost_val = manager->cost_cache_[i];
+      if (cost_val != cur->cost_) {
+        ++cur;
+        // Initialize an interval.
+        cur->start_ = i;
+        cur->cost_ = cost_val;
+      }
+      cur->end_ = i + 1;
+    }
+    assert((size_t)(cur - manager->cache_intervals_) + 1 ==
+           manager->cache_intervals_size_);
+  }
+
+  manager->costs_ = (float*)WebPSafeMalloc(pix_count, sizeof(*manager->costs_));
+  if (manager->costs_ == NULL) {
+    CostManagerClear(manager);
+    return 0;
+  }
+  // Set the initial costs_ high for every pixel as we will keep the minimum.
+  for (i = 0; i < pix_count; ++i) manager->costs_[i] = FLT_MAX;
+
+  return 1;
+}
+
+// Given the cost and the position that define an interval, update the cost at
+// pixel 'i' if it is smaller than the previously computed value.
+static WEBP_INLINE void UpdateCost(CostManager* const manager, int i,
+                                   int position, float cost) {
+  const int k = i - position;
+  assert(k >= 0 && k < MAX_LENGTH);
+
+  if (manager->costs_[i] > cost) {
+    manager->costs_[i] = cost;
+    manager->dist_array_[i] = k + 1;
+  }
+}
+
+// Given the cost and the position that define an interval, update the cost for
+// all the pixels between 'start' and 'end' excluded.
+static WEBP_INLINE void UpdateCostPerInterval(CostManager* const manager,
+                                              int start, int end, int position,
+                                              float cost) {
+  int i;
+  for (i = start; i < end; ++i) UpdateCost(manager, i, position, cost);
+}
+
+// Given two intervals, make 'prev' be the previous one of 'next' in 'manager'.
+static WEBP_INLINE void ConnectIntervals(CostManager* const manager,
+                                         CostInterval* const prev,
+                                         CostInterval* const next) {
+  if (prev != NULL) {
+    prev->next_ = next;
+  } else {
+    manager->head_ = next;
+  }
+
+  if (next != NULL) next->previous_ = prev;
+}
+
+// Pop an interval in the manager.
+static WEBP_INLINE void PopInterval(CostManager* const manager,
+                                    CostInterval* const interval) {
+  if (interval == NULL) return;
+
+  ConnectIntervals(manager, interval->previous_, interval->next_);
+  if (CostIntervalIsInFreeList(manager, interval)) {
+    CostIntervalAddToFreeList(manager, interval);
+  } else {  // recycle regularly malloc'd intervals too
+    interval->next_ = manager->recycled_intervals_;
+    manager->recycled_intervals_ = interval;
+  }
+  --manager->count_;
+  assert(manager->count_ >= 0);
+}
+
+// Update the cost at index i by going over all the stored intervals that
+// overlap with i.
+// If 'do_clean_intervals' is set to something different than 0, intervals that
+// end before 'i' will be popped.
+static WEBP_INLINE void UpdateCostAtIndex(CostManager* const manager, int i,
+                                          int do_clean_intervals) {
+  CostInterval* current = manager->head_;
+
+  while (current != NULL && current->start_ <= i) {
+    CostInterval* const next = current->next_;
+    if (current->end_ <= i) {
+      if (do_clean_intervals) {
+        // We have an outdated interval, remove it.
+        PopInterval(manager, current);
+      }
+    } else {
+      UpdateCost(manager, i, current->index_, current->cost_);
+    }
+    current = next;
+  }
+}
+
+// Given a current orphan interval and its previous interval, before
+// it was orphaned (which can be NULL), set it at the right place in the list
+// of intervals using the start_ ordering and the previous interval as a hint.
+static WEBP_INLINE void PositionOrphanInterval(CostManager* const manager,
+                                               CostInterval* const current,
+                                               CostInterval* previous) {
+  assert(current != NULL);
+
+  if (previous == NULL) previous = manager->head_;
+  while (previous != NULL && current->start_ < previous->start_) {
+    previous = previous->previous_;
+  }
+  while (previous != NULL && previous->next_ != NULL &&
+         previous->next_->start_ < current->start_) {
+    previous = previous->next_;
+  }
+
+  if (previous != NULL) {
+    ConnectIntervals(manager, current, previous->next_);
+  } else {
+    ConnectIntervals(manager, current, manager->head_);
+  }
+  ConnectIntervals(manager, previous, current);
+}
+
+// Insert an interval in the list contained in the manager by starting at
+// interval_in as a hint. The intervals are sorted by start_ value.
+static WEBP_INLINE void InsertInterval(CostManager* const manager,
+                                       CostInterval* const interval_in,
+                                       float cost, int position, int start,
+                                       int end) {
+  CostInterval* interval_new;
+
+  if (start >= end) return;
+  if (manager->count_ >= COST_CACHE_INTERVAL_SIZE_MAX) {
+    // Serialize the interval if we cannot store it.
+    UpdateCostPerInterval(manager, start, end, position, cost);
+    return;
+  }
+  if (manager->free_intervals_ != NULL) {
+    interval_new = manager->free_intervals_;
+    manager->free_intervals_ = interval_new->next_;
+  } else if (manager->recycled_intervals_ != NULL) {
+    interval_new = manager->recycled_intervals_;
+    manager->recycled_intervals_ = interval_new->next_;
+  } else {  // malloc for good
+    interval_new = (CostInterval*)WebPSafeMalloc(1, sizeof(*interval_new));
+    if (interval_new == NULL) {
+      // Write down the interval if we cannot create it.
+      UpdateCostPerInterval(manager, start, end, position, cost);
+      return;
+    }
+  }
+
+  interval_new->cost_ = cost;
+  interval_new->index_ = position;
+  interval_new->start_ = start;
+  interval_new->end_ = end;
+  PositionOrphanInterval(manager, interval_new, interval_in);
+
+  ++manager->count_;
+}
+
+// Given a new cost interval defined by its start at position, its length value
+// and distance_cost, add its contributions to the previous intervals and costs.
+// If handling the interval or one of its subintervals becomes to heavy, its
+// contribution is added to the costs right away.
+static WEBP_INLINE void PushInterval(CostManager* const manager,
+                                     float distance_cost, int position,
+                                     int len) {
+  size_t i;
+  CostInterval* interval = manager->head_;
+  CostInterval* interval_next;
+  const CostCacheInterval* const cost_cache_intervals =
+      manager->cache_intervals_;
+  // If the interval is small enough, no need to deal with the heavy
+  // interval logic, just serialize it right away. This constant is empirical.
+  const int kSkipDistance = 10;
+
+  if (len < kSkipDistance) {
+    int j;
+    for (j = position; j < position + len; ++j) {
+      const int k = j - position;
+      float cost_tmp;
+      assert(k >= 0 && k < MAX_LENGTH);
+      cost_tmp = distance_cost + manager->cost_cache_[k];
+
+      if (manager->costs_[j] > cost_tmp) {
+        manager->costs_[j] = cost_tmp;
+        manager->dist_array_[j] = k + 1;
+      }
+    }
+    return;
+  }
+
+  for (i = 0; i < manager->cache_intervals_size_ &&
+              cost_cache_intervals[i].start_ < len;
+       ++i) {
+    // Define the intersection of the ith interval with the new one.
+    int start = position + cost_cache_intervals[i].start_;
+    const int end = position + (cost_cache_intervals[i].end_ > len
+                                 ? len
+                                 : cost_cache_intervals[i].end_);
+    const float cost = distance_cost + cost_cache_intervals[i].cost_;
+
+    for (; interval != NULL && interval->start_ < end;
+         interval = interval_next) {
+      interval_next = interval->next_;
+
+      // Make sure we have some overlap
+      if (start >= interval->end_) continue;
+
+      if (cost >= interval->cost_) {
+        // When intervals are represented, the lower, the better.
+        // [**********************************************************[
+        // start                                                    end
+        //                   [----------------------------------[
+        //                   interval->start_       interval->end_
+        // If we are worse than what we already have, add whatever we have so
+        // far up to interval.
+        const int start_new = interval->end_;
+        InsertInterval(manager, interval, cost, position, start,
+                       interval->start_);
+        start = start_new;
+        if (start >= end) break;
+        continue;
+      }
+
+      if (start <= interval->start_) {
+        if (interval->end_ <= end) {
+          //                   [----------------------------------[
+          //                   interval->start_       interval->end_
+          // [**************************************************************[
+          // start                                                        end
+          // We can safely remove the old interval as it is fully included.
+          PopInterval(manager, interval);
+        } else {
+          //              [------------------------------------[
+          //              interval->start_        interval->end_
+          // [*****************************[
+          // start                       end
+          interval->start_ = end;
+          break;
+        }
+      } else {
+        if (end < interval->end_) {
+          // [--------------------------------------------------------------[
+          // interval->start_                                  interval->end_
+          //                     [*****************************[
+          //                     start                       end
+          // We have to split the old interval as it fully contains the new one.
+          const int end_original = interval->end_;
+          interval->end_ = start;
+          InsertInterval(manager, interval, interval->cost_, interval->index_,
+                         end, end_original);
+          interval = interval->next_;
+          break;
+        } else {
+          // [------------------------------------[
+          // interval->start_        interval->end_
+          //                     [*****************************[
+          //                     start                       end
+          interval->end_ = start;
+        }
+      }
+    }
+    // Insert the remaining interval from start to end.
+    InsertInterval(manager, interval, cost, position, start, end);
+  }
+}
+
+static int BackwardReferencesHashChainDistanceOnly(
+    int xsize, int ysize, const uint32_t* const argb, int cache_bits,
+    const VP8LHashChain* const hash_chain, const VP8LBackwardRefs* const refs,
+    uint16_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);
+  const size_t literal_array_size =
+      sizeof(float) * (VP8LHistogramNumCodes(cache_bits));
+  const size_t cost_model_size = sizeof(CostModel) + literal_array_size;
+  CostModel* const cost_model =
+      (CostModel*)WebPSafeCalloc(1ULL, cost_model_size);
+  VP8LColorCache hashers;
+  CostManager* cost_manager =
+      (CostManager*)WebPSafeCalloc(1ULL, sizeof(*cost_manager));
+  int offset_prev = -1, len_prev = -1;
+  float offset_cost = -1.f;
+  int first_offset_is_constant = -1;  // initialized with 'impossible' value
+  int reach = 0;
+
+  if (cost_model == NULL || cost_manager == NULL) goto Error;
+
+  cost_model->literal_ = (float*)(cost_model + 1);
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+
+  if (!CostModelBuild(cost_model, xsize, cache_bits, refs)) {
+    goto Error;
+  }
+
+  if (!CostManagerInit(cost_manager, dist_array, pix_count, cost_model)) {
+    goto Error;
+  }
+
+  // We loop one pixel at a time, but store all currently best points to
+  // non-processed locations from this point.
+  dist_array[0] = 0;
+  // Add first pixel as literal.
+  AddSingleLiteralWithCostModel(argb, &hashers, cost_model, 0, use_color_cache,
+                                0.f, cost_manager->costs_, dist_array);
+
+  for (i = 1; i < pix_count; ++i) {
+    const float prev_cost = cost_manager->costs_[i - 1];
+    int offset, len;
+    VP8LHashChainFindCopy(hash_chain, i, &offset, &len);
+
+    // Try adding the pixel as a literal.
+    AddSingleLiteralWithCostModel(argb, &hashers, cost_model, i,
+                                  use_color_cache, prev_cost,
+                                  cost_manager->costs_, dist_array);
+
+    // If we are dealing with a non-literal.
+    if (len >= 2) {
+      if (offset != offset_prev) {
+        const int code = VP8LDistanceToPlaneCode(xsize, offset);
+        offset_cost = GetDistanceCost(cost_model, code);
+        first_offset_is_constant = 1;
+        PushInterval(cost_manager, prev_cost + offset_cost, i, len);
+      } else {
+        assert(offset_cost >= 0);
+        assert(len_prev >= 0);
+        assert(first_offset_is_constant == 0 || first_offset_is_constant == 1);
+        // Instead of considering all contributions from a pixel i by calling:
+        //         PushInterval(cost_manager, prev_cost + offset_cost, i, len);
+        // we optimize these contributions in case offset_cost stays the same
+        // for consecutive pixels. This describes a set of pixels similar to a
+        // previous set (e.g. constant color regions).
+        if (first_offset_is_constant) {
+          reach = i - 1 + len_prev - 1;
+          first_offset_is_constant = 0;
+        }
+
+        if (i + len - 1 > reach) {
+          // We can only be go further with the same offset if the previous
+          // length was maxed, hence len_prev == len == MAX_LENGTH.
+          // TODO(vrabaud), bump i to the end right away (insert cache and
+          // update cost).
+          // TODO(vrabaud), check if one of the points in between does not have
+          // a lower cost.
+          // Already consider the pixel at "reach" to add intervals that are
+          // better than whatever we add.
+          int offset_j, len_j = 0;
+          int j;
+          assert(len == MAX_LENGTH || len == pix_count - i);
+          // Figure out the last consecutive pixel within [i, reach + 1] with
+          // the same offset.
+          for (j = i; j <= reach; ++j) {
+            VP8LHashChainFindCopy(hash_chain, j + 1, &offset_j, &len_j);
+            if (offset_j != offset) {
+              VP8LHashChainFindCopy(hash_chain, j, &offset_j, &len_j);
+              break;
+            }
+          }
+          // Update the cost at j - 1 and j.
+          UpdateCostAtIndex(cost_manager, j - 1, 0);
+          UpdateCostAtIndex(cost_manager, j, 0);
+
+          PushInterval(cost_manager, cost_manager->costs_[j - 1] + offset_cost,
+                       j, len_j);
+          reach = j + len_j - 1;
+        }
+      }
+    }
+
+    UpdateCostAtIndex(cost_manager, i, 1);
+    offset_prev = offset;
+    len_prev = len;
+  }
+
+  ok = !refs->error_;
+ Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  CostManagerClear(cost_manager);
+  WebPSafeFree(cost_model);
+  WebPSafeFree(cost_manager);
+  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(uint16_t* const dist_array,
+                           int dist_array_size,
+                           uint16_t** const chosen_path,
+                           int* const chosen_path_size) {
+  uint16_t* path = dist_array + dist_array_size;
+  uint16_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(
+    const uint32_t* const argb, int cache_bits,
+    const uint16_t* const chosen_path, int chosen_path_size,
+    const VP8LHashChain* const hash_chain, VP8LBackwardRefs* const refs) {
+  const int use_color_cache = (cache_bits > 0);
+  int ix;
+  int i = 0;
+  int ok = 0;
+  int cc_init = 0;
+  VP8LColorCache hashers;
+
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+
+  VP8LClearBackwardRefs(refs);
+  for (ix = 0; ix < chosen_path_size; ++ix) {
+    const int len = chosen_path[ix];
+    if (len != 1) {
+      int k;
+      const int offset = VP8LHashChainFindOffset(hash_chain, i);
+      VP8LBackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(offset, len));
+      if (use_color_cache) {
+        for (k = 0; k < len; ++k) {
+          VP8LColorCacheInsert(&hashers, argb[i + k]);
+        }
+      }
+      i += len;
+    } else {
+      PixOrCopy v;
+      const int idx =
+          use_color_cache ? VP8LColorCacheContains(&hashers, argb[i]) : -1;
+      if (idx >= 0) {
+        // use_color_cache is true and hashers contains argb[i]
+        // push pixel as a color cache index
+        v = PixOrCopyCreateCacheIdx(idx);
+      } else {
+        if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
+        v = PixOrCopyCreateLiteral(argb[i]);
+      }
+      VP8LBackwardRefsCursorAdd(refs, v);
+      ++i;
+    }
+  }
+  ok = !refs->error_;
+ Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  return ok;
+}
+
+// Returns 1 on success.
+extern int VP8LBackwardReferencesTraceBackwards(
+    int xsize, int ysize, const uint32_t* const argb, int cache_bits,
+    const VP8LHashChain* const hash_chain,
+    const VP8LBackwardRefs* const refs_src, VP8LBackwardRefs* const refs_dst);
+int VP8LBackwardReferencesTraceBackwards(int xsize, int ysize,
+                                         const uint32_t* const argb,
+                                         int cache_bits,
+                                         const VP8LHashChain* const hash_chain,
+                                         const VP8LBackwardRefs* const refs_src,
+                                         VP8LBackwardRefs* const refs_dst) {
+  int ok = 0;
+  const int dist_array_size = xsize * ysize;
+  uint16_t* chosen_path = NULL;
+  int chosen_path_size = 0;
+  uint16_t* dist_array =
+      (uint16_t*)WebPSafeMalloc(dist_array_size, sizeof(*dist_array));
+
+  if (dist_array == NULL) goto Error;
+
+  if (!BackwardReferencesHashChainDistanceOnly(
+          xsize, ysize, argb, cache_bits, hash_chain, refs_src, dist_array)) {
+    goto Error;
+  }
+  TraceBackwards(dist_array, dist_array_size, &chosen_path, &chosen_path_size);
+  if (!BackwardReferencesHashChainFollowChosenPath(
+          argb, cache_bits, chosen_path, chosen_path_size, hash_chain,
+          refs_dst)) {
+    goto Error;
+  }
+  ok = 1;
+ Error:
+  WebPSafeFree(dist_array);
+  return ok;
+}
diff --git a/media/libwebp/src/enc/backward_references_enc.c b/media/libwebp/src/enc/backward_references_enc.c
new file mode 100644
index 0000000000..49a0fac034
--- /dev/null
+++ b/media/libwebp/src/enc/backward_references_enc.c
@@ -0,0 +1,1068 @@
+// 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 "src/enc/backward_references_enc.h"
+
+#include <assert.h>
+#include <float.h>
+#include <math.h>
+
+#include "src/dsp/dsp.h"
+#include "src/dsp/lossless.h"
+#include "src/dsp/lossless_common.h"
+#include "src/enc/histogram_enc.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/utils/color_cache_utils.h"
+#include "src/utils/utils.h"
+#include "src/webp/encode.h"
+
+#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 << WINDOW_SIZE_BITS) - 120)
+
+// Minimum number of pixels for which it is cheaper to encode a
+// distance + length instead of each pixel as a literal.
+#define MIN_LENGTH 4
+
+// -----------------------------------------------------------------------------
+
+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
+};
+
+extern int VP8LDistanceToPlaneCode(int xsize, int dist);
+int VP8LDistanceToPlaneCode(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;
+}
+
+// Returns the exact index where array1 and array2 are different. For an index
+// inferior or equal to best_len_match, the return value just has to be strictly
+// inferior to best_len_match. The current behavior is to return 0 if this index
+// is best_len_match, and the index itself otherwise.
+// If no two elements are the same, it returns max_limit.
+static WEBP_INLINE int FindMatchLength(const uint32_t* const array1,
+                                       const uint32_t* const array2,
+                                       int best_len_match, int max_limit) {
+  // Before 'expensive' linear match, check if the two arrays match at the
+  // current best length index.
+  if (array1[best_len_match] != array2[best_len_match]) return 0;
+
+  return VP8LVectorMismatch(array1, array2, max_limit);
+}
+
+// -----------------------------------------------------------------------------
+//  VP8LBackwardRefs
+
+struct PixOrCopyBlock {
+  PixOrCopyBlock* next_;   // next block (or NULL)
+  PixOrCopy* start_;       // data start
+  int size_;               // currently used size
+};
+
+extern void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs);
+void VP8LClearBackwardRefs(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);
+  VP8LClearBackwardRefs(refs);
+  while (refs->free_blocks_ != NULL) {
+    PixOrCopyBlock* const next = refs->free_blocks_->next_;
+    WebPSafeFree(refs->free_blocks_);
+    refs->free_blocks_ = next;
+  }
+}
+
+// Swaps the content of two VP8LBackwardRefs.
+static void BackwardRefsSwap(VP8LBackwardRefs* const refs1,
+                             VP8LBackwardRefs* const refs2) {
+  const int point_to_refs1 =
+      (refs1->tail_ != NULL && refs1->tail_ == &refs1->refs_);
+  const int point_to_refs2 =
+      (refs2->tail_ != NULL && refs2->tail_ == &refs2->refs_);
+  const VP8LBackwardRefs tmp = *refs1;
+  *refs1 = *refs2;
+  *refs2 = tmp;
+  if (point_to_refs2) refs1->tail_ = &refs1->refs_;
+  if (point_to_refs1) refs2->tail_ = &refs2->refs_;
+}
+
+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;
+}
+
+// Return 1 on success, 0 on error.
+static int BackwardRefsClone(const VP8LBackwardRefs* const from,
+                             VP8LBackwardRefs* const to) {
+  const PixOrCopyBlock* block_from = from->refs_;
+  VP8LClearBackwardRefs(to);
+  while (block_from != NULL) {
+    PixOrCopyBlock* const block_to = BackwardRefsNewBlock(to);
+    if (block_to == NULL) return 0;
+    memcpy(block_to->start_, block_from->start_,
+           block_from->size_ * sizeof(PixOrCopy));
+    block_to->size_ = block_from->size_;
+    block_from = block_from->next_;
+  }
+  return 1;
+}
+
+extern void VP8LBackwardRefsCursorAdd(VP8LBackwardRefs* const refs,
+                                      const PixOrCopy v);
+void VP8LBackwardRefsCursorAdd(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;
+}
+
+// -----------------------------------------------------------------------------
+// Hash chains
+
+int VP8LHashChainInit(VP8LHashChain* const p, int size) {
+  assert(p->size_ == 0);
+  assert(p->offset_length_ == NULL);
+  assert(size > 0);
+  p->offset_length_ =
+      (uint32_t*)WebPSafeMalloc(size, sizeof(*p->offset_length_));
+  if (p->offset_length_ == NULL) return 0;
+  p->size_ = size;
+
+  return 1;
+}
+
+void VP8LHashChainClear(VP8LHashChain* const p) {
+  assert(p != NULL);
+  WebPSafeFree(p->offset_length_);
+
+  p->size_ = 0;
+  p->offset_length_ = NULL;
+}
+
+// -----------------------------------------------------------------------------
+
+static const uint32_t kHashMultiplierHi = 0xc6a4a793u;
+static const uint32_t kHashMultiplierLo = 0x5bd1e996u;
+
+static WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW WEBP_INLINE
+uint32_t GetPixPairHash64(const uint32_t* const argb) {
+  uint32_t key;
+  key  = argb[1] * kHashMultiplierHi;
+  key += argb[0] * kHashMultiplierLo;
+  key = key >> (32 - HASH_BITS);
+  return key;
+}
+
+// Returns the maximum number of hash chain lookups to do for a
+// given compression quality. Return value in range [8, 86].
+static int GetMaxItersForQuality(int quality) {
+  return 8 + (quality * quality) / 128;
+}
+
+static int GetWindowSizeForHashChain(int quality, int xsize) {
+  const int max_window_size = (quality > 75) ? WINDOW_SIZE
+                            : (quality > 50) ? (xsize << 8)
+                            : (quality > 25) ? (xsize << 6)
+                            : (xsize << 4);
+  assert(xsize > 0);
+  return (max_window_size > WINDOW_SIZE) ? WINDOW_SIZE : max_window_size;
+}
+
+static WEBP_INLINE int MaxFindCopyLength(int len) {
+  return (len < MAX_LENGTH) ? len : MAX_LENGTH;
+}
+
+int VP8LHashChainFill(VP8LHashChain* const p, int quality,
+                      const uint32_t* const argb, int xsize, int ysize,
+                      int low_effort, const WebPPicture* const pic,
+                      int percent_range, int* const percent) {
+  const int size = xsize * ysize;
+  const int iter_max = GetMaxItersForQuality(quality);
+  const uint32_t window_size = GetWindowSizeForHashChain(quality, xsize);
+  int remaining_percent = percent_range;
+  int percent_start = *percent;
+  int pos;
+  int argb_comp;
+  uint32_t base_position;
+  int32_t* hash_to_first_index;
+  // Temporarily use the p->offset_length_ as a hash chain.
+  int32_t* chain = (int32_t*)p->offset_length_;
+  assert(size > 0);
+  assert(p->size_ != 0);
+  assert(p->offset_length_ != NULL);
+
+  if (size <= 2) {
+    p->offset_length_[0] = p->offset_length_[size - 1] = 0;
+    return 1;
+  }
+
+  hash_to_first_index =
+      (int32_t*)WebPSafeMalloc(HASH_SIZE, sizeof(*hash_to_first_index));
+  if (hash_to_first_index == NULL) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return 0;
+  }
+
+  percent_range = remaining_percent / 2;
+  remaining_percent -= percent_range;
+
+  // Set the int32_t array to -1.
+  memset(hash_to_first_index, 0xff, HASH_SIZE * sizeof(*hash_to_first_index));
+  // Fill the chain linking pixels with the same hash.
+  argb_comp = (argb[0] == argb[1]);
+  for (pos = 0; pos < size - 2;) {
+    uint32_t hash_code;
+    const int argb_comp_next = (argb[pos + 1] == argb[pos + 2]);
+    if (argb_comp && argb_comp_next) {
+      // Consecutive pixels with the same color will share the same hash.
+      // We therefore use a different hash: the color and its repetition
+      // length.
+      uint32_t tmp[2];
+      uint32_t len = 1;
+      tmp[0] = argb[pos];
+      // Figure out how far the pixels are the same.
+      // The last pixel has a different 64 bit hash, as its next pixel does
+      // not have the same color, so we just need to get to the last pixel equal
+      // to its follower.
+      while (pos + (int)len + 2 < size && argb[pos + len + 2] == argb[pos]) {
+        ++len;
+      }
+      if (len > MAX_LENGTH) {
+        // Skip the pixels that match for distance=1 and length>MAX_LENGTH
+        // because they are linked to their predecessor and we automatically
+        // check that in the main for loop below. Skipping means setting no
+        // predecessor in the chain, hence -1.
+        memset(chain + pos, 0xff, (len - MAX_LENGTH) * sizeof(*chain));
+        pos += len - MAX_LENGTH;
+        len = MAX_LENGTH;
+      }
+      // Process the rest of the hash chain.
+      while (len) {
+        tmp[1] = len--;
+        hash_code = GetPixPairHash64(tmp);
+        chain[pos] = hash_to_first_index[hash_code];
+        hash_to_first_index[hash_code] = pos++;
+      }
+      argb_comp = 0;
+    } else {
+      // Just move one pixel forward.
+      hash_code = GetPixPairHash64(argb + pos);
+      chain[pos] = hash_to_first_index[hash_code];
+      hash_to_first_index[hash_code] = pos++;
+      argb_comp = argb_comp_next;
+    }
+
+    if (!WebPReportProgress(
+            pic, percent_start + percent_range * pos / (size - 2), percent)) {
+      WebPSafeFree(hash_to_first_index);
+      return 0;
+    }
+  }
+  // Process the penultimate pixel.
+  chain[pos] = hash_to_first_index[GetPixPairHash64(argb + pos)];
+
+  WebPSafeFree(hash_to_first_index);
+
+  percent_start += percent_range;
+  if (!WebPReportProgress(pic, percent_start, percent)) return 0;
+  percent_range = remaining_percent;
+
+  // Find the best match interval at each pixel, defined by an offset to the
+  // pixel and a length. The right-most pixel cannot match anything to the right
+  // (hence a best length of 0) and the left-most pixel nothing to the left
+  // (hence an offset of 0).
+  assert(size > 2);
+  p->offset_length_[0] = p->offset_length_[size - 1] = 0;
+  for (base_position = size - 2; base_position > 0;) {
+    const int max_len = MaxFindCopyLength(size - 1 - base_position);
+    const uint32_t* const argb_start = argb + base_position;
+    int iter = iter_max;
+    int best_length = 0;
+    uint32_t best_distance = 0;
+    uint32_t best_argb;
+    const int min_pos =
+        (base_position > window_size) ? base_position - window_size : 0;
+    const int length_max = (max_len < 256) ? max_len : 256;
+    uint32_t max_base_position;
+
+    pos = chain[base_position];
+    if (!low_effort) {
+      int curr_length;
+      // Heuristic: use the comparison with the above line as an initialization.
+      if (base_position >= (uint32_t)xsize) {
+        curr_length = FindMatchLength(argb_start - xsize, argb_start,
+                                      best_length, max_len);
+        if (curr_length > best_length) {
+          best_length = curr_length;
+          best_distance = xsize;
+        }
+        --iter;
+      }
+      // Heuristic: compare to the previous pixel.
+      curr_length =
+          FindMatchLength(argb_start - 1, argb_start, best_length, max_len);
+      if (curr_length > best_length) {
+        best_length = curr_length;
+        best_distance = 1;
+      }
+      --iter;
+      // Skip the for loop if we already have the maximum.
+      if (best_length == MAX_LENGTH) pos = min_pos - 1;
+    }
+    best_argb = argb_start[best_length];
+
+    for (; pos >= min_pos && --iter; pos = chain[pos]) {
+      int curr_length;
+      assert(base_position > (uint32_t)pos);
+
+      if (argb[pos + best_length] != best_argb) continue;
+
+      curr_length = VP8LVectorMismatch(argb + pos, argb_start, max_len);
+      if (best_length < curr_length) {
+        best_length = curr_length;
+        best_distance = base_position - pos;
+        best_argb = argb_start[best_length];
+        // Stop if we have reached a good enough length.
+        if (best_length >= length_max) break;
+      }
+    }
+    // We have the best match but in case the two intervals continue matching
+    // to the left, we have the best matches for the left-extended pixels.
+    max_base_position = base_position;
+    while (1) {
+      assert(best_length <= MAX_LENGTH);
+      assert(best_distance <= WINDOW_SIZE);
+      p->offset_length_[base_position] =
+          (best_distance << MAX_LENGTH_BITS) | (uint32_t)best_length;
+      --base_position;
+      // Stop if we don't have a match or if we are out of bounds.
+      if (best_distance == 0 || base_position == 0) break;
+      // Stop if we cannot extend the matching intervals to the left.
+      if (base_position < best_distance ||
+          argb[base_position - best_distance] != argb[base_position]) {
+        break;
+      }
+      // Stop if we are matching at its limit because there could be a closer
+      // matching interval with the same maximum length. Then again, if the
+      // matching interval is as close as possible (best_distance == 1), we will
+      // never find anything better so let's continue.
+      if (best_length == MAX_LENGTH && best_distance != 1 &&
+          base_position + MAX_LENGTH < max_base_position) {
+        break;
+      }
+      if (best_length < MAX_LENGTH) {
+        ++best_length;
+        max_base_position = base_position;
+      }
+    }
+
+    if (!WebPReportProgress(pic,
+                            percent_start + percent_range *
+                                                (size - 2 - base_position) /
+                                                (size - 2),
+                            percent)) {
+      return 0;
+    }
+  }
+
+  return WebPReportProgress(pic, percent_start + percent_range, percent);
+}
+
+static WEBP_INLINE void AddSingleLiteral(uint32_t pixel, int use_color_cache,
+                                         VP8LColorCache* const hashers,
+                                         VP8LBackwardRefs* const refs) {
+  PixOrCopy v;
+  if (use_color_cache) {
+    const uint32_t key = VP8LColorCacheGetIndex(hashers, pixel);
+    if (VP8LColorCacheLookup(hashers, key) == pixel) {
+      v = PixOrCopyCreateCacheIdx(key);
+    } else {
+      v = PixOrCopyCreateLiteral(pixel);
+      VP8LColorCacheSet(hashers, key, pixel);
+    }
+  } else {
+    v = PixOrCopyCreateLiteral(pixel);
+  }
+  VP8LBackwardRefsCursorAdd(refs, v);
+}
+
+static int BackwardReferencesRle(int xsize, int ysize,
+                                 const uint32_t* const argb,
+                                 int cache_bits, VP8LBackwardRefs* const refs) {
+  const int pix_count = xsize * ysize;
+  int i, k;
+  const int use_color_cache = (cache_bits > 0);
+  VP8LColorCache hashers;
+
+  if (use_color_cache && !VP8LColorCacheInit(&hashers, cache_bits)) {
+    return 0;
+  }
+  VP8LClearBackwardRefs(refs);
+  // Add first pixel as literal.
+  AddSingleLiteral(argb[0], use_color_cache, &hashers, refs);
+  i = 1;
+  while (i < pix_count) {
+    const int max_len = MaxFindCopyLength(pix_count - i);
+    const int rle_len = FindMatchLength(argb + i, argb + i - 1, 0, max_len);
+    const int prev_row_len = (i < xsize) ? 0 :
+        FindMatchLength(argb + i, argb + i - xsize, 0, max_len);
+    if (rle_len >= prev_row_len && rle_len >= MIN_LENGTH) {
+      VP8LBackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(1, rle_len));
+      // We don't need to update the color cache here since it is always the
+      // same pixel being copied, and that does not change the color cache
+      // state.
+      i += rle_len;
+    } else if (prev_row_len >= MIN_LENGTH) {
+      VP8LBackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(xsize, prev_row_len));
+      if (use_color_cache) {
+        for (k = 0; k < prev_row_len; ++k) {
+          VP8LColorCacheInsert(&hashers, argb[i + k]);
+        }
+      }
+      i += prev_row_len;
+    } else {
+      AddSingleLiteral(argb[i], use_color_cache, &hashers, refs);
+      i++;
+    }
+  }
+  if (use_color_cache) VP8LColorCacheClear(&hashers);
+  return !refs->error_;
+}
+
+static int BackwardReferencesLz77(int xsize, int ysize,
+                                  const uint32_t* const argb, int cache_bits,
+                                  const VP8LHashChain* const hash_chain,
+                                  VP8LBackwardRefs* const refs) {
+  int i;
+  int i_last_check = -1;
+  int ok = 0;
+  int cc_init = 0;
+  const int use_color_cache = (cache_bits > 0);
+  const int pix_count = xsize * ysize;
+  VP8LColorCache hashers;
+
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+  VP8LClearBackwardRefs(refs);
+  for (i = 0; i < pix_count;) {
+    // Alternative#1: Code the pixels starting at 'i' using backward reference.
+    int offset = 0;
+    int len = 0;
+    int j;
+    VP8LHashChainFindCopy(hash_chain, i, &offset, &len);
+    if (len >= MIN_LENGTH) {
+      const int len_ini = len;
+      int max_reach = 0;
+      const int j_max =
+          (i + len_ini >= pix_count) ? pix_count - 1 : i + len_ini;
+      // Only start from what we have not checked already.
+      i_last_check = (i > i_last_check) ? i : i_last_check;
+      // We know the best match for the current pixel but we try to find the
+      // best matches for the current pixel AND the next one combined.
+      // The naive method would use the intervals:
+      // [i,i+len) + [i+len, length of best match at i+len)
+      // while we check if we can use:
+      // [i,j) (where j<=i+len) + [j, length of best match at j)
+      for (j = i_last_check + 1; j <= j_max; ++j) {
+        const int len_j = VP8LHashChainFindLength(hash_chain, j);
+        const int reach =
+            j + (len_j >= MIN_LENGTH ? len_j : 1);  // 1 for single literal.
+        if (reach > max_reach) {
+          len = j - i;
+          max_reach = reach;
+          if (max_reach >= pix_count) break;
+        }
+      }
+    } else {
+      len = 1;
+    }
+    // Go with literal or backward reference.
+    assert(len > 0);
+    if (len == 1) {
+      AddSingleLiteral(argb[i], use_color_cache, &hashers, refs);
+    } else {
+      VP8LBackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(offset, len));
+      if (use_color_cache) {
+        for (j = i; j < i + len; ++j) VP8LColorCacheInsert(&hashers, argb[j]);
+      }
+    }
+    i += len;
+  }
+
+  ok = !refs->error_;
+ Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  return ok;
+}
+
+// Compute an LZ77 by forcing matches to happen within a given distance cost.
+// We therefore limit the algorithm to the lowest 32 values in the PlaneCode
+// definition.
+#define WINDOW_OFFSETS_SIZE_MAX 32
+static int BackwardReferencesLz77Box(int xsize, int ysize,
+                                     const uint32_t* const argb, int cache_bits,
+                                     const VP8LHashChain* const hash_chain_best,
+                                     VP8LHashChain* hash_chain,
+                                     VP8LBackwardRefs* const refs) {
+  int i;
+  const int pix_count = xsize * ysize;
+  uint16_t* counts;
+  int window_offsets[WINDOW_OFFSETS_SIZE_MAX] = {0};
+  int window_offsets_new[WINDOW_OFFSETS_SIZE_MAX] = {0};
+  int window_offsets_size = 0;
+  int window_offsets_new_size = 0;
+  uint16_t* const counts_ini =
+      (uint16_t*)WebPSafeMalloc(xsize * ysize, sizeof(*counts_ini));
+  int best_offset_prev = -1, best_length_prev = -1;
+  if (counts_ini == NULL) return 0;
+
+  // counts[i] counts how many times a pixel is repeated starting at position i.
+  i = pix_count - 2;
+  counts = counts_ini + i;
+  counts[1] = 1;
+  for (; i >= 0; --i, --counts) {
+    if (argb[i] == argb[i + 1]) {
+      // Max out the counts to MAX_LENGTH.
+      counts[0] = counts[1] + (counts[1] != MAX_LENGTH);
+    } else {
+      counts[0] = 1;
+    }
+  }
+
+  // Figure out the window offsets around a pixel. They are stored in a
+  // spiraling order around the pixel as defined by VP8LDistanceToPlaneCode.
+  {
+    int x, y;
+    for (y = 0; y <= 6; ++y) {
+      for (x = -6; x <= 6; ++x) {
+        const int offset = y * xsize + x;
+        int plane_code;
+        // Ignore offsets that bring us after the pixel.
+        if (offset <= 0) continue;
+        plane_code = VP8LDistanceToPlaneCode(xsize, offset) - 1;
+        if (plane_code >= WINDOW_OFFSETS_SIZE_MAX) continue;
+        window_offsets[plane_code] = offset;
+      }
+    }
+    // For narrow images, not all plane codes are reached, so remove those.
+    for (i = 0; i < WINDOW_OFFSETS_SIZE_MAX; ++i) {
+      if (window_offsets[i] == 0) continue;
+      window_offsets[window_offsets_size++] = window_offsets[i];
+    }
+    // Given a pixel P, find the offsets that reach pixels unreachable from P-1
+    // with any of the offsets in window_offsets[].
+    for (i = 0; i < window_offsets_size; ++i) {
+      int j;
+      int is_reachable = 0;
+      for (j = 0; j < window_offsets_size && !is_reachable; ++j) {
+        is_reachable |= (window_offsets[i] == window_offsets[j] + 1);
+      }
+      if (!is_reachable) {
+        window_offsets_new[window_offsets_new_size] = window_offsets[i];
+        ++window_offsets_new_size;
+      }
+    }
+  }
+
+  hash_chain->offset_length_[0] = 0;
+  for (i = 1; i < pix_count; ++i) {
+    int ind;
+    int best_length = VP8LHashChainFindLength(hash_chain_best, i);
+    int best_offset;
+    int do_compute = 1;
+
+    if (best_length >= MAX_LENGTH) {
+      // Do not recompute the best match if we already have a maximal one in the
+      // window.
+      best_offset = VP8LHashChainFindOffset(hash_chain_best, i);
+      for (ind = 0; ind < window_offsets_size; ++ind) {
+        if (best_offset == window_offsets[ind]) {
+          do_compute = 0;
+          break;
+        }
+      }
+    }
+    if (do_compute) {
+      // Figure out if we should use the offset/length from the previous pixel
+      // as an initial guess and therefore only inspect the offsets in
+      // window_offsets_new[].
+      const int use_prev =
+          (best_length_prev > 1) && (best_length_prev < MAX_LENGTH);
+      const int num_ind =
+          use_prev ? window_offsets_new_size : window_offsets_size;
+      best_length = use_prev ? best_length_prev - 1 : 0;
+      best_offset = use_prev ? best_offset_prev : 0;
+      // Find the longest match in a window around the pixel.
+      for (ind = 0; ind < num_ind; ++ind) {
+        int curr_length = 0;
+        int j = i;
+        int j_offset =
+            use_prev ? i - window_offsets_new[ind] : i - window_offsets[ind];
+        if (j_offset < 0 || argb[j_offset] != argb[i]) continue;
+        // The longest match is the sum of how many times each pixel is
+        // repeated.
+        do {
+          const int counts_j_offset = counts_ini[j_offset];
+          const int counts_j = counts_ini[j];
+          if (counts_j_offset != counts_j) {
+            curr_length +=
+                (counts_j_offset < counts_j) ? counts_j_offset : counts_j;
+            break;
+          }
+          // The same color is repeated counts_pos times at j_offset and j.
+          curr_length += counts_j_offset;
+          j_offset += counts_j_offset;
+          j += counts_j_offset;
+        } while (curr_length <= MAX_LENGTH && j < pix_count &&
+                 argb[j_offset] == argb[j]);
+        if (best_length < curr_length) {
+          best_offset =
+              use_prev ? window_offsets_new[ind] : window_offsets[ind];
+          if (curr_length >= MAX_LENGTH) {
+            best_length = MAX_LENGTH;
+            break;
+          } else {
+            best_length = curr_length;
+          }
+        }
+      }
+    }
+
+    assert(i + best_length <= pix_count);
+    assert(best_length <= MAX_LENGTH);
+    if (best_length <= MIN_LENGTH) {
+      hash_chain->offset_length_[i] = 0;
+      best_offset_prev = 0;
+      best_length_prev = 0;
+    } else {
+      hash_chain->offset_length_[i] =
+          (best_offset << MAX_LENGTH_BITS) | (uint32_t)best_length;
+      best_offset_prev = best_offset;
+      best_length_prev = best_length;
+    }
+  }
+  hash_chain->offset_length_[0] = 0;
+  WebPSafeFree(counts_ini);
+
+  return BackwardReferencesLz77(xsize, ysize, argb, cache_bits, hash_chain,
+                                refs);
+}
+
+// -----------------------------------------------------------------------------
+
+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 = VP8LDistanceToPlaneCode(xsize, dist);
+      c.cur_pos->argb_or_distance = transformed_dist;
+    }
+    VP8LRefsCursorNext(&c);
+  }
+}
+
+// Evaluate optimal cache bits for the local color cache.
+// The input *best_cache_bits sets the maximum cache bits to use (passing 0
+// implies disabling the local color cache). The local color cache is also
+// disabled for the lower (<= 25) quality.
+// Returns 0 in case of memory error.
+static int CalculateBestCacheSize(const uint32_t* argb, int quality,
+                                  const VP8LBackwardRefs* const refs,
+                                  int* const best_cache_bits) {
+  int i;
+  const int cache_bits_max = (quality <= 25) ? 0 : *best_cache_bits;
+  float entropy_min = MAX_ENTROPY;
+  int cc_init[MAX_COLOR_CACHE_BITS + 1] = { 0 };
+  VP8LColorCache hashers[MAX_COLOR_CACHE_BITS + 1];
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  VP8LHistogram* histos[MAX_COLOR_CACHE_BITS + 1] = { NULL };
+  int ok = 0;
+
+  assert(cache_bits_max >= 0 && cache_bits_max <= MAX_COLOR_CACHE_BITS);
+
+  if (cache_bits_max == 0) {
+    *best_cache_bits = 0;
+    // Local color cache is disabled.
+    return 1;
+  }
+
+  // Allocate data.
+  for (i = 0; i <= cache_bits_max; ++i) {
+    histos[i] = VP8LAllocateHistogram(i);
+    if (histos[i] == NULL) goto Error;
+    VP8LHistogramInit(histos[i], i, /*init_arrays=*/ 1);
+    if (i == 0) continue;
+    cc_init[i] = VP8LColorCacheInit(&hashers[i], i);
+    if (!cc_init[i]) goto Error;
+  }
+
+  // Find the cache_bits giving the lowest entropy. The search is done in a
+  // brute-force way as the function (entropy w.r.t cache_bits) can be
+  // anything in practice.
+  while (VP8LRefsCursorOk(&c)) {
+    const PixOrCopy* const v = c.cur_pos;
+    if (PixOrCopyIsLiteral(v)) {
+      const uint32_t pix = *argb++;
+      const uint32_t a = (pix >> 24) & 0xff;
+      const uint32_t r = (pix >> 16) & 0xff;
+      const uint32_t g = (pix >>  8) & 0xff;
+      const uint32_t b = (pix >>  0) & 0xff;
+      // The keys of the caches can be derived from the longest one.
+      int key = VP8LHashPix(pix, 32 - cache_bits_max);
+      // Do not use the color cache for cache_bits = 0.
+      ++histos[0]->blue_[b];
+      ++histos[0]->literal_[g];
+      ++histos[0]->red_[r];
+      ++histos[0]->alpha_[a];
+      // Deal with cache_bits > 0.
+      for (i = cache_bits_max; i >= 1; --i, key >>= 1) {
+        if (VP8LColorCacheLookup(&hashers[i], key) == pix) {
+          ++histos[i]->literal_[NUM_LITERAL_CODES + NUM_LENGTH_CODES + key];
+        } else {
+          VP8LColorCacheSet(&hashers[i], key, pix);
+          ++histos[i]->blue_[b];
+          ++histos[i]->literal_[g];
+          ++histos[i]->red_[r];
+          ++histos[i]->alpha_[a];
+        }
+      }
+    } else {
+      int code, extra_bits, extra_bits_value;
+      // We should compute the contribution of the (distance,length)
+      // histograms but those are the same independently from the cache size.
+      // As those constant contributions are in the end added to the other
+      // histogram contributions, we can ignore them, except for the length
+      // prefix that is part of the literal_ histogram.
+      int len = PixOrCopyLength(v);
+      uint32_t argb_prev = *argb ^ 0xffffffffu;
+      VP8LPrefixEncode(len, &code, &extra_bits, &extra_bits_value);
+      for (i = 0; i <= cache_bits_max; ++i) {
+        ++histos[i]->literal_[NUM_LITERAL_CODES + code];
+      }
+      // Update the color caches.
+      do {
+        if (*argb != argb_prev) {
+          // Efficiency: insert only if the color changes.
+          int key = VP8LHashPix(*argb, 32 - cache_bits_max);
+          for (i = cache_bits_max; i >= 1; --i, key >>= 1) {
+            hashers[i].colors_[key] = *argb;
+          }
+          argb_prev = *argb;
+        }
+        argb++;
+      } while (--len != 0);
+    }
+    VP8LRefsCursorNext(&c);
+  }
+
+  for (i = 0; i <= cache_bits_max; ++i) {
+    const float entropy = VP8LHistogramEstimateBits(histos[i]);
+    if (i == 0 || entropy < entropy_min) {
+      entropy_min = entropy;
+      *best_cache_bits = i;
+    }
+  }
+  ok = 1;
+ Error:
+  for (i = 0; i <= cache_bits_max; ++i) {
+    if (cc_init[i]) VP8LColorCacheClear(&hashers[i]);
+    VP8LFreeHistogram(histos[i]);
+  }
+  return ok;
+}
+
+// Update (in-place) backward references for specified cache_bits.
+static int BackwardRefsWithLocalCache(const uint32_t* const argb,
+                                      int cache_bits,
+                                      VP8LBackwardRefs* const refs) {
+  int pixel_index = 0;
+  VP8LColorCache hashers;
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  if (!VP8LColorCacheInit(&hashers, cache_bits)) return 0;
+
+  while (VP8LRefsCursorOk(&c)) {
+    PixOrCopy* const v = c.cur_pos;
+    if (PixOrCopyIsLiteral(v)) {
+      const uint32_t argb_literal = v->argb_or_distance;
+      const int ix = VP8LColorCacheContains(&hashers, argb_literal);
+      if (ix >= 0) {
+        // hashers contains argb_literal
+        *v = PixOrCopyCreateCacheIdx(ix);
+      } else {
+        VP8LColorCacheInsert(&hashers, argb_literal);
+      }
+      ++pixel_index;
+    } else {
+      // refs was created without local cache, so it can not have cache indexes.
+      int k;
+      assert(PixOrCopyIsCopy(v));
+      for (k = 0; k < v->len; ++k) {
+        VP8LColorCacheInsert(&hashers, argb[pixel_index++]);
+      }
+    }
+    VP8LRefsCursorNext(&c);
+  }
+  VP8LColorCacheClear(&hashers);
+  return 1;
+}
+
+static VP8LBackwardRefs* GetBackwardReferencesLowEffort(
+    int width, int height, const uint32_t* const argb,
+    int* const cache_bits, const VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs* const refs_lz77) {
+  *cache_bits = 0;
+  if (!BackwardReferencesLz77(width, height, argb, 0, hash_chain, refs_lz77)) {
+    return NULL;
+  }
+  BackwardReferences2DLocality(width, refs_lz77);
+  return refs_lz77;
+}
+
+extern int VP8LBackwardReferencesTraceBackwards(
+    int xsize, int ysize, const uint32_t* const argb, int cache_bits,
+    const VP8LHashChain* const hash_chain,
+    const VP8LBackwardRefs* const refs_src, VP8LBackwardRefs* const refs_dst);
+static int GetBackwardReferences(int width, int height,
+                                 const uint32_t* const argb, int quality,
+                                 int lz77_types_to_try, int cache_bits_max,
+                                 int do_no_cache,
+                                 const VP8LHashChain* const hash_chain,
+                                 VP8LBackwardRefs* const refs,
+                                 int* const cache_bits_best) {
+  VP8LHistogram* histo = NULL;
+  int i, lz77_type;
+  // Index 0 is for a color cache, index 1 for no cache (if needed).
+  int lz77_types_best[2] = {0, 0};
+  float bit_costs_best[2] = {FLT_MAX, FLT_MAX};
+  VP8LHashChain hash_chain_box;
+  VP8LBackwardRefs* const refs_tmp = &refs[do_no_cache ? 2 : 1];
+  int status = 0;
+  memset(&hash_chain_box, 0, sizeof(hash_chain_box));
+
+  histo = VP8LAllocateHistogram(MAX_COLOR_CACHE_BITS);
+  if (histo == NULL) goto Error;
+
+  for (lz77_type = 1; lz77_types_to_try;
+       lz77_types_to_try &= ~lz77_type, lz77_type <<= 1) {
+    int res = 0;
+    float bit_cost = 0.f;
+    if ((lz77_types_to_try & lz77_type) == 0) continue;
+    switch (lz77_type) {
+      case kLZ77RLE:
+        res = BackwardReferencesRle(width, height, argb, 0, refs_tmp);
+        break;
+      case kLZ77Standard:
+        // Compute LZ77 with no cache (0 bits), as the ideal LZ77 with a color
+        // cache is not that different in practice.
+        res = BackwardReferencesLz77(width, height, argb, 0, hash_chain,
+                                     refs_tmp);
+        break;
+      case kLZ77Box:
+        if (!VP8LHashChainInit(&hash_chain_box, width * height)) goto Error;
+        res = BackwardReferencesLz77Box(width, height, argb, 0, hash_chain,
+                                        &hash_chain_box, refs_tmp);
+        break;
+      default:
+        assert(0);
+    }
+    if (!res) goto Error;
+
+    // Start with the no color cache case.
+    for (i = 1; i >= 0; --i) {
+      int cache_bits = (i == 1) ? 0 : cache_bits_max;
+
+      if (i == 1 && !do_no_cache) continue;
+
+      if (i == 0) {
+        // Try with a color cache.
+        if (!CalculateBestCacheSize(argb, quality, refs_tmp, &cache_bits)) {
+          goto Error;
+        }
+        if (cache_bits > 0) {
+          if (!BackwardRefsWithLocalCache(argb, cache_bits, refs_tmp)) {
+            goto Error;
+          }
+        }
+      }
+
+      if (i == 0 && do_no_cache && cache_bits == 0) {
+        // No need to re-compute bit_cost as it was computed at i == 1.
+      } else {
+        VP8LHistogramCreate(histo, refs_tmp, cache_bits);
+        bit_cost = VP8LHistogramEstimateBits(histo);
+      }
+
+      if (bit_cost < bit_costs_best[i]) {
+        if (i == 1) {
+          // Do not swap as the full cache analysis would have the wrong
+          // VP8LBackwardRefs to start with.
+          if (!BackwardRefsClone(refs_tmp, &refs[1])) goto Error;
+        } else {
+          BackwardRefsSwap(refs_tmp, &refs[0]);
+        }
+        bit_costs_best[i] = bit_cost;
+        lz77_types_best[i] = lz77_type;
+        if (i == 0) *cache_bits_best = cache_bits;
+      }
+    }
+  }
+  assert(lz77_types_best[0] > 0);
+  assert(!do_no_cache || lz77_types_best[1] > 0);
+
+  // Improve on simple LZ77 but only for high quality (TraceBackwards is
+  // costly).
+  for (i = 1; i >= 0; --i) {
+    if (i == 1 && !do_no_cache) continue;
+    if ((lz77_types_best[i] == kLZ77Standard ||
+         lz77_types_best[i] == kLZ77Box) &&
+        quality >= 25) {
+      const VP8LHashChain* const hash_chain_tmp =
+          (lz77_types_best[i] == kLZ77Standard) ? hash_chain : &hash_chain_box;
+      const int cache_bits = (i == 1) ? 0 : *cache_bits_best;
+      float bit_cost_trace;
+      if (!VP8LBackwardReferencesTraceBackwards(width, height, argb, cache_bits,
+                                                hash_chain_tmp, &refs[i],
+                                                refs_tmp)) {
+        goto Error;
+      }
+      VP8LHistogramCreate(histo, refs_tmp, cache_bits);
+      bit_cost_trace = VP8LHistogramEstimateBits(histo);
+      if (bit_cost_trace < bit_costs_best[i]) {
+        BackwardRefsSwap(refs_tmp, &refs[i]);
+      }
+    }
+
+    BackwardReferences2DLocality(width, &refs[i]);
+
+    if (i == 1 && lz77_types_best[0] == lz77_types_best[1] &&
+        *cache_bits_best == 0) {
+      // If the best cache size is 0 and we have the same best LZ77, just copy
+      // the data over and stop here.
+      if (!BackwardRefsClone(&refs[1], &refs[0])) goto Error;
+      break;
+    }
+  }
+  status = 1;
+
+ Error:
+  VP8LHashChainClear(&hash_chain_box);
+  VP8LFreeHistogram(histo);
+  return status;
+}
+
+int VP8LGetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int low_effort, int lz77_types_to_try, int cache_bits_max, int do_no_cache,
+    const VP8LHashChain* const hash_chain, VP8LBackwardRefs* const refs,
+    int* const cache_bits_best, const WebPPicture* const pic, int percent_range,
+    int* const percent) {
+  if (low_effort) {
+    VP8LBackwardRefs* refs_best;
+    *cache_bits_best = cache_bits_max;
+    refs_best = GetBackwardReferencesLowEffort(
+        width, height, argb, cache_bits_best, hash_chain, refs);
+    if (refs_best == NULL) {
+      WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+      return 0;
+    }
+    // Set it in first position.
+    BackwardRefsSwap(refs_best, &refs[0]);
+  } else {
+    if (!GetBackwardReferences(width, height, argb, quality, lz77_types_to_try,
+                               cache_bits_max, do_no_cache, hash_chain, refs,
+                               cache_bits_best)) {
+      WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+      return 0;
+    }
+  }
+
+  return WebPReportProgress(pic, *percent + percent_range, percent);
+}
diff --git a/media/libwebp/src/enc/backward_references_enc.h b/media/libwebp/src/enc/backward_references_enc.h
new file mode 100644
index 0000000000..4dff1c27b5
--- /dev/null
+++ b/media/libwebp/src/enc/backward_references_enc.h
@@ -0,0 +1,244 @@
+// 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)
+//
+
+#ifndef WEBP_ENC_BACKWARD_REFERENCES_ENC_H_
+#define WEBP_ENC_BACKWARD_REFERENCES_ENC_H_
+
+#include <assert.h>
+#include <stdlib.h>
+#include "src/webp/types.h"
+#include "src/webp/encode.h"
+#include "src/webp/format_constants.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The maximum allowed limit is 11.
+#define MAX_COLOR_CACHE_BITS 10
+
+// -----------------------------------------------------------------------------
+// PixOrCopy
+
+enum Mode {
+  kLiteral,
+  kCacheIdx,
+  kCopy,
+  kNone
+};
+
+typedef struct {
+  // mode as uint8_t to make the memory layout to be exactly 8 bytes.
+  uint8_t mode;
+  uint16_t len;
+  uint32_t argb_or_distance;
+} PixOrCopy;
+
+static WEBP_INLINE PixOrCopy PixOrCopyCreateCopy(uint32_t distance,
+                                                 uint16_t len) {
+  PixOrCopy retval;
+  retval.mode = kCopy;
+  retval.argb_or_distance = distance;
+  retval.len = len;
+  return retval;
+}
+
+static WEBP_INLINE PixOrCopy PixOrCopyCreateCacheIdx(int idx) {
+  PixOrCopy retval;
+  assert(idx >= 0);
+  assert(idx < (1 << MAX_COLOR_CACHE_BITS));
+  retval.mode = kCacheIdx;
+  retval.argb_or_distance = idx;
+  retval.len = 1;
+  return retval;
+}
+
+static WEBP_INLINE PixOrCopy PixOrCopyCreateLiteral(uint32_t argb) {
+  PixOrCopy retval;
+  retval.mode = kLiteral;
+  retval.argb_or_distance = argb;
+  retval.len = 1;
+  return retval;
+}
+
+static WEBP_INLINE int PixOrCopyIsLiteral(const PixOrCopy* const p) {
+  return (p->mode == kLiteral);
+}
+
+static WEBP_INLINE int PixOrCopyIsCacheIdx(const PixOrCopy* const p) {
+  return (p->mode == kCacheIdx);
+}
+
+static WEBP_INLINE int PixOrCopyIsCopy(const PixOrCopy* const p) {
+  return (p->mode == kCopy);
+}
+
+static WEBP_INLINE uint32_t PixOrCopyLiteral(const PixOrCopy* const p,
+                                             int component) {
+  assert(p->mode == kLiteral);
+  return (p->argb_or_distance >> (component * 8)) & 0xff;
+}
+
+static WEBP_INLINE uint32_t PixOrCopyLength(const PixOrCopy* const p) {
+  return p->len;
+}
+
+static WEBP_INLINE uint32_t PixOrCopyCacheIdx(const PixOrCopy* const p) {
+  assert(p->mode == kCacheIdx);
+  assert(p->argb_or_distance < (1U << MAX_COLOR_CACHE_BITS));
+  return p->argb_or_distance;
+}
+
+static WEBP_INLINE uint32_t PixOrCopyDistance(const PixOrCopy* const p) {
+  assert(p->mode == kCopy);
+  return p->argb_or_distance;
+}
+
+// -----------------------------------------------------------------------------
+// VP8LHashChain
+
+#define HASH_BITS 18
+#define HASH_SIZE (1 << HASH_BITS)
+
+// If you change this, you need MAX_LENGTH_BITS + WINDOW_SIZE_BITS <= 32 as it
+// is used in VP8LHashChain.
+#define MAX_LENGTH_BITS 12
+#define WINDOW_SIZE_BITS 20
+// We want the max value to be attainable and stored in MAX_LENGTH_BITS bits.
+#define MAX_LENGTH ((1 << MAX_LENGTH_BITS) - 1)
+#if MAX_LENGTH_BITS + WINDOW_SIZE_BITS > 32
+#error "MAX_LENGTH_BITS + WINDOW_SIZE_BITS > 32"
+#endif
+
+typedef struct VP8LHashChain VP8LHashChain;
+struct VP8LHashChain {
+  // The 20 most significant bits contain the offset at which the best match
+  // is found. These 20 bits are the limit defined by GetWindowSizeForHashChain
+  // (through WINDOW_SIZE = 1<<20).
+  // The lower 12 bits contain the length of the match. The 12 bit limit is
+  // defined in MaxFindCopyLength with MAX_LENGTH=4096.
+  uint32_t* offset_length_;
+  // This is the maximum size of the hash_chain that can be constructed.
+  // Typically this is the pixel count (width x height) for a given image.
+  int size_;
+};
+
+// Must be called first, to set size.
+int VP8LHashChainInit(VP8LHashChain* const p, int size);
+// Pre-compute the best matches for argb. pic and percent are for progress.
+int VP8LHashChainFill(VP8LHashChain* const p, int quality,
+                      const uint32_t* const argb, int xsize, int ysize,
+                      int low_effort, const WebPPicture* const pic,
+                      int percent_range, int* const percent);
+void VP8LHashChainClear(VP8LHashChain* const p);  // release memory
+
+static WEBP_INLINE int VP8LHashChainFindOffset(const VP8LHashChain* const p,
+                                               const int base_position) {
+  return p->offset_length_[base_position] >> MAX_LENGTH_BITS;
+}
+
+static WEBP_INLINE int VP8LHashChainFindLength(const VP8LHashChain* const p,
+                                               const int base_position) {
+  return p->offset_length_[base_position] & ((1U << MAX_LENGTH_BITS) - 1);
+}
+
+static WEBP_INLINE void VP8LHashChainFindCopy(const VP8LHashChain* const p,
+                                              int base_position,
+                                              int* const offset_ptr,
+                                              int* const length_ptr) {
+  *offset_ptr = VP8LHashChainFindOffset(p, base_position);
+  *length_ptr = VP8LHashChainFindLength(p, base_position);
+}
+
+// -----------------------------------------------------------------------------
+// VP8LBackwardRefs (block-based backward-references storage)
+
+// maximum number of reference blocks the image will be segmented into
+#define MAX_REFS_BLOCK_PER_IMAGE 16
+
+typedef struct PixOrCopyBlock PixOrCopyBlock;   // forward declaration
+typedef struct VP8LBackwardRefs VP8LBackwardRefs;
+
+// Container for blocks chain
+struct VP8LBackwardRefs {
+  int block_size_;               // common block-size
+  int error_;                    // set to true if some memory error occurred
+  PixOrCopyBlock* refs_;         // list of currently used blocks
+  PixOrCopyBlock** tail_;        // for list recycling
+  PixOrCopyBlock* free_blocks_;  // free-list
+  PixOrCopyBlock* last_block_;   // used for adding new refs (internal)
+};
+
+// Initialize the object. 'block_size' is the common block size to store
+// references (typically, width * height / MAX_REFS_BLOCK_PER_IMAGE).
+void VP8LBackwardRefsInit(VP8LBackwardRefs* const refs, int block_size);
+// Release memory for backward references.
+void VP8LBackwardRefsClear(VP8LBackwardRefs* const refs);
+
+// Cursor for iterating on references content
+typedef struct {
+  // public:
+  PixOrCopy* cur_pos;           // current position
+  // private:
+  PixOrCopyBlock* cur_block_;   // current block in the refs list
+  const PixOrCopy* last_pos_;   // sentinel for switching to next block
+} VP8LRefsCursor;
+
+// Returns a cursor positioned at the beginning of the references list.
+VP8LRefsCursor VP8LRefsCursorInit(const VP8LBackwardRefs* const refs);
+// Returns true if cursor is pointing at a valid position.
+static WEBP_INLINE int VP8LRefsCursorOk(const VP8LRefsCursor* const c) {
+  return (c->cur_pos != NULL);
+}
+// Move to next block of references. Internal, not to be called directly.
+void VP8LRefsCursorNextBlock(VP8LRefsCursor* const c);
+// Move to next position, or NULL. Should not be called if !VP8LRefsCursorOk().
+static WEBP_INLINE void VP8LRefsCursorNext(VP8LRefsCursor* const c) {
+  assert(c != NULL);
+  assert(VP8LRefsCursorOk(c));
+  if (++c->cur_pos == c->last_pos_) VP8LRefsCursorNextBlock(c);
+}
+
+// -----------------------------------------------------------------------------
+// Main entry points
+
+enum VP8LLZ77Type {
+  kLZ77Standard = 1,
+  kLZ77RLE = 2,
+  kLZ77Box = 4
+};
+
+// Evaluates best possible backward references for specified quality.
+// The input cache_bits to 'VP8LGetBackwardReferences' sets the maximum cache
+// bits to use (passing 0 implies disabling the local color cache).
+// The optimal cache bits is evaluated and set for the *cache_bits_best
+// parameter with the matching refs_best.
+// If do_no_cache == 0, refs is an array of 2 values and the best
+// VP8LBackwardRefs is put in the first element.
+// If do_no_cache != 0, refs is an array of 3 values and the best
+// VP8LBackwardRefs is put in the first element, the best value with no-cache in
+// the second element.
+// In both cases, the last element is used as temporary internally.
+// pic and percent are for progress.
+// Returns false in case of error (stored in pic->error_code).
+int VP8LGetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int low_effort, int lz77_types_to_try, int cache_bits_max, int do_no_cache,
+    const VP8LHashChain* const hash_chain, VP8LBackwardRefs* const refs,
+    int* const cache_bits_best, const WebPPicture* const pic, int percent_range,
+    int* const percent);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // WEBP_ENC_BACKWARD_REFERENCES_ENC_H_
diff --git a/media/libwebp/src/enc/config_enc.c b/media/libwebp/src/enc/config_enc.c
new file mode 100644
index 0000000000..3518b41403
--- /dev/null
+++ b/media/libwebp/src/enc/config_enc.c
@@ -0,0 +1,157 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Coding tools configuration
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifdef HAVE_CONFIG_H
+#include "src/webp/config.h"
+#endif
+
+#include "src/webp/encode.h"
+
+//------------------------------------------------------------------------------
+// WebPConfig
+//------------------------------------------------------------------------------
+
+int WebPConfigInitInternal(WebPConfig* config,
+                           WebPPreset preset, float quality, int version) {
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_ENCODER_ABI_VERSION)) {
+    return 0;   // caller/system version mismatch!
+  }
+  if (config == NULL) return 0;
+
+  config->quality = quality;
+  config->target_size = 0;
+  config->target_PSNR = 0.;
+  config->method = 4;
+  config->sns_strength = 50;
+  config->filter_strength = 60;   // mid-filtering
+  config->filter_sharpness = 0;
+  config->filter_type = 1;        // default: strong (so U/V is filtered too)
+  config->partitions = 0;
+  config->segments = 4;
+  config->pass = 1;
+  config->qmin = 0;
+  config->qmax = 100;
+  config->show_compressed = 0;
+  config->preprocessing = 0;
+  config->autofilter = 0;
+  config->partition_limit = 0;
+  config->alpha_compression = 1;
+  config->alpha_filtering = 1;
+  config->alpha_quality = 100;
+  config->lossless = 0;
+  config->exact = 0;
+  config->image_hint = WEBP_HINT_DEFAULT;
+  config->emulate_jpeg_size = 0;
+  config->thread_level = 0;
+  config->low_memory = 0;
+  config->near_lossless = 100;
+  config->use_delta_palette = 0;
+  config->use_sharp_yuv = 0;
+
+  // TODO(skal): tune.
+  switch (preset) {
+    case WEBP_PRESET_PICTURE:
+      config->sns_strength = 80;
+      config->filter_sharpness = 4;
+      config->filter_strength = 35;
+      config->preprocessing &= ~2;   // no dithering
+      break;
+    case WEBP_PRESET_PHOTO:
+      config->sns_strength = 80;
+      config->filter_sharpness = 3;
+      config->filter_strength = 30;
+      config->preprocessing |= 2;
+      break;
+    case WEBP_PRESET_DRAWING:
+      config->sns_strength = 25;
+      config->filter_sharpness = 6;
+      config->filter_strength = 10;
+      break;
+    case WEBP_PRESET_ICON:
+      config->sns_strength = 0;
+      config->filter_strength = 0;   // disable filtering to retain sharpness
+      config->preprocessing &= ~2;   // no dithering
+      break;
+    case WEBP_PRESET_TEXT:
+      config->sns_strength = 0;
+      config->filter_strength = 0;   // disable filtering to retain sharpness
+      config->preprocessing &= ~2;   // no dithering
+      config->segments = 2;
+      break;
+    case WEBP_PRESET_DEFAULT:
+    default:
+      break;
+  }
+  return WebPValidateConfig(config);
+}
+
+int WebPValidateConfig(const WebPConfig* config) {
+  if (config == NULL) return 0;
+  if (config->quality < 0 || config->quality > 100) return 0;
+  if (config->target_size < 0) return 0;
+  if (config->target_PSNR < 0) return 0;
+  if (config->method < 0 || config->method > 6) return 0;
+  if (config->segments < 1 || config->segments > 4) return 0;
+  if (config->sns_strength < 0 || config->sns_strength > 100) return 0;
+  if (config->filter_strength < 0 || config->filter_strength > 100) return 0;
+  if (config->filter_sharpness < 0 || config->filter_sharpness > 7) return 0;
+  if (config->filter_type < 0 || config->filter_type > 1) return 0;
+  if (config->autofilter < 0 || config->autofilter > 1) return 0;
+  if (config->pass < 1 || config->pass > 10) return 0;
+  if (config->qmin < 0 || config->qmax > 100 || config->qmin > config->qmax) {
+    return 0;
+  }
+  if (config->show_compressed < 0 || config->show_compressed > 1) return 0;
+  if (config->preprocessing < 0 || config->preprocessing > 7) return 0;
+  if (config->partitions < 0 || config->partitions > 3) return 0;
+  if (config->partition_limit < 0 || config->partition_limit > 100) return 0;
+  if (config->alpha_compression < 0) return 0;
+  if (config->alpha_filtering < 0) return 0;
+  if (config->alpha_quality < 0 || config->alpha_quality > 100) return 0;
+  if (config->lossless < 0 || config->lossless > 1) return 0;
+  if (config->near_lossless < 0 || config->near_lossless > 100) return 0;
+  if (config->image_hint >= WEBP_HINT_LAST) return 0;
+  if (config->emulate_jpeg_size < 0 || config->emulate_jpeg_size > 1) return 0;
+  if (config->thread_level < 0 || config->thread_level > 1) return 0;
+  if (config->low_memory < 0 || config->low_memory > 1) return 0;
+  if (config->exact < 0 || config->exact > 1) return 0;
+  if (config->use_delta_palette < 0 || config->use_delta_palette > 1) {
+    return 0;
+  }
+  if (config->use_sharp_yuv < 0 || config->use_sharp_yuv > 1) return 0;
+
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+#define MAX_LEVEL 9
+
+// Mapping between -z level and -m / -q parameter settings.
+static const struct {
+  uint8_t method_;
+  uint8_t quality_;
+} kLosslessPresets[MAX_LEVEL + 1] = {
+  { 0,  0 }, { 1, 20 }, { 2, 25 }, { 3, 30 }, { 3, 50 },
+  { 4, 50 }, { 4, 75 }, { 4, 90 }, { 5, 90 }, { 6, 100 }
+};
+
+int WebPConfigLosslessPreset(WebPConfig* config, int level) {
+  if (config == NULL || level < 0 || level > MAX_LEVEL) return 0;
+  config->lossless = 1;
+  config->method = kLosslessPresets[level].method_;
+  config->quality = kLosslessPresets[level].quality_;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/cost_enc.c b/media/libwebp/src/enc/cost_enc.c
new file mode 100644
index 0000000000..48fd9bc347
--- /dev/null
+++ b/media/libwebp/src/enc/cost_enc.c
@@ -0,0 +1,342 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Cost tables for level and modes
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "src/enc/cost_enc.h"
+
+//------------------------------------------------------------------------------
+// Level cost tables
+
+// For each given level, the following table gives the pattern of contexts to
+// use for coding it (in [][0]) as well as the bit value to use for each
+// context (in [][1]).
+const uint16_t VP8LevelCodes[MAX_VARIABLE_LEVEL][2] = {
+                  {0x001, 0x000}, {0x007, 0x001}, {0x00f, 0x005},
+  {0x00f, 0x00d}, {0x033, 0x003}, {0x033, 0x003}, {0x033, 0x023},
+  {0x033, 0x023}, {0x033, 0x023}, {0x033, 0x023}, {0x0d3, 0x013},
+  {0x0d3, 0x013}, {0x0d3, 0x013}, {0x0d3, 0x013}, {0x0d3, 0x013},
+  {0x0d3, 0x013}, {0x0d3, 0x013}, {0x0d3, 0x013}, {0x0d3, 0x093},
+  {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093},
+  {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093},
+  {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093},
+  {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x153}
+};
+
+static int VariableLevelCost(int level, const uint8_t probas[NUM_PROBAS]) {
+  int pattern = VP8LevelCodes[level - 1][0];
+  int bits = VP8LevelCodes[level - 1][1];
+  int cost = 0;
+  int i;
+  for (i = 2; pattern; ++i) {
+    if (pattern & 1) {
+      cost += VP8BitCost(bits & 1, probas[i]);
+    }
+    bits >>= 1;
+    pattern >>= 1;
+  }
+  return cost;
+}
+
+//------------------------------------------------------------------------------
+// Pre-calc level costs once for all
+
+void VP8CalculateLevelCosts(VP8EncProba* const proba) {
+  int ctype, band, ctx;
+
+  if (!proba->dirty_) return;  // nothing to do.
+
+  for (ctype = 0; ctype < NUM_TYPES; ++ctype) {
+    int n;
+    for (band = 0; band < NUM_BANDS; ++band) {
+      for (ctx = 0; ctx < NUM_CTX; ++ctx) {
+        const uint8_t* const p = proba->coeffs_[ctype][band][ctx];
+        uint16_t* const table = proba->level_cost_[ctype][band][ctx];
+        const int cost0 = (ctx > 0) ? VP8BitCost(1, p[0]) : 0;
+        const int cost_base = VP8BitCost(1, p[1]) + cost0;
+        int v;
+        table[0] = VP8BitCost(0, p[1]) + cost0;
+        for (v = 1; v <= MAX_VARIABLE_LEVEL; ++v) {
+          table[v] = cost_base + VariableLevelCost(v, p);
+        }
+        // Starting at level 67 and up, the variable part of the cost is
+        // actually constant.
+      }
+    }
+    for (n = 0; n < 16; ++n) {    // replicate bands. We don't need to sentinel.
+      for (ctx = 0; ctx < NUM_CTX; ++ctx) {
+        proba->remapped_costs_[ctype][n][ctx] =
+            proba->level_cost_[ctype][VP8EncBands[n]][ctx];
+      }
+    }
+  }
+  proba->dirty_ = 0;
+}
+
+//------------------------------------------------------------------------------
+// Mode cost tables.
+
+// These are the fixed probabilities (in the coding trees) turned into bit-cost
+// by calling VP8BitCost().
+const uint16_t VP8FixedCostsUV[4] = { 302, 984, 439, 642 };
+// note: these values include the fixed VP8BitCost(1, 145) mode selection cost.
+const uint16_t VP8FixedCostsI16[4] = { 663, 919, 872, 919 };
+const uint16_t VP8FixedCostsI4[NUM_BMODES][NUM_BMODES][NUM_BMODES] = {
+  { {   40, 1151, 1723, 1874, 2103, 2019, 1628, 1777, 2226, 2137 },
+    {  192,  469, 1296, 1308, 1849, 1794, 1781, 1703, 1713, 1522 },
+    {  142,  910,  762, 1684, 1849, 1576, 1460, 1305, 1801, 1657 },
+    {  559,  641, 1370,  421, 1182, 1569, 1612, 1725,  863, 1007 },
+    {  299, 1059, 1256, 1108,  636, 1068, 1581, 1883,  869, 1142 },
+    {  277, 1111,  707, 1362, 1089,  672, 1603, 1541, 1545, 1291 },
+    {  214,  781, 1609, 1303, 1632, 2229,  726, 1560, 1713,  918 },
+    {  152, 1037, 1046, 1759, 1983, 2174, 1358,  742, 1740, 1390 },
+    {  512, 1046, 1420,  753,  752, 1297, 1486, 1613,  460, 1207 },
+    {  424,  827, 1362,  719, 1462, 1202, 1199, 1476, 1199,  538 } },
+  { {  240,  402, 1134, 1491, 1659, 1505, 1517, 1555, 1979, 2099 },
+    {  467,  242,  960, 1232, 1714, 1620, 1834, 1570, 1676, 1391 },
+    {  500,  455,  463, 1507, 1699, 1282, 1564,  982, 2114, 2114 },
+    {  672,  643, 1372,  331, 1589, 1667, 1453, 1938,  996,  876 },
+    {  458,  783, 1037,  911,  738,  968, 1165, 1518,  859, 1033 },
+    {  504,  815,  504, 1139, 1219,  719, 1506, 1085, 1268, 1268 },
+    {  333,  630, 1445, 1239, 1883, 3672,  799, 1548, 1865,  598 },
+    {  399,  644,  746, 1342, 1856, 1350, 1493,  613, 1855, 1015 },
+    {  622,  749, 1205,  608, 1066, 1408, 1290, 1406,  546,  971 },
+    {  500,  753, 1041,  668, 1230, 1617, 1297, 1425, 1383,  523 } },
+  { {  394,  553,  523, 1502, 1536,  981, 1608, 1142, 1666, 2181 },
+    {  655,  430,  375, 1411, 1861, 1220, 1677, 1135, 1978, 1553 },
+    {  690,  640,  245, 1954, 2070, 1194, 1528,  982, 1972, 2232 },
+    {  559,  834,  741,  867, 1131,  980, 1225,  852, 1092,  784 },
+    {  690,  875,  516,  959,  673,  894, 1056, 1190, 1528, 1126 },
+    {  740,  951,  384, 1277, 1177,  492, 1579, 1155, 1846, 1513 },
+    {  323,  775, 1062, 1776, 3062, 1274,  813, 1188, 1372,  655 },
+    {  488,  971,  484, 1767, 1515, 1775, 1115,  503, 1539, 1461 },
+    {  740, 1006,  998,  709,  851, 1230, 1337,  788,  741,  721 },
+    {  522, 1073,  573, 1045, 1346,  887, 1046, 1146, 1203,  697 } },
+  { {  105,  864, 1442, 1009, 1934, 1840, 1519, 1920, 1673, 1579 },
+    {  534,  305, 1193,  683, 1388, 2164, 1802, 1894, 1264, 1170 },
+    {  305,  518,  877, 1108, 1426, 3215, 1425, 1064, 1320, 1242 },
+    {  683,  732, 1927,  257, 1493, 2048, 1858, 1552, 1055,  947 },
+    {  394,  814, 1024,  660,  959, 1556, 1282, 1289,  893, 1047 },
+    {  528,  615,  996,  940, 1201,  635, 1094, 2515,  803, 1358 },
+    {  347,  614, 1609, 1187, 3133, 1345, 1007, 1339, 1017,  667 },
+    {  218,  740,  878, 1605, 3650, 3650, 1345,  758, 1357, 1617 },
+    {  672,  750, 1541,  558, 1257, 1599, 1870, 2135,  402, 1087 },
+    {  592,  684, 1161,  430, 1092, 1497, 1475, 1489, 1095,  822 } },
+  { {  228, 1056, 1059, 1368,  752,  982, 1512, 1518,  987, 1782 },
+    {  494,  514,  818,  942,  965,  892, 1610, 1356, 1048, 1363 },
+    {  512,  648,  591, 1042,  761,  991, 1196, 1454, 1309, 1463 },
+    {  683,  749, 1043,  676,  841, 1396, 1133, 1138,  654,  939 },
+    {  622, 1101, 1126,  994,  361, 1077, 1203, 1318,  877, 1219 },
+    {  631, 1068,  857, 1650,  651,  477, 1650, 1419,  828, 1170 },
+    {  555,  727, 1068, 1335, 3127, 1339,  820, 1331, 1077,  429 },
+    {  504,  879,  624, 1398,  889,  889, 1392,  808,  891, 1406 },
+    {  683, 1602, 1289,  977,  578,  983, 1280, 1708,  406, 1122 },
+    {  399,  865, 1433, 1070, 1072,  764,  968, 1477, 1223,  678 } },
+  { {  333,  760,  935, 1638, 1010,  529, 1646, 1410, 1472, 2219 },
+    {  512,  494,  750, 1160, 1215,  610, 1870, 1868, 1628, 1169 },
+    {  572,  646,  492, 1934, 1208,  603, 1580, 1099, 1398, 1995 },
+    {  786,  789,  942,  581, 1018,  951, 1599, 1207,  731,  768 },
+    {  690, 1015,  672, 1078,  582,  504, 1693, 1438, 1108, 2897 },
+    {  768, 1267,  571, 2005, 1243,  244, 2881, 1380, 1786, 1453 },
+    {  452,  899, 1293,  903, 1311, 3100,  465, 1311, 1319,  813 },
+    {  394,  927,  942, 1103, 1358, 1104,  946,  593, 1363, 1109 },
+    {  559, 1005, 1007, 1016,  658, 1173, 1021, 1164,  623, 1028 },
+    {  564,  796,  632, 1005, 1014,  863, 2316, 1268,  938,  764 } },
+  { {  266,  606, 1098, 1228, 1497, 1243,  948, 1030, 1734, 1461 },
+    {  366,  585,  901, 1060, 1407, 1247,  876, 1134, 1620, 1054 },
+    {  452,  565,  542, 1729, 1479, 1479, 1016,  886, 2938, 1150 },
+    {  555, 1088, 1533,  950, 1354,  895,  834, 1019, 1021,  496 },
+    {  704,  815, 1193,  971,  973,  640, 1217, 2214,  832,  578 },
+    {  672, 1245,  579,  871,  875,  774,  872, 1273, 1027,  949 },
+    {  296, 1134, 2050, 1784, 1636, 3425,  442, 1550, 2076,  722 },
+    {  342,  982, 1259, 1846, 1848, 1848,  622,  568, 1847, 1052 },
+    {  555, 1064, 1304,  828,  746, 1343, 1075, 1329, 1078,  494 },
+    {  288, 1167, 1285, 1174, 1639, 1639,  833, 2254, 1304,  509 } },
+  { {  342,  719,  767, 1866, 1757, 1270, 1246,  550, 1746, 2151 },
+    {  483,  653,  694, 1509, 1459, 1410, 1218,  507, 1914, 1266 },
+    {  488,  757,  447, 2979, 1813, 1268, 1654,  539, 1849, 2109 },
+    {  522, 1097, 1085,  851, 1365, 1111,  851,  901,  961,  605 },
+    {  709,  716,  841,  728,  736,  945,  941,  862, 2845, 1057 },
+    {  512, 1323,  500, 1336, 1083,  681, 1342,  717, 1604, 1350 },
+    {  452, 1155, 1372, 1900, 1501, 3290,  311,  944, 1919,  922 },
+    {  403, 1520,  977, 2132, 1733, 3522, 1076,  276, 3335, 1547 },
+    {  559, 1374, 1101,  615,  673, 2462,  974,  795,  984,  984 },
+    {  547, 1122, 1062,  812, 1410,  951, 1140,  622, 1268,  651 } },
+  { {  165,  982, 1235,  938, 1334, 1366, 1659, 1578,  964, 1612 },
+    {  592,  422,  925,  847, 1139, 1112, 1387, 2036,  861, 1041 },
+    {  403,  837,  732,  770,  941, 1658, 1250,  809, 1407, 1407 },
+    {  896,  874, 1071,  381, 1568, 1722, 1437, 2192,  480, 1035 },
+    {  640, 1098, 1012, 1032,  684, 1382, 1581, 2106,  416,  865 },
+    {  559, 1005,  819,  914,  710,  770, 1418,  920,  838, 1435 },
+    {  415, 1258, 1245,  870, 1278, 3067,  770, 1021, 1287,  522 },
+    {  406,  990,  601, 1009, 1265, 1265, 1267,  759, 1017, 1277 },
+    {  968, 1182, 1329,  788, 1032, 1292, 1705, 1714,  203, 1403 },
+    {  732,  877, 1279,  471,  901, 1161, 1545, 1294,  755,  755 } },
+  { {  111,  931, 1378, 1185, 1933, 1648, 1148, 1714, 1873, 1307 },
+    {  406,  414, 1030, 1023, 1910, 1404, 1313, 1647, 1509,  793 },
+    {  342,  640,  575, 1088, 1241, 1349, 1161, 1350, 1756, 1502 },
+    {  559,  766, 1185,  357, 1682, 1428, 1329, 1897, 1219,  802 },
+    {  473,  909, 1164,  771,  719, 2508, 1427, 1432,  722,  782 },
+    {  342,  892,  785, 1145, 1150,  794, 1296, 1550,  973, 1057 },
+    {  208, 1036, 1326, 1343, 1606, 3395,  815, 1455, 1618,  712 },
+    {  228,  928,  890, 1046, 3499, 1711,  994,  829, 1720, 1318 },
+    {  768,  724, 1058,  636,  991, 1075, 1319, 1324,  616,  825 },
+    {  305, 1167, 1358,  899, 1587, 1587,  987, 1988, 1332,  501 } }
+};
+
+//------------------------------------------------------------------------------
+// helper functions for residuals struct VP8Residual.
+
+void VP8InitResidual(int first, int coeff_type,
+                     VP8Encoder* const enc, VP8Residual* const res) {
+  res->coeff_type = coeff_type;
+  res->prob  = enc->proba_.coeffs_[coeff_type];
+  res->stats = enc->proba_.stats_[coeff_type];
+  res->costs = enc->proba_.remapped_costs_[coeff_type];
+  res->first = first;
+}
+
+//------------------------------------------------------------------------------
+// Mode costs
+
+int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]) {
+  const int x = (it->i4_ & 3), y = (it->i4_ >> 2);
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int R = 0;
+  int ctx;
+
+  VP8InitResidual(0, 3, enc, &res);
+  ctx = it->top_nz_[x] + it->left_nz_[y];
+  VP8SetResidualCoeffs(levels, &res);
+  R += VP8GetResidualCost(ctx, &res);
+  return R;
+}
+
+int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd) {
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int x, y;
+  int R = 0;
+
+  VP8IteratorNzToBytes(it);   // re-import the non-zero context
+
+  // DC
+  VP8InitResidual(0, 1, enc, &res);
+  VP8SetResidualCoeffs(rd->y_dc_levels, &res);
+  R += VP8GetResidualCost(it->top_nz_[8] + it->left_nz_[8], &res);
+
+  // AC
+  VP8InitResidual(1, 0, enc, &res);
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      R += VP8GetResidualCost(ctx, &res);
+      it->top_nz_[x] = it->left_nz_[y] = (res.last >= 0);
+    }
+  }
+  return R;
+}
+
+int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int ch, x, y;
+  int R = 0;
+
+  VP8IteratorNzToBytes(it);  // re-import the non-zero context
+
+  VP8InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        R += VP8GetResidualCost(ctx, &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = (res.last >= 0);
+      }
+    }
+  }
+  return R;
+}
+
+
+//------------------------------------------------------------------------------
+// Recording of token probabilities.
+
+// We keep the table-free variant around for reference, in case.
+#define USE_LEVEL_CODE_TABLE
+
+// Simulate block coding, but only record statistics.
+// Note: no need to record the fixed probas.
+int VP8RecordCoeffs(int ctx, const VP8Residual* const res) {
+  int n = res->first;
+  // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  proba_t* s = res->stats[n][ctx];
+  if (res->last  < 0) {
+    VP8RecordStats(0, s + 0);
+    return 0;
+  }
+  while (n <= res->last) {
+    int v;
+    VP8RecordStats(1, s + 0);  // order of record doesn't matter
+    while ((v = res->coeffs[n++]) == 0) {
+      VP8RecordStats(0, s + 1);
+      s = res->stats[VP8EncBands[n]][0];
+    }
+    VP8RecordStats(1, s + 1);
+    if (!VP8RecordStats(2u < (unsigned int)(v + 1), s + 2)) {  // v = -1 or 1
+      s = res->stats[VP8EncBands[n]][1];
+    } else {
+      v = abs(v);
+#if !defined(USE_LEVEL_CODE_TABLE)
+      if (!VP8RecordStats(v > 4, s + 3)) {
+        if (VP8RecordStats(v != 2, s + 4))
+          VP8RecordStats(v == 4, s + 5);
+      } else if (!VP8RecordStats(v > 10, s + 6)) {
+        VP8RecordStats(v > 6, s + 7);
+      } else if (!VP8RecordStats((v >= 3 + (8 << 2)), s + 8)) {
+        VP8RecordStats((v >= 3 + (8 << 1)), s + 9);
+      } else {
+        VP8RecordStats((v >= 3 + (8 << 3)), s + 10);
+      }
+#else
+      if (v > MAX_VARIABLE_LEVEL) {
+        v = MAX_VARIABLE_LEVEL;
+      }
+
+      {
+        const int bits = VP8LevelCodes[v - 1][1];
+        int pattern = VP8LevelCodes[v - 1][0];
+        int i;
+        for (i = 0; (pattern >>= 1) != 0; ++i) {
+          const int mask = 2 << i;
+          if (pattern & 1) VP8RecordStats(!!(bits & mask), s + 3 + i);
+        }
+      }
+#endif
+      s = res->stats[VP8EncBands[n]][2];
+    }
+  }
+  if (n < 16) VP8RecordStats(0, s + 0);
+  return 1;
+}
+
+//------------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/cost_enc.h b/media/libwebp/src/enc/cost_enc.h
new file mode 100644
index 0000000000..a4b177b342
--- /dev/null
+++ b/media/libwebp/src/enc/cost_enc.h
@@ -0,0 +1,82 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Cost tables for level and modes.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_ENC_COST_ENC_H_
+#define WEBP_ENC_COST_ENC_H_
+
+#include <assert.h>
+#include <stdlib.h>
+#include "src/enc/vp8i_enc.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// On-the-fly info about the current set of residuals. Handy to avoid
+// passing zillions of params.
+typedef struct VP8Residual VP8Residual;
+struct VP8Residual {
+  int first;
+  int last;
+  const int16_t* coeffs;
+
+  int coeff_type;
+  ProbaArray*   prob;
+  StatsArray*   stats;
+  CostArrayPtr  costs;
+};
+
+void VP8InitResidual(int first, int coeff_type,
+                     VP8Encoder* const enc, VP8Residual* const res);
+
+int VP8RecordCoeffs(int ctx, const VP8Residual* const res);
+
+// Record proba context used.
+static WEBP_INLINE int VP8RecordStats(int bit, proba_t* const stats) {
+  proba_t p = *stats;
+  // An overflow is inbound. Note we handle this at 0xfffe0000u instead of
+  // 0xffff0000u to make sure p + 1u does not overflow.
+  if (p >= 0xfffe0000u) {
+    p = ((p + 1u) >> 1) & 0x7fff7fffu;  // -> divide the stats by 2.
+  }
+  // record bit count (lower 16 bits) and increment total count (upper 16 bits).
+  p += 0x00010000u + bit;
+  *stats = p;
+  return bit;
+}
+
+// Cost of coding one event with probability 'proba'.
+static WEBP_INLINE int VP8BitCost(int bit, uint8_t proba) {
+  return !bit ? VP8EntropyCost[proba] : VP8EntropyCost[255 - proba];
+}
+
+// Level cost calculations
+extern const uint16_t VP8LevelCodes[MAX_VARIABLE_LEVEL][2];
+void VP8CalculateLevelCosts(VP8EncProba* const proba);
+static WEBP_INLINE int VP8LevelCost(const uint16_t* const table, int level) {
+  return VP8LevelFixedCosts[level]
+       + table[(level > MAX_VARIABLE_LEVEL) ? MAX_VARIABLE_LEVEL : level];
+}
+
+// Mode costs
+extern const uint16_t VP8FixedCostsUV[4];
+extern const uint16_t VP8FixedCostsI16[4];
+extern const uint16_t VP8FixedCostsI4[NUM_BMODES][NUM_BMODES][NUM_BMODES];
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // WEBP_ENC_COST_ENC_H_
diff --git a/media/libwebp/src/enc/filter_enc.c b/media/libwebp/src/enc/filter_enc.c
new file mode 100644
index 0000000000..580800bfb8
--- /dev/null
+++ b/media/libwebp/src/enc/filter_enc.c
@@ -0,0 +1,235 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Selecting filter level
+//
+// Author: somnath@google.com (Somnath Banerjee)
+
+#include <assert.h>
+#include "src/enc/vp8i_enc.h"
+#include "src/dsp/dsp.h"
+
+// This table gives, for a given sharpness, the filtering strength to be
+// used (at least) in order to filter a given edge step delta.
+// This is constructed by brute force inspection: for all delta, we iterate
+// over all possible filtering strength / thresh until needs_filter() returns
+// true.
+#define MAX_DELTA_SIZE 64
+static const uint8_t kLevelsFromDelta[8][MAX_DELTA_SIZE] = {
+  { 0,   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 },
+  { 0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 14, 15, 17, 18,
+    20, 21, 23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42,
+    44, 45, 47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 14, 16, 17, 19,
+    20, 22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43,
+    44, 46, 47, 49, 50, 52, 53, 55, 56, 58, 59, 61, 62, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 15, 16, 18, 19,
+    21, 22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43,
+    45, 46, 48, 49, 51, 52, 54, 55, 57, 58, 60, 61, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 14, 15, 17, 18, 20,
+    21, 23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42, 44,
+    45, 47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 13, 15, 16, 17, 19, 20,
+    22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43, 44,
+    46, 47, 49, 50, 52, 53, 55, 56, 58, 59, 61, 62, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 13, 15, 16, 18, 19, 21,
+    22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43, 45,
+    46, 48, 49, 51, 52, 54, 55, 57, 58, 60, 61, 63, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 14, 15, 17, 18, 20, 21,
+    23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42, 44, 45,
+    47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 }
+};
+
+int VP8FilterStrengthFromDelta(int sharpness, int delta) {
+  const int pos = (delta < MAX_DELTA_SIZE) ? delta : MAX_DELTA_SIZE - 1;
+  assert(sharpness >= 0 && sharpness <= 7);
+  return kLevelsFromDelta[sharpness][pos];
+}
+
+//------------------------------------------------------------------------------
+// Paragraph 15.4: compute the inner-edge filtering strength
+
+#if !defined(WEBP_REDUCE_SIZE)
+
+static int GetILevel(int sharpness, int level) {
+  if (sharpness > 0) {
+    if (sharpness > 4) {
+      level >>= 2;
+    } else {
+      level >>= 1;
+    }
+    if (level > 9 - sharpness) {
+      level = 9 - sharpness;
+    }
+  }
+  if (level < 1) level = 1;
+  return level;
+}
+
+static void DoFilter(const VP8EncIterator* const it, int level) {
+  const VP8Encoder* const enc = it->enc_;
+  const int ilevel = GetILevel(enc->config_->filter_sharpness, level);
+  const int limit = 2 * level + ilevel;
+
+  uint8_t* const y_dst = it->yuv_out2_ + Y_OFF_ENC;
+  uint8_t* const u_dst = it->yuv_out2_ + U_OFF_ENC;
+  uint8_t* const v_dst = it->yuv_out2_ + V_OFF_ENC;
+
+  // copy current block to yuv_out2_
+  memcpy(y_dst, it->yuv_out_, YUV_SIZE_ENC * sizeof(uint8_t));
+
+  if (enc->filter_hdr_.simple_ == 1) {   // simple
+    VP8SimpleHFilter16i(y_dst, BPS, limit);
+    VP8SimpleVFilter16i(y_dst, BPS, limit);
+  } else {    // complex
+    const int hev_thresh = (level >= 40) ? 2 : (level >= 15) ? 1 : 0;
+    VP8HFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8HFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+    VP8VFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8VFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+  }
+}
+
+//------------------------------------------------------------------------------
+// SSIM metric for one macroblock
+
+static double GetMBSSIM(const uint8_t* yuv1, const uint8_t* yuv2) {
+  int x, y;
+  double sum = 0.;
+
+  // compute SSIM in a 10 x 10 window
+  for (y = VP8_SSIM_KERNEL; y < 16 - VP8_SSIM_KERNEL; y++) {
+    for (x = VP8_SSIM_KERNEL; x < 16 - VP8_SSIM_KERNEL; x++) {
+      sum += VP8SSIMGetClipped(yuv1 + Y_OFF_ENC, BPS, yuv2 + Y_OFF_ENC, BPS,
+                               x, y, 16, 16);
+    }
+  }
+  for (x = 1; x < 7; x++) {
+    for (y = 1; y < 7; y++) {
+      sum += VP8SSIMGetClipped(yuv1 + U_OFF_ENC, BPS, yuv2 + U_OFF_ENC, BPS,
+                               x, y, 8, 8);
+      sum += VP8SSIMGetClipped(yuv1 + V_OFF_ENC, BPS, yuv2 + V_OFF_ENC, BPS,
+                               x, y, 8, 8);
+    }
+  }
+  return sum;
+}
+
+#endif  // !defined(WEBP_REDUCE_SIZE)
+
+//------------------------------------------------------------------------------
+// Exposed APIs: Encoder should call the following 3 functions to adjust
+// loop filter strength
+
+void VP8InitFilter(VP8EncIterator* const it) {
+#if !defined(WEBP_REDUCE_SIZE)
+  if (it->lf_stats_ != NULL) {
+    int s, i;
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      for (i = 0; i < MAX_LF_LEVELS; i++) {
+        (*it->lf_stats_)[s][i] = 0;
+      }
+    }
+    VP8SSIMDspInit();
+  }
+#else
+  (void)it;
+#endif
+}
+
+void VP8StoreFilterStats(VP8EncIterator* const it) {
+#if !defined(WEBP_REDUCE_SIZE)
+  int d;
+  VP8Encoder* const enc = it->enc_;
+  const int s = it->mb_->segment_;
+  const int level0 = enc->dqm_[s].fstrength_;
+
+  // explore +/-quant range of values around level0
+  const int delta_min = -enc->dqm_[s].quant_;
+  const int delta_max = enc->dqm_[s].quant_;
+  const int step_size = (delta_max - delta_min >= 4) ? 4 : 1;
+
+  if (it->lf_stats_ == NULL) return;
+
+  // NOTE: Currently we are applying filter only across the sublock edges
+  // There are two reasons for that.
+  // 1. Applying filter on macro block edges will change the pixels in
+  // the left and top macro blocks. That will be hard to restore
+  // 2. Macro Blocks on the bottom and right are not yet compressed. So we
+  // cannot apply filter on the right and bottom macro block edges.
+  if (it->mb_->type_ == 1 && it->mb_->skip_) return;
+
+  // Always try filter level  zero
+  (*it->lf_stats_)[s][0] += GetMBSSIM(it->yuv_in_, it->yuv_out_);
+
+  for (d = delta_min; d <= delta_max; d += step_size) {
+    const int level = level0 + d;
+    if (level <= 0 || level >= MAX_LF_LEVELS) {
+      continue;
+    }
+    DoFilter(it, level);
+    (*it->lf_stats_)[s][level] += GetMBSSIM(it->yuv_in_, it->yuv_out2_);
+  }
+#else  // defined(WEBP_REDUCE_SIZE)
+  (void)it;
+#endif  // !defined(WEBP_REDUCE_SIZE)
+}
+
+void VP8AdjustFilterStrength(VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+#if !defined(WEBP_REDUCE_SIZE)
+  if (it->lf_stats_ != NULL) {
+    int s;
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      int i, best_level = 0;
+      // Improvement over filter level 0 should be at least 1e-5 (relatively)
+      double best_v = 1.00001 * (*it->lf_stats_)[s][0];
+      for (i = 1; i < MAX_LF_LEVELS; i++) {
+        const double v = (*it->lf_stats_)[s][i];
+        if (v > best_v) {
+          best_v = v;
+          best_level = i;
+        }
+      }
+      enc->dqm_[s].fstrength_ = best_level;
+    }
+    return;
+  }
+#endif  // !defined(WEBP_REDUCE_SIZE)
+  if (enc->config_->filter_strength > 0) {
+    int max_level = 0;
+    int s;
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      VP8SegmentInfo* const dqm = &enc->dqm_[s];
+      // this '>> 3' accounts for some inverse WHT scaling
+      const int delta = (dqm->max_edge_ * dqm->y2_.q_[1]) >> 3;
+      const int level =
+          VP8FilterStrengthFromDelta(enc->filter_hdr_.sharpness_, delta);
+      if (level > dqm->fstrength_) {
+        dqm->fstrength_ = level;
+      }
+      if (max_level < dqm->fstrength_) {
+        max_level = dqm->fstrength_;
+      }
+    }
+    enc->filter_hdr_.level_ = max_level;
+  }
+}
+
+// -----------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/frame_enc.c b/media/libwebp/src/enc/frame_enc.c
new file mode 100644
index 0000000000..b93d9e5b99
--- /dev/null
+++ b/media/libwebp/src/enc/frame_enc.c
@@ -0,0 +1,899 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+//   frame coding and analysis
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <string.h>
+#include <math.h>
+
+#include "src/enc/cost_enc.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/dsp/dsp.h"
+#include "src/webp/format_constants.h"  // RIFF constants
+
+#define SEGMENT_VISU 0
+#define DEBUG_SEARCH 0    // useful to track search convergence
+
+//------------------------------------------------------------------------------
+// multi-pass convergence
+
+#define HEADER_SIZE_ESTIMATE (RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE +  \
+                              VP8_FRAME_HEADER_SIZE)
+#define DQ_LIMIT 0.4  // convergence is considered reached if dq < DQ_LIMIT
+// we allow 2k of extra head-room in PARTITION0 limit.
+#define PARTITION0_SIZE_LIMIT ((VP8_MAX_PARTITION0_SIZE - 2048ULL) << 11)
+
+static float Clamp(float v, float min, float max) {
+  return (v < min) ? min : (v > max) ? max : v;
+}
+
+typedef struct {  // struct for organizing convergence in either size or PSNR
+  int is_first;
+  float dq;
+  float q, last_q;
+  float qmin, qmax;
+  double value, last_value;   // PSNR or size
+  double target;
+  int do_size_search;
+} PassStats;
+
+static int InitPassStats(const VP8Encoder* const enc, PassStats* const s) {
+  const uint64_t target_size = (uint64_t)enc->config_->target_size;
+  const int do_size_search = (target_size != 0);
+  const float target_PSNR = enc->config_->target_PSNR;
+
+  s->is_first = 1;
+  s->dq = 10.f;
+  s->qmin = 1.f * enc->config_->qmin;
+  s->qmax = 1.f * enc->config_->qmax;
+  s->q = s->last_q = Clamp(enc->config_->quality, s->qmin, s->qmax);
+  s->target = do_size_search ? (double)target_size
+            : (target_PSNR > 0.) ? target_PSNR
+            : 40.;   // default, just in case
+  s->value = s->last_value = 0.;
+  s->do_size_search = do_size_search;
+  return do_size_search;
+}
+
+static float ComputeNextQ(PassStats* const s) {
+  float dq;
+  if (s->is_first) {
+    dq = (s->value > s->target) ? -s->dq : s->dq;
+    s->is_first = 0;
+  } else if (s->value != s->last_value) {
+    const double slope = (s->target - s->value) / (s->last_value - s->value);
+    dq = (float)(slope * (s->last_q - s->q));
+  } else {
+    dq = 0.;  // we're done?!
+  }
+  // Limit variable to avoid large swings.
+  s->dq = Clamp(dq, -30.f, 30.f);
+  s->last_q = s->q;
+  s->last_value = s->value;
+  s->q = Clamp(s->q + s->dq, s->qmin, s->qmax);
+  return s->q;
+}
+
+//------------------------------------------------------------------------------
+// Tables for level coding
+
+const uint8_t VP8Cat3[] = { 173, 148, 140 };
+const uint8_t VP8Cat4[] = { 176, 155, 140, 135 };
+const uint8_t VP8Cat5[] = { 180, 157, 141, 134, 130 };
+const uint8_t VP8Cat6[] =
+    { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129 };
+
+//------------------------------------------------------------------------------
+// Reset the statistics about: number of skips, token proba, level cost,...
+
+static void ResetStats(VP8Encoder* const enc) {
+  VP8EncProba* const proba = &enc->proba_;
+  VP8CalculateLevelCosts(proba);
+  proba->nb_skip_ = 0;
+}
+
+//------------------------------------------------------------------------------
+// Skip decision probability
+
+#define SKIP_PROBA_THRESHOLD 250  // value below which using skip_proba is OK.
+
+static int CalcSkipProba(uint64_t nb, uint64_t total) {
+  return (int)(total ? (total - nb) * 255 / total : 255);
+}
+
+// Returns the bit-cost for coding the skip probability.
+static int FinalizeSkipProba(VP8Encoder* const enc) {
+  VP8EncProba* const proba = &enc->proba_;
+  const int nb_mbs = enc->mb_w_ * enc->mb_h_;
+  const int nb_events = proba->nb_skip_;
+  int size;
+  proba->skip_proba_ = CalcSkipProba(nb_events, nb_mbs);
+  proba->use_skip_proba_ = (proba->skip_proba_ < SKIP_PROBA_THRESHOLD);
+  size = 256;   // 'use_skip_proba' bit
+  if (proba->use_skip_proba_) {
+    size +=  nb_events * VP8BitCost(1, proba->skip_proba_)
+         + (nb_mbs - nb_events) * VP8BitCost(0, proba->skip_proba_);
+    size += 8 * 256;   // cost of signaling the skip_proba_ itself.
+  }
+  return size;
+}
+
+// Collect statistics and deduce probabilities for next coding pass.
+// Return the total bit-cost for coding the probability updates.
+static int CalcTokenProba(int nb, int total) {
+  assert(nb <= total);
+  return nb ? (255 - nb * 255 / total) : 255;
+}
+
+// Cost of coding 'nb' 1's and 'total-nb' 0's using 'proba' probability.
+static int BranchCost(int nb, int total, int proba) {
+  return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba);
+}
+
+static void ResetTokenStats(VP8Encoder* const enc) {
+  VP8EncProba* const proba = &enc->proba_;
+  memset(proba->stats_, 0, sizeof(proba->stats_));
+}
+
+static int FinalizeTokenProbas(VP8EncProba* const proba) {
+  int has_changed = 0;
+  int size = 0;
+  int t, b, c, p;
+  for (t = 0; t < NUM_TYPES; ++t) {
+    for (b = 0; b < NUM_BANDS; ++b) {
+      for (c = 0; c < NUM_CTX; ++c) {
+        for (p = 0; p < NUM_PROBAS; ++p) {
+          const proba_t stats = proba->stats_[t][b][c][p];
+          const int nb = (stats >> 0) & 0xffff;
+          const int total = (stats >> 16) & 0xffff;
+          const int update_proba = VP8CoeffsUpdateProba[t][b][c][p];
+          const int old_p = VP8CoeffsProba0[t][b][c][p];
+          const int new_p = CalcTokenProba(nb, total);
+          const int old_cost = BranchCost(nb, total, old_p)
+                             + VP8BitCost(0, update_proba);
+          const int new_cost = BranchCost(nb, total, new_p)
+                             + VP8BitCost(1, update_proba)
+                             + 8 * 256;
+          const int use_new_p = (old_cost > new_cost);
+          size += VP8BitCost(use_new_p, update_proba);
+          if (use_new_p) {  // only use proba that seem meaningful enough.
+            proba->coeffs_[t][b][c][p] = new_p;
+            has_changed |= (new_p != old_p);
+            size += 8 * 256;
+          } else {
+            proba->coeffs_[t][b][c][p] = old_p;
+          }
+        }
+      }
+    }
+  }
+  proba->dirty_ = has_changed;
+  return size;
+}
+
+//------------------------------------------------------------------------------
+// Finalize Segment probability based on the coding tree
+
+static int GetProba(int a, int b) {
+  const int total = a + b;
+  return (total == 0) ? 255     // that's the default probability.
+                      : (255 * a + total / 2) / total;  // rounded proba
+}
+
+static void ResetSegments(VP8Encoder* const enc) {
+  int n;
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    enc->mb_info_[n].segment_ = 0;
+  }
+}
+
+static void SetSegmentProbas(VP8Encoder* const enc) {
+  int p[NUM_MB_SEGMENTS] = { 0 };
+  int n;
+
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    const VP8MBInfo* const mb = &enc->mb_info_[n];
+    ++p[mb->segment_];
+  }
+#if !defined(WEBP_DISABLE_STATS)
+  if (enc->pic_->stats != NULL) {
+    for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
+      enc->pic_->stats->segment_size[n] = p[n];
+    }
+  }
+#endif
+  if (enc->segment_hdr_.num_segments_ > 1) {
+    uint8_t* const probas = enc->proba_.segments_;
+    probas[0] = GetProba(p[0] + p[1], p[2] + p[3]);
+    probas[1] = GetProba(p[0], p[1]);
+    probas[2] = GetProba(p[2], p[3]);
+
+    enc->segment_hdr_.update_map_ =
+        (probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255);
+    if (!enc->segment_hdr_.update_map_) ResetSegments(enc);
+    enc->segment_hdr_.size_ =
+        p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) +
+        p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) +
+        p[2] * (VP8BitCost(1, probas[0]) + VP8BitCost(0, probas[2])) +
+        p[3] * (VP8BitCost(1, probas[0]) + VP8BitCost(1, probas[2]));
+  } else {
+    enc->segment_hdr_.update_map_ = 0;
+    enc->segment_hdr_.size_ = 0;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Coefficient coding
+
+static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
+  int n = res->first;
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  const uint8_t* p = res->prob[n][ctx];
+  if (!VP8PutBit(bw, res->last >= 0, p[0])) {
+    return 0;
+  }
+
+  while (n < 16) {
+    const int c = res->coeffs[n++];
+    const int sign = c < 0;
+    int v = sign ? -c : c;
+    if (!VP8PutBit(bw, v != 0, p[1])) {
+      p = res->prob[VP8EncBands[n]][0];
+      continue;
+    }
+    if (!VP8PutBit(bw, v > 1, p[2])) {
+      p = res->prob[VP8EncBands[n]][1];
+    } else {
+      if (!VP8PutBit(bw, v > 4, p[3])) {
+        if (VP8PutBit(bw, v != 2, p[4])) {
+          VP8PutBit(bw, v == 4, p[5]);
+        }
+      } else if (!VP8PutBit(bw, v > 10, p[6])) {
+        if (!VP8PutBit(bw, v > 6, p[7])) {
+          VP8PutBit(bw, v == 6, 159);
+        } else {
+          VP8PutBit(bw, v >= 9, 165);
+          VP8PutBit(bw, !(v & 1), 145);
+        }
+      } else {
+        int mask;
+        const uint8_t* tab;
+        if (v < 3 + (8 << 1)) {          // VP8Cat3  (3b)
+          VP8PutBit(bw, 0, p[8]);
+          VP8PutBit(bw, 0, p[9]);
+          v -= 3 + (8 << 0);
+          mask = 1 << 2;
+          tab = VP8Cat3;
+        } else if (v < 3 + (8 << 2)) {   // VP8Cat4  (4b)
+          VP8PutBit(bw, 0, p[8]);
+          VP8PutBit(bw, 1, p[9]);
+          v -= 3 + (8 << 1);
+          mask = 1 << 3;
+          tab = VP8Cat4;
+        } else if (v < 3 + (8 << 3)) {   // VP8Cat5  (5b)
+          VP8PutBit(bw, 1, p[8]);
+          VP8PutBit(bw, 0, p[10]);
+          v -= 3 + (8 << 2);
+          mask = 1 << 4;
+          tab = VP8Cat5;
+        } else {                         // VP8Cat6 (11b)
+          VP8PutBit(bw, 1, p[8]);
+          VP8PutBit(bw, 1, p[10]);
+          v -= 3 + (8 << 3);
+          mask = 1 << 10;
+          tab = VP8Cat6;
+        }
+        while (mask) {
+          VP8PutBit(bw, !!(v & mask), *tab++);
+          mask >>= 1;
+        }
+      }
+      p = res->prob[VP8EncBands[n]][2];
+    }
+    VP8PutBitUniform(bw, sign);
+    if (n == 16 || !VP8PutBit(bw, n <= res->last, p[0])) {
+      return 1;   // EOB
+    }
+  }
+  return 1;
+}
+
+static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it,
+                          const VP8ModeScore* const rd) {
+  int x, y, ch;
+  VP8Residual res;
+  uint64_t pos1, pos2, pos3;
+  const int i16 = (it->mb_->type_ == 1);
+  const int segment = it->mb_->segment_;
+  VP8Encoder* const enc = it->enc_;
+
+  VP8IteratorNzToBytes(it);
+
+  pos1 = VP8BitWriterPos(bw);
+  if (i16) {
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+      PutCoeffs(bw, it->top_nz_[8] + it->left_nz_[8], &res);
+    VP8InitResidual(1, 0, enc, &res);
+  } else {
+    VP8InitResidual(0, 3, enc, &res);
+  }
+
+  // luma-AC
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] = PutCoeffs(bw, ctx, &res);
+    }
+  }
+  pos2 = VP8BitWriterPos(bw);
+
+  // U/V
+  VP8InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
+            PutCoeffs(bw, ctx, &res);
+      }
+    }
+  }
+  pos3 = VP8BitWriterPos(bw);
+  it->luma_bits_ = pos2 - pos1;
+  it->uv_bits_ = pos3 - pos2;
+  it->bit_count_[segment][i16] += it->luma_bits_;
+  it->bit_count_[segment][2] += it->uv_bits_;
+  VP8IteratorBytesToNz(it);
+}
+
+// Same as CodeResiduals, but doesn't actually write anything.
+// Instead, it just records the event distribution.
+static void RecordResiduals(VP8EncIterator* const it,
+                            const VP8ModeScore* const rd) {
+  int x, y, ch;
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+
+  VP8IteratorNzToBytes(it);
+
+  if (it->mb_->type_ == 1) {   // i16x16
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+      VP8RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res);
+    VP8InitResidual(1, 0, enc, &res);
+  } else {
+    VP8InitResidual(0, 3, enc, &res);
+  }
+
+  // luma-AC
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] = VP8RecordCoeffs(ctx, &res);
+    }
+  }
+
+  // U/V
+  VP8InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
+            VP8RecordCoeffs(ctx, &res);
+      }
+    }
+  }
+
+  VP8IteratorBytesToNz(it);
+}
+
+//------------------------------------------------------------------------------
+// Token buffer
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
+                        VP8TBuffer* const tokens) {
+  int x, y, ch;
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+
+  VP8IteratorNzToBytes(it);
+  if (it->mb_->type_ == 1) {   // i16x16
+    const int ctx = it->top_nz_[8] + it->left_nz_[8];
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+        VP8RecordCoeffTokens(ctx, &res, tokens);
+    VP8InitResidual(1, 0, enc, &res);
+  } else {
+    VP8InitResidual(0, 3, enc, &res);
+  }
+
+  // luma-AC
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] =
+          VP8RecordCoeffTokens(ctx, &res, tokens);
+    }
+  }
+
+  // U/V
+  VP8InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
+            VP8RecordCoeffTokens(ctx, &res, tokens);
+      }
+    }
+  }
+  VP8IteratorBytesToNz(it);
+  return !tokens->error_;
+}
+
+#endif    // !DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+// ExtraInfo map / Debug function
+
+#if !defined(WEBP_DISABLE_STATS)
+
+#if SEGMENT_VISU
+static void SetBlock(uint8_t* p, int value, int size) {
+  int y;
+  for (y = 0; y < size; ++y) {
+    memset(p, value, size);
+    p += BPS;
+  }
+}
+#endif
+
+static void ResetSSE(VP8Encoder* const enc) {
+  enc->sse_[0] = 0;
+  enc->sse_[1] = 0;
+  enc->sse_[2] = 0;
+  // Note: enc->sse_[3] is managed by alpha.c
+  enc->sse_count_ = 0;
+}
+
+static void StoreSSE(const VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  const uint8_t* const in = it->yuv_in_;
+  const uint8_t* const out = it->yuv_out_;
+  // Note: not totally accurate at boundary. And doesn't include in-loop filter.
+  enc->sse_[0] += VP8SSE16x16(in + Y_OFF_ENC, out + Y_OFF_ENC);
+  enc->sse_[1] += VP8SSE8x8(in + U_OFF_ENC, out + U_OFF_ENC);
+  enc->sse_[2] += VP8SSE8x8(in + V_OFF_ENC, out + V_OFF_ENC);
+  enc->sse_count_ += 16 * 16;
+}
+
+static void StoreSideInfo(const VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  const VP8MBInfo* const mb = it->mb_;
+  WebPPicture* const pic = enc->pic_;
+
+  if (pic->stats != NULL) {
+    StoreSSE(it);
+    enc->block_count_[0] += (mb->type_ == 0);
+    enc->block_count_[1] += (mb->type_ == 1);
+    enc->block_count_[2] += (mb->skip_ != 0);
+  }
+
+  if (pic->extra_info != NULL) {
+    uint8_t* const info = &pic->extra_info[it->x_ + it->y_ * enc->mb_w_];
+    switch (pic->extra_info_type) {
+      case 1: *info = mb->type_; break;
+      case 2: *info = mb->segment_; break;
+      case 3: *info = enc->dqm_[mb->segment_].quant_; break;
+      case 4: *info = (mb->type_ == 1) ? it->preds_[0] : 0xff; break;
+      case 5: *info = mb->uv_mode_; break;
+      case 6: {
+        const int b = (int)((it->luma_bits_ + it->uv_bits_ + 7) >> 3);
+        *info = (b > 255) ? 255 : b; break;
+      }
+      case 7: *info = mb->alpha_; break;
+      default: *info = 0; break;
+    }
+  }
+#if SEGMENT_VISU  // visualize segments and prediction modes
+  SetBlock(it->yuv_out_ + Y_OFF_ENC, mb->segment_ * 64, 16);
+  SetBlock(it->yuv_out_ + U_OFF_ENC, it->preds_[0] * 64, 8);
+  SetBlock(it->yuv_out_ + V_OFF_ENC, mb->uv_mode_ * 64, 8);
+#endif
+}
+
+static void ResetSideInfo(const VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  WebPPicture* const pic = enc->pic_;
+  if (pic->stats != NULL) {
+    memset(enc->block_count_, 0, sizeof(enc->block_count_));
+  }
+  ResetSSE(enc);
+}
+#else  // defined(WEBP_DISABLE_STATS)
+static void ResetSSE(VP8Encoder* const enc) {
+  (void)enc;
+}
+static void StoreSideInfo(const VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  WebPPicture* const pic = enc->pic_;
+  if (pic->extra_info != NULL) {
+    if (it->x_ == 0 && it->y_ == 0) {   // only do it once, at start
+      memset(pic->extra_info, 0,
+             enc->mb_w_ * enc->mb_h_ * sizeof(*pic->extra_info));
+    }
+  }
+}
+
+static void ResetSideInfo(const VP8EncIterator* const it) {
+  (void)it;
+}
+#endif  // !defined(WEBP_DISABLE_STATS)
+
+static double GetPSNR(uint64_t mse, uint64_t size) {
+  return (mse > 0 && size > 0) ? 10. * log10(255. * 255. * size / mse) : 99;
+}
+
+//------------------------------------------------------------------------------
+//  StatLoop(): only collect statistics (number of skips, token usage, ...).
+//  This is used for deciding optimal probabilities. It also modifies the
+//  quantizer value if some target (size, PSNR) was specified.
+
+static void SetLoopParams(VP8Encoder* const enc, float q) {
+  // Make sure the quality parameter is inside valid bounds
+  q = Clamp(q, 0.f, 100.f);
+
+  VP8SetSegmentParams(enc, q);      // setup segment quantizations and filters
+  SetSegmentProbas(enc);            // compute segment probabilities
+
+  ResetStats(enc);
+  ResetSSE(enc);
+}
+
+static uint64_t OneStatPass(VP8Encoder* const enc, VP8RDLevel rd_opt,
+                            int nb_mbs, int percent_delta,
+                            PassStats* const s) {
+  VP8EncIterator it;
+  uint64_t size = 0;
+  uint64_t size_p0 = 0;
+  uint64_t distortion = 0;
+  const uint64_t pixel_count = nb_mbs * 384;
+
+  VP8IteratorInit(enc, &it);
+  SetLoopParams(enc, s->q);
+  do {
+    VP8ModeScore info;
+    VP8IteratorImport(&it, NULL);
+    if (VP8Decimate(&it, &info, rd_opt)) {
+      // Just record the number of skips and act like skip_proba is not used.
+      ++enc->proba_.nb_skip_;
+    }
+    RecordResiduals(&it, &info);
+    size += info.R + info.H;
+    size_p0 += info.H;
+    distortion += info.D;
+    if (percent_delta && !VP8IteratorProgress(&it, percent_delta)) {
+      return 0;
+    }
+    VP8IteratorSaveBoundary(&it);
+  } while (VP8IteratorNext(&it) && --nb_mbs > 0);
+
+  size_p0 += enc->segment_hdr_.size_;
+  if (s->do_size_search) {
+    size += FinalizeSkipProba(enc);
+    size += FinalizeTokenProbas(&enc->proba_);
+    size = ((size + size_p0 + 1024) >> 11) + HEADER_SIZE_ESTIMATE;
+    s->value = (double)size;
+  } else {
+    s->value = GetPSNR(distortion, pixel_count);
+  }
+  return size_p0;
+}
+
+static int StatLoop(VP8Encoder* const enc) {
+  const int method = enc->method_;
+  const int do_search = enc->do_search_;
+  const int fast_probe = ((method == 0 || method == 3) && !do_search);
+  int num_pass_left = enc->config_->pass;
+  const int task_percent = 20;
+  const int percent_per_pass =
+      (task_percent + num_pass_left / 2) / num_pass_left;
+  const int final_percent = enc->percent_ + task_percent;
+  const VP8RDLevel rd_opt =
+      (method >= 3 || do_search) ? RD_OPT_BASIC : RD_OPT_NONE;
+  int nb_mbs = enc->mb_w_ * enc->mb_h_;
+  PassStats stats;
+
+  InitPassStats(enc, &stats);
+  ResetTokenStats(enc);
+
+  // Fast mode: quick analysis pass over few mbs. Better than nothing.
+  if (fast_probe) {
+    if (method == 3) {  // we need more stats for method 3 to be reliable.
+      nb_mbs = (nb_mbs > 200) ? nb_mbs >> 1 : 100;
+    } else {
+      nb_mbs = (nb_mbs > 200) ? nb_mbs >> 2 : 50;
+    }
+  }
+
+  while (num_pass_left-- > 0) {
+    const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
+                             (num_pass_left == 0) ||
+                             (enc->max_i4_header_bits_ == 0);
+    const uint64_t size_p0 =
+        OneStatPass(enc, rd_opt, nb_mbs, percent_per_pass, &stats);
+    if (size_p0 == 0) return 0;
+#if (DEBUG_SEARCH > 0)
+    printf("#%d value:%.1lf -> %.1lf   q:%.2f -> %.2f\n",
+           num_pass_left, stats.last_value, stats.value, stats.last_q, stats.q);
+#endif
+    if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
+      ++num_pass_left;
+      enc->max_i4_header_bits_ >>= 1;  // strengthen header bit limitation...
+      continue;                        // ...and start over
+    }
+    if (is_last_pass) {
+      break;
+    }
+    // If no target size: just do several pass without changing 'q'
+    if (do_search) {
+      ComputeNextQ(&stats);
+      if (fabs(stats.dq) <= DQ_LIMIT) break;
+    }
+  }
+  if (!do_search || !stats.do_size_search) {
+    // Need to finalize probas now, since it wasn't done during the search.
+    FinalizeSkipProba(enc);
+    FinalizeTokenProbas(&enc->proba_);
+  }
+  VP8CalculateLevelCosts(&enc->proba_);  // finalize costs
+  return WebPReportProgress(enc->pic_, final_percent, &enc->percent_);
+}
+
+//------------------------------------------------------------------------------
+// Main loops
+//
+
+static const uint8_t kAverageBytesPerMB[8] = { 50, 24, 16, 9, 7, 5, 3, 2 };
+
+static int PreLoopInitialize(VP8Encoder* const enc) {
+  int p;
+  int ok = 1;
+  const int average_bytes_per_MB = kAverageBytesPerMB[enc->base_quant_ >> 4];
+  const int bytes_per_parts =
+      enc->mb_w_ * enc->mb_h_ * average_bytes_per_MB / enc->num_parts_;
+  // Initialize the bit-writers
+  for (p = 0; ok && p < enc->num_parts_; ++p) {
+    ok = VP8BitWriterInit(enc->parts_ + p, bytes_per_parts);
+  }
+  if (!ok) {
+    VP8EncFreeBitWriters(enc);  // malloc error occurred
+    WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  return ok;
+}
+
+static int PostLoopFinalize(VP8EncIterator* const it, int ok) {
+  VP8Encoder* const enc = it->enc_;
+  if (ok) {      // Finalize the partitions, check for extra errors.
+    int p;
+    for (p = 0; p < enc->num_parts_; ++p) {
+      VP8BitWriterFinish(enc->parts_ + p);
+      ok &= !enc->parts_[p].error_;
+    }
+  }
+
+  if (ok) {      // All good. Finish up.
+#if !defined(WEBP_DISABLE_STATS)
+    if (enc->pic_->stats != NULL) {  // finalize byte counters...
+      int i, s;
+      for (i = 0; i <= 2; ++i) {
+        for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+          enc->residual_bytes_[i][s] = (int)((it->bit_count_[s][i] + 7) >> 3);
+        }
+      }
+    }
+#endif
+    VP8AdjustFilterStrength(it);     // ...and store filter stats.
+  } else {
+    // Something bad happened -> need to do some memory cleanup.
+    VP8EncFreeBitWriters(enc);
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+//  VP8EncLoop(): does the final bitstream coding.
+
+static void ResetAfterSkip(VP8EncIterator* const it) {
+  if (it->mb_->type_ == 1) {
+    *it->nz_ = 0;  // reset all predictors
+    it->left_nz_[8] = 0;
+  } else {
+    *it->nz_ &= (1 << 24);  // preserve the dc_nz bit
+  }
+}
+
+int VP8EncLoop(VP8Encoder* const enc) {
+  VP8EncIterator it;
+  int ok = PreLoopInitialize(enc);
+  if (!ok) return 0;
+
+  StatLoop(enc);  // stats-collection loop
+
+  VP8IteratorInit(enc, &it);
+  VP8InitFilter(&it);
+  do {
+    VP8ModeScore info;
+    const int dont_use_skip = !enc->proba_.use_skip_proba_;
+    const VP8RDLevel rd_opt = enc->rd_opt_level_;
+
+    VP8IteratorImport(&it, NULL);
+    // Warning! order is important: first call VP8Decimate() and
+    // *then* decide how to code the skip decision if there's one.
+    if (!VP8Decimate(&it, &info, rd_opt) || dont_use_skip) {
+      CodeResiduals(it.bw_, &it, &info);
+    } else {   // reset predictors after a skip
+      ResetAfterSkip(&it);
+    }
+    StoreSideInfo(&it);
+    VP8StoreFilterStats(&it);
+    VP8IteratorExport(&it);
+    ok = VP8IteratorProgress(&it, 20);
+    VP8IteratorSaveBoundary(&it);
+  } while (ok && VP8IteratorNext(&it));
+
+  return PostLoopFinalize(&it, ok);
+}
+
+//------------------------------------------------------------------------------
+// Single pass using Token Buffer.
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+#define MIN_COUNT 96  // minimum number of macroblocks before updating stats
+
+int VP8EncTokenLoop(VP8Encoder* const enc) {
+  // Roughly refresh the proba eight times per pass
+  int max_count = (enc->mb_w_ * enc->mb_h_) >> 3;
+  int num_pass_left = enc->config_->pass;
+  int remaining_progress = 40;  // percents
+  const int do_search = enc->do_search_;
+  VP8EncIterator it;
+  VP8EncProba* const proba = &enc->proba_;
+  const VP8RDLevel rd_opt = enc->rd_opt_level_;
+  const uint64_t pixel_count = enc->mb_w_ * enc->mb_h_ * 384;
+  PassStats stats;
+  int ok;
+
+  InitPassStats(enc, &stats);
+  ok = PreLoopInitialize(enc);
+  if (!ok) return 0;
+
+  if (max_count < MIN_COUNT) max_count = MIN_COUNT;
+
+  assert(enc->num_parts_ == 1);
+  assert(enc->use_tokens_);
+  assert(proba->use_skip_proba_ == 0);
+  assert(rd_opt >= RD_OPT_BASIC);   // otherwise, token-buffer won't be useful
+  assert(num_pass_left > 0);
+
+  while (ok && num_pass_left-- > 0) {
+    const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
+                             (num_pass_left == 0) ||
+                             (enc->max_i4_header_bits_ == 0);
+    uint64_t size_p0 = 0;
+    uint64_t distortion = 0;
+    int cnt = max_count;
+    // The final number of passes is not trivial to know in advance.
+    const int pass_progress = remaining_progress / (2 + num_pass_left);
+    remaining_progress -= pass_progress;
+    VP8IteratorInit(enc, &it);
+    SetLoopParams(enc, stats.q);
+    if (is_last_pass) {
+      ResetTokenStats(enc);
+      VP8InitFilter(&it);  // don't collect stats until last pass (too costly)
+    }
+    VP8TBufferClear(&enc->tokens_);
+    do {
+      VP8ModeScore info;
+      VP8IteratorImport(&it, NULL);
+      if (--cnt < 0) {
+        FinalizeTokenProbas(proba);
+        VP8CalculateLevelCosts(proba);  // refresh cost tables for rd-opt
+        cnt = max_count;
+      }
+      VP8Decimate(&it, &info, rd_opt);
+      ok = RecordTokens(&it, &info, &enc->tokens_);
+      if (!ok) {
+        WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+        break;
+      }
+      size_p0 += info.H;
+      distortion += info.D;
+      if (is_last_pass) {
+        StoreSideInfo(&it);
+        VP8StoreFilterStats(&it);
+        VP8IteratorExport(&it);
+        ok = VP8IteratorProgress(&it, pass_progress);
+      }
+      VP8IteratorSaveBoundary(&it);
+    } while (ok && VP8IteratorNext(&it));
+    if (!ok) break;
+
+    size_p0 += enc->segment_hdr_.size_;
+    if (stats.do_size_search) {
+      uint64_t size = FinalizeTokenProbas(&enc->proba_);
+      size += VP8EstimateTokenSize(&enc->tokens_,
+                                   (const uint8_t*)proba->coeffs_);
+      size = (size + size_p0 + 1024) >> 11;  // -> size in bytes
+      size += HEADER_SIZE_ESTIMATE;
+      stats.value = (double)size;
+    } else {  // compute and store PSNR
+      stats.value = GetPSNR(distortion, pixel_count);
+    }
+
+#if (DEBUG_SEARCH > 0)
+    printf("#%2d metric:%.1lf -> %.1lf   last_q=%.2lf q=%.2lf dq=%.2lf "
+           " range:[%.1f, %.1f]\n",
+           num_pass_left, stats.last_value, stats.value,
+           stats.last_q, stats.q, stats.dq, stats.qmin, stats.qmax);
+#endif
+    if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
+      ++num_pass_left;
+      enc->max_i4_header_bits_ >>= 1;  // strengthen header bit limitation...
+      if (is_last_pass) {
+        ResetSideInfo(&it);
+      }
+      continue;                        // ...and start over
+    }
+    if (is_last_pass) {
+      break;   // done
+    }
+    if (do_search) {
+      ComputeNextQ(&stats);  // Adjust q
+    }
+  }
+  if (ok) {
+    if (!stats.do_size_search) {
+      FinalizeTokenProbas(&enc->proba_);
+    }
+    ok = VP8EmitTokens(&enc->tokens_, enc->parts_ + 0,
+                       (const uint8_t*)proba->coeffs_, 1);
+  }
+  ok = ok && WebPReportProgress(enc->pic_, enc->percent_ + remaining_progress,
+                                &enc->percent_);
+  return PostLoopFinalize(&it, ok);
+}
+
+#else
+
+int VP8EncTokenLoop(VP8Encoder* const enc) {
+  (void)enc;
+  return 0;   // we shouldn't be here.
+}
+
+#endif    // DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/histogram_enc.c b/media/libwebp/src/enc/histogram_enc.c
new file mode 100644
index 0000000000..8418def2e1
--- /dev/null
+++ b/media/libwebp/src/enc/histogram_enc.c
@@ -0,0 +1,1248 @@
+// 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)
+//
+#ifdef HAVE_CONFIG_H
+#include "src/webp/config.h"
+#endif
+
+#include <float.h>
+#include <math.h>
+
+#include "src/dsp/lossless.h"
+#include "src/dsp/lossless_common.h"
+#include "src/enc/backward_references_enc.h"
+#include "src/enc/histogram_enc.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/utils/utils.h"
+
+#define MAX_BIT_COST FLT_MAX
+
+// Number of partitions for the three dominant (literal, red and blue) symbol
+// costs.
+#define NUM_PARTITIONS 4
+// The size of the bin-hash corresponding to the three dominant costs.
+#define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS)
+// Maximum number of histograms allowed in greedy combining algorithm.
+#define MAX_HISTO_GREEDY 100
+
+static void HistogramClear(VP8LHistogram* const p) {
+  uint32_t* const literal = p->literal_;
+  const int cache_bits = p->palette_code_bits_;
+  const int histo_size = VP8LGetHistogramSize(cache_bits);
+  memset(p, 0, histo_size);
+  p->palette_code_bits_ = cache_bits;
+  p->literal_ = literal;
+}
+
+// Swap two histogram pointers.
+static void HistogramSwap(VP8LHistogram** const A, VP8LHistogram** const B) {
+  VP8LHistogram* const tmp = *A;
+  *A = *B;
+  *B = tmp;
+}
+
+static void HistogramCopy(const VP8LHistogram* const src,
+                          VP8LHistogram* const dst) {
+  uint32_t* const dst_literal = dst->literal_;
+  const int dst_cache_bits = dst->palette_code_bits_;
+  const int literal_size = VP8LHistogramNumCodes(dst_cache_bits);
+  const int histo_size = VP8LGetHistogramSize(dst_cache_bits);
+  assert(src->palette_code_bits_ == dst_cache_bits);
+  memcpy(dst, src, histo_size);
+  dst->literal_ = dst_literal;
+  memcpy(dst->literal_, src->literal_, literal_size * sizeof(*dst->literal_));
+}
+
+int VP8LGetHistogramSize(int cache_bits) {
+  const int literal_size = VP8LHistogramNumCodes(cache_bits);
+  const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size;
+  assert(total_size <= (size_t)0x7fffffff);
+  return (int)total_size;
+}
+
+void VP8LFreeHistogram(VP8LHistogram* const histo) {
+  WebPSafeFree(histo);
+}
+
+void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) {
+  WebPSafeFree(histo);
+}
+
+void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
+                            VP8LHistogram* const histo) {
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  while (VP8LRefsCursorOk(&c)) {
+    VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos, NULL, 0);
+    VP8LRefsCursorNext(&c);
+  }
+}
+
+void VP8LHistogramCreate(VP8LHistogram* const p,
+                         const VP8LBackwardRefs* const refs,
+                         int palette_code_bits) {
+  if (palette_code_bits >= 0) {
+    p->palette_code_bits_ = palette_code_bits;
+  }
+  HistogramClear(p);
+  VP8LHistogramStoreRefs(refs, p);
+}
+
+void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits,
+                       int init_arrays) {
+  p->palette_code_bits_ = palette_code_bits;
+  if (init_arrays) {
+    HistogramClear(p);
+  } else {
+    p->trivial_symbol_ = 0;
+    p->bit_cost_ = 0.;
+    p->literal_cost_ = 0.;
+    p->red_cost_ = 0.;
+    p->blue_cost_ = 0.;
+    memset(p->is_used_, 0, sizeof(p->is_used_));
+  }
+}
+
+VP8LHistogram* VP8LAllocateHistogram(int cache_bits) {
+  VP8LHistogram* histo = NULL;
+  const int total_size = VP8LGetHistogramSize(cache_bits);
+  uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
+  if (memory == NULL) return NULL;
+  histo = (VP8LHistogram*)memory;
+  // literal_ won't necessary be aligned.
+  histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
+  VP8LHistogramInit(histo, cache_bits, /*init_arrays=*/ 0);
+  return histo;
+}
+
+// Resets the pointers of the histograms to point to the bit buffer in the set.
+static void HistogramSetResetPointers(VP8LHistogramSet* const set,
+                                      int cache_bits) {
+  int i;
+  const int histo_size = VP8LGetHistogramSize(cache_bits);
+  uint8_t* memory = (uint8_t*) (set->histograms);
+  memory += set->max_size * sizeof(*set->histograms);
+  for (i = 0; i < set->max_size; ++i) {
+    memory = (uint8_t*) WEBP_ALIGN(memory);
+    set->histograms[i] = (VP8LHistogram*) memory;
+    // literal_ won't necessary be aligned.
+    set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
+    memory += histo_size;
+  }
+}
+
+// Returns the total size of the VP8LHistogramSet.
+static size_t HistogramSetTotalSize(int size, int cache_bits) {
+  const int histo_size = VP8LGetHistogramSize(cache_bits);
+  return (sizeof(VP8LHistogramSet) + size * (sizeof(VP8LHistogram*) +
+          histo_size + WEBP_ALIGN_CST));
+}
+
+VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
+  int i;
+  VP8LHistogramSet* set;
+  const size_t total_size = HistogramSetTotalSize(size, cache_bits);
+  uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
+  if (memory == NULL) return NULL;
+
+  set = (VP8LHistogramSet*)memory;
+  memory += sizeof(*set);
+  set->histograms = (VP8LHistogram**)memory;
+  set->max_size = size;
+  set->size = size;
+  HistogramSetResetPointers(set, cache_bits);
+  for (i = 0; i < size; ++i) {
+    VP8LHistogramInit(set->histograms[i], cache_bits, /*init_arrays=*/ 0);
+  }
+  return set;
+}
+
+void VP8LHistogramSetClear(VP8LHistogramSet* const set) {
+  int i;
+  const int cache_bits = set->histograms[0]->palette_code_bits_;
+  const int size = set->max_size;
+  const size_t total_size = HistogramSetTotalSize(size, cache_bits);
+  uint8_t* memory = (uint8_t*)set;
+
+  memset(memory, 0, total_size);
+  memory += sizeof(*set);
+  set->histograms = (VP8LHistogram**)memory;
+  set->max_size = size;
+  set->size = size;
+  HistogramSetResetPointers(set, cache_bits);
+  for (i = 0; i < size; ++i) {
+    set->histograms[i]->palette_code_bits_ = cache_bits;
+  }
+}
+
+// Removes the histogram 'i' from 'set' by setting it to NULL.
+static void HistogramSetRemoveHistogram(VP8LHistogramSet* const set, int i,
+                                        int* const num_used) {
+  assert(set->histograms[i] != NULL);
+  set->histograms[i] = NULL;
+  --*num_used;
+  // If we remove the last valid one, shrink until the next valid one.
+  if (i == set->size - 1) {
+    while (set->size >= 1 && set->histograms[set->size - 1] == NULL) {
+      --set->size;
+    }
+  }
+}
+
+// -----------------------------------------------------------------------------
+
+void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
+                                     const PixOrCopy* const v,
+                                     int (*const distance_modifier)(int, int),
+                                     int distance_modifier_arg0) {
+  if (PixOrCopyIsLiteral(v)) {
+    ++histo->alpha_[PixOrCopyLiteral(v, 3)];
+    ++histo->red_[PixOrCopyLiteral(v, 2)];
+    ++histo->literal_[PixOrCopyLiteral(v, 1)];
+    ++histo->blue_[PixOrCopyLiteral(v, 0)];
+  } else if (PixOrCopyIsCacheIdx(v)) {
+    const int literal_ix =
+        NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
+    assert(histo->palette_code_bits_ != 0);
+    ++histo->literal_[literal_ix];
+  } else {
+    int code, extra_bits;
+    VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
+    ++histo->literal_[NUM_LITERAL_CODES + code];
+    if (distance_modifier == NULL) {
+      VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
+    } else {
+      VP8LPrefixEncodeBits(
+          distance_modifier(distance_modifier_arg0, PixOrCopyDistance(v)),
+          &code, &extra_bits);
+    }
+    ++histo->distance_[code];
+  }
+}
+
+// -----------------------------------------------------------------------------
+// Entropy-related functions.
+
+static WEBP_INLINE float BitsEntropyRefine(const VP8LBitEntropy* entropy) {
+  float mix;
+  if (entropy->nonzeros < 5) {
+    if (entropy->nonzeros <= 1) {
+      return 0;
+    }
+    // Two symbols, they will be 0 and 1 in a Huffman code.
+    // Let's mix in a bit of entropy to favor good clustering when
+    // distributions of these are combined.
+    if (entropy->nonzeros == 2) {
+      return 0.99f * entropy->sum + 0.01f * entropy->entropy;
+    }
+    // No matter what the entropy says, we cannot be better than min_limit
+    // with Huffman coding. I am mixing a bit of entropy into the
+    // min_limit since it produces much better (~0.5 %) compression results
+    // perhaps because of better entropy clustering.
+    if (entropy->nonzeros == 3) {
+      mix = 0.95f;
+    } else {
+      mix = 0.7f;  // nonzeros == 4.
+    }
+  } else {
+    mix = 0.627f;
+  }
+
+  {
+    float min_limit = 2.f * entropy->sum - entropy->max_val;
+    min_limit = mix * min_limit + (1.f - mix) * entropy->entropy;
+    return (entropy->entropy < min_limit) ? min_limit : entropy->entropy;
+  }
+}
+
+float VP8LBitsEntropy(const uint32_t* const array, int n) {
+  VP8LBitEntropy entropy;
+  VP8LBitsEntropyUnrefined(array, n, &entropy);
+
+  return BitsEntropyRefine(&entropy);
+}
+
+static float InitialHuffmanCost(void) {
+  // Small bias because Huffman code length is typically not stored in
+  // full length.
+  static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3;
+  static const float kSmallBias = 9.1f;
+  return kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
+}
+
+// Finalize the Huffman cost based on streak numbers and length type (<3 or >=3)
+static float FinalHuffmanCost(const VP8LStreaks* const stats) {
+  // The constants in this function are experimental and got rounded from
+  // their original values in 1/8 when switched to 1/1024.
+  float retval = InitialHuffmanCost();
+  // Second coefficient: Many zeros in the histogram are covered efficiently
+  // by a run-length encode. Originally 2/8.
+  retval += stats->counts[0] * 1.5625f + 0.234375f * stats->streaks[0][1];
+  // Second coefficient: Constant values are encoded less efficiently, but still
+  // RLE'ed. Originally 6/8.
+  retval += stats->counts[1] * 2.578125f + 0.703125f * stats->streaks[1][1];
+  // 0s are usually encoded more efficiently than non-0s.
+  // Originally 15/8.
+  retval += 1.796875f * stats->streaks[0][0];
+  // Originally 26/8.
+  retval += 3.28125f * stats->streaks[1][0];
+  return retval;
+}
+
+// Get the symbol entropy for the distribution 'population'.
+// Set 'trivial_sym', if there's only one symbol present in the distribution.
+static float PopulationCost(const uint32_t* const population, int length,
+                            uint32_t* const trivial_sym,
+                            uint8_t* const is_used) {
+  VP8LBitEntropy bit_entropy;
+  VP8LStreaks stats;
+  VP8LGetEntropyUnrefined(population, length, &bit_entropy, &stats);
+  if (trivial_sym != NULL) {
+    *trivial_sym = (bit_entropy.nonzeros == 1) ? bit_entropy.nonzero_code
+                                               : VP8L_NON_TRIVIAL_SYM;
+  }
+  // The histogram is used if there is at least one non-zero streak.
+  *is_used = (stats.streaks[1][0] != 0 || stats.streaks[1][1] != 0);
+
+  return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
+}
+
+// trivial_at_end is 1 if the two histograms only have one element that is
+// non-zero: both the zero-th one, or both the last one.
+static WEBP_INLINE float GetCombinedEntropy(const uint32_t* const X,
+                                            const uint32_t* const Y, int length,
+                                            int is_X_used, int is_Y_used,
+                                            int trivial_at_end) {
+  VP8LStreaks stats;
+  if (trivial_at_end) {
+    // This configuration is due to palettization that transforms an indexed
+    // pixel into 0xff000000 | (pixel << 8) in VP8LBundleColorMap.
+    // BitsEntropyRefine is 0 for histograms with only one non-zero value.
+    // Only FinalHuffmanCost needs to be evaluated.
+    memset(&stats, 0, sizeof(stats));
+    // Deal with the non-zero value at index 0 or length-1.
+    stats.streaks[1][0] = 1;
+    // Deal with the following/previous zero streak.
+    stats.counts[0] = 1;
+    stats.streaks[0][1] = length - 1;
+    return FinalHuffmanCost(&stats);
+  } else {
+    VP8LBitEntropy bit_entropy;
+    if (is_X_used) {
+      if (is_Y_used) {
+        VP8LGetCombinedEntropyUnrefined(X, Y, length, &bit_entropy, &stats);
+      } else {
+        VP8LGetEntropyUnrefined(X, length, &bit_entropy, &stats);
+      }
+    } else {
+      if (is_Y_used) {
+        VP8LGetEntropyUnrefined(Y, length, &bit_entropy, &stats);
+      } else {
+        memset(&stats, 0, sizeof(stats));
+        stats.counts[0] = 1;
+        stats.streaks[0][length > 3] = length;
+        VP8LBitEntropyInit(&bit_entropy);
+      }
+    }
+
+    return BitsEntropyRefine(&bit_entropy) + FinalHuffmanCost(&stats);
+  }
+}
+
+// Estimates the Entropy + Huffman + other block overhead size cost.
+float VP8LHistogramEstimateBits(VP8LHistogram* const p) {
+  return
+      PopulationCost(p->literal_, VP8LHistogramNumCodes(p->palette_code_bits_),
+                     NULL, &p->is_used_[0])
+      + PopulationCost(p->red_, NUM_LITERAL_CODES, NULL, &p->is_used_[1])
+      + PopulationCost(p->blue_, NUM_LITERAL_CODES, NULL, &p->is_used_[2])
+      + PopulationCost(p->alpha_, NUM_LITERAL_CODES, NULL, &p->is_used_[3])
+      + PopulationCost(p->distance_, NUM_DISTANCE_CODES, NULL, &p->is_used_[4])
+      + VP8LExtraCost(p->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES)
+      + VP8LExtraCost(p->distance_, NUM_DISTANCE_CODES);
+}
+
+// -----------------------------------------------------------------------------
+// Various histogram combine/cost-eval functions
+
+static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
+                                       const VP8LHistogram* const b,
+                                       float cost_threshold, float* cost) {
+  const int palette_code_bits = a->palette_code_bits_;
+  int trivial_at_end = 0;
+  assert(a->palette_code_bits_ == b->palette_code_bits_);
+  *cost += GetCombinedEntropy(a->literal_, b->literal_,
+                              VP8LHistogramNumCodes(palette_code_bits),
+                              a->is_used_[0], b->is_used_[0], 0);
+  *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
+                                 b->literal_ + NUM_LITERAL_CODES,
+                                 NUM_LENGTH_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  if (a->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM &&
+      a->trivial_symbol_ == b->trivial_symbol_) {
+    // A, R and B are all 0 or 0xff.
+    const uint32_t color_a = (a->trivial_symbol_ >> 24) & 0xff;
+    const uint32_t color_r = (a->trivial_symbol_ >> 16) & 0xff;
+    const uint32_t color_b = (a->trivial_symbol_ >> 0) & 0xff;
+    if ((color_a == 0 || color_a == 0xff) &&
+        (color_r == 0 || color_r == 0xff) &&
+        (color_b == 0 || color_b == 0xff)) {
+      trivial_at_end = 1;
+    }
+  }
+
+  *cost +=
+      GetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES, a->is_used_[1],
+                         b->is_used_[1], trivial_at_end);
+  if (*cost > cost_threshold) return 0;
+
+  *cost +=
+      GetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES, a->is_used_[2],
+                         b->is_used_[2], trivial_at_end);
+  if (*cost > cost_threshold) return 0;
+
+  *cost +=
+      GetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES,
+                         a->is_used_[3], b->is_used_[3], trivial_at_end);
+  if (*cost > cost_threshold) return 0;
+
+  *cost +=
+      GetCombinedEntropy(a->distance_, b->distance_, NUM_DISTANCE_CODES,
+                         a->is_used_[4], b->is_used_[4], 0);
+  *cost +=
+      VP8LExtraCostCombined(a->distance_, b->distance_, NUM_DISTANCE_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  return 1;
+}
+
+static WEBP_INLINE void HistogramAdd(const VP8LHistogram* const a,
+                                     const VP8LHistogram* const b,
+                                     VP8LHistogram* const out) {
+  VP8LHistogramAdd(a, b, out);
+  out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_)
+                       ? a->trivial_symbol_
+                       : VP8L_NON_TRIVIAL_SYM;
+}
+
+// Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing
+// to the threshold value 'cost_threshold'. The score returned is
+//  Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed.
+// Since the previous score passed is 'cost_threshold', we only need to compare
+// the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out
+// early.
+static float HistogramAddEval(const VP8LHistogram* const a,
+                              const VP8LHistogram* const b,
+                              VP8LHistogram* const out, float cost_threshold) {
+  float cost = 0;
+  const float sum_cost = a->bit_cost_ + b->bit_cost_;
+  cost_threshold += sum_cost;
+
+  if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
+    HistogramAdd(a, b, out);
+    out->bit_cost_ = cost;
+    out->palette_code_bits_ = a->palette_code_bits_;
+  }
+
+  return cost - sum_cost;
+}
+
+// Same as HistogramAddEval(), except that the resulting histogram
+// is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit
+// the term C(b) which is constant over all the evaluations.
+static float HistogramAddThresh(const VP8LHistogram* const a,
+                                const VP8LHistogram* const b,
+                                float cost_threshold) {
+  float cost;
+  assert(a != NULL && b != NULL);
+  cost = -a->bit_cost_;
+  GetCombinedHistogramEntropy(a, b, cost_threshold, &cost);
+  return cost;
+}
+
+// -----------------------------------------------------------------------------
+
+// The structure to keep track of cost range for the three dominant entropy
+// symbols.
+typedef struct {
+  float literal_max_;
+  float literal_min_;
+  float red_max_;
+  float red_min_;
+  float blue_max_;
+  float blue_min_;
+} DominantCostRange;
+
+static void DominantCostRangeInit(DominantCostRange* const c) {
+  c->literal_max_ = 0.;
+  c->literal_min_ = MAX_BIT_COST;
+  c->red_max_ = 0.;
+  c->red_min_ = MAX_BIT_COST;
+  c->blue_max_ = 0.;
+  c->blue_min_ = MAX_BIT_COST;
+}
+
+static void UpdateDominantCostRange(
+    const VP8LHistogram* const h, DominantCostRange* const c) {
+  if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_;
+  if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_;
+  if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_;
+  if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_;
+  if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_;
+  if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_;
+}
+
+static void UpdateHistogramCost(VP8LHistogram* const h) {
+  uint32_t alpha_sym, red_sym, blue_sym;
+  const float alpha_cost =
+      PopulationCost(h->alpha_, NUM_LITERAL_CODES, &alpha_sym, &h->is_used_[3]);
+  const float distance_cost =
+      PopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL, &h->is_used_[4]) +
+      VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
+  const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
+  h->literal_cost_ =
+      PopulationCost(h->literal_, num_codes, NULL, &h->is_used_[0]) +
+          VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES, NUM_LENGTH_CODES);
+  h->red_cost_ =
+      PopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym, &h->is_used_[1]);
+  h->blue_cost_ =
+      PopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym, &h->is_used_[2]);
+  h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ +
+                 alpha_cost + distance_cost;
+  if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) {
+    h->trivial_symbol_ = VP8L_NON_TRIVIAL_SYM;
+  } else {
+    h->trivial_symbol_ =
+        ((uint32_t)alpha_sym << 24) | (red_sym << 16) | (blue_sym << 0);
+  }
+}
+
+static int GetBinIdForEntropy(float min, float max, float val) {
+  const float range = max - min;
+  if (range > 0.) {
+    const float delta = val - min;
+    return (int)((NUM_PARTITIONS - 1e-6) * delta / range);
+  } else {
+    return 0;
+  }
+}
+
+static int GetHistoBinIndex(const VP8LHistogram* const h,
+                            const DominantCostRange* const c, int low_effort) {
+  int bin_id = GetBinIdForEntropy(c->literal_min_, c->literal_max_,
+                                  h->literal_cost_);
+  assert(bin_id < NUM_PARTITIONS);
+  if (!low_effort) {
+    bin_id = bin_id * NUM_PARTITIONS
+           + GetBinIdForEntropy(c->red_min_, c->red_max_, h->red_cost_);
+    bin_id = bin_id * NUM_PARTITIONS
+           + GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_);
+    assert(bin_id < BIN_SIZE);
+  }
+  return bin_id;
+}
+
+// Construct the histograms from backward references.
+static void HistogramBuild(
+    int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs,
+    VP8LHistogramSet* const image_histo) {
+  int x = 0, y = 0;
+  const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
+  VP8LHistogram** const histograms = image_histo->histograms;
+  VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs);
+  assert(histo_bits > 0);
+  VP8LHistogramSetClear(image_histo);
+  while (VP8LRefsCursorOk(&c)) {
+    const PixOrCopy* const v = c.cur_pos;
+    const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
+    VP8LHistogramAddSinglePixOrCopy(histograms[ix], v, NULL, 0);
+    x += PixOrCopyLength(v);
+    while (x >= xsize) {
+      x -= xsize;
+      ++y;
+    }
+    VP8LRefsCursorNext(&c);
+  }
+}
+
+// Copies the histograms and computes its bit_cost.
+static const uint16_t kInvalidHistogramSymbol = (uint16_t)(-1);
+static void HistogramCopyAndAnalyze(VP8LHistogramSet* const orig_histo,
+                                    VP8LHistogramSet* const image_histo,
+                                    int* const num_used,
+                                    uint16_t* const histogram_symbols) {
+  int i, cluster_id;
+  int num_used_orig = *num_used;
+  VP8LHistogram** const orig_histograms = orig_histo->histograms;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  assert(image_histo->max_size == orig_histo->max_size);
+  for (cluster_id = 0, i = 0; i < orig_histo->max_size; ++i) {
+    VP8LHistogram* const histo = orig_histograms[i];
+    UpdateHistogramCost(histo);
+
+    // Skip the histogram if it is completely empty, which can happen for tiles
+    // with no information (when they are skipped because of LZ77).
+    if (!histo->is_used_[0] && !histo->is_used_[1] && !histo->is_used_[2]
+        && !histo->is_used_[3] && !histo->is_used_[4]) {
+      // The first histogram is always used. If an histogram is empty, we set
+      // its id to be the same as the previous one: this will improve
+      // compressibility for later LZ77.
+      assert(i > 0);
+      HistogramSetRemoveHistogram(image_histo, i, num_used);
+      HistogramSetRemoveHistogram(orig_histo, i, &num_used_orig);
+      histogram_symbols[i] = kInvalidHistogramSymbol;
+    } else {
+      // Copy histograms from orig_histo[] to image_histo[].
+      HistogramCopy(histo, histograms[i]);
+      histogram_symbols[i] = cluster_id++;
+      assert(cluster_id <= image_histo->max_size);
+    }
+  }
+}
+
+// Partition histograms to different entropy bins for three dominant (literal,
+// red and blue) symbol costs and compute the histogram aggregate bit_cost.
+static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo,
+                                       uint16_t* const bin_map,
+                                       int low_effort) {
+  int i;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  const int histo_size = image_histo->size;
+  DominantCostRange cost_range;
+  DominantCostRangeInit(&cost_range);
+
+  // Analyze the dominant (literal, red and blue) entropy costs.
+  for (i = 0; i < histo_size; ++i) {
+    if (histograms[i] == NULL) continue;
+    UpdateDominantCostRange(histograms[i], &cost_range);
+  }
+
+  // bin-hash histograms on three of the dominant (literal, red and blue)
+  // symbol costs and store the resulting bin_id for each histogram.
+  for (i = 0; i < histo_size; ++i) {
+    // bin_map[i] is not set to a special value as its use will later be guarded
+    // by another (histograms[i] == NULL).
+    if (histograms[i] == NULL) continue;
+    bin_map[i] = GetHistoBinIndex(histograms[i], &cost_range, low_effort);
+  }
+}
+
+// Merges some histograms with same bin_id together if it's advantageous.
+// Sets the remaining histograms to NULL.
+static void HistogramCombineEntropyBin(
+    VP8LHistogramSet* const image_histo, int* num_used,
+    const uint16_t* const clusters, uint16_t* const cluster_mappings,
+    VP8LHistogram* cur_combo, const uint16_t* const bin_map, int num_bins,
+    float combine_cost_factor, int low_effort) {
+  VP8LHistogram** const histograms = image_histo->histograms;
+  int idx;
+  struct {
+    int16_t first;    // position of the histogram that accumulates all
+                      // histograms with the same bin_id
+    uint16_t num_combine_failures;   // number of combine failures per bin_id
+  } bin_info[BIN_SIZE];
+
+  assert(num_bins <= BIN_SIZE);
+  for (idx = 0; idx < num_bins; ++idx) {
+    bin_info[idx].first = -1;
+    bin_info[idx].num_combine_failures = 0;
+  }
+
+  // By default, a cluster matches itself.
+  for (idx = 0; idx < *num_used; ++idx) cluster_mappings[idx] = idx;
+  for (idx = 0; idx < image_histo->size; ++idx) {
+    int bin_id, first;
+    if (histograms[idx] == NULL) continue;
+    bin_id = bin_map[idx];
+    first = bin_info[bin_id].first;
+    if (first == -1) {
+      bin_info[bin_id].first = idx;
+    } else if (low_effort) {
+      HistogramAdd(histograms[idx], histograms[first], histograms[first]);
+      HistogramSetRemoveHistogram(image_histo, idx, num_used);
+      cluster_mappings[clusters[idx]] = clusters[first];
+    } else {
+      // try to merge #idx into #first (both share the same bin_id)
+      const float bit_cost = histograms[idx]->bit_cost_;
+      const float bit_cost_thresh = -bit_cost * combine_cost_factor;
+      const float curr_cost_diff = HistogramAddEval(
+          histograms[first], histograms[idx], cur_combo, bit_cost_thresh);
+      if (curr_cost_diff < bit_cost_thresh) {
+        // Try to merge two histograms only if the combo is a trivial one or
+        // the two candidate histograms are already non-trivial.
+        // For some images, 'try_combine' turns out to be false for a lot of
+        // histogram pairs. In that case, we fallback to combining
+        // histograms as usual to avoid increasing the header size.
+        const int try_combine =
+            (cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) ||
+            ((histograms[idx]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) &&
+             (histograms[first]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM));
+        const int max_combine_failures = 32;
+        if (try_combine ||
+            bin_info[bin_id].num_combine_failures >= max_combine_failures) {
+          // move the (better) merged histogram to its final slot
+          HistogramSwap(&cur_combo, &histograms[first]);
+          HistogramSetRemoveHistogram(image_histo, idx, num_used);
+          cluster_mappings[clusters[idx]] = clusters[first];
+        } else {
+          ++bin_info[bin_id].num_combine_failures;
+        }
+      }
+    }
+  }
+  if (low_effort) {
+    // for low_effort case, update the final cost when everything is merged
+    for (idx = 0; idx < image_histo->size; ++idx) {
+      if (histograms[idx] == NULL) continue;
+      UpdateHistogramCost(histograms[idx]);
+    }
+  }
+}
+
+// Implement a Lehmer random number generator with a multiplicative constant of
+// 48271 and a modulo constant of 2^31 - 1.
+static uint32_t MyRand(uint32_t* const seed) {
+  *seed = (uint32_t)(((uint64_t)(*seed) * 48271u) % 2147483647u);
+  assert(*seed > 0);
+  return *seed;
+}
+
+// -----------------------------------------------------------------------------
+// Histogram pairs priority queue
+
+// Pair of histograms. Negative idx1 value means that pair is out-of-date.
+typedef struct {
+  int idx1;
+  int idx2;
+  float cost_diff;
+  float cost_combo;
+} HistogramPair;
+
+typedef struct {
+  HistogramPair* queue;
+  int size;
+  int max_size;
+} HistoQueue;
+
+static int HistoQueueInit(HistoQueue* const histo_queue, const int max_size) {
+  histo_queue->size = 0;
+  histo_queue->max_size = max_size;
+  // We allocate max_size + 1 because the last element at index "size" is
+  // used as temporary data (and it could be up to max_size).
+  histo_queue->queue = (HistogramPair*)WebPSafeMalloc(
+      histo_queue->max_size + 1, sizeof(*histo_queue->queue));
+  return histo_queue->queue != NULL;
+}
+
+static void HistoQueueClear(HistoQueue* const histo_queue) {
+  assert(histo_queue != NULL);
+  WebPSafeFree(histo_queue->queue);
+  histo_queue->size = 0;
+  histo_queue->max_size = 0;
+}
+
+// Pop a specific pair in the queue by replacing it with the last one
+// and shrinking the queue.
+static void HistoQueuePopPair(HistoQueue* const histo_queue,
+                              HistogramPair* const pair) {
+  assert(pair >= histo_queue->queue &&
+         pair < (histo_queue->queue + histo_queue->size));
+  assert(histo_queue->size > 0);
+  *pair = histo_queue->queue[histo_queue->size - 1];
+  --histo_queue->size;
+}
+
+// Check whether a pair in the queue should be updated as head or not.
+static void HistoQueueUpdateHead(HistoQueue* const histo_queue,
+                                 HistogramPair* const pair) {
+  assert(pair->cost_diff < 0.);
+  assert(pair >= histo_queue->queue &&
+         pair < (histo_queue->queue + histo_queue->size));
+  assert(histo_queue->size > 0);
+  if (pair->cost_diff < histo_queue->queue[0].cost_diff) {
+    // Replace the best pair.
+    const HistogramPair tmp = histo_queue->queue[0];
+    histo_queue->queue[0] = *pair;
+    *pair = tmp;
+  }
+}
+
+// Update the cost diff and combo of a pair of histograms. This needs to be
+// called when the the histograms have been merged with a third one.
+static void HistoQueueUpdatePair(const VP8LHistogram* const h1,
+                                 const VP8LHistogram* const h2, float threshold,
+                                 HistogramPair* const pair) {
+  const float sum_cost = h1->bit_cost_ + h2->bit_cost_;
+  pair->cost_combo = 0.;
+  GetCombinedHistogramEntropy(h1, h2, sum_cost + threshold, &pair->cost_combo);
+  pair->cost_diff = pair->cost_combo - sum_cost;
+}
+
+// Create a pair from indices "idx1" and "idx2" provided its cost
+// is inferior to "threshold", a negative entropy.
+// It returns the cost of the pair, or 0. if it superior to threshold.
+static float HistoQueuePush(HistoQueue* const histo_queue,
+                            VP8LHistogram** const histograms, int idx1,
+                            int idx2, float threshold) {
+  const VP8LHistogram* h1;
+  const VP8LHistogram* h2;
+  HistogramPair pair;
+
+  // Stop here if the queue is full.
+  if (histo_queue->size == histo_queue->max_size) return 0.;
+  assert(threshold <= 0.);
+  if (idx1 > idx2) {
+    const int tmp = idx2;
+    idx2 = idx1;
+    idx1 = tmp;
+  }
+  pair.idx1 = idx1;
+  pair.idx2 = idx2;
+  h1 = histograms[idx1];
+  h2 = histograms[idx2];
+
+  HistoQueueUpdatePair(h1, h2, threshold, &pair);
+
+  // Do not even consider the pair if it does not improve the entropy.
+  if (pair.cost_diff >= threshold) return 0.;
+
+  histo_queue->queue[histo_queue->size++] = pair;
+  HistoQueueUpdateHead(histo_queue, &histo_queue->queue[histo_queue->size - 1]);
+
+  return pair.cost_diff;
+}
+
+// -----------------------------------------------------------------------------
+
+// Combines histograms by continuously choosing the one with the highest cost
+// reduction.
+static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo,
+                                  int* const num_used) {
+  int ok = 0;
+  const int image_histo_size = image_histo->size;
+  int i, j;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  // Priority queue of histogram pairs.
+  HistoQueue histo_queue;
+
+  // image_histo_size^2 for the queue size is safe. If you look at
+  // HistogramCombineGreedy, and imagine that UpdateQueueFront always pushes
+  // data to the queue, you insert at most:
+  // - image_histo_size*(image_histo_size-1)/2 (the first two for loops)
+  // - image_histo_size - 1 in the last for loop at the first iteration of
+  //   the while loop, image_histo_size - 2 at the second iteration ...
+  //   therefore image_histo_size*(image_histo_size-1)/2 overall too
+  if (!HistoQueueInit(&histo_queue, image_histo_size * image_histo_size)) {
+    goto End;
+  }
+
+  for (i = 0; i < image_histo_size; ++i) {
+    if (image_histo->histograms[i] == NULL) continue;
+    for (j = i + 1; j < image_histo_size; ++j) {
+      // Initialize queue.
+      if (image_histo->histograms[j] == NULL) continue;
+      HistoQueuePush(&histo_queue, histograms, i, j, 0.);
+    }
+  }
+
+  while (histo_queue.size > 0) {
+    const int idx1 = histo_queue.queue[0].idx1;
+    const int idx2 = histo_queue.queue[0].idx2;
+    HistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]);
+    histograms[idx1]->bit_cost_ = histo_queue.queue[0].cost_combo;
+
+    // Remove merged histogram.
+    HistogramSetRemoveHistogram(image_histo, idx2, num_used);
+
+    // Remove pairs intersecting the just combined best pair.
+    for (i = 0; i < histo_queue.size;) {
+      HistogramPair* const p = histo_queue.queue + i;
+      if (p->idx1 == idx1 || p->idx2 == idx1 ||
+          p->idx1 == idx2 || p->idx2 == idx2) {
+        HistoQueuePopPair(&histo_queue, p);
+      } else {
+        HistoQueueUpdateHead(&histo_queue, p);
+        ++i;
+      }
+    }
+
+    // Push new pairs formed with combined histogram to the queue.
+    for (i = 0; i < image_histo->size; ++i) {
+      if (i == idx1 || image_histo->histograms[i] == NULL) continue;
+      HistoQueuePush(&histo_queue, image_histo->histograms, idx1, i, 0.);
+    }
+  }
+
+  ok = 1;
+
+ End:
+  HistoQueueClear(&histo_queue);
+  return ok;
+}
+
+// Perform histogram aggregation using a stochastic approach.
+// 'do_greedy' is set to 1 if a greedy approach needs to be performed
+// afterwards, 0 otherwise.
+static int PairComparison(const void* idx1, const void* idx2) {
+  // To be used with bsearch: <0 when *idx1<*idx2, >0 if >, 0 when ==.
+  return (*(int*) idx1 - *(int*) idx2);
+}
+static int HistogramCombineStochastic(VP8LHistogramSet* const image_histo,
+                                      int* const num_used, int min_cluster_size,
+                                      int* const do_greedy) {
+  int j, iter;
+  uint32_t seed = 1;
+  int tries_with_no_success = 0;
+  const int outer_iters = *num_used;
+  const int num_tries_no_success = outer_iters / 2;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  // Priority queue of histogram pairs. Its size of 'kHistoQueueSize'
+  // impacts the quality of the compression and the speed: the smaller the
+  // faster but the worse for the compression.
+  HistoQueue histo_queue;
+  const int kHistoQueueSize = 9;
+  int ok = 0;
+  // mapping from an index in image_histo with no NULL histogram to the full
+  // blown image_histo.
+  int* mappings;
+
+  if (*num_used < min_cluster_size) {
+    *do_greedy = 1;
+    return 1;
+  }
+
+  mappings = (int*) WebPSafeMalloc(*num_used, sizeof(*mappings));
+  if (mappings == NULL) return 0;
+  if (!HistoQueueInit(&histo_queue, kHistoQueueSize)) goto End;
+  // Fill the initial mapping.
+  for (j = 0, iter = 0; iter < image_histo->size; ++iter) {
+    if (histograms[iter] == NULL) continue;
+    mappings[j++] = iter;
+  }
+  assert(j == *num_used);
+
+  // Collapse similar histograms in 'image_histo'.
+  for (iter = 0;
+       iter < outer_iters && *num_used >= min_cluster_size &&
+           ++tries_with_no_success < num_tries_no_success;
+       ++iter) {
+    int* mapping_index;
+    float best_cost =
+        (histo_queue.size == 0) ? 0.f : histo_queue.queue[0].cost_diff;
+    int best_idx1 = -1, best_idx2 = 1;
+    const uint32_t rand_range = (*num_used - 1) * (*num_used);
+    // (*num_used) / 2 was chosen empirically. Less means faster but worse
+    // compression.
+    const int num_tries = (*num_used) / 2;
+
+    // Pick random samples.
+    for (j = 0; *num_used >= 2 && j < num_tries; ++j) {
+      float curr_cost;
+      // Choose two different histograms at random and try to combine them.
+      const uint32_t tmp = MyRand(&seed) % rand_range;
+      uint32_t idx1 = tmp / (*num_used - 1);
+      uint32_t idx2 = tmp % (*num_used - 1);
+      if (idx2 >= idx1) ++idx2;
+      idx1 = mappings[idx1];
+      idx2 = mappings[idx2];
+
+      // Calculate cost reduction on combination.
+      curr_cost =
+          HistoQueuePush(&histo_queue, histograms, idx1, idx2, best_cost);
+      if (curr_cost < 0) {  // found a better pair?
+        best_cost = curr_cost;
+        // Empty the queue if we reached full capacity.
+        if (histo_queue.size == histo_queue.max_size) break;
+      }
+    }
+    if (histo_queue.size == 0) continue;
+
+    // Get the best histograms.
+    best_idx1 = histo_queue.queue[0].idx1;
+    best_idx2 = histo_queue.queue[0].idx2;
+    assert(best_idx1 < best_idx2);
+    // Pop best_idx2 from mappings.
+    mapping_index = (int*) bsearch(&best_idx2, mappings, *num_used,
+                                   sizeof(best_idx2), &PairComparison);
+    assert(mapping_index != NULL);
+    memmove(mapping_index, mapping_index + 1, sizeof(*mapping_index) *
+        ((*num_used) - (mapping_index - mappings) - 1));
+    // Merge the histograms and remove best_idx2 from the queue.
+    HistogramAdd(histograms[best_idx2], histograms[best_idx1],
+                 histograms[best_idx1]);
+    histograms[best_idx1]->bit_cost_ = histo_queue.queue[0].cost_combo;
+    HistogramSetRemoveHistogram(image_histo, best_idx2, num_used);
+    // Parse the queue and update each pair that deals with best_idx1,
+    // best_idx2 or image_histo_size.
+    for (j = 0; j < histo_queue.size;) {
+      HistogramPair* const p = histo_queue.queue + j;
+      const int is_idx1_best = p->idx1 == best_idx1 || p->idx1 == best_idx2;
+      const int is_idx2_best = p->idx2 == best_idx1 || p->idx2 == best_idx2;
+      int do_eval = 0;
+      // The front pair could have been duplicated by a random pick so
+      // check for it all the time nevertheless.
+      if (is_idx1_best && is_idx2_best) {
+        HistoQueuePopPair(&histo_queue, p);
+        continue;
+      }
+      // Any pair containing one of the two best indices should only refer to
+      // best_idx1. Its cost should also be updated.
+      if (is_idx1_best) {
+        p->idx1 = best_idx1;
+        do_eval = 1;
+      } else if (is_idx2_best) {
+        p->idx2 = best_idx1;
+        do_eval = 1;
+      }
+      // Make sure the index order is respected.
+      if (p->idx1 > p->idx2) {
+        const int tmp = p->idx2;
+        p->idx2 = p->idx1;
+        p->idx1 = tmp;
+      }
+      if (do_eval) {
+        // Re-evaluate the cost of an updated pair.
+        HistoQueueUpdatePair(histograms[p->idx1], histograms[p->idx2], 0., p);
+        if (p->cost_diff >= 0.) {
+          HistoQueuePopPair(&histo_queue, p);
+          continue;
+        }
+      }
+      HistoQueueUpdateHead(&histo_queue, p);
+      ++j;
+    }
+    tries_with_no_success = 0;
+  }
+  *do_greedy = (*num_used <= min_cluster_size);
+  ok = 1;
+
+ End:
+  HistoQueueClear(&histo_queue);
+  WebPSafeFree(mappings);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+// Histogram refinement
+
+// Find the best 'out' histogram for each of the 'in' histograms.
+// At call-time, 'out' contains the histograms of the clusters.
+// Note: we assume that out[]->bit_cost_ is already up-to-date.
+static void HistogramRemap(const VP8LHistogramSet* const in,
+                           VP8LHistogramSet* const out,
+                           uint16_t* const symbols) {
+  int i;
+  VP8LHistogram** const in_histo = in->histograms;
+  VP8LHistogram** const out_histo = out->histograms;
+  const int in_size = out->max_size;
+  const int out_size = out->size;
+  if (out_size > 1) {
+    for (i = 0; i < in_size; ++i) {
+      int best_out = 0;
+      float best_bits = MAX_BIT_COST;
+      int k;
+      if (in_histo[i] == NULL) {
+        // Arbitrarily set to the previous value if unused to help future LZ77.
+        symbols[i] = symbols[i - 1];
+        continue;
+      }
+      for (k = 0; k < out_size; ++k) {
+        float cur_bits;
+        cur_bits = HistogramAddThresh(out_histo[k], in_histo[i], best_bits);
+        if (k == 0 || cur_bits < best_bits) {
+          best_bits = cur_bits;
+          best_out = k;
+        }
+      }
+      symbols[i] = best_out;
+    }
+  } else {
+    assert(out_size == 1);
+    for (i = 0; i < in_size; ++i) {
+      symbols[i] = 0;
+    }
+  }
+
+  // Recompute each out based on raw and symbols.
+  VP8LHistogramSetClear(out);
+  out->size = out_size;
+
+  for (i = 0; i < in_size; ++i) {
+    int idx;
+    if (in_histo[i] == NULL) continue;
+    idx = symbols[i];
+    HistogramAdd(in_histo[i], out_histo[idx], out_histo[idx]);
+  }
+}
+
+static float GetCombineCostFactor(int histo_size, int quality) {
+  float combine_cost_factor = 0.16f;
+  if (quality < 90) {
+    if (histo_size > 256) combine_cost_factor /= 2.f;
+    if (histo_size > 512) combine_cost_factor /= 2.f;
+    if (histo_size > 1024) combine_cost_factor /= 2.f;
+    if (quality <= 50) combine_cost_factor /= 2.f;
+  }
+  return combine_cost_factor;
+}
+
+// Given a HistogramSet 'set', the mapping of clusters 'cluster_mapping' and the
+// current assignment of the cells in 'symbols', merge the clusters and
+// assign the smallest possible clusters values.
+static void OptimizeHistogramSymbols(const VP8LHistogramSet* const set,
+                                     uint16_t* const cluster_mappings,
+                                     int num_clusters,
+                                     uint16_t* const cluster_mappings_tmp,
+                                     uint16_t* const symbols) {
+  int i, cluster_max;
+  int do_continue = 1;
+  // First, assign the lowest cluster to each pixel.
+  while (do_continue) {
+    do_continue = 0;
+    for (i = 0; i < num_clusters; ++i) {
+      int k;
+      k = cluster_mappings[i];
+      while (k != cluster_mappings[k]) {
+        cluster_mappings[k] = cluster_mappings[cluster_mappings[k]];
+        k = cluster_mappings[k];
+      }
+      if (k != cluster_mappings[i]) {
+        do_continue = 1;
+        cluster_mappings[i] = k;
+      }
+    }
+  }
+  // Create a mapping from a cluster id to its minimal version.
+  cluster_max = 0;
+  memset(cluster_mappings_tmp, 0,
+         set->max_size * sizeof(*cluster_mappings_tmp));
+  assert(cluster_mappings[0] == 0);
+  // Re-map the ids.
+  for (i = 0; i < set->max_size; ++i) {
+    int cluster;
+    if (symbols[i] == kInvalidHistogramSymbol) continue;
+    cluster = cluster_mappings[symbols[i]];
+    assert(symbols[i] < num_clusters);
+    if (cluster > 0 && cluster_mappings_tmp[cluster] == 0) {
+      ++cluster_max;
+      cluster_mappings_tmp[cluster] = cluster_max;
+    }
+    symbols[i] = cluster_mappings_tmp[cluster];
+  }
+
+  // Make sure all cluster values are used.
+  cluster_max = 0;
+  for (i = 0; i < set->max_size; ++i) {
+    if (symbols[i] == kInvalidHistogramSymbol) continue;
+    if (symbols[i] <= cluster_max) continue;
+    ++cluster_max;
+    assert(symbols[i] == cluster_max);
+  }
+}
+
+static void RemoveEmptyHistograms(VP8LHistogramSet* const image_histo) {
+  uint32_t size;
+  int i;
+  for (i = 0, size = 0; i < image_histo->size; ++i) {
+    if (image_histo->histograms[i] == NULL) continue;
+    image_histo->histograms[size++] = image_histo->histograms[i];
+  }
+  image_histo->size = size;
+}
+
+int VP8LGetHistoImageSymbols(int xsize, int ysize,
+                             const VP8LBackwardRefs* const refs, int quality,
+                             int low_effort, int histogram_bits, int cache_bits,
+                             VP8LHistogramSet* const image_histo,
+                             VP8LHistogram* const tmp_histo,
+                             uint16_t* const histogram_symbols,
+                             const WebPPicture* const pic, int percent_range,
+                             int* const percent) {
+  const int histo_xsize =
+      histogram_bits ? VP8LSubSampleSize(xsize, histogram_bits) : 1;
+  const int histo_ysize =
+      histogram_bits ? VP8LSubSampleSize(ysize, histogram_bits) : 1;
+  const int image_histo_raw_size = histo_xsize * histo_ysize;
+  VP8LHistogramSet* const orig_histo =
+      VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
+  // Don't attempt linear bin-partition heuristic for
+  // histograms of small sizes (as bin_map will be very sparse) and
+  // maximum quality q==100 (to preserve the compression gains at that level).
+  const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE;
+  int entropy_combine;
+  uint16_t* const map_tmp =
+      WebPSafeMalloc(2 * image_histo_raw_size, sizeof(map_tmp));
+  uint16_t* const cluster_mappings = map_tmp + image_histo_raw_size;
+  int num_used = image_histo_raw_size;
+  if (orig_histo == NULL || map_tmp == NULL) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  // Construct the histograms from backward references.
+  HistogramBuild(xsize, histogram_bits, refs, orig_histo);
+  // Copies the histograms and computes its bit_cost.
+  // histogram_symbols is optimized
+  HistogramCopyAndAnalyze(orig_histo, image_histo, &num_used,
+                          histogram_symbols);
+
+  entropy_combine =
+      (num_used > entropy_combine_num_bins * 2) && (quality < 100);
+
+  if (entropy_combine) {
+    uint16_t* const bin_map = map_tmp;
+    const float combine_cost_factor =
+        GetCombineCostFactor(image_histo_raw_size, quality);
+    const uint32_t num_clusters = num_used;
+
+    HistogramAnalyzeEntropyBin(image_histo, bin_map, low_effort);
+    // Collapse histograms with similar entropy.
+    HistogramCombineEntropyBin(
+        image_histo, &num_used, histogram_symbols, cluster_mappings, tmp_histo,
+        bin_map, entropy_combine_num_bins, combine_cost_factor, low_effort);
+    OptimizeHistogramSymbols(image_histo, cluster_mappings, num_clusters,
+                             map_tmp, histogram_symbols);
+  }
+
+  // Don't combine the histograms using stochastic and greedy heuristics for
+  // low-effort compression mode.
+  if (!low_effort || !entropy_combine) {
+    const float x = quality / 100.f;
+    // cubic ramp between 1 and MAX_HISTO_GREEDY:
+    const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1));
+    int do_greedy;
+    if (!HistogramCombineStochastic(image_histo, &num_used, threshold_size,
+                                    &do_greedy)) {
+      WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+      goto Error;
+    }
+    if (do_greedy) {
+      RemoveEmptyHistograms(image_histo);
+      if (!HistogramCombineGreedy(image_histo, &num_used)) {
+        WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+        goto Error;
+      }
+    }
+  }
+
+  // Find the optimal map from original histograms to the final ones.
+  RemoveEmptyHistograms(image_histo);
+  HistogramRemap(orig_histo, image_histo, histogram_symbols);
+
+  if (!WebPReportProgress(pic, *percent + percent_range, percent)) {
+    goto Error;
+  }
+
+ Error:
+  VP8LFreeHistogramSet(orig_histo);
+  WebPSafeFree(map_tmp);
+  return (pic->error_code == VP8_ENC_OK);
+}
diff --git a/media/libwebp/src/enc/histogram_enc.h b/media/libwebp/src/enc/histogram_enc.h
new file mode 100644
index 0000000000..4c0bb97464
--- /dev/null
+++ b/media/libwebp/src/enc/histogram_enc.h
@@ -0,0 +1,130 @@
+// 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)
+//
+// Models the histograms of literal and distance codes.
+
+#ifndef WEBP_ENC_HISTOGRAM_ENC_H_
+#define WEBP_ENC_HISTOGRAM_ENC_H_
+
+#include <string.h>
+
+#include "src/enc/backward_references_enc.h"
+#include "src/webp/format_constants.h"
+#include "src/webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Not a trivial literal symbol.
+#define VP8L_NON_TRIVIAL_SYM (0xffffffff)
+
+// A simple container for histograms of data.
+typedef struct {
+  // literal_ contains green literal, palette-code and
+  // copy-length-prefix histogram
+  uint32_t* literal_;         // Pointer to the allocated buffer for literal.
+  uint32_t red_[NUM_LITERAL_CODES];
+  uint32_t blue_[NUM_LITERAL_CODES];
+  uint32_t alpha_[NUM_LITERAL_CODES];
+  // Backward reference prefix-code histogram.
+  uint32_t distance_[NUM_DISTANCE_CODES];
+  int palette_code_bits_;
+  uint32_t trivial_symbol_;  // True, if histograms for Red, Blue & Alpha
+                             // literal symbols are single valued.
+  float bit_cost_;           // cached value of bit cost.
+  float literal_cost_;       // Cached values of dominant entropy costs:
+  float red_cost_;           // literal, red & blue.
+  float blue_cost_;
+  uint8_t is_used_[5];       // 5 for literal, red, blue, alpha, distance
+} VP8LHistogram;
+
+// Collection of histograms with fixed capacity, allocated as one
+// big memory chunk. Can be destroyed by calling WebPSafeFree().
+typedef struct {
+  int size;         // number of slots currently in use
+  int max_size;     // maximum capacity
+  VP8LHistogram** histograms;
+} VP8LHistogramSet;
+
+// Create the histogram.
+//
+// The input data is the PixOrCopy data, which models the literals, stop
+// codes and backward references (both distances and lengths).  Also: if
+// palette_code_bits is >= 0, initialize the histogram with this value.
+void VP8LHistogramCreate(VP8LHistogram* const p,
+                         const VP8LBackwardRefs* const refs,
+                         int palette_code_bits);
+
+// Return the size of the histogram for a given cache_bits.
+int VP8LGetHistogramSize(int cache_bits);
+
+// Set the palette_code_bits and reset the stats.
+// If init_arrays is true, the arrays are also filled with 0's.
+void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits,
+                       int init_arrays);
+
+// Collect all the references into a histogram (without reset)
+void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
+                            VP8LHistogram* const histo);
+
+// Free the memory allocated for the histogram.
+void VP8LFreeHistogram(VP8LHistogram* const histo);
+
+// Free the memory allocated for the histogram set.
+void VP8LFreeHistogramSet(VP8LHistogramSet* const histo);
+
+// Allocate an array of pointer to histograms, allocated and initialized
+// using 'cache_bits'. Return NULL in case of memory error.
+VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits);
+
+// Set the histograms in set to 0.
+void VP8LHistogramSetClear(VP8LHistogramSet* const set);
+
+// Allocate and initialize histogram object with specified 'cache_bits'.
+// Returns NULL in case of memory error.
+// Special case of VP8LAllocateHistogramSet, with size equals 1.
+VP8LHistogram* VP8LAllocateHistogram(int cache_bits);
+
+// Accumulate a token 'v' into a histogram.
+void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
+                                     const PixOrCopy* const v,
+                                     int (*const distance_modifier)(int, int),
+                                     int distance_modifier_arg0);
+
+static WEBP_INLINE int VP8LHistogramNumCodes(int palette_code_bits) {
+  return NUM_LITERAL_CODES + NUM_LENGTH_CODES +
+      ((palette_code_bits > 0) ? (1 << palette_code_bits) : 0);
+}
+
+// Builds the histogram image. pic and percent are for progress.
+// Returns false in case of error (stored in pic->error_code).
+int VP8LGetHistoImageSymbols(int xsize, int ysize,
+                             const VP8LBackwardRefs* const refs, int quality,
+                             int low_effort, int histogram_bits, int cache_bits,
+                             VP8LHistogramSet* const image_histo,
+                             VP8LHistogram* const tmp_histo,
+                             uint16_t* const histogram_symbols,
+                             const WebPPicture* const pic, int percent_range,
+                             int* const percent);
+
+// Returns the entropy for the symbols in the input array.
+float VP8LBitsEntropy(const uint32_t* const array, int n);
+
+// Estimate how many bits the combined entropy of literals and distance
+// approximately maps to.
+float VP8LHistogramEstimateBits(VP8LHistogram* const p);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // WEBP_ENC_HISTOGRAM_ENC_H_
diff --git a/media/libwebp/src/enc/iterator_enc.c b/media/libwebp/src/enc/iterator_enc.c
new file mode 100644
index 0000000000..29f91d8315
--- /dev/null
+++ b/media/libwebp/src/enc/iterator_enc.c
@@ -0,0 +1,459 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// VP8Iterator: block iterator
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <string.h>
+
+#include "src/enc/vp8i_enc.h"
+
+//------------------------------------------------------------------------------
+// VP8Iterator
+//------------------------------------------------------------------------------
+
+static void InitLeft(VP8EncIterator* const it) {
+  it->y_left_[-1] = it->u_left_[-1] = it->v_left_[-1] =
+      (it->y_ > 0) ? 129 : 127;
+  memset(it->y_left_, 129, 16);
+  memset(it->u_left_, 129, 8);
+  memset(it->v_left_, 129, 8);
+  it->left_nz_[8] = 0;
+  if (it->top_derr_ != NULL) {
+    memset(&it->left_derr_, 0, sizeof(it->left_derr_));
+  }
+}
+
+static void InitTop(VP8EncIterator* const it) {
+  const VP8Encoder* const enc = it->enc_;
+  const size_t top_size = enc->mb_w_ * 16;
+  memset(enc->y_top_, 127, 2 * top_size);
+  memset(enc->nz_, 0, enc->mb_w_ * sizeof(*enc->nz_));
+  if (enc->top_derr_ != NULL) {
+    memset(enc->top_derr_, 0, enc->mb_w_ * sizeof(*enc->top_derr_));
+  }
+}
+
+void VP8IteratorSetRow(VP8EncIterator* const it, int y) {
+  VP8Encoder* const enc = it->enc_;
+  it->x_ = 0;
+  it->y_ = y;
+  it->bw_ = &enc->parts_[y & (enc->num_parts_ - 1)];
+  it->preds_ = enc->preds_ + y * 4 * enc->preds_w_;
+  it->nz_ = enc->nz_;
+  it->mb_ = enc->mb_info_ + y * enc->mb_w_;
+  it->y_top_ = enc->y_top_;
+  it->uv_top_ = enc->uv_top_;
+  InitLeft(it);
+}
+
+void VP8IteratorReset(VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  VP8IteratorSetRow(it, 0);
+  VP8IteratorSetCountDown(it, enc->mb_w_ * enc->mb_h_);  // default
+  InitTop(it);
+  memset(it->bit_count_, 0, sizeof(it->bit_count_));
+  it->do_trellis_ = 0;
+}
+
+void VP8IteratorSetCountDown(VP8EncIterator* const it, int count_down) {
+  it->count_down_ = it->count_down0_ = count_down;
+}
+
+int VP8IteratorIsDone(const VP8EncIterator* const it) {
+  return (it->count_down_ <= 0);
+}
+
+void VP8IteratorInit(VP8Encoder* const enc, VP8EncIterator* const it) {
+  it->enc_ = enc;
+  it->yuv_in_   = (uint8_t*)WEBP_ALIGN(it->yuv_mem_);
+  it->yuv_out_  = it->yuv_in_ + YUV_SIZE_ENC;
+  it->yuv_out2_ = it->yuv_out_ + YUV_SIZE_ENC;
+  it->yuv_p_    = it->yuv_out2_ + YUV_SIZE_ENC;
+  it->lf_stats_ = enc->lf_stats_;
+  it->percent0_ = enc->percent_;
+  it->y_left_ = (uint8_t*)WEBP_ALIGN(it->yuv_left_mem_ + 1);
+  it->u_left_ = it->y_left_ + 16 + 16;
+  it->v_left_ = it->u_left_ + 16;
+  it->top_derr_ = enc->top_derr_;
+  VP8IteratorReset(it);
+}
+
+int VP8IteratorProgress(const VP8EncIterator* const it, int delta) {
+  VP8Encoder* const enc = it->enc_;
+  if (delta && enc->pic_->progress_hook != NULL) {
+    const int done = it->count_down0_ - it->count_down_;
+    const int percent = (it->count_down0_ <= 0)
+                      ? it->percent0_
+                      : it->percent0_ + delta * done / it->count_down0_;
+    return WebPReportProgress(enc->pic_, percent, &enc->percent_);
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Import the source samples into the cache. Takes care of replicating
+// boundary pixels if necessary.
+
+static WEBP_INLINE int MinSize(int a, int b) { return (a < b) ? a : b; }
+
+static void ImportBlock(const uint8_t* src, int src_stride,
+                        uint8_t* dst, int w, int h, int size) {
+  int i;
+  for (i = 0; i < h; ++i) {
+    memcpy(dst, src, w);
+    if (w < size) {
+      memset(dst + w, dst[w - 1], size - w);
+    }
+    dst += BPS;
+    src += src_stride;
+  }
+  for (i = h; i < size; ++i) {
+    memcpy(dst, dst - BPS, size);
+    dst += BPS;
+  }
+}
+
+static void ImportLine(const uint8_t* src, int src_stride,
+                       uint8_t* dst, int len, int total_len) {
+  int i;
+  for (i = 0; i < len; ++i, src += src_stride) dst[i] = *src;
+  for (; i < total_len; ++i) dst[i] = dst[len - 1];
+}
+
+void VP8IteratorImport(VP8EncIterator* const it, uint8_t* const tmp_32) {
+  const VP8Encoder* const enc = it->enc_;
+  const int x = it->x_, y = it->y_;
+  const WebPPicture* const pic = enc->pic_;
+  const uint8_t* const ysrc = pic->y + (y * pic->y_stride  + x) * 16;
+  const uint8_t* const usrc = pic->u + (y * pic->uv_stride + x) * 8;
+  const uint8_t* const vsrc = pic->v + (y * pic->uv_stride + x) * 8;
+  const int w = MinSize(pic->width - x * 16, 16);
+  const int h = MinSize(pic->height - y * 16, 16);
+  const int uv_w = (w + 1) >> 1;
+  const int uv_h = (h + 1) >> 1;
+
+  ImportBlock(ysrc, pic->y_stride,  it->yuv_in_ + Y_OFF_ENC, w, h, 16);
+  ImportBlock(usrc, pic->uv_stride, it->yuv_in_ + U_OFF_ENC, uv_w, uv_h, 8);
+  ImportBlock(vsrc, pic->uv_stride, it->yuv_in_ + V_OFF_ENC, uv_w, uv_h, 8);
+
+  if (tmp_32 == NULL) return;
+
+  // Import source (uncompressed) samples into boundary.
+  if (x == 0) {
+    InitLeft(it);
+  } else {
+    if (y == 0) {
+      it->y_left_[-1] = it->u_left_[-1] = it->v_left_[-1] = 127;
+    } else {
+      it->y_left_[-1] = ysrc[- 1 - pic->y_stride];
+      it->u_left_[-1] = usrc[- 1 - pic->uv_stride];
+      it->v_left_[-1] = vsrc[- 1 - pic->uv_stride];
+    }
+    ImportLine(ysrc - 1, pic->y_stride,  it->y_left_, h,   16);
+    ImportLine(usrc - 1, pic->uv_stride, it->u_left_, uv_h, 8);
+    ImportLine(vsrc - 1, pic->uv_stride, it->v_left_, uv_h, 8);
+  }
+
+  it->y_top_  = tmp_32 + 0;
+  it->uv_top_ = tmp_32 + 16;
+  if (y == 0) {
+    memset(tmp_32, 127, 32 * sizeof(*tmp_32));
+  } else {
+    ImportLine(ysrc - pic->y_stride,  1, tmp_32,          w,   16);
+    ImportLine(usrc - pic->uv_stride, 1, tmp_32 + 16,     uv_w, 8);
+    ImportLine(vsrc - pic->uv_stride, 1, tmp_32 + 16 + 8, uv_w, 8);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Copy back the compressed samples into user space if requested.
+
+static void ExportBlock(const uint8_t* src, uint8_t* dst, int dst_stride,
+                        int w, int h) {
+  while (h-- > 0) {
+    memcpy(dst, src, w);
+    dst += dst_stride;
+    src += BPS;
+  }
+}
+
+void VP8IteratorExport(const VP8EncIterator* const it) {
+  const VP8Encoder* const enc = it->enc_;
+  if (enc->config_->show_compressed) {
+    const int x = it->x_, y = it->y_;
+    const uint8_t* const ysrc = it->yuv_out_ + Y_OFF_ENC;
+    const uint8_t* const usrc = it->yuv_out_ + U_OFF_ENC;
+    const uint8_t* const vsrc = it->yuv_out_ + V_OFF_ENC;
+    const WebPPicture* const pic = enc->pic_;
+    uint8_t* const ydst = pic->y + (y * pic->y_stride + x) * 16;
+    uint8_t* const udst = pic->u + (y * pic->uv_stride + x) * 8;
+    uint8_t* const vdst = pic->v + (y * pic->uv_stride + x) * 8;
+    int w = (pic->width - x * 16);
+    int h = (pic->height - y * 16);
+
+    if (w > 16) w = 16;
+    if (h > 16) h = 16;
+
+    // Luma plane
+    ExportBlock(ysrc, ydst, pic->y_stride, w, h);
+
+    {   // U/V planes
+      const int uv_w = (w + 1) >> 1;
+      const int uv_h = (h + 1) >> 1;
+      ExportBlock(usrc, udst, pic->uv_stride, uv_w, uv_h);
+      ExportBlock(vsrc, vdst, pic->uv_stride, uv_w, uv_h);
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+// Non-zero contexts setup/teardown
+
+// Nz bits:
+//  0  1  2  3  Y
+//  4  5  6  7
+//  8  9 10 11
+// 12 13 14 15
+// 16 17        U
+// 18 19
+// 20 21        V
+// 22 23
+// 24           DC-intra16
+
+// Convert packed context to byte array
+#define BIT(nz, n) (!!((nz) & (1 << (n))))
+
+void VP8IteratorNzToBytes(VP8EncIterator* const it) {
+  const int tnz = it->nz_[0], lnz = it->nz_[-1];
+  int* const top_nz = it->top_nz_;
+  int* const left_nz = it->left_nz_;
+
+  // Top-Y
+  top_nz[0] = BIT(tnz, 12);
+  top_nz[1] = BIT(tnz, 13);
+  top_nz[2] = BIT(tnz, 14);
+  top_nz[3] = BIT(tnz, 15);
+  // Top-U
+  top_nz[4] = BIT(tnz, 18);
+  top_nz[5] = BIT(tnz, 19);
+  // Top-V
+  top_nz[6] = BIT(tnz, 22);
+  top_nz[7] = BIT(tnz, 23);
+  // DC
+  top_nz[8] = BIT(tnz, 24);
+
+  // left-Y
+  left_nz[0] = BIT(lnz,  3);
+  left_nz[1] = BIT(lnz,  7);
+  left_nz[2] = BIT(lnz, 11);
+  left_nz[3] = BIT(lnz, 15);
+  // left-U
+  left_nz[4] = BIT(lnz, 17);
+  left_nz[5] = BIT(lnz, 19);
+  // left-V
+  left_nz[6] = BIT(lnz, 21);
+  left_nz[7] = BIT(lnz, 23);
+  // left-DC is special, iterated separately
+}
+
+void VP8IteratorBytesToNz(VP8EncIterator* const it) {
+  uint32_t nz = 0;
+  const int* const top_nz = it->top_nz_;
+  const int* const left_nz = it->left_nz_;
+  // top
+  nz |= (top_nz[0] << 12) | (top_nz[1] << 13);
+  nz |= (top_nz[2] << 14) | (top_nz[3] << 15);
+  nz |= (top_nz[4] << 18) | (top_nz[5] << 19);
+  nz |= (top_nz[6] << 22) | (top_nz[7] << 23);
+  nz |= (top_nz[8] << 24);  // we propagate the _top_ bit, esp. for intra4
+  // left
+  nz |= (left_nz[0] << 3) | (left_nz[1] << 7);
+  nz |= (left_nz[2] << 11);
+  nz |= (left_nz[4] << 17) | (left_nz[6] << 21);
+
+  *it->nz_ = nz;
+}
+
+#undef BIT
+
+//------------------------------------------------------------------------------
+// Advance to the next position, doing the bookkeeping.
+
+void VP8IteratorSaveBoundary(VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  const int x = it->x_, y = it->y_;
+  const uint8_t* const ysrc = it->yuv_out_ + Y_OFF_ENC;
+  const uint8_t* const uvsrc = it->yuv_out_ + U_OFF_ENC;
+  if (x < enc->mb_w_ - 1) {   // left
+    int i;
+    for (i = 0; i < 16; ++i) {
+      it->y_left_[i] = ysrc[15 + i * BPS];
+    }
+    for (i = 0; i < 8; ++i) {
+      it->u_left_[i] = uvsrc[7 + i * BPS];
+      it->v_left_[i] = uvsrc[15 + i * BPS];
+    }
+    // top-left (before 'top'!)
+    it->y_left_[-1] = it->y_top_[15];
+    it->u_left_[-1] = it->uv_top_[0 + 7];
+    it->v_left_[-1] = it->uv_top_[8 + 7];
+  }
+  if (y < enc->mb_h_ - 1) {  // top
+    memcpy(it->y_top_, ysrc + 15 * BPS, 16);
+    memcpy(it->uv_top_, uvsrc + 7 * BPS, 8 + 8);
+  }
+}
+
+int VP8IteratorNext(VP8EncIterator* const it) {
+  if (++it->x_ == it->enc_->mb_w_) {
+    VP8IteratorSetRow(it, ++it->y_);
+  } else {
+    it->preds_ += 4;
+    it->mb_ += 1;
+    it->nz_ += 1;
+    it->y_top_ += 16;
+    it->uv_top_ += 16;
+  }
+  return (0 < --it->count_down_);
+}
+
+//------------------------------------------------------------------------------
+// Helper function to set mode properties
+
+void VP8SetIntra16Mode(const VP8EncIterator* const it, int mode) {
+  uint8_t* preds = it->preds_;
+  int y;
+  for (y = 0; y < 4; ++y) {
+    memset(preds, mode, 4);
+    preds += it->enc_->preds_w_;
+  }
+  it->mb_->type_ = 1;
+}
+
+void VP8SetIntra4Mode(const VP8EncIterator* const it, const uint8_t* modes) {
+  uint8_t* preds = it->preds_;
+  int y;
+  for (y = 4; y > 0; --y) {
+    memcpy(preds, modes, 4 * sizeof(*modes));
+    preds += it->enc_->preds_w_;
+    modes += 4;
+  }
+  it->mb_->type_ = 0;
+}
+
+void VP8SetIntraUVMode(const VP8EncIterator* const it, int mode) {
+  it->mb_->uv_mode_ = mode;
+}
+
+void VP8SetSkip(const VP8EncIterator* const it, int skip) {
+  it->mb_->skip_ = skip;
+}
+
+void VP8SetSegment(const VP8EncIterator* const it, int segment) {
+  it->mb_->segment_ = segment;
+}
+
+//------------------------------------------------------------------------------
+// Intra4x4 sub-blocks iteration
+//
+//  We store and update the boundary samples into an array of 37 pixels. They
+//  are updated as we iterate and reconstructs each intra4x4 blocks in turn.
+//  The position of the samples has the following snake pattern:
+//
+// 16|17 18 19 20|21 22 23 24|25 26 27 28|29 30 31 32|33 34 35 36  <- Top-right
+// --+-----------+-----------+-----------+-----------+
+// 15|         19|         23|         27|         31|
+// 14|         18|         22|         26|         30|
+// 13|         17|         21|         25|         29|
+// 12|13 14 15 16|17 18 19 20|21 22 23 24|25 26 27 28|
+// --+-----------+-----------+-----------+-----------+
+// 11|         15|         19|         23|         27|
+// 10|         14|         18|         22|         26|
+//  9|         13|         17|         21|         25|
+//  8| 9 10 11 12|13 14 15 16|17 18 19 20|21 22 23 24|
+// --+-----------+-----------+-----------+-----------+
+//  7|         11|         15|         19|         23|
+//  6|         10|         14|         18|         22|
+//  5|          9|         13|         17|         21|
+//  4| 5  6  7  8| 9 10 11 12|13 14 15 16|17 18 19 20|
+// --+-----------+-----------+-----------+-----------+
+//  3|          7|         11|         15|         19|
+//  2|          6|         10|         14|         18|
+//  1|          5|          9|         13|         17|
+//  0| 1  2  3  4| 5  6  7  8| 9 10 11 12|13 14 15 16|
+// --+-----------+-----------+-----------+-----------+
+
+// Array to record the position of the top sample to pass to the prediction
+// functions in dsp.c.
+static const uint8_t VP8TopLeftI4[16] = {
+  17, 21, 25, 29,
+  13, 17, 21, 25,
+  9,  13, 17, 21,
+  5,   9, 13, 17
+};
+
+void VP8IteratorStartI4(VP8EncIterator* const it) {
+  const VP8Encoder* const enc = it->enc_;
+  int i;
+
+  it->i4_ = 0;    // first 4x4 sub-block
+  it->i4_top_ = it->i4_boundary_ + VP8TopLeftI4[0];
+
+  // Import the boundary samples
+  for (i = 0; i < 17; ++i) {    // left
+    it->i4_boundary_[i] = it->y_left_[15 - i];
+  }
+  for (i = 0; i < 16; ++i) {    // top
+    it->i4_boundary_[17 + i] = it->y_top_[i];
+  }
+  // top-right samples have a special case on the far right of the picture
+  if (it->x_ < enc->mb_w_ - 1) {
+    for (i = 16; i < 16 + 4; ++i) {
+      it->i4_boundary_[17 + i] = it->y_top_[i];
+    }
+  } else {    // else, replicate the last valid pixel four times
+    for (i = 16; i < 16 + 4; ++i) {
+      it->i4_boundary_[17 + i] = it->i4_boundary_[17 + 15];
+    }
+  }
+  VP8IteratorNzToBytes(it);  // import the non-zero context
+}
+
+int VP8IteratorRotateI4(VP8EncIterator* const it,
+                        const uint8_t* const yuv_out) {
+  const uint8_t* const blk = yuv_out + VP8Scan[it->i4_];
+  uint8_t* const top = it->i4_top_;
+  int i;
+
+  // Update the cache with 7 fresh samples
+  for (i = 0; i <= 3; ++i) {
+    top[-4 + i] = blk[i + 3 * BPS];   // store future top samples
+  }
+  if ((it->i4_ & 3) != 3) {  // if not on the right sub-blocks #3, #7, #11, #15
+    for (i = 0; i <= 2; ++i) {        // store future left samples
+      top[i] = blk[3 + (2 - i) * BPS];
+    }
+  } else {  // else replicate top-right samples, as says the specs.
+    for (i = 0; i <= 3; ++i) {
+      top[i] = top[i + 4];
+    }
+  }
+  // move pointers to next sub-block
+  ++it->i4_;
+  if (it->i4_ == 16) {    // we're done
+    return 0;
+  }
+
+  it->i4_top_ = it->i4_boundary_ + VP8TopLeftI4[it->i4_];
+  return 1;
+}
+
+//------------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/moz.build b/media/libwebp/src/enc/moz.build
new file mode 100644
index 0000000000..d10abfb032
--- /dev/null
+++ b/media/libwebp/src/enc/moz.build
@@ -0,0 +1,43 @@
+# -*- Mode: python; indent-tabs-mode: nil; tab-width: 40 -*-
+# vim: set filetype=python:
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+SOURCES += [
+    'alpha_enc.c',
+    'analysis_enc.c',
+    'backward_references_cost_enc.c',
+    'backward_references_enc.c',
+    'config_enc.c',
+    'cost_enc.c',
+    'filter_enc.c',
+    'frame_enc.c',
+    'histogram_enc.c',
+    'iterator_enc.c',
+    'near_lossless_enc.c',
+    'picture_csp_enc.c',
+    'picture_enc.c',
+    # 'picture_psnr_enc.c', # Unused file.
+    'picture_rescale_enc.c',
+    'picture_tools_enc.c',
+    'predictor_enc.c',
+    'quant_enc.c',
+    'syntax_enc.c',
+    'token_enc.c',
+    'tree_enc.c',
+    'vp8l_enc.c',
+    'webp_enc.c',
+]
+
+LOCAL_INCLUDES += [
+    '/media/libwebp',
+]
+
+# Add libFuzzer configuration directives
+include('/tools/fuzzing/libfuzzer-config.mozbuild')
+
+FINAL_LIBRARY = 'gkmedias'
+
+# We allow warnings for third-party code that can be updated from upstream.
+AllowCompilerWarnings()
diff --git a/media/libwebp/src/enc/near_lossless_enc.c b/media/libwebp/src/enc/near_lossless_enc.c
new file mode 100644
index 0000000000..5517a7e271
--- /dev/null
+++ b/media/libwebp/src/enc/near_lossless_enc.c
@@ -0,0 +1,151 @@
+// Copyright 2014 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.
+// -----------------------------------------------------------------------------
+//
+// Near-lossless image preprocessing adjusts pixel values to help
+// compressibility with a guarantee of maximum deviation between original and
+// resulting pixel values.
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+// Converted to C by Aleksander Kramarz (akramarz@google.com)
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include "src/dsp/lossless_common.h"
+#include "src/utils/utils.h"
+#include "src/enc/vp8li_enc.h"
+
+#if (WEBP_NEAR_LOSSLESS == 1)
+
+#define MIN_DIM_FOR_NEAR_LOSSLESS 64
+#define MAX_LIMIT_BITS             5
+
+// Quantizes the value up or down to a multiple of 1<<bits (or to 255),
+// choosing the closer one, resolving ties using bankers' rounding.
+static uint32_t FindClosestDiscretized(uint32_t a, int bits) {
+  const uint32_t mask = (1u << bits) - 1;
+  const uint32_t biased = a + (mask >> 1) + ((a >> bits) & 1);
+  assert(bits > 0);
+  if (biased > 0xff) return 0xff;
+  return biased & ~mask;
+}
+
+// Applies FindClosestDiscretized to all channels of pixel.
+static uint32_t ClosestDiscretizedArgb(uint32_t a, int bits) {
+  return
+      (FindClosestDiscretized(a >> 24, bits) << 24) |
+      (FindClosestDiscretized((a >> 16) & 0xff, bits) << 16) |
+      (FindClosestDiscretized((a >> 8) & 0xff, bits) << 8) |
+      (FindClosestDiscretized(a & 0xff, bits));
+}
+
+// Checks if distance between corresponding channel values of pixels a and b
+// is within the given limit.
+static int IsNear(uint32_t a, uint32_t b, int limit) {
+  int k;
+  for (k = 0; k < 4; ++k) {
+    const int delta =
+        (int)((a >> (k * 8)) & 0xff) - (int)((b >> (k * 8)) & 0xff);
+    if (delta >= limit || delta <= -limit) {
+      return 0;
+    }
+  }
+  return 1;
+}
+
+static int IsSmooth(const uint32_t* const prev_row,
+                    const uint32_t* const curr_row,
+                    const uint32_t* const next_row,
+                    int ix, int limit) {
+  // Check that all pixels in 4-connected neighborhood are smooth.
+  return (IsNear(curr_row[ix], curr_row[ix - 1], limit) &&
+          IsNear(curr_row[ix], curr_row[ix + 1], limit) &&
+          IsNear(curr_row[ix], prev_row[ix], limit) &&
+          IsNear(curr_row[ix], next_row[ix], limit));
+}
+
+// Adjusts pixel values of image with given maximum error.
+static void NearLossless(int xsize, int ysize, const uint32_t* argb_src,
+                         int stride, int limit_bits, uint32_t* copy_buffer,
+                         uint32_t* argb_dst) {
+  int x, y;
+  const int limit = 1 << limit_bits;
+  uint32_t* prev_row = copy_buffer;
+  uint32_t* curr_row = prev_row + xsize;
+  uint32_t* next_row = curr_row + xsize;
+  memcpy(curr_row, argb_src, xsize * sizeof(argb_src[0]));
+  memcpy(next_row, argb_src + stride, xsize * sizeof(argb_src[0]));
+
+  for (y = 0; y < ysize; ++y, argb_src += stride, argb_dst += xsize) {
+    if (y == 0 || y == ysize - 1) {
+      memcpy(argb_dst, argb_src, xsize * sizeof(argb_src[0]));
+    } else {
+      memcpy(next_row, argb_src + stride, xsize * sizeof(argb_src[0]));
+      argb_dst[0] = argb_src[0];
+      argb_dst[xsize - 1] = argb_src[xsize - 1];
+      for (x = 1; x < xsize - 1; ++x) {
+        if (IsSmooth(prev_row, curr_row, next_row, x, limit)) {
+          argb_dst[x] = curr_row[x];
+        } else {
+          argb_dst[x] = ClosestDiscretizedArgb(curr_row[x], limit_bits);
+        }
+      }
+    }
+    {
+      // Three-way swap.
+      uint32_t* const temp = prev_row;
+      prev_row = curr_row;
+      curr_row = next_row;
+      next_row = temp;
+    }
+  }
+}
+
+int VP8ApplyNearLossless(const WebPPicture* const picture, int quality,
+                         uint32_t* const argb_dst) {
+  int i;
+  const int xsize = picture->width;
+  const int ysize = picture->height;
+  const int stride = picture->argb_stride;
+  uint32_t* const copy_buffer =
+      (uint32_t*)WebPSafeMalloc(xsize * 3, sizeof(*copy_buffer));
+  const int limit_bits = VP8LNearLosslessBits(quality);
+  assert(argb_dst != NULL);
+  assert(limit_bits > 0);
+  assert(limit_bits <= MAX_LIMIT_BITS);
+  if (copy_buffer == NULL) {
+    return 0;
+  }
+  // For small icon images, don't attempt to apply near-lossless compression.
+  if ((xsize < MIN_DIM_FOR_NEAR_LOSSLESS &&
+       ysize < MIN_DIM_FOR_NEAR_LOSSLESS) ||
+      ysize < 3) {
+    for (i = 0; i < ysize; ++i) {
+      memcpy(argb_dst + i * xsize, picture->argb + i * picture->argb_stride,
+             xsize * sizeof(*argb_dst));
+    }
+    WebPSafeFree(copy_buffer);
+    return 1;
+  }
+
+  NearLossless(xsize, ysize, picture->argb, stride, limit_bits, copy_buffer,
+               argb_dst);
+  for (i = limit_bits - 1; i != 0; --i) {
+    NearLossless(xsize, ysize, argb_dst, xsize, i, copy_buffer, argb_dst);
+  }
+  WebPSafeFree(copy_buffer);
+  return 1;
+}
+#else  // (WEBP_NEAR_LOSSLESS == 1)
+
+// Define a stub to suppress compiler warnings.
+extern void VP8LNearLosslessStub(void);
+void VP8LNearLosslessStub(void) {}
+
+#endif  // (WEBP_NEAR_LOSSLESS == 1)
diff --git a/media/libwebp/src/enc/picture_csp_enc.c b/media/libwebp/src/enc/picture_csp_enc.c
new file mode 100644
index 0000000000..fabebcf202
--- /dev/null
+++ b/media/libwebp/src/enc/picture_csp_enc.c
@@ -0,0 +1,854 @@
+// Copyright 2014 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.
+// -----------------------------------------------------------------------------
+//
+// WebPPicture utils for colorspace conversion
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <math.h>
+
+#include "sharpyuv/sharpyuv.h"
+#include "sharpyuv/sharpyuv_csp.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/utils/random_utils.h"
+#include "src/utils/utils.h"
+#include "src/dsp/dsp.h"
+#include "src/dsp/lossless.h"
+#include "src/dsp/yuv.h"
+#include "src/dsp/cpu.h"
+
+#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
+#include <pthread.h>
+#endif
+
+// Uncomment to disable gamma-compression during RGB->U/V averaging
+#define USE_GAMMA_COMPRESSION
+
+// If defined, use table to compute x / alpha.
+#define USE_INVERSE_ALPHA_TABLE
+
+#ifdef WORDS_BIGENDIAN
+// uint32_t 0xff000000 is 0xff,00,00,00 in memory
+#define CHANNEL_OFFSET(i) (i)
+#else
+// uint32_t 0xff000000 is 0x00,00,00,ff in memory
+#define CHANNEL_OFFSET(i) (3-(i))
+#endif
+
+#define ALPHA_OFFSET CHANNEL_OFFSET(0)
+
+//------------------------------------------------------------------------------
+// Detection of non-trivial transparency
+
+// Returns true if alpha[] has non-0xff values.
+static int CheckNonOpaque(const uint8_t* alpha, int width, int height,
+                          int x_step, int y_step) {
+  if (alpha == NULL) return 0;
+  WebPInitAlphaProcessing();
+  if (x_step == 1) {
+    for (; height-- > 0; alpha += y_step) {
+      if (WebPHasAlpha8b(alpha, width)) return 1;
+    }
+  } else {
+    for (; height-- > 0; alpha += y_step) {
+      if (WebPHasAlpha32b(alpha, width)) return 1;
+    }
+  }
+  return 0;
+}
+
+// Checking for the presence of non-opaque alpha.
+int WebPPictureHasTransparency(const WebPPicture* picture) {
+  if (picture == NULL) return 0;
+  if (picture->use_argb) {
+    const int alpha_offset = ALPHA_OFFSET;
+    return CheckNonOpaque((const uint8_t*)picture->argb + alpha_offset,
+                          picture->width, picture->height,
+                          4, picture->argb_stride * sizeof(*picture->argb));
+  }
+  return CheckNonOpaque(picture->a, picture->width, picture->height,
+                        1, picture->a_stride);
+}
+
+//------------------------------------------------------------------------------
+// Code for gamma correction
+
+#if defined(USE_GAMMA_COMPRESSION)
+
+// Gamma correction compensates loss of resolution during chroma subsampling.
+#define GAMMA_FIX 12      // fixed-point precision for linear values
+#define GAMMA_TAB_FIX 7   // fixed-point fractional bits precision
+#define GAMMA_TAB_SIZE (1 << (GAMMA_FIX - GAMMA_TAB_FIX))
+static const double kGamma = 0.80;
+static const int kGammaScale = ((1 << GAMMA_FIX) - 1);
+static const int kGammaTabScale = (1 << GAMMA_TAB_FIX);
+static const int kGammaTabRounder = (1 << GAMMA_TAB_FIX >> 1);
+
+static int kLinearToGammaTab[GAMMA_TAB_SIZE + 1];
+static uint16_t kGammaToLinearTab[256];
+static volatile int kGammaTablesOk = 0;
+static void InitGammaTables(void);
+
+WEBP_DSP_INIT_FUNC(InitGammaTables) {
+  if (!kGammaTablesOk) {
+    int v;
+    const double scale = (double)(1 << GAMMA_TAB_FIX) / kGammaScale;
+    const double norm = 1. / 255.;
+    for (v = 0; v <= 255; ++v) {
+      kGammaToLinearTab[v] =
+          (uint16_t)(pow(norm * v, kGamma) * kGammaScale + .5);
+    }
+    for (v = 0; v <= GAMMA_TAB_SIZE; ++v) {
+      kLinearToGammaTab[v] = (int)(255. * pow(scale * v, 1. / kGamma) + .5);
+    }
+    kGammaTablesOk = 1;
+  }
+}
+
+static WEBP_INLINE uint32_t GammaToLinear(uint8_t v) {
+  return kGammaToLinearTab[v];
+}
+
+static WEBP_INLINE int Interpolate(int v) {
+  const int tab_pos = v >> (GAMMA_TAB_FIX + 2);    // integer part
+  const int x = v & ((kGammaTabScale << 2) - 1);  // fractional part
+  const int v0 = kLinearToGammaTab[tab_pos];
+  const int v1 = kLinearToGammaTab[tab_pos + 1];
+  const int y = v1 * x + v0 * ((kGammaTabScale << 2) - x);   // interpolate
+  assert(tab_pos + 1 < GAMMA_TAB_SIZE + 1);
+  return y;
+}
+
+// Convert a linear value 'v' to YUV_FIX+2 fixed-point precision
+// U/V value, suitable for RGBToU/V calls.
+static WEBP_INLINE int LinearToGamma(uint32_t base_value, int shift) {
+  const int y = Interpolate(base_value << shift);   // final uplifted value
+  return (y + kGammaTabRounder) >> GAMMA_TAB_FIX;    // descale
+}
+
+#else
+
+static void InitGammaTables(void) {}
+static WEBP_INLINE uint32_t GammaToLinear(uint8_t v) { return v; }
+static WEBP_INLINE int LinearToGamma(uint32_t base_value, int shift) {
+  return (int)(base_value << shift);
+}
+
+#endif    // USE_GAMMA_COMPRESSION
+
+//------------------------------------------------------------------------------
+// RGB -> YUV conversion
+
+static int RGBToY(int r, int g, int b, VP8Random* const rg) {
+  return (rg == NULL) ? VP8RGBToY(r, g, b, YUV_HALF)
+                      : VP8RGBToY(r, g, b, VP8RandomBits(rg, YUV_FIX));
+}
+
+static int RGBToU(int r, int g, int b, VP8Random* const rg) {
+  return (rg == NULL) ? VP8RGBToU(r, g, b, YUV_HALF << 2)
+                      : VP8RGBToU(r, g, b, VP8RandomBits(rg, YUV_FIX + 2));
+}
+
+static int RGBToV(int r, int g, int b, VP8Random* const rg) {
+  return (rg == NULL) ? VP8RGBToV(r, g, b, YUV_HALF << 2)
+                      : VP8RGBToV(r, g, b, VP8RandomBits(rg, YUV_FIX + 2));
+}
+
+//------------------------------------------------------------------------------
+// Sharp RGB->YUV conversion
+
+static const int kMinDimensionIterativeConversion = 4;
+
+//------------------------------------------------------------------------------
+// Main function
+
+extern void SharpYuvInit(VP8CPUInfo cpu_info_func);
+
+static void SafeInitSharpYuv(void) {
+#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
+  static pthread_mutex_t initsharpyuv_lock = PTHREAD_MUTEX_INITIALIZER;
+  if (pthread_mutex_lock(&initsharpyuv_lock)) return;
+#endif
+
+  SharpYuvInit(VP8GetCPUInfo);
+
+#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
+  (void)pthread_mutex_unlock(&initsharpyuv_lock);
+#endif
+}
+
+static int PreprocessARGB(const uint8_t* r_ptr,
+                          const uint8_t* g_ptr,
+                          const uint8_t* b_ptr,
+                          int step, int rgb_stride,
+                          WebPPicture* const picture) {
+  const int ok = SharpYuvConvert(
+      r_ptr, g_ptr, b_ptr, step, rgb_stride, /*rgb_bit_depth=*/8,
+      picture->y, picture->y_stride, picture->u, picture->uv_stride, picture->v,
+      picture->uv_stride, /*yuv_bit_depth=*/8, picture->width,
+      picture->height, SharpYuvGetConversionMatrix(kSharpYuvMatrixWebp));
+  if (!ok) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+// "Fast" regular RGB->YUV
+
+#define SUM4(ptr, step) LinearToGamma(                     \
+    GammaToLinear((ptr)[0]) +                              \
+    GammaToLinear((ptr)[(step)]) +                         \
+    GammaToLinear((ptr)[rgb_stride]) +                     \
+    GammaToLinear((ptr)[rgb_stride + (step)]), 0)          \
+
+#define SUM2(ptr) \
+    LinearToGamma(GammaToLinear((ptr)[0]) + GammaToLinear((ptr)[rgb_stride]), 1)
+
+#define SUM2ALPHA(ptr) ((ptr)[0] + (ptr)[rgb_stride])
+#define SUM4ALPHA(ptr) (SUM2ALPHA(ptr) + SUM2ALPHA((ptr) + 4))
+
+#if defined(USE_INVERSE_ALPHA_TABLE)
+
+static const int kAlphaFix = 19;
+// Following table is (1 << kAlphaFix) / a. The (v * kInvAlpha[a]) >> kAlphaFix
+// formula is then equal to v / a in most (99.6%) cases. Note that this table
+// and constant are adjusted very tightly to fit 32b arithmetic.
+// In particular, they use the fact that the operands for 'v / a' are actually
+// derived as v = (a0.p0 + a1.p1 + a2.p2 + a3.p3) and a = a0 + a1 + a2 + a3
+// with ai in [0..255] and pi in [0..1<<GAMMA_FIX). The constraint to avoid
+// overflow is: GAMMA_FIX + kAlphaFix <= 31.
+static const uint32_t kInvAlpha[4 * 0xff + 1] = {
+  0,  /* alpha = 0 */
+  524288, 262144, 174762, 131072, 104857, 87381, 74898, 65536,
+  58254, 52428, 47662, 43690, 40329, 37449, 34952, 32768,
+  30840, 29127, 27594, 26214, 24966, 23831, 22795, 21845,
+  20971, 20164, 19418, 18724, 18078, 17476, 16912, 16384,
+  15887, 15420, 14979, 14563, 14169, 13797, 13443, 13107,
+  12787, 12483, 12192, 11915, 11650, 11397, 11155, 10922,
+  10699, 10485, 10280, 10082, 9892, 9709, 9532, 9362,
+  9198, 9039, 8886, 8738, 8594, 8456, 8322, 8192,
+  8065, 7943, 7825, 7710, 7598, 7489, 7384, 7281,
+  7182, 7084, 6990, 6898, 6808, 6721, 6636, 6553,
+  6472, 6393, 6316, 6241, 6168, 6096, 6026, 5957,
+  5890, 5825, 5761, 5698, 5637, 5577, 5518, 5461,
+  5405, 5349, 5295, 5242, 5190, 5140, 5090, 5041,
+  4993, 4946, 4899, 4854, 4809, 4766, 4723, 4681,
+  4639, 4599, 4559, 4519, 4481, 4443, 4405, 4369,
+  4332, 4297, 4262, 4228, 4194, 4161, 4128, 4096,
+  4064, 4032, 4002, 3971, 3942, 3912, 3883, 3855,
+  3826, 3799, 3771, 3744, 3718, 3692, 3666, 3640,
+  3615, 3591, 3566, 3542, 3518, 3495, 3472, 3449,
+  3426, 3404, 3382, 3360, 3339, 3318, 3297, 3276,
+  3256, 3236, 3216, 3196, 3177, 3158, 3139, 3120,
+  3102, 3084, 3066, 3048, 3030, 3013, 2995, 2978,
+  2962, 2945, 2928, 2912, 2896, 2880, 2864, 2849,
+  2833, 2818, 2803, 2788, 2774, 2759, 2744, 2730,
+  2716, 2702, 2688, 2674, 2661, 2647, 2634, 2621,
+  2608, 2595, 2582, 2570, 2557, 2545, 2532, 2520,
+  2508, 2496, 2484, 2473, 2461, 2449, 2438, 2427,
+  2416, 2404, 2394, 2383, 2372, 2361, 2351, 2340,
+  2330, 2319, 2309, 2299, 2289, 2279, 2269, 2259,
+  2250, 2240, 2231, 2221, 2212, 2202, 2193, 2184,
+  2175, 2166, 2157, 2148, 2139, 2131, 2122, 2114,
+  2105, 2097, 2088, 2080, 2072, 2064, 2056, 2048,
+  2040, 2032, 2024, 2016, 2008, 2001, 1993, 1985,
+  1978, 1971, 1963, 1956, 1949, 1941, 1934, 1927,
+  1920, 1913, 1906, 1899, 1892, 1885, 1879, 1872,
+  1865, 1859, 1852, 1846, 1839, 1833, 1826, 1820,
+  1814, 1807, 1801, 1795, 1789, 1783, 1777, 1771,
+  1765, 1759, 1753, 1747, 1741, 1736, 1730, 1724,
+  1718, 1713, 1707, 1702, 1696, 1691, 1685, 1680,
+  1675, 1669, 1664, 1659, 1653, 1648, 1643, 1638,
+  1633, 1628, 1623, 1618, 1613, 1608, 1603, 1598,
+  1593, 1588, 1583, 1579, 1574, 1569, 1565, 1560,
+  1555, 1551, 1546, 1542, 1537, 1533, 1528, 1524,
+  1519, 1515, 1510, 1506, 1502, 1497, 1493, 1489,
+  1485, 1481, 1476, 1472, 1468, 1464, 1460, 1456,
+  1452, 1448, 1444, 1440, 1436, 1432, 1428, 1424,
+  1420, 1416, 1413, 1409, 1405, 1401, 1398, 1394,
+  1390, 1387, 1383, 1379, 1376, 1372, 1368, 1365,
+  1361, 1358, 1354, 1351, 1347, 1344, 1340, 1337,
+  1334, 1330, 1327, 1323, 1320, 1317, 1314, 1310,
+  1307, 1304, 1300, 1297, 1294, 1291, 1288, 1285,
+  1281, 1278, 1275, 1272, 1269, 1266, 1263, 1260,
+  1257, 1254, 1251, 1248, 1245, 1242, 1239, 1236,
+  1233, 1230, 1227, 1224, 1222, 1219, 1216, 1213,
+  1210, 1208, 1205, 1202, 1199, 1197, 1194, 1191,
+  1188, 1186, 1183, 1180, 1178, 1175, 1172, 1170,
+  1167, 1165, 1162, 1159, 1157, 1154, 1152, 1149,
+  1147, 1144, 1142, 1139, 1137, 1134, 1132, 1129,
+  1127, 1125, 1122, 1120, 1117, 1115, 1113, 1110,
+  1108, 1106, 1103, 1101, 1099, 1096, 1094, 1092,
+  1089, 1087, 1085, 1083, 1081, 1078, 1076, 1074,
+  1072, 1069, 1067, 1065, 1063, 1061, 1059, 1057,
+  1054, 1052, 1050, 1048, 1046, 1044, 1042, 1040,
+  1038, 1036, 1034, 1032, 1030, 1028, 1026, 1024,
+  1022, 1020, 1018, 1016, 1014, 1012, 1010, 1008,
+  1006, 1004, 1002, 1000, 998, 996, 994, 992,
+  991, 989, 987, 985, 983, 981, 979, 978,
+  976, 974, 972, 970, 969, 967, 965, 963,
+  961, 960, 958, 956, 954, 953, 951, 949,
+  948, 946, 944, 942, 941, 939, 937, 936,
+  934, 932, 931, 929, 927, 926, 924, 923,
+  921, 919, 918, 916, 914, 913, 911, 910,
+  908, 907, 905, 903, 902, 900, 899, 897,
+  896, 894, 893, 891, 890, 888, 887, 885,
+  884, 882, 881, 879, 878, 876, 875, 873,
+  872, 870, 869, 868, 866, 865, 863, 862,
+  860, 859, 858, 856, 855, 853, 852, 851,
+  849, 848, 846, 845, 844, 842, 841, 840,
+  838, 837, 836, 834, 833, 832, 830, 829,
+  828, 826, 825, 824, 823, 821, 820, 819,
+  817, 816, 815, 814, 812, 811, 810, 809,
+  807, 806, 805, 804, 802, 801, 800, 799,
+  798, 796, 795, 794, 793, 791, 790, 789,
+  788, 787, 786, 784, 783, 782, 781, 780,
+  779, 777, 776, 775, 774, 773, 772, 771,
+  769, 768, 767, 766, 765, 764, 763, 762,
+  760, 759, 758, 757, 756, 755, 754, 753,
+  752, 751, 750, 748, 747, 746, 745, 744,
+  743, 742, 741, 740, 739, 738, 737, 736,
+  735, 734, 733, 732, 731, 730, 729, 728,
+  727, 726, 725, 724, 723, 722, 721, 720,
+  719, 718, 717, 716, 715, 714, 713, 712,
+  711, 710, 709, 708, 707, 706, 705, 704,
+  703, 702, 701, 700, 699, 699, 698, 697,
+  696, 695, 694, 693, 692, 691, 690, 689,
+  688, 688, 687, 686, 685, 684, 683, 682,
+  681, 680, 680, 679, 678, 677, 676, 675,
+  674, 673, 673, 672, 671, 670, 669, 668,
+  667, 667, 666, 665, 664, 663, 662, 661,
+  661, 660, 659, 658, 657, 657, 656, 655,
+  654, 653, 652, 652, 651, 650, 649, 648,
+  648, 647, 646, 645, 644, 644, 643, 642,
+  641, 640, 640, 639, 638, 637, 637, 636,
+  635, 634, 633, 633, 632, 631, 630, 630,
+  629, 628, 627, 627, 626, 625, 624, 624,
+  623, 622, 621, 621, 620, 619, 618, 618,
+  617, 616, 616, 615, 614, 613, 613, 612,
+  611, 611, 610, 609, 608, 608, 607, 606,
+  606, 605, 604, 604, 603, 602, 601, 601,
+  600, 599, 599, 598, 597, 597, 596, 595,
+  595, 594, 593, 593, 592, 591, 591, 590,
+  589, 589, 588, 587, 587, 586, 585, 585,
+  584, 583, 583, 582, 581, 581, 580, 579,
+  579, 578, 578, 577, 576, 576, 575, 574,
+  574, 573, 572, 572, 571, 571, 570, 569,
+  569, 568, 568, 567, 566, 566, 565, 564,
+  564, 563, 563, 562, 561, 561, 560, 560,
+  559, 558, 558, 557, 557, 556, 555, 555,
+  554, 554, 553, 553, 552, 551, 551, 550,
+  550, 549, 548, 548, 547, 547, 546, 546,
+  545, 544, 544, 543, 543, 542, 542, 541,
+  541, 540, 539, 539, 538, 538, 537, 537,
+  536, 536, 535, 534, 534, 533, 533, 532,
+  532, 531, 531, 530, 530, 529, 529, 528,
+  527, 527, 526, 526, 525, 525, 524, 524,
+  523, 523, 522, 522, 521, 521, 520, 520,
+  519, 519, 518, 518, 517, 517, 516, 516,
+  515, 515, 514, 514
+};
+
+// Note that LinearToGamma() expects the values to be premultiplied by 4,
+// so we incorporate this factor 4 inside the DIVIDE_BY_ALPHA macro directly.
+#define DIVIDE_BY_ALPHA(sum, a)  (((sum) * kInvAlpha[(a)]) >> (kAlphaFix - 2))
+
+#else
+
+#define DIVIDE_BY_ALPHA(sum, a) (4 * (sum) / (a))
+
+#endif  // USE_INVERSE_ALPHA_TABLE
+
+static WEBP_INLINE int LinearToGammaWeighted(const uint8_t* src,
+                                             const uint8_t* a_ptr,
+                                             uint32_t total_a, int step,
+                                             int rgb_stride) {
+  const uint32_t sum =
+      a_ptr[0] * GammaToLinear(src[0]) +
+      a_ptr[step] * GammaToLinear(src[step]) +
+      a_ptr[rgb_stride] * GammaToLinear(src[rgb_stride]) +
+      a_ptr[rgb_stride + step] * GammaToLinear(src[rgb_stride + step]);
+  assert(total_a > 0 && total_a <= 4 * 0xff);
+#if defined(USE_INVERSE_ALPHA_TABLE)
+  assert((uint64_t)sum * kInvAlpha[total_a] < ((uint64_t)1 << 32));
+#endif
+  return LinearToGamma(DIVIDE_BY_ALPHA(sum, total_a), 0);
+}
+
+static WEBP_INLINE void ConvertRowToY(const uint8_t* const r_ptr,
+                                      const uint8_t* const g_ptr,
+                                      const uint8_t* const b_ptr,
+                                      int step,
+                                      uint8_t* const dst_y,
+                                      int width,
+                                      VP8Random* const rg) {
+  int i, j;
+  for (i = 0, j = 0; i < width; i += 1, j += step) {
+    dst_y[i] = RGBToY(r_ptr[j], g_ptr[j], b_ptr[j], rg);
+  }
+}
+
+static WEBP_INLINE void AccumulateRGBA(const uint8_t* const r_ptr,
+                                       const uint8_t* const g_ptr,
+                                       const uint8_t* const b_ptr,
+                                       const uint8_t* const a_ptr,
+                                       int rgb_stride,
+                                       uint16_t* dst, int width) {
+  int i, j;
+  // we loop over 2x2 blocks and produce one R/G/B/A value for each.
+  for (i = 0, j = 0; i < (width >> 1); i += 1, j += 2 * 4, dst += 4) {
+    const uint32_t a = SUM4ALPHA(a_ptr + j);
+    int r, g, b;
+    if (a == 4 * 0xff || a == 0) {
+      r = SUM4(r_ptr + j, 4);
+      g = SUM4(g_ptr + j, 4);
+      b = SUM4(b_ptr + j, 4);
+    } else {
+      r = LinearToGammaWeighted(r_ptr + j, a_ptr + j, a, 4, rgb_stride);
+      g = LinearToGammaWeighted(g_ptr + j, a_ptr + j, a, 4, rgb_stride);
+      b = LinearToGammaWeighted(b_ptr + j, a_ptr + j, a, 4, rgb_stride);
+    }
+    dst[0] = r;
+    dst[1] = g;
+    dst[2] = b;
+    dst[3] = a;
+  }
+  if (width & 1) {
+    const uint32_t a = 2u * SUM2ALPHA(a_ptr + j);
+    int r, g, b;
+    if (a == 4 * 0xff || a == 0) {
+      r = SUM2(r_ptr + j);
+      g = SUM2(g_ptr + j);
+      b = SUM2(b_ptr + j);
+    } else {
+      r = LinearToGammaWeighted(r_ptr + j, a_ptr + j, a, 0, rgb_stride);
+      g = LinearToGammaWeighted(g_ptr + j, a_ptr + j, a, 0, rgb_stride);
+      b = LinearToGammaWeighted(b_ptr + j, a_ptr + j, a, 0, rgb_stride);
+    }
+    dst[0] = r;
+    dst[1] = g;
+    dst[2] = b;
+    dst[3] = a;
+  }
+}
+
+static WEBP_INLINE void AccumulateRGB(const uint8_t* const r_ptr,
+                                      const uint8_t* const g_ptr,
+                                      const uint8_t* const b_ptr,
+                                      int step, int rgb_stride,
+                                      uint16_t* dst, int width) {
+  int i, j;
+  for (i = 0, j = 0; i < (width >> 1); i += 1, j += 2 * step, dst += 4) {
+    dst[0] = SUM4(r_ptr + j, step);
+    dst[1] = SUM4(g_ptr + j, step);
+    dst[2] = SUM4(b_ptr + j, step);
+    // MemorySanitizer may raise false positives with data that passes through
+    // RGBA32PackedToPlanar_16b_SSE41() due to incorrect modeling of shuffles.
+    // See https://crbug.com/webp/573.
+#ifdef WEBP_MSAN
+    dst[3] = 0;
+#endif
+  }
+  if (width & 1) {
+    dst[0] = SUM2(r_ptr + j);
+    dst[1] = SUM2(g_ptr + j);
+    dst[2] = SUM2(b_ptr + j);
+#ifdef WEBP_MSAN
+    dst[3] = 0;
+#endif
+  }
+}
+
+static WEBP_INLINE void ConvertRowsToUV(const uint16_t* rgb,
+                                        uint8_t* const dst_u,
+                                        uint8_t* const dst_v,
+                                        int width,
+                                        VP8Random* const rg) {
+  int i;
+  for (i = 0; i < width; i += 1, rgb += 4) {
+    const int r = rgb[0], g = rgb[1], b = rgb[2];
+    dst_u[i] = RGBToU(r, g, b, rg);
+    dst_v[i] = RGBToV(r, g, b, rg);
+  }
+}
+
+static int ImportYUVAFromRGBA(const uint8_t* r_ptr,
+                              const uint8_t* g_ptr,
+                              const uint8_t* b_ptr,
+                              const uint8_t* a_ptr,
+                              int step,         // bytes per pixel
+                              int rgb_stride,   // bytes per scanline
+                              float dithering,
+                              int use_iterative_conversion,
+                              WebPPicture* const picture) {
+  int y;
+  const int width = picture->width;
+  const int height = picture->height;
+  const int has_alpha = CheckNonOpaque(a_ptr, width, height, step, rgb_stride);
+  const int is_rgb = (r_ptr < b_ptr);  // otherwise it's bgr
+
+  picture->colorspace = has_alpha ? WEBP_YUV420A : WEBP_YUV420;
+  picture->use_argb = 0;
+
+  // disable smart conversion if source is too small (overkill).
+  if (width < kMinDimensionIterativeConversion ||
+      height < kMinDimensionIterativeConversion) {
+    use_iterative_conversion = 0;
+  }
+
+  if (!WebPPictureAllocYUVA(picture)) {
+    return 0;
+  }
+  if (has_alpha) {
+    assert(step == 4);
+#if defined(USE_GAMMA_COMPRESSION) && defined(USE_INVERSE_ALPHA_TABLE)
+    assert(kAlphaFix + GAMMA_FIX <= 31);
+#endif
+  }
+
+  if (use_iterative_conversion) {
+    SafeInitSharpYuv();
+    if (!PreprocessARGB(r_ptr, g_ptr, b_ptr, step, rgb_stride, picture)) {
+      return 0;
+    }
+    if (has_alpha) {
+      WebPExtractAlpha(a_ptr, rgb_stride, width, height,
+                       picture->a, picture->a_stride);
+    }
+  } else {
+    const int uv_width = (width + 1) >> 1;
+    int use_dsp = (step == 3);  // use special function in this case
+    // temporary storage for accumulated R/G/B values during conversion to U/V
+    uint16_t* const tmp_rgb =
+        (uint16_t*)WebPSafeMalloc(4 * uv_width, sizeof(*tmp_rgb));
+    uint8_t* dst_y = picture->y;
+    uint8_t* dst_u = picture->u;
+    uint8_t* dst_v = picture->v;
+    uint8_t* dst_a = picture->a;
+
+    VP8Random base_rg;
+    VP8Random* rg = NULL;
+    if (dithering > 0.) {
+      VP8InitRandom(&base_rg, dithering);
+      rg = &base_rg;
+      use_dsp = 0;   // can't use dsp in this case
+    }
+    WebPInitConvertARGBToYUV();
+    InitGammaTables();
+
+    if (tmp_rgb == NULL) return 0;  // malloc error
+
+    // Downsample Y/U/V planes, two rows at a time
+    for (y = 0; y < (height >> 1); ++y) {
+      int rows_have_alpha = has_alpha;
+      if (use_dsp) {
+        if (is_rgb) {
+          WebPConvertRGB24ToY(r_ptr, dst_y, width);
+          WebPConvertRGB24ToY(r_ptr + rgb_stride,
+                              dst_y + picture->y_stride, width);
+        } else {
+          WebPConvertBGR24ToY(b_ptr, dst_y, width);
+          WebPConvertBGR24ToY(b_ptr + rgb_stride,
+                              dst_y + picture->y_stride, width);
+        }
+      } else {
+        ConvertRowToY(r_ptr, g_ptr, b_ptr, step, dst_y, width, rg);
+        ConvertRowToY(r_ptr + rgb_stride,
+                      g_ptr + rgb_stride,
+                      b_ptr + rgb_stride, step,
+                      dst_y + picture->y_stride, width, rg);
+      }
+      dst_y += 2 * picture->y_stride;
+      if (has_alpha) {
+        rows_have_alpha &= !WebPExtractAlpha(a_ptr, rgb_stride, width, 2,
+                                             dst_a, picture->a_stride);
+        dst_a += 2 * picture->a_stride;
+      }
+      // Collect averaged R/G/B(/A)
+      if (!rows_have_alpha) {
+        AccumulateRGB(r_ptr, g_ptr, b_ptr, step, rgb_stride, tmp_rgb, width);
+      } else {
+        AccumulateRGBA(r_ptr, g_ptr, b_ptr, a_ptr, rgb_stride, tmp_rgb, width);
+      }
+      // Convert to U/V
+      if (rg == NULL) {
+        WebPConvertRGBA32ToUV(tmp_rgb, dst_u, dst_v, uv_width);
+      } else {
+        ConvertRowsToUV(tmp_rgb, dst_u, dst_v, uv_width, rg);
+      }
+      dst_u += picture->uv_stride;
+      dst_v += picture->uv_stride;
+      r_ptr += 2 * rgb_stride;
+      b_ptr += 2 * rgb_stride;
+      g_ptr += 2 * rgb_stride;
+      if (has_alpha) a_ptr += 2 * rgb_stride;
+    }
+    if (height & 1) {    // extra last row
+      int row_has_alpha = has_alpha;
+      if (use_dsp) {
+        if (r_ptr < b_ptr) {
+          WebPConvertRGB24ToY(r_ptr, dst_y, width);
+        } else {
+          WebPConvertBGR24ToY(b_ptr, dst_y, width);
+        }
+      } else {
+        ConvertRowToY(r_ptr, g_ptr, b_ptr, step, dst_y, width, rg);
+      }
+      if (row_has_alpha) {
+        row_has_alpha &= !WebPExtractAlpha(a_ptr, 0, width, 1, dst_a, 0);
+      }
+      // Collect averaged R/G/B(/A)
+      if (!row_has_alpha) {
+        // Collect averaged R/G/B
+        AccumulateRGB(r_ptr, g_ptr, b_ptr, step, /* rgb_stride = */ 0,
+                      tmp_rgb, width);
+      } else {
+        AccumulateRGBA(r_ptr, g_ptr, b_ptr, a_ptr, /* rgb_stride = */ 0,
+                       tmp_rgb, width);
+      }
+      if (rg == NULL) {
+        WebPConvertRGBA32ToUV(tmp_rgb, dst_u, dst_v, uv_width);
+      } else {
+        ConvertRowsToUV(tmp_rgb, dst_u, dst_v, uv_width, rg);
+      }
+    }
+    WebPSafeFree(tmp_rgb);
+  }
+  return 1;
+}
+
+#undef SUM4
+#undef SUM2
+#undef SUM4ALPHA
+#undef SUM2ALPHA
+
+//------------------------------------------------------------------------------
+// call for ARGB->YUVA conversion
+
+static int PictureARGBToYUVA(WebPPicture* picture, WebPEncCSP colorspace,
+                             float dithering, int use_iterative_conversion) {
+  if (picture == NULL) return 0;
+  if (picture->argb == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_NULL_PARAMETER);
+  } else if ((colorspace & WEBP_CSP_UV_MASK) != WEBP_YUV420) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_INVALID_CONFIGURATION);
+  } else {
+    const uint8_t* const argb = (const uint8_t*)picture->argb;
+    const uint8_t* const a = argb + CHANNEL_OFFSET(0);
+    const uint8_t* const r = argb + CHANNEL_OFFSET(1);
+    const uint8_t* const g = argb + CHANNEL_OFFSET(2);
+    const uint8_t* const b = argb + CHANNEL_OFFSET(3);
+
+    picture->colorspace = WEBP_YUV420;
+    return ImportYUVAFromRGBA(r, g, b, a, 4, 4 * picture->argb_stride,
+                              dithering, use_iterative_conversion, picture);
+  }
+}
+
+int WebPPictureARGBToYUVADithered(WebPPicture* picture, WebPEncCSP colorspace,
+                                  float dithering) {
+  return PictureARGBToYUVA(picture, colorspace, dithering, 0);
+}
+
+int WebPPictureARGBToYUVA(WebPPicture* picture, WebPEncCSP colorspace) {
+  return PictureARGBToYUVA(picture, colorspace, 0.f, 0);
+}
+
+int WebPPictureSharpARGBToYUVA(WebPPicture* picture) {
+  return PictureARGBToYUVA(picture, WEBP_YUV420, 0.f, 1);
+}
+// for backward compatibility
+int WebPPictureSmartARGBToYUVA(WebPPicture* picture) {
+  return WebPPictureSharpARGBToYUVA(picture);
+}
+
+//------------------------------------------------------------------------------
+// call for YUVA -> ARGB conversion
+
+int WebPPictureYUVAToARGB(WebPPicture* picture) {
+  if (picture == NULL) return 0;
+  if (picture->y == NULL || picture->u == NULL || picture->v == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_NULL_PARAMETER);
+  }
+  if ((picture->colorspace & WEBP_CSP_ALPHA_BIT) && picture->a == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_NULL_PARAMETER);
+  }
+  if ((picture->colorspace & WEBP_CSP_UV_MASK) != WEBP_YUV420) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_INVALID_CONFIGURATION);
+  }
+  // Allocate a new argb buffer (discarding the previous one).
+  if (!WebPPictureAllocARGB(picture)) return 0;
+  picture->use_argb = 1;
+
+  // Convert
+  {
+    int y;
+    const int width = picture->width;
+    const int height = picture->height;
+    const int argb_stride = 4 * picture->argb_stride;
+    uint8_t* dst = (uint8_t*)picture->argb;
+    const uint8_t* cur_u = picture->u, *cur_v = picture->v, *cur_y = picture->y;
+    WebPUpsampleLinePairFunc upsample =
+        WebPGetLinePairConverter(ALPHA_OFFSET > 0);
+
+    // First row, with replicated top samples.
+    upsample(cur_y, NULL, cur_u, cur_v, cur_u, cur_v, dst, NULL, width);
+    cur_y += picture->y_stride;
+    dst += argb_stride;
+    // Center rows.
+    for (y = 1; y + 1 < height; y += 2) {
+      const uint8_t* const top_u = cur_u;
+      const uint8_t* const top_v = cur_v;
+      cur_u += picture->uv_stride;
+      cur_v += picture->uv_stride;
+      upsample(cur_y, cur_y + picture->y_stride, top_u, top_v, cur_u, cur_v,
+               dst, dst + argb_stride, width);
+      cur_y += 2 * picture->y_stride;
+      dst += 2 * argb_stride;
+    }
+    // Last row (if needed), with replicated bottom samples.
+    if (height > 1 && !(height & 1)) {
+      upsample(cur_y, NULL, cur_u, cur_v, cur_u, cur_v, dst, NULL, width);
+    }
+    // Insert alpha values if needed, in replacement for the default 0xff ones.
+    if (picture->colorspace & WEBP_CSP_ALPHA_BIT) {
+      for (y = 0; y < height; ++y) {
+        uint32_t* const argb_dst = picture->argb + y * picture->argb_stride;
+        const uint8_t* const src = picture->a + y * picture->a_stride;
+        int x;
+        for (x = 0; x < width; ++x) {
+          argb_dst[x] = (argb_dst[x] & 0x00ffffffu) | ((uint32_t)src[x] << 24);
+        }
+      }
+    }
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// automatic import / conversion
+
+static int Import(WebPPicture* const picture,
+                  const uint8_t* rgb, int rgb_stride,
+                  int step, int swap_rb, int import_alpha) {
+  int y;
+  // swap_rb -> b,g,r,a , !swap_rb -> r,g,b,a
+  const uint8_t* r_ptr = rgb + (swap_rb ? 2 : 0);
+  const uint8_t* g_ptr = rgb + 1;
+  const uint8_t* b_ptr = rgb + (swap_rb ? 0 : 2);
+  const int width = picture->width;
+  const int height = picture->height;
+
+  if (abs(rgb_stride) < (import_alpha ? 4 : 3) * width) return 0;
+
+  if (!picture->use_argb) {
+    const uint8_t* a_ptr = import_alpha ? rgb + 3 : NULL;
+    return ImportYUVAFromRGBA(r_ptr, g_ptr, b_ptr, a_ptr, step, rgb_stride,
+                              0.f /* no dithering */, 0, picture);
+  }
+  if (!WebPPictureAlloc(picture)) return 0;
+
+  VP8LDspInit();
+  WebPInitAlphaProcessing();
+
+  if (import_alpha) {
+    // dst[] byte order is {a,r,g,b} for big-endian, {b,g,r,a} for little endian
+    uint32_t* dst = picture->argb;
+    const int do_copy = (ALPHA_OFFSET == 3) && swap_rb;
+    assert(step == 4);
+    if (do_copy) {
+      for (y = 0; y < height; ++y) {
+        memcpy(dst, rgb, width * 4);
+        rgb += rgb_stride;
+        dst += picture->argb_stride;
+      }
+    } else {
+      for (y = 0; y < height; ++y) {
+#ifdef WORDS_BIGENDIAN
+        // BGRA or RGBA input order.
+        const uint8_t* a_ptr = rgb + 3;
+        WebPPackARGB(a_ptr, r_ptr, g_ptr, b_ptr, width, dst);
+        r_ptr += rgb_stride;
+        g_ptr += rgb_stride;
+        b_ptr += rgb_stride;
+#else
+        // RGBA input order. Need to swap R and B.
+        VP8LConvertBGRAToRGBA((const uint32_t*)rgb, width, (uint8_t*)dst);
+#endif
+        rgb += rgb_stride;
+        dst += picture->argb_stride;
+      }
+    }
+  } else {
+    uint32_t* dst = picture->argb;
+    assert(step >= 3);
+    for (y = 0; y < height; ++y) {
+      WebPPackRGB(r_ptr, g_ptr, b_ptr, width, step, dst);
+      r_ptr += rgb_stride;
+      g_ptr += rgb_stride;
+      b_ptr += rgb_stride;
+      dst += picture->argb_stride;
+    }
+  }
+  return 1;
+}
+
+// Public API
+
+#if !defined(WEBP_REDUCE_CSP)
+
+int WebPPictureImportBGR(WebPPicture* picture,
+                         const uint8_t* bgr, int bgr_stride) {
+  return (picture != NULL && bgr != NULL)
+             ? Import(picture, bgr, bgr_stride, 3, 1, 0)
+             : 0;
+}
+
+int WebPPictureImportBGRA(WebPPicture* picture,
+                          const uint8_t* bgra, int bgra_stride) {
+  return (picture != NULL && bgra != NULL)
+             ? Import(picture, bgra, bgra_stride, 4, 1, 1)
+             : 0;
+}
+
+
+int WebPPictureImportBGRX(WebPPicture* picture,
+                          const uint8_t* bgrx, int bgrx_stride) {
+  return (picture != NULL && bgrx != NULL)
+             ? Import(picture, bgrx, bgrx_stride, 4, 1, 0)
+             : 0;
+}
+
+#endif   // WEBP_REDUCE_CSP
+
+int WebPPictureImportRGB(WebPPicture* picture,
+                         const uint8_t* rgb, int rgb_stride) {
+  return (picture != NULL && rgb != NULL)
+             ? Import(picture, rgb, rgb_stride, 3, 0, 0)
+             : 0;
+}
+
+int WebPPictureImportRGBA(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return (picture != NULL && rgba != NULL)
+             ? Import(picture, rgba, rgba_stride, 4, 0, 1)
+             : 0;
+}
+
+int WebPPictureImportRGBX(WebPPicture* picture,
+                          const uint8_t* rgbx, int rgbx_stride) {
+  return (picture != NULL && rgbx != NULL)
+             ? Import(picture, rgbx, rgbx_stride, 4, 0, 0)
+             : 0;
+}
+
+//------------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/picture_enc.c b/media/libwebp/src/enc/picture_enc.c
new file mode 100644
index 0000000000..3af6383d38
--- /dev/null
+++ b/media/libwebp/src/enc/picture_enc.c
@@ -0,0 +1,304 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// WebPPicture class basis
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include "src/enc/vp8i_enc.h"
+#include "src/dsp/dsp.h"
+#include "src/utils/utils.h"
+
+//------------------------------------------------------------------------------
+// WebPPicture
+//------------------------------------------------------------------------------
+
+static int DummyWriter(const uint8_t* data, size_t data_size,
+                       const WebPPicture* const picture) {
+  // The following are to prevent 'unused variable' error message.
+  (void)data;
+  (void)data_size;
+  (void)picture;
+  return 1;
+}
+
+int WebPPictureInitInternal(WebPPicture* picture, int version) {
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_ENCODER_ABI_VERSION)) {
+    return 0;   // caller/system version mismatch!
+  }
+  if (picture != NULL) {
+    memset(picture, 0, sizeof(*picture));
+    picture->writer = DummyWriter;
+    WebPEncodingSetError(picture, VP8_ENC_OK);
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+int WebPValidatePicture(const WebPPicture* const picture) {
+  if (picture == NULL) return 0;
+  if (picture->width <= 0 || picture->height <= 0) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_BAD_DIMENSION);
+  }
+  if (picture->width <= 0 || picture->width / 4 > INT_MAX / 4 ||
+      picture->height <= 0 || picture->height / 4 > INT_MAX / 4) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_BAD_DIMENSION);
+  }
+  if (picture->colorspace != WEBP_YUV420 &&
+      picture->colorspace != WEBP_YUV420A) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_INVALID_CONFIGURATION);
+  }
+  return 1;
+}
+
+static void WebPPictureResetBufferARGB(WebPPicture* const picture) {
+  picture->memory_argb_ = NULL;
+  picture->argb = NULL;
+  picture->argb_stride = 0;
+}
+
+static void WebPPictureResetBufferYUVA(WebPPicture* const picture) {
+  picture->memory_ = NULL;
+  picture->y = picture->u = picture->v = picture->a = NULL;
+  picture->y_stride = picture->uv_stride = 0;
+  picture->a_stride = 0;
+}
+
+void WebPPictureResetBuffers(WebPPicture* const picture) {
+  WebPPictureResetBufferARGB(picture);
+  WebPPictureResetBufferYUVA(picture);
+}
+
+int WebPPictureAllocARGB(WebPPicture* const picture) {
+  void* memory;
+  const int width = picture->width;
+  const int height = picture->height;
+  const uint64_t argb_size = (uint64_t)width * height;
+
+  if (!WebPValidatePicture(picture)) return 0;
+
+  WebPSafeFree(picture->memory_argb_);
+  WebPPictureResetBufferARGB(picture);
+
+  // allocate a new buffer.
+  memory = WebPSafeMalloc(argb_size + WEBP_ALIGN_CST, sizeof(*picture->argb));
+  if (memory == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  picture->memory_argb_ = memory;
+  picture->argb = (uint32_t*)WEBP_ALIGN(memory);
+  picture->argb_stride = width;
+  return 1;
+}
+
+int WebPPictureAllocYUVA(WebPPicture* const picture) {
+  const int has_alpha = (int)picture->colorspace & WEBP_CSP_ALPHA_BIT;
+  const int width = picture->width;
+  const int height = picture->height;
+  const int y_stride = width;
+  const int uv_width = (int)(((int64_t)width + 1) >> 1);
+  const int uv_height = (int)(((int64_t)height + 1) >> 1);
+  const int uv_stride = uv_width;
+  int a_width, a_stride;
+  uint64_t y_size, uv_size, a_size, total_size;
+  uint8_t* mem;
+
+  if (!WebPValidatePicture(picture)) return 0;
+
+  WebPSafeFree(picture->memory_);
+  WebPPictureResetBufferYUVA(picture);
+
+  // alpha
+  a_width = has_alpha ? width : 0;
+  a_stride = a_width;
+  y_size = (uint64_t)y_stride * height;
+  uv_size = (uint64_t)uv_stride * uv_height;
+  a_size =  (uint64_t)a_stride * height;
+
+  total_size = y_size + a_size + 2 * uv_size;
+
+  // Security and validation checks
+  if (width <= 0 || height <= 0 ||           // luma/alpha param error
+      uv_width <= 0 || uv_height <= 0) {     // u/v param error
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_BAD_DIMENSION);
+  }
+  // allocate a new buffer.
+  mem = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*mem));
+  if (mem == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+
+  // From now on, we're in the clear, we can no longer fail...
+  picture->memory_ = (void*)mem;
+  picture->y_stride  = y_stride;
+  picture->uv_stride = uv_stride;
+  picture->a_stride  = a_stride;
+
+  // TODO(skal): we could align the y/u/v planes and adjust stride.
+  picture->y = mem;
+  mem += y_size;
+
+  picture->u = mem;
+  mem += uv_size;
+  picture->v = mem;
+  mem += uv_size;
+
+  if (a_size > 0) {
+    picture->a = mem;
+    mem += a_size;
+  }
+  (void)mem;  // makes the static analyzer happy
+  return 1;
+}
+
+int WebPPictureAlloc(WebPPicture* picture) {
+  if (picture != NULL) {
+    WebPPictureFree(picture);   // erase previous buffer
+
+    if (!picture->use_argb) {
+      return WebPPictureAllocYUVA(picture);
+    } else {
+      return WebPPictureAllocARGB(picture);
+    }
+  }
+  return 1;
+}
+
+void WebPPictureFree(WebPPicture* picture) {
+  if (picture != NULL) {
+    WebPSafeFree(picture->memory_);
+    WebPSafeFree(picture->memory_argb_);
+    WebPPictureResetBuffers(picture);
+  }
+}
+
+//------------------------------------------------------------------------------
+// WebPMemoryWriter: Write-to-memory
+
+void WebPMemoryWriterInit(WebPMemoryWriter* writer) {
+  writer->mem = NULL;
+  writer->size = 0;
+  writer->max_size = 0;
+}
+
+int WebPMemoryWrite(const uint8_t* data, size_t data_size,
+                    const WebPPicture* picture) {
+  WebPMemoryWriter* const w = (WebPMemoryWriter*)picture->custom_ptr;
+  uint64_t next_size;
+  if (w == NULL) {
+    return 1;
+  }
+  next_size = (uint64_t)w->size + data_size;
+  if (next_size > w->max_size) {
+    uint8_t* new_mem;
+    uint64_t next_max_size = 2ULL * w->max_size;
+    if (next_max_size < next_size) next_max_size = next_size;
+    if (next_max_size < 8192ULL) next_max_size = 8192ULL;
+    new_mem = (uint8_t*)WebPSafeMalloc(next_max_size, 1);
+    if (new_mem == NULL) {
+      return 0;
+    }
+    if (w->size > 0) {
+      memcpy(new_mem, w->mem, w->size);
+    }
+    WebPSafeFree(w->mem);
+    w->mem = new_mem;
+    // down-cast is ok, thanks to WebPSafeMalloc
+    w->max_size = (size_t)next_max_size;
+  }
+  if (data_size > 0) {
+    memcpy(w->mem + w->size, data, data_size);
+    w->size += data_size;
+  }
+  return 1;
+}
+
+void WebPMemoryWriterClear(WebPMemoryWriter* writer) {
+  if (writer != NULL) {
+    WebPSafeFree(writer->mem);
+    writer->mem = NULL;
+    writer->size = 0;
+    writer->max_size = 0;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Simplest high-level calls:
+
+typedef int (*Importer)(WebPPicture* const, const uint8_t* const, int);
+
+static size_t Encode(const uint8_t* rgba, int width, int height, int stride,
+                     Importer import, float quality_factor, int lossless,
+                     uint8_t** output) {
+  WebPPicture pic;
+  WebPConfig config;
+  WebPMemoryWriter wrt;
+  int ok;
+
+  if (output == NULL) return 0;
+
+  if (!WebPConfigPreset(&config, WEBP_PRESET_DEFAULT, quality_factor) ||
+      !WebPPictureInit(&pic)) {
+    return 0;  // shouldn't happen, except if system installation is broken
+  }
+
+  config.lossless = !!lossless;
+  pic.use_argb = !!lossless;
+  pic.width = width;
+  pic.height = height;
+  pic.writer = WebPMemoryWrite;
+  pic.custom_ptr = &wrt;
+  WebPMemoryWriterInit(&wrt);
+
+  ok = import(&pic, rgba, stride) && WebPEncode(&config, &pic);
+  WebPPictureFree(&pic);
+  if (!ok) {
+    WebPMemoryWriterClear(&wrt);
+    *output = NULL;
+    return 0;
+  }
+  *output = wrt.mem;
+  return wrt.size;
+}
+
+#define ENCODE_FUNC(NAME, IMPORTER)                                     \
+size_t NAME(const uint8_t* in, int w, int h, int bps, float q,          \
+            uint8_t** out) {                                            \
+  return Encode(in, w, h, bps, IMPORTER, q, 0, out);                    \
+}
+
+ENCODE_FUNC(WebPEncodeRGB, WebPPictureImportRGB)
+ENCODE_FUNC(WebPEncodeRGBA, WebPPictureImportRGBA)
+#if !defined(WEBP_REDUCE_CSP)
+ENCODE_FUNC(WebPEncodeBGR, WebPPictureImportBGR)
+ENCODE_FUNC(WebPEncodeBGRA, WebPPictureImportBGRA)
+#endif  // WEBP_REDUCE_CSP
+
+#undef ENCODE_FUNC
+
+#define LOSSLESS_DEFAULT_QUALITY 70.
+#define LOSSLESS_ENCODE_FUNC(NAME, IMPORTER)                                 \
+size_t NAME(const uint8_t* in, int w, int h, int bps, uint8_t** out) {       \
+  return Encode(in, w, h, bps, IMPORTER, LOSSLESS_DEFAULT_QUALITY, 1, out);  \
+}
+
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGB, WebPPictureImportRGB)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGBA, WebPPictureImportRGBA)
+#if !defined(WEBP_REDUCE_CSP)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGR, WebPPictureImportBGR)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGRA, WebPPictureImportBGRA)
+#endif  // WEBP_REDUCE_CSP
+
+#undef LOSSLESS_ENCODE_FUNC
+
+//------------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/picture_psnr_enc.c b/media/libwebp/src/enc/picture_psnr_enc.c
new file mode 100644
index 0000000000..1a2f0bef3e
--- /dev/null
+++ b/media/libwebp/src/enc/picture_psnr_enc.c
@@ -0,0 +1,258 @@
+// Copyright 2014 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.
+// -----------------------------------------------------------------------------
+//
+// WebPPicture tools for measuring distortion
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "src/webp/encode.h"
+
+#if !(defined(WEBP_DISABLE_STATS) || defined(WEBP_REDUCE_SIZE))
+
+#include <math.h>
+#include <stdlib.h>
+
+#include "src/dsp/dsp.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/utils/utils.h"
+
+typedef double (*AccumulateFunc)(const uint8_t* src, int src_stride,
+                                 const uint8_t* ref, int ref_stride,
+                                 int w, int h);
+
+//------------------------------------------------------------------------------
+// local-min distortion
+//
+// For every pixel in the *reference* picture, we search for the local best
+// match in the compressed image. This is not a symmetrical measure.
+
+#define RADIUS 2  // search radius. Shouldn't be too large.
+
+static double AccumulateLSIM(const uint8_t* src, int src_stride,
+                             const uint8_t* ref, int ref_stride,
+                             int w, int h) {
+  int x, y;
+  double total_sse = 0.;
+  for (y = 0; y < h; ++y) {
+    const int y_0 = (y - RADIUS < 0) ? 0 : y - RADIUS;
+    const int y_1 = (y + RADIUS + 1 >= h) ? h : y + RADIUS + 1;
+    for (x = 0; x < w; ++x) {
+      const int x_0 = (x - RADIUS < 0) ? 0 : x - RADIUS;
+      const int x_1 = (x + RADIUS + 1 >= w) ? w : x + RADIUS + 1;
+      double best_sse = 255. * 255.;
+      const double value = (double)ref[y * ref_stride + x];
+      int i, j;
+      for (j = y_0; j < y_1; ++j) {
+        const uint8_t* const s = src + j * src_stride;
+        for (i = x_0; i < x_1; ++i) {
+          const double diff = s[i] - value;
+          const double sse = diff * diff;
+          if (sse < best_sse) best_sse = sse;
+        }
+      }
+      total_sse += best_sse;
+    }
+  }
+  return total_sse;
+}
+#undef RADIUS
+
+static double AccumulateSSE(const uint8_t* src, int src_stride,
+                            const uint8_t* ref, int ref_stride,
+                            int w, int h) {
+  int y;
+  double total_sse = 0.;
+  for (y = 0; y < h; ++y) {
+    total_sse += VP8AccumulateSSE(src, ref, w);
+    src += src_stride;
+    ref += ref_stride;
+  }
+  return total_sse;
+}
+
+//------------------------------------------------------------------------------
+
+static double AccumulateSSIM(const uint8_t* src, int src_stride,
+                             const uint8_t* ref, int ref_stride,
+                             int w, int h) {
+  const int w0 = (w < VP8_SSIM_KERNEL) ? w : VP8_SSIM_KERNEL;
+  const int w1 = w - VP8_SSIM_KERNEL - 1;
+  const int h0 = (h < VP8_SSIM_KERNEL) ? h : VP8_SSIM_KERNEL;
+  const int h1 = h - VP8_SSIM_KERNEL - 1;
+  int x, y;
+  double sum = 0.;
+  for (y = 0; y < h0; ++y) {
+    for (x = 0; x < w; ++x) {
+      sum += VP8SSIMGetClipped(src, src_stride, ref, ref_stride, x, y, w, h);
+    }
+  }
+  for (; y < h1; ++y) {
+    for (x = 0; x < w0; ++x) {
+      sum += VP8SSIMGetClipped(src, src_stride, ref, ref_stride, x, y, w, h);
+    }
+    for (; x < w1; ++x) {
+      const int off1 = x - VP8_SSIM_KERNEL + (y - VP8_SSIM_KERNEL) * src_stride;
+      const int off2 = x - VP8_SSIM_KERNEL + (y - VP8_SSIM_KERNEL) * ref_stride;
+      sum += VP8SSIMGet(src + off1, src_stride, ref + off2, ref_stride);
+    }
+    for (; x < w; ++x) {
+      sum += VP8SSIMGetClipped(src, src_stride, ref, ref_stride, x, y, w, h);
+    }
+  }
+  for (; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      sum += VP8SSIMGetClipped(src, src_stride, ref, ref_stride, x, y, w, h);
+    }
+  }
+  return sum;
+}
+
+//------------------------------------------------------------------------------
+// Distortion
+
+// Max value returned in case of exact similarity.
+static const double kMinDistortion_dB = 99.;
+
+static double GetPSNR(double v, double size) {
+  return (v > 0. && size > 0.) ? -4.3429448 * log(v / (size * 255 * 255.))
+                               : kMinDistortion_dB;
+}
+
+static double GetLogSSIM(double v, double size) {
+  v = (size > 0.) ? v / size : 1.;
+  return (v < 1.) ? -10.0 * log10(1. - v) : kMinDistortion_dB;
+}
+
+int WebPPlaneDistortion(const uint8_t* src, size_t src_stride,
+                        const uint8_t* ref, size_t ref_stride,
+                        int width, int height, size_t x_step,
+                        int type, float* distortion, float* result) {
+  uint8_t* allocated = NULL;
+  const AccumulateFunc metric = (type == 0) ? AccumulateSSE :
+                                (type == 1) ? AccumulateSSIM :
+                                              AccumulateLSIM;
+  if (src == NULL || ref == NULL ||
+      src_stride < x_step * width || ref_stride < x_step * width ||
+      result == NULL || distortion == NULL) {
+    return 0;
+  }
+
+  VP8SSIMDspInit();
+  if (x_step != 1) {   // extract a packed plane if needed
+    int x, y;
+    uint8_t* tmp1;
+    uint8_t* tmp2;
+    allocated =
+        (uint8_t*)WebPSafeMalloc(2ULL * width * height, sizeof(*allocated));
+    if (allocated == NULL) return 0;
+    tmp1 = allocated;
+    tmp2 = tmp1 + (size_t)width * height;
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        tmp1[x + y * width] = src[x * x_step + y * src_stride];
+        tmp2[x + y * width] = ref[x * x_step + y * ref_stride];
+      }
+    }
+    src = tmp1;
+    ref = tmp2;
+  }
+  *distortion = (float)metric(src, width, ref, width, width, height);
+  WebPSafeFree(allocated);
+
+  *result = (type == 1) ? (float)GetLogSSIM(*distortion, (double)width * height)
+                        : (float)GetPSNR(*distortion, (double)width * height);
+  return 1;
+}
+
+#ifdef WORDS_BIGENDIAN
+#define BLUE_OFFSET 3   // uint32_t 0x000000ff is 0x00,00,00,ff in memory
+#else
+#define BLUE_OFFSET 0   // uint32_t 0x000000ff is 0xff,00,00,00 in memory
+#endif
+
+int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
+                          int type, float results[5]) {
+  int w, h, c;
+  int ok = 0;
+  WebPPicture p0, p1;
+  double total_size = 0., total_distortion = 0.;
+  if (src == NULL || ref == NULL ||
+      src->width != ref->width || src->height != ref->height ||
+      results == NULL) {
+    return 0;
+  }
+
+  VP8SSIMDspInit();
+  if (!WebPPictureInit(&p0) || !WebPPictureInit(&p1)) return 0;
+  w = src->width;
+  h = src->height;
+  if (!WebPPictureView(src, 0, 0, w, h, &p0)) goto Error;
+  if (!WebPPictureView(ref, 0, 0, w, h, &p1)) goto Error;
+
+  // We always measure distortion in ARGB space.
+  if (p0.use_argb == 0 && !WebPPictureYUVAToARGB(&p0)) goto Error;
+  if (p1.use_argb == 0 && !WebPPictureYUVAToARGB(&p1)) goto Error;
+  for (c = 0; c < 4; ++c) {
+    float distortion;
+    const size_t stride0 = 4 * (size_t)p0.argb_stride;
+    const size_t stride1 = 4 * (size_t)p1.argb_stride;
+    // results are reported as BGRA
+    const int offset = c ^ BLUE_OFFSET;
+    if (!WebPPlaneDistortion((const uint8_t*)p0.argb + offset, stride0,
+                             (const uint8_t*)p1.argb + offset, stride1,
+                             w, h, 4, type, &distortion, results + c)) {
+      goto Error;
+    }
+    total_distortion += distortion;
+    total_size += w * h;
+  }
+
+  results[4] = (type == 1) ? (float)GetLogSSIM(total_distortion, total_size)
+                           : (float)GetPSNR(total_distortion, total_size);
+  ok = 1;
+
+ Error:
+  WebPPictureFree(&p0);
+  WebPPictureFree(&p1);
+  return ok;
+}
+
+#undef BLUE_OFFSET
+
+#else  // defined(WEBP_DISABLE_STATS)
+int WebPPlaneDistortion(const uint8_t* src, size_t src_stride,
+                        const uint8_t* ref, size_t ref_stride,
+                        int width, int height, size_t x_step,
+                        int type, float* distortion, float* result) {
+  (void)src;
+  (void)src_stride;
+  (void)ref;
+  (void)ref_stride;
+  (void)width;
+  (void)height;
+  (void)x_step;
+  (void)type;
+  if (distortion == NULL || result == NULL) return 0;
+  *distortion = 0.f;
+  *result = 0.f;
+  return 1;
+}
+
+int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
+                          int type, float results[5]) {
+  int i;
+  (void)src;
+  (void)ref;
+  (void)type;
+  if (results == NULL) return 0;
+  for (i = 0; i < 5; ++i) results[i] = 0.f;
+  return 1;
+}
+
+#endif  // !defined(WEBP_DISABLE_STATS)
diff --git a/media/libwebp/src/enc/picture_rescale_enc.c b/media/libwebp/src/enc/picture_rescale_enc.c
new file mode 100644
index 0000000000..839f91cacc
--- /dev/null
+++ b/media/libwebp/src/enc/picture_rescale_enc.c
@@ -0,0 +1,300 @@
+// Copyright 2014 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.
+// -----------------------------------------------------------------------------
+//
+// WebPPicture tools: copy, crop, rescaling and view.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "src/webp/encode.h"
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include "src/enc/vp8i_enc.h"
+
+#if !defined(WEBP_REDUCE_SIZE)
+#include "src/utils/rescaler_utils.h"
+#include "src/utils/utils.h"
+#endif  // !defined(WEBP_REDUCE_SIZE)
+
+#define HALVE(x) (((x) + 1) >> 1)
+
+// Grab the 'specs' (writer, *opaque, width, height...) from 'src' and copy them
+// into 'dst'. Mark 'dst' as not owning any memory.
+static void PictureGrabSpecs(const WebPPicture* const src,
+                             WebPPicture* const dst) {
+  assert(src != NULL && dst != NULL);
+  *dst = *src;
+  WebPPictureResetBuffers(dst);
+}
+
+//------------------------------------------------------------------------------
+
+// Adjust top-left corner to chroma sample position.
+static void SnapTopLeftPosition(const WebPPicture* const pic,
+                                int* const left, int* const top) {
+  if (!pic->use_argb) {
+    *left &= ~1;
+    *top &= ~1;
+  }
+}
+
+// Adjust top-left corner and verify that the sub-rectangle is valid.
+static int AdjustAndCheckRectangle(const WebPPicture* const pic,
+                                   int* const left, int* const top,
+                                   int width, int height) {
+  SnapTopLeftPosition(pic, left, top);
+  if ((*left) < 0 || (*top) < 0) return 0;
+  if (width <= 0 || height <= 0) return 0;
+  if ((*left) + width > pic->width) return 0;
+  if ((*top) + height > pic->height) return 0;
+  return 1;
+}
+
+#if !defined(WEBP_REDUCE_SIZE)
+int WebPPictureCopy(const WebPPicture* src, WebPPicture* dst) {
+  if (src == NULL || dst == NULL) return 0;
+  if (src == dst) return 1;
+
+  PictureGrabSpecs(src, dst);
+  if (!WebPPictureAlloc(dst)) return 0;
+
+  if (!src->use_argb) {
+    WebPCopyPlane(src->y, src->y_stride,
+                  dst->y, dst->y_stride, dst->width, dst->height);
+    WebPCopyPlane(src->u, src->uv_stride, dst->u, dst->uv_stride,
+                  HALVE(dst->width), HALVE(dst->height));
+    WebPCopyPlane(src->v, src->uv_stride, dst->v, dst->uv_stride,
+                  HALVE(dst->width), HALVE(dst->height));
+    if (dst->a != NULL)  {
+      WebPCopyPlane(src->a, src->a_stride,
+                    dst->a, dst->a_stride, dst->width, dst->height);
+    }
+  } else {
+    WebPCopyPlane((const uint8_t*)src->argb, 4 * src->argb_stride,
+                  (uint8_t*)dst->argb, 4 * dst->argb_stride,
+                  4 * dst->width, dst->height);
+  }
+  return 1;
+}
+#endif  // !defined(WEBP_REDUCE_SIZE)
+
+int WebPPictureIsView(const WebPPicture* picture) {
+  if (picture == NULL) return 0;
+  if (picture->use_argb) {
+    return (picture->memory_argb_ == NULL);
+  }
+  return (picture->memory_ == NULL);
+}
+
+int WebPPictureView(const WebPPicture* src,
+                    int left, int top, int width, int height,
+                    WebPPicture* dst) {
+  if (src == NULL || dst == NULL) return 0;
+
+  // verify rectangle position.
+  if (!AdjustAndCheckRectangle(src, &left, &top, width, height)) return 0;
+
+  if (src != dst) {  // beware of aliasing! We don't want to leak 'memory_'.
+    PictureGrabSpecs(src, dst);
+  }
+  dst->width = width;
+  dst->height = height;
+  if (!src->use_argb) {
+    dst->y = src->y + top * src->y_stride + left;
+    dst->u = src->u + (top >> 1) * src->uv_stride + (left >> 1);
+    dst->v = src->v + (top >> 1) * src->uv_stride + (left >> 1);
+    dst->y_stride = src->y_stride;
+    dst->uv_stride = src->uv_stride;
+    if (src->a != NULL) {
+      dst->a = src->a + top * src->a_stride + left;
+      dst->a_stride = src->a_stride;
+    }
+  } else {
+    dst->argb = src->argb + top * src->argb_stride + left;
+    dst->argb_stride = src->argb_stride;
+  }
+  return 1;
+}
+
+#if !defined(WEBP_REDUCE_SIZE)
+//------------------------------------------------------------------------------
+// Picture cropping
+
+int WebPPictureCrop(WebPPicture* pic,
+                    int left, int top, int width, int height) {
+  WebPPicture tmp;
+
+  if (pic == NULL) return 0;
+  if (!AdjustAndCheckRectangle(pic, &left, &top, width, height)) return 0;
+
+  PictureGrabSpecs(pic, &tmp);
+  tmp.width = width;
+  tmp.height = height;
+  if (!WebPPictureAlloc(&tmp)) return 0;
+
+  if (!pic->use_argb) {
+    const int y_offset = top * pic->y_stride + left;
+    const int uv_offset = (top / 2) * pic->uv_stride + left / 2;
+    WebPCopyPlane(pic->y + y_offset, pic->y_stride,
+                  tmp.y, tmp.y_stride, width, height);
+    WebPCopyPlane(pic->u + uv_offset, pic->uv_stride,
+                  tmp.u, tmp.uv_stride, HALVE(width), HALVE(height));
+    WebPCopyPlane(pic->v + uv_offset, pic->uv_stride,
+                  tmp.v, tmp.uv_stride, HALVE(width), HALVE(height));
+
+    if (tmp.a != NULL) {
+      const int a_offset = top * pic->a_stride + left;
+      WebPCopyPlane(pic->a + a_offset, pic->a_stride,
+                    tmp.a, tmp.a_stride, width, height);
+    }
+  } else {
+    const uint8_t* const src =
+        (const uint8_t*)(pic->argb + top * pic->argb_stride + left);
+    WebPCopyPlane(src, pic->argb_stride * 4, (uint8_t*)tmp.argb,
+                  tmp.argb_stride * 4, width * 4, height);
+  }
+  WebPPictureFree(pic);
+  *pic = tmp;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Simple picture rescaler
+
+static int RescalePlane(const uint8_t* src,
+                        int src_width, int src_height, int src_stride,
+                        uint8_t* dst,
+                        int dst_width, int dst_height, int dst_stride,
+                        rescaler_t* const work,
+                        int num_channels) {
+  WebPRescaler rescaler;
+  int y = 0;
+  if (!WebPRescalerInit(&rescaler, src_width, src_height,
+                        dst, dst_width, dst_height, dst_stride,
+                        num_channels, work)) {
+    return 0;
+  }
+  while (y < src_height) {
+    y += WebPRescalerImport(&rescaler, src_height - y,
+                            src + y * src_stride, src_stride);
+    WebPRescalerExport(&rescaler);
+  }
+  return 1;
+}
+
+static void AlphaMultiplyARGB(WebPPicture* const pic, int inverse) {
+  assert(pic->argb != NULL);
+  WebPMultARGBRows((uint8_t*)pic->argb, pic->argb_stride * sizeof(*pic->argb),
+                   pic->width, pic->height, inverse);
+}
+
+static void AlphaMultiplyY(WebPPicture* const pic, int inverse) {
+  if (pic->a != NULL) {
+    WebPMultRows(pic->y, pic->y_stride, pic->a, pic->a_stride,
+                 pic->width, pic->height, inverse);
+  }
+}
+
+int WebPPictureRescale(WebPPicture* picture, int width, int height) {
+  WebPPicture tmp;
+  int prev_width, prev_height;
+  rescaler_t* work;
+
+  if (picture == NULL) return 0;
+  prev_width = picture->width;
+  prev_height = picture->height;
+  if (!WebPRescalerGetScaledDimensions(
+          prev_width, prev_height, &width, &height)) {
+    return 0;
+  }
+
+  PictureGrabSpecs(picture, &tmp);
+  tmp.width = width;
+  tmp.height = height;
+  if (!WebPPictureAlloc(&tmp)) return 0;
+
+  if (!picture->use_argb) {
+    work = (rescaler_t*)WebPSafeMalloc(2ULL * width, sizeof(*work));
+    if (work == NULL) {
+      WebPPictureFree(&tmp);
+      return 0;
+    }
+    // If present, we need to rescale alpha first (for AlphaMultiplyY).
+    if (picture->a != NULL) {
+      WebPInitAlphaProcessing();
+      if (!RescalePlane(picture->a, prev_width, prev_height, picture->a_stride,
+                        tmp.a, width, height, tmp.a_stride, work, 1)) {
+        return 0;
+      }
+    }
+
+    // We take transparency into account on the luma plane only. That's not
+    // totally exact blending, but still is a good approximation.
+    AlphaMultiplyY(picture, 0);
+    if (!RescalePlane(picture->y, prev_width, prev_height, picture->y_stride,
+                      tmp.y, width, height, tmp.y_stride, work, 1) ||
+        !RescalePlane(picture->u, HALVE(prev_width), HALVE(prev_height),
+                      picture->uv_stride, tmp.u, HALVE(width), HALVE(height),
+                      tmp.uv_stride, work, 1) ||
+        !RescalePlane(picture->v, HALVE(prev_width), HALVE(prev_height),
+                      picture->uv_stride, tmp.v, HALVE(width), HALVE(height),
+                      tmp.uv_stride, work, 1)) {
+      return 0;
+    }
+    AlphaMultiplyY(&tmp, 1);
+  } else {
+    work = (rescaler_t*)WebPSafeMalloc(2ULL * width * 4, sizeof(*work));
+    if (work == NULL) {
+      WebPPictureFree(&tmp);
+      return 0;
+    }
+    // In order to correctly interpolate colors, we need to apply the alpha
+    // weighting first (black-matting), scale the RGB values, and remove
+    // the premultiplication afterward (while preserving the alpha channel).
+    WebPInitAlphaProcessing();
+    AlphaMultiplyARGB(picture, 0);
+    if (!RescalePlane((const uint8_t*)picture->argb, prev_width, prev_height,
+                      picture->argb_stride * 4, (uint8_t*)tmp.argb, width,
+                      height, tmp.argb_stride * 4, work, 4)) {
+      return 0;
+    }
+    AlphaMultiplyARGB(&tmp, 1);
+  }
+  WebPPictureFree(picture);
+  WebPSafeFree(work);
+  *picture = tmp;
+  return 1;
+}
+
+#else  // defined(WEBP_REDUCE_SIZE)
+
+int WebPPictureCopy(const WebPPicture* src, WebPPicture* dst) {
+  (void)src;
+  (void)dst;
+  return 0;
+}
+
+int WebPPictureCrop(WebPPicture* pic,
+                    int left, int top, int width, int height) {
+  (void)pic;
+  (void)left;
+  (void)top;
+  (void)width;
+  (void)height;
+  return 0;
+}
+
+int WebPPictureRescale(WebPPicture* pic, int width, int height) {
+  (void)pic;
+  (void)width;
+  (void)height;
+  return 0;
+}
+#endif  // !defined(WEBP_REDUCE_SIZE)
diff --git a/media/libwebp/src/enc/picture_tools_enc.c b/media/libwebp/src/enc/picture_tools_enc.c
new file mode 100644
index 0000000000..147cc18608
--- /dev/null
+++ b/media/libwebp/src/enc/picture_tools_enc.c
@@ -0,0 +1,274 @@
+// Copyright 2014 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.
+// -----------------------------------------------------------------------------
+//
+// WebPPicture tools: alpha handling, etc.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+
+#include "src/enc/vp8i_enc.h"
+#include "src/dsp/yuv.h"
+
+//------------------------------------------------------------------------------
+// Helper: clean up fully transparent area to help compressibility.
+
+#define SIZE 8
+#define SIZE2 (SIZE / 2)
+static int IsTransparentARGBArea(const uint32_t* ptr, int stride, int size) {
+  int y, x;
+  for (y = 0; y < size; ++y) {
+    for (x = 0; x < size; ++x) {
+      if (ptr[x] & 0xff000000u) {
+        return 0;
+      }
+    }
+    ptr += stride;
+  }
+  return 1;
+}
+
+static void Flatten(uint8_t* ptr, int v, int stride, int size) {
+  int y;
+  for (y = 0; y < size; ++y) {
+    memset(ptr, v, size);
+    ptr += stride;
+  }
+}
+
+static void FlattenARGB(uint32_t* ptr, uint32_t v, int stride, int size) {
+  int x, y;
+  for (y = 0; y < size; ++y) {
+    for (x = 0; x < size; ++x) ptr[x] = v;
+    ptr += stride;
+  }
+}
+
+// Smoothen the luma components of transparent pixels. Return true if the whole
+// block is transparent.
+static int SmoothenBlock(const uint8_t* a_ptr, int a_stride, uint8_t* y_ptr,
+                         int y_stride, int width, int height) {
+  int sum = 0, count = 0;
+  int x, y;
+  const uint8_t* alpha_ptr = a_ptr;
+  uint8_t* luma_ptr = y_ptr;
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      if (alpha_ptr[x] != 0) {
+        ++count;
+        sum += luma_ptr[x];
+      }
+    }
+    alpha_ptr += a_stride;
+    luma_ptr += y_stride;
+  }
+  if (count > 0 && count < width * height) {
+    const uint8_t avg_u8 = (uint8_t)(sum / count);
+    alpha_ptr = a_ptr;
+    luma_ptr = y_ptr;
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        if (alpha_ptr[x] == 0) luma_ptr[x] = avg_u8;
+      }
+      alpha_ptr += a_stride;
+      luma_ptr += y_stride;
+    }
+  }
+  return (count == 0);
+}
+
+void WebPReplaceTransparentPixels(WebPPicture* const pic, uint32_t color) {
+  if (pic != NULL && pic->use_argb) {
+    int y = pic->height;
+    uint32_t* argb = pic->argb;
+    color &= 0xffffffu;   // force alpha=0
+    WebPInitAlphaProcessing();
+    while (y-- > 0) {
+      WebPAlphaReplace(argb, pic->width, color);
+      argb += pic->argb_stride;
+    }
+  }
+}
+
+void WebPCleanupTransparentArea(WebPPicture* pic) {
+  int x, y, w, h;
+  if (pic == NULL) return;
+  w = pic->width / SIZE;
+  h = pic->height / SIZE;
+
+  // note: we ignore the left-overs on right/bottom, except for SmoothenBlock().
+  if (pic->use_argb) {
+    uint32_t argb_value = 0;
+    for (y = 0; y < h; ++y) {
+      int need_reset = 1;
+      for (x = 0; x < w; ++x) {
+        const int off = (y * pic->argb_stride + x) * SIZE;
+        if (IsTransparentARGBArea(pic->argb + off, pic->argb_stride, SIZE)) {
+          if (need_reset) {
+            argb_value = pic->argb[off];
+            need_reset = 0;
+          }
+          FlattenARGB(pic->argb + off, argb_value, pic->argb_stride, SIZE);
+        } else {
+          need_reset = 1;
+        }
+      }
+    }
+  } else {
+    const int width = pic->width;
+    const int height = pic->height;
+    const int y_stride = pic->y_stride;
+    const int uv_stride = pic->uv_stride;
+    const int a_stride = pic->a_stride;
+    uint8_t* y_ptr = pic->y;
+    uint8_t* u_ptr = pic->u;
+    uint8_t* v_ptr = pic->v;
+    const uint8_t* a_ptr = pic->a;
+    int values[3] = { 0 };
+    if (a_ptr == NULL || y_ptr == NULL || u_ptr == NULL || v_ptr == NULL) {
+      return;
+    }
+    for (y = 0; y + SIZE <= height; y += SIZE) {
+      int need_reset = 1;
+      for (x = 0; x + SIZE <= width; x += SIZE) {
+        if (SmoothenBlock(a_ptr + x, a_stride, y_ptr + x, y_stride,
+                          SIZE, SIZE)) {
+          if (need_reset) {
+            values[0] = y_ptr[x];
+            values[1] = u_ptr[x >> 1];
+            values[2] = v_ptr[x >> 1];
+            need_reset = 0;
+          }
+          Flatten(y_ptr + x,        values[0], y_stride,  SIZE);
+          Flatten(u_ptr + (x >> 1), values[1], uv_stride, SIZE2);
+          Flatten(v_ptr + (x >> 1), values[2], uv_stride, SIZE2);
+        } else {
+          need_reset = 1;
+        }
+      }
+      if (x < width) {
+        SmoothenBlock(a_ptr + x, a_stride, y_ptr + x, y_stride,
+                      width - x, SIZE);
+      }
+      a_ptr += SIZE * a_stride;
+      y_ptr += SIZE * y_stride;
+      u_ptr += SIZE2 * uv_stride;
+      v_ptr += SIZE2 * uv_stride;
+    }
+    if (y < height) {
+      const int sub_height = height - y;
+      for (x = 0; x + SIZE <= width; x += SIZE) {
+        SmoothenBlock(a_ptr + x, a_stride, y_ptr + x, y_stride,
+                      SIZE, sub_height);
+      }
+      if (x < width) {
+        SmoothenBlock(a_ptr + x, a_stride, y_ptr + x, y_stride,
+                      width - x, sub_height);
+      }
+    }
+  }
+}
+
+#undef SIZE
+#undef SIZE2
+
+//------------------------------------------------------------------------------
+// Blend color and remove transparency info
+
+#define BLEND(V0, V1, ALPHA) \
+    ((((V0) * (255 - (ALPHA)) + (V1) * (ALPHA)) * 0x101 + 256) >> 16)
+#define BLEND_10BIT(V0, V1, ALPHA) \
+    ((((V0) * (1020 - (ALPHA)) + (V1) * (ALPHA)) * 0x101 + 1024) >> 18)
+
+static WEBP_INLINE uint32_t MakeARGB32(int r, int g, int b) {
+  return (0xff000000u | (r << 16) | (g << 8) | b);
+}
+
+void WebPBlendAlpha(WebPPicture* picture, uint32_t background_rgb) {
+  const int red = (background_rgb >> 16) & 0xff;
+  const int green = (background_rgb >> 8) & 0xff;
+  const int blue = (background_rgb >> 0) & 0xff;
+  int x, y;
+  if (picture == NULL) return;
+  if (!picture->use_argb) {
+    // omit last pixel during u/v loop
+    const int uv_width = (picture->width >> 1);
+    const int Y0 = VP8RGBToY(red, green, blue, YUV_HALF);
+    // VP8RGBToU/V expects the u/v values summed over four pixels
+    const int U0 = VP8RGBToU(4 * red, 4 * green, 4 * blue, 4 * YUV_HALF);
+    const int V0 = VP8RGBToV(4 * red, 4 * green, 4 * blue, 4 * YUV_HALF);
+    const int has_alpha = picture->colorspace & WEBP_CSP_ALPHA_BIT;
+    uint8_t* y_ptr = picture->y;
+    uint8_t* u_ptr = picture->u;
+    uint8_t* v_ptr = picture->v;
+    uint8_t* a_ptr = picture->a;
+    if (!has_alpha || a_ptr == NULL) return;    // nothing to do
+    for (y = 0; y < picture->height; ++y) {
+      // Luma blending
+      for (x = 0; x < picture->width; ++x) {
+        const uint8_t alpha = a_ptr[x];
+        if (alpha < 0xff) {
+          y_ptr[x] = BLEND(Y0, y_ptr[x], alpha);
+        }
+      }
+      // Chroma blending every even line
+      if ((y & 1) == 0) {
+        uint8_t* const a_ptr2 =
+            (y + 1 == picture->height) ? a_ptr : a_ptr + picture->a_stride;
+        for (x = 0; x < uv_width; ++x) {
+          // Average four alpha values into a single blending weight.
+          // TODO(skal): might lead to visible contouring. Can we do better?
+          const uint32_t alpha =
+              a_ptr[2 * x + 0] + a_ptr[2 * x + 1] +
+              a_ptr2[2 * x + 0] + a_ptr2[2 * x + 1];
+          u_ptr[x] = BLEND_10BIT(U0, u_ptr[x], alpha);
+          v_ptr[x] = BLEND_10BIT(V0, v_ptr[x], alpha);
+        }
+        if (picture->width & 1) {  // rightmost pixel
+          const uint32_t alpha = 2 * (a_ptr[2 * x + 0] + a_ptr2[2 * x + 0]);
+          u_ptr[x] = BLEND_10BIT(U0, u_ptr[x], alpha);
+          v_ptr[x] = BLEND_10BIT(V0, v_ptr[x], alpha);
+        }
+      } else {
+        u_ptr += picture->uv_stride;
+        v_ptr += picture->uv_stride;
+      }
+      memset(a_ptr, 0xff, picture->width);  // reset alpha value to opaque
+      a_ptr += picture->a_stride;
+      y_ptr += picture->y_stride;
+    }
+  } else {
+    uint32_t* argb = picture->argb;
+    const uint32_t background = MakeARGB32(red, green, blue);
+    for (y = 0; y < picture->height; ++y) {
+      for (x = 0; x < picture->width; ++x) {
+        const int alpha = (argb[x] >> 24) & 0xff;
+        if (alpha != 0xff) {
+          if (alpha > 0) {
+            int r = (argb[x] >> 16) & 0xff;
+            int g = (argb[x] >>  8) & 0xff;
+            int b = (argb[x] >>  0) & 0xff;
+            r = BLEND(red, r, alpha);
+            g = BLEND(green, g, alpha);
+            b = BLEND(blue, b, alpha);
+            argb[x] = MakeARGB32(r, g, b);
+          } else {
+            argb[x] = background;
+          }
+        }
+      }
+      argb += picture->argb_stride;
+    }
+  }
+}
+
+#undef BLEND
+#undef BLEND_10BIT
+
+//------------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/predictor_enc.c b/media/libwebp/src/enc/predictor_enc.c
new file mode 100644
index 0000000000..b3d44b59d5
--- /dev/null
+++ b/media/libwebp/src/enc/predictor_enc.c
@@ -0,0 +1,792 @@
+// Copyright 2016 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.
+// -----------------------------------------------------------------------------
+//
+// Image transform methods for lossless encoder.
+//
+// Authors: Vikas Arora (vikaas.arora@gmail.com)
+//          Jyrki Alakuijala (jyrki@google.com)
+//          Urvang Joshi (urvang@google.com)
+//          Vincent Rabaud (vrabaud@google.com)
+
+#include "src/dsp/lossless.h"
+#include "src/dsp/lossless_common.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/enc/vp8li_enc.h"
+
+#define MAX_DIFF_COST (1e30f)
+
+static const float kSpatialPredictorBias = 15.f;
+static const int kPredLowEffort = 11;
+static const uint32_t kMaskAlpha = 0xff000000;
+
+// Mostly used to reduce code size + readability
+static WEBP_INLINE int GetMin(int a, int b) { return (a > b) ? b : a; }
+
+//------------------------------------------------------------------------------
+// Methods to calculate Entropy (Shannon).
+
+static float PredictionCostSpatial(const int counts[256], int weight_0,
+                                   float exp_val) {
+  const int significant_symbols = 256 >> 4;
+  const float exp_decay_factor = 0.6f;
+  float bits = (float)weight_0 * counts[0];
+  int i;
+  for (i = 1; i < significant_symbols; ++i) {
+    bits += exp_val * (counts[i] + counts[256 - i]);
+    exp_val *= exp_decay_factor;
+  }
+  return (float)(-0.1 * bits);
+}
+
+static float PredictionCostSpatialHistogram(const int accumulated[4][256],
+                                            const int tile[4][256]) {
+  int i;
+  float retval = 0.f;
+  for (i = 0; i < 4; ++i) {
+    const float kExpValue = 0.94f;
+    retval += PredictionCostSpatial(tile[i], 1, kExpValue);
+    retval += VP8LCombinedShannonEntropy(tile[i], accumulated[i]);
+  }
+  return (float)retval;
+}
+
+static WEBP_INLINE void UpdateHisto(int histo_argb[4][256], uint32_t argb) {
+  ++histo_argb[0][argb >> 24];
+  ++histo_argb[1][(argb >> 16) & 0xff];
+  ++histo_argb[2][(argb >> 8) & 0xff];
+  ++histo_argb[3][argb & 0xff];
+}
+
+//------------------------------------------------------------------------------
+// Spatial transform functions.
+
+static WEBP_INLINE void PredictBatch(int mode, int x_start, int y,
+                                     int num_pixels, const uint32_t* current,
+                                     const uint32_t* upper, uint32_t* out) {
+  if (x_start == 0) {
+    if (y == 0) {
+      // ARGB_BLACK.
+      VP8LPredictorsSub[0](current, NULL, 1, out);
+    } else {
+      // Top one.
+      VP8LPredictorsSub[2](current, upper, 1, out);
+    }
+    ++x_start;
+    ++out;
+    --num_pixels;
+  }
+  if (y == 0) {
+    // Left one.
+    VP8LPredictorsSub[1](current + x_start, NULL, num_pixels, out);
+  } else {
+    VP8LPredictorsSub[mode](current + x_start, upper + x_start, num_pixels,
+                            out);
+  }
+}
+
+#if (WEBP_NEAR_LOSSLESS == 1)
+static WEBP_INLINE int GetMax(int a, int b) { return (a < b) ? b : a; }
+
+static int MaxDiffBetweenPixels(uint32_t p1, uint32_t p2) {
+  const int diff_a = abs((int)(p1 >> 24) - (int)(p2 >> 24));
+  const int diff_r = abs((int)((p1 >> 16) & 0xff) - (int)((p2 >> 16) & 0xff));
+  const int diff_g = abs((int)((p1 >> 8) & 0xff) - (int)((p2 >> 8) & 0xff));
+  const int diff_b = abs((int)(p1 & 0xff) - (int)(p2 & 0xff));
+  return GetMax(GetMax(diff_a, diff_r), GetMax(diff_g, diff_b));
+}
+
+static int MaxDiffAroundPixel(uint32_t current, uint32_t up, uint32_t down,
+                              uint32_t left, uint32_t right) {
+  const int diff_up = MaxDiffBetweenPixels(current, up);
+  const int diff_down = MaxDiffBetweenPixels(current, down);
+  const int diff_left = MaxDiffBetweenPixels(current, left);
+  const int diff_right = MaxDiffBetweenPixels(current, right);
+  return GetMax(GetMax(diff_up, diff_down), GetMax(diff_left, diff_right));
+}
+
+static uint32_t AddGreenToBlueAndRed(uint32_t argb) {
+  const uint32_t green = (argb >> 8) & 0xff;
+  uint32_t red_blue = argb & 0x00ff00ffu;
+  red_blue += (green << 16) | green;
+  red_blue &= 0x00ff00ffu;
+  return (argb & 0xff00ff00u) | red_blue;
+}
+
+static void MaxDiffsForRow(int width, int stride, const uint32_t* const argb,
+                           uint8_t* const max_diffs, int used_subtract_green) {
+  uint32_t current, up, down, left, right;
+  int x;
+  if (width <= 2) return;
+  current = argb[0];
+  right = argb[1];
+  if (used_subtract_green) {
+    current = AddGreenToBlueAndRed(current);
+    right = AddGreenToBlueAndRed(right);
+  }
+  // max_diffs[0] and max_diffs[width - 1] are never used.
+  for (x = 1; x < width - 1; ++x) {
+    up = argb[-stride + x];
+    down = argb[stride + x];
+    left = current;
+    current = right;
+    right = argb[x + 1];
+    if (used_subtract_green) {
+      up = AddGreenToBlueAndRed(up);
+      down = AddGreenToBlueAndRed(down);
+      right = AddGreenToBlueAndRed(right);
+    }
+    max_diffs[x] = MaxDiffAroundPixel(current, up, down, left, right);
+  }
+}
+
+// Quantize the difference between the actual component value and its prediction
+// to a multiple of quantization, working modulo 256, taking care not to cross
+// a boundary (inclusive upper limit).
+static uint8_t NearLosslessComponent(uint8_t value, uint8_t predict,
+                                     uint8_t boundary, int quantization) {
+  const int residual = (value - predict) & 0xff;
+  const int boundary_residual = (boundary - predict) & 0xff;
+  const int lower = residual & ~(quantization - 1);
+  const int upper = lower + quantization;
+  // Resolve ties towards a value closer to the prediction (i.e. towards lower
+  // if value comes after prediction and towards upper otherwise).
+  const int bias = ((boundary - value) & 0xff) < boundary_residual;
+  if (residual - lower < upper - residual + bias) {
+    // lower is closer to residual than upper.
+    if (residual > boundary_residual && lower <= boundary_residual) {
+      // Halve quantization step to avoid crossing boundary. This midpoint is
+      // on the same side of boundary as residual because midpoint >= residual
+      // (since lower is closer than upper) and residual is above the boundary.
+      return lower + (quantization >> 1);
+    }
+    return lower;
+  } else {
+    // upper is closer to residual than lower.
+    if (residual <= boundary_residual && upper > boundary_residual) {
+      // Halve quantization step to avoid crossing boundary. This midpoint is
+      // on the same side of boundary as residual because midpoint <= residual
+      // (since upper is closer than lower) and residual is below the boundary.
+      return lower + (quantization >> 1);
+    }
+    return upper & 0xff;
+  }
+}
+
+static WEBP_INLINE uint8_t NearLosslessDiff(uint8_t a, uint8_t b) {
+  return (uint8_t)((((int)(a) - (int)(b))) & 0xff);
+}
+
+// Quantize every component of the difference between the actual pixel value and
+// its prediction to a multiple of a quantization (a power of 2, not larger than
+// max_quantization which is a power of 2, smaller than max_diff). Take care if
+// value and predict have undergone subtract green, which means that red and
+// blue are represented as offsets from green.
+static uint32_t NearLossless(uint32_t value, uint32_t predict,
+                             int max_quantization, int max_diff,
+                             int used_subtract_green) {
+  int quantization;
+  uint8_t new_green = 0;
+  uint8_t green_diff = 0;
+  uint8_t a, r, g, b;
+  if (max_diff <= 2) {
+    return VP8LSubPixels(value, predict);
+  }
+  quantization = max_quantization;
+  while (quantization >= max_diff) {
+    quantization >>= 1;
+  }
+  if ((value >> 24) == 0 || (value >> 24) == 0xff) {
+    // Preserve transparency of fully transparent or fully opaque pixels.
+    a = NearLosslessDiff((value >> 24) & 0xff, (predict >> 24) & 0xff);
+  } else {
+    a = NearLosslessComponent(value >> 24, predict >> 24, 0xff, quantization);
+  }
+  g = NearLosslessComponent((value >> 8) & 0xff, (predict >> 8) & 0xff, 0xff,
+                            quantization);
+  if (used_subtract_green) {
+    // The green offset will be added to red and blue components during decoding
+    // to obtain the actual red and blue values.
+    new_green = ((predict >> 8) + g) & 0xff;
+    // The amount by which green has been adjusted during quantization. It is
+    // subtracted from red and blue for compensation, to avoid accumulating two
+    // quantization errors in them.
+    green_diff = NearLosslessDiff(new_green, (value >> 8) & 0xff);
+  }
+  r = NearLosslessComponent(NearLosslessDiff((value >> 16) & 0xff, green_diff),
+                            (predict >> 16) & 0xff, 0xff - new_green,
+                            quantization);
+  b = NearLosslessComponent(NearLosslessDiff(value & 0xff, green_diff),
+                            predict & 0xff, 0xff - new_green, quantization);
+  return ((uint32_t)a << 24) | ((uint32_t)r << 16) | ((uint32_t)g << 8) | b;
+}
+#endif  // (WEBP_NEAR_LOSSLESS == 1)
+
+// Stores the difference between the pixel and its prediction in "out".
+// In case of a lossy encoding, updates the source image to avoid propagating
+// the deviation further to pixels which depend on the current pixel for their
+// predictions.
+static WEBP_INLINE void GetResidual(
+    int width, int height, uint32_t* const upper_row,
+    uint32_t* const current_row, const uint8_t* const max_diffs, int mode,
+    int x_start, int x_end, int y, int max_quantization, int exact,
+    int used_subtract_green, uint32_t* const out) {
+  if (exact) {
+    PredictBatch(mode, x_start, y, x_end - x_start, current_row, upper_row,
+                 out);
+  } else {
+    const VP8LPredictorFunc pred_func = VP8LPredictors[mode];
+    int x;
+    for (x = x_start; x < x_end; ++x) {
+      uint32_t predict;
+      uint32_t residual;
+      if (y == 0) {
+        predict = (x == 0) ? ARGB_BLACK : current_row[x - 1];  // Left.
+      } else if (x == 0) {
+        predict = upper_row[x];  // Top.
+      } else {
+        predict = pred_func(&current_row[x - 1], upper_row + x);
+      }
+#if (WEBP_NEAR_LOSSLESS == 1)
+      if (max_quantization == 1 || mode == 0 || y == 0 || y == height - 1 ||
+          x == 0 || x == width - 1) {
+        residual = VP8LSubPixels(current_row[x], predict);
+      } else {
+        residual = NearLossless(current_row[x], predict, max_quantization,
+                                max_diffs[x], used_subtract_green);
+        // Update the source image.
+        current_row[x] = VP8LAddPixels(predict, residual);
+        // x is never 0 here so we do not need to update upper_row like below.
+      }
+#else
+      (void)max_diffs;
+      (void)height;
+      (void)max_quantization;
+      (void)used_subtract_green;
+      residual = VP8LSubPixels(current_row[x], predict);
+#endif
+      if ((current_row[x] & kMaskAlpha) == 0) {
+        // If alpha is 0, cleanup RGB. We can choose the RGB values of the
+        // residual for best compression. The prediction of alpha itself can be
+        // non-zero and must be kept though. We choose RGB of the residual to be
+        // 0.
+        residual &= kMaskAlpha;
+        // Update the source image.
+        current_row[x] = predict & ~kMaskAlpha;
+        // The prediction for the rightmost pixel in a row uses the leftmost
+        // pixel
+        // in that row as its top-right context pixel. Hence if we change the
+        // leftmost pixel of current_row, the corresponding change must be
+        // applied
+        // to upper_row as well where top-right context is being read from.
+        if (x == 0 && y != 0) upper_row[width] = current_row[0];
+      }
+      out[x - x_start] = residual;
+    }
+  }
+}
+
+// Returns best predictor and updates the accumulated histogram.
+// If max_quantization > 1, assumes that near lossless processing will be
+// applied, quantizing residuals to multiples of quantization levels up to
+// max_quantization (the actual quantization level depends on smoothness near
+// the given pixel).
+static int GetBestPredictorForTile(int width, int height,
+                                   int tile_x, int tile_y, int bits,
+                                   int accumulated[4][256],
+                                   uint32_t* const argb_scratch,
+                                   const uint32_t* const argb,
+                                   int max_quantization,
+                                   int exact, int used_subtract_green,
+                                   const uint32_t* const modes) {
+  const int kNumPredModes = 14;
+  const int start_x = tile_x << bits;
+  const int start_y = tile_y << bits;
+  const int tile_size = 1 << bits;
+  const int max_y = GetMin(tile_size, height - start_y);
+  const int max_x = GetMin(tile_size, width - start_x);
+  // Whether there exist columns just outside the tile.
+  const int have_left = (start_x > 0);
+  // Position and size of the strip covering the tile and adjacent columns if
+  // they exist.
+  const int context_start_x = start_x - have_left;
+#if (WEBP_NEAR_LOSSLESS == 1)
+  const int context_width = max_x + have_left + (max_x < width - start_x);
+#endif
+  const int tiles_per_row = VP8LSubSampleSize(width, bits);
+  // Prediction modes of the left and above neighbor tiles.
+  const int left_mode = (tile_x > 0) ?
+      (modes[tile_y * tiles_per_row + tile_x - 1] >> 8) & 0xff : 0xff;
+  const int above_mode = (tile_y > 0) ?
+      (modes[(tile_y - 1) * tiles_per_row + tile_x] >> 8) & 0xff : 0xff;
+  // The width of upper_row and current_row is one pixel larger than image width
+  // to allow the top right pixel to point to the leftmost pixel of the next row
+  // when at the right edge.
+  uint32_t* upper_row = argb_scratch;
+  uint32_t* current_row = upper_row + width + 1;
+  uint8_t* const max_diffs = (uint8_t*)(current_row + width + 1);
+  float best_diff = MAX_DIFF_COST;
+  int best_mode = 0;
+  int mode;
+  int histo_stack_1[4][256];
+  int histo_stack_2[4][256];
+  // Need pointers to be able to swap arrays.
+  int (*histo_argb)[256] = histo_stack_1;
+  int (*best_histo)[256] = histo_stack_2;
+  int i, j;
+  uint32_t residuals[1 << MAX_TRANSFORM_BITS];
+  assert(bits <= MAX_TRANSFORM_BITS);
+  assert(max_x <= (1 << MAX_TRANSFORM_BITS));
+
+  for (mode = 0; mode < kNumPredModes; ++mode) {
+    float cur_diff;
+    int relative_y;
+    memset(histo_argb, 0, sizeof(histo_stack_1));
+    if (start_y > 0) {
+      // Read the row above the tile which will become the first upper_row.
+      // Include a pixel to the left if it exists; include a pixel to the right
+      // in all cases (wrapping to the leftmost pixel of the next row if it does
+      // not exist).
+      memcpy(current_row + context_start_x,
+             argb + (start_y - 1) * width + context_start_x,
+             sizeof(*argb) * (max_x + have_left + 1));
+    }
+    for (relative_y = 0; relative_y < max_y; ++relative_y) {
+      const int y = start_y + relative_y;
+      int relative_x;
+      uint32_t* tmp = upper_row;
+      upper_row = current_row;
+      current_row = tmp;
+      // Read current_row. Include a pixel to the left if it exists; include a
+      // pixel to the right in all cases except at the bottom right corner of
+      // the image (wrapping to the leftmost pixel of the next row if it does
+      // not exist in the current row).
+      memcpy(current_row + context_start_x,
+             argb + y * width + context_start_x,
+             sizeof(*argb) * (max_x + have_left + (y + 1 < height)));
+#if (WEBP_NEAR_LOSSLESS == 1)
+      if (max_quantization > 1 && y >= 1 && y + 1 < height) {
+        MaxDiffsForRow(context_width, width, argb + y * width + context_start_x,
+                       max_diffs + context_start_x, used_subtract_green);
+      }
+#endif
+
+      GetResidual(width, height, upper_row, current_row, max_diffs, mode,
+                  start_x, start_x + max_x, y, max_quantization, exact,
+                  used_subtract_green, residuals);
+      for (relative_x = 0; relative_x < max_x; ++relative_x) {
+        UpdateHisto(histo_argb, residuals[relative_x]);
+      }
+    }
+    cur_diff = PredictionCostSpatialHistogram(
+        (const int (*)[256])accumulated, (const int (*)[256])histo_argb);
+    // Favor keeping the areas locally similar.
+    if (mode == left_mode) cur_diff -= kSpatialPredictorBias;
+    if (mode == above_mode) cur_diff -= kSpatialPredictorBias;
+
+    if (cur_diff < best_diff) {
+      int (*tmp)[256] = histo_argb;
+      histo_argb = best_histo;
+      best_histo = tmp;
+      best_diff = cur_diff;
+      best_mode = mode;
+    }
+  }
+
+  for (i = 0; i < 4; i++) {
+    for (j = 0; j < 256; j++) {
+      accumulated[i][j] += best_histo[i][j];
+    }
+  }
+
+  return best_mode;
+}
+
+// Converts pixels of the image to residuals with respect to predictions.
+// If max_quantization > 1, applies near lossless processing, quantizing
+// residuals to multiples of quantization levels up to max_quantization
+// (the actual quantization level depends on smoothness near the given pixel).
+static void CopyImageWithPrediction(int width, int height,
+                                    int bits, uint32_t* const modes,
+                                    uint32_t* const argb_scratch,
+                                    uint32_t* const argb,
+                                    int low_effort, int max_quantization,
+                                    int exact, int used_subtract_green) {
+  const int tiles_per_row = VP8LSubSampleSize(width, bits);
+  // The width of upper_row and current_row is one pixel larger than image width
+  // to allow the top right pixel to point to the leftmost pixel of the next row
+  // when at the right edge.
+  uint32_t* upper_row = argb_scratch;
+  uint32_t* current_row = upper_row + width + 1;
+  uint8_t* current_max_diffs = (uint8_t*)(current_row + width + 1);
+#if (WEBP_NEAR_LOSSLESS == 1)
+  uint8_t* lower_max_diffs = current_max_diffs + width;
+#endif
+  int y;
+
+  for (y = 0; y < height; ++y) {
+    int x;
+    uint32_t* const tmp32 = upper_row;
+    upper_row = current_row;
+    current_row = tmp32;
+    memcpy(current_row, argb + y * width,
+           sizeof(*argb) * (width + (y + 1 < height)));
+
+    if (low_effort) {
+      PredictBatch(kPredLowEffort, 0, y, width, current_row, upper_row,
+                   argb + y * width);
+    } else {
+#if (WEBP_NEAR_LOSSLESS == 1)
+      if (max_quantization > 1) {
+        // Compute max_diffs for the lower row now, because that needs the
+        // contents of argb for the current row, which we will overwrite with
+        // residuals before proceeding with the next row.
+        uint8_t* const tmp8 = current_max_diffs;
+        current_max_diffs = lower_max_diffs;
+        lower_max_diffs = tmp8;
+        if (y + 2 < height) {
+          MaxDiffsForRow(width, width, argb + (y + 1) * width, lower_max_diffs,
+                         used_subtract_green);
+        }
+      }
+#endif
+      for (x = 0; x < width;) {
+        const int mode =
+            (modes[(y >> bits) * tiles_per_row + (x >> bits)] >> 8) & 0xff;
+        int x_end = x + (1 << bits);
+        if (x_end > width) x_end = width;
+        GetResidual(width, height, upper_row, current_row, current_max_diffs,
+                    mode, x, x_end, y, max_quantization, exact,
+                    used_subtract_green, argb + y * width + x);
+        x = x_end;
+      }
+    }
+  }
+}
+
+// Finds the best predictor for each tile, and converts the image to residuals
+// with respect to predictions. If near_lossless_quality < 100, applies
+// near lossless processing, shaving off more bits of residuals for lower
+// qualities.
+int VP8LResidualImage(int width, int height, int bits, int low_effort,
+                      uint32_t* const argb, uint32_t* const argb_scratch,
+                      uint32_t* const image, int near_lossless_quality,
+                      int exact, int used_subtract_green,
+                      const WebPPicture* const pic, int percent_range,
+                      int* const percent) {
+  const int tiles_per_row = VP8LSubSampleSize(width, bits);
+  const int tiles_per_col = VP8LSubSampleSize(height, bits);
+  int percent_start = *percent;
+  int tile_y;
+  int histo[4][256];
+  const int max_quantization = 1 << VP8LNearLosslessBits(near_lossless_quality);
+  if (low_effort) {
+    int i;
+    for (i = 0; i < tiles_per_row * tiles_per_col; ++i) {
+      image[i] = ARGB_BLACK | (kPredLowEffort << 8);
+    }
+  } else {
+    memset(histo, 0, sizeof(histo));
+    for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) {
+      int tile_x;
+      for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) {
+        const int pred = GetBestPredictorForTile(
+            width, height, tile_x, tile_y, bits, histo, argb_scratch, argb,
+            max_quantization, exact, used_subtract_green, image);
+        image[tile_y * tiles_per_row + tile_x] = ARGB_BLACK | (pred << 8);
+      }
+
+      if (!WebPReportProgress(
+              pic, percent_start + percent_range * tile_y / tiles_per_col,
+              percent)) {
+        return 0;
+      }
+    }
+  }
+
+  CopyImageWithPrediction(width, height, bits, image, argb_scratch, argb,
+                          low_effort, max_quantization, exact,
+                          used_subtract_green);
+  return WebPReportProgress(pic, percent_start + percent_range, percent);
+}
+
+//------------------------------------------------------------------------------
+// Color transform functions.
+
+static WEBP_INLINE void MultipliersClear(VP8LMultipliers* const m) {
+  m->green_to_red_ = 0;
+  m->green_to_blue_ = 0;
+  m->red_to_blue_ = 0;
+}
+
+static WEBP_INLINE void ColorCodeToMultipliers(uint32_t color_code,
+                                               VP8LMultipliers* const m) {
+  m->green_to_red_  = (color_code >>  0) & 0xff;
+  m->green_to_blue_ = (color_code >>  8) & 0xff;
+  m->red_to_blue_   = (color_code >> 16) & 0xff;
+}
+
+static WEBP_INLINE uint32_t MultipliersToColorCode(
+    const VP8LMultipliers* const m) {
+  return 0xff000000u |
+         ((uint32_t)(m->red_to_blue_) << 16) |
+         ((uint32_t)(m->green_to_blue_) << 8) |
+         m->green_to_red_;
+}
+
+static float PredictionCostCrossColor(const int accumulated[256],
+                                      const int counts[256]) {
+  // Favor low entropy, locally and globally.
+  // Favor small absolute values for PredictionCostSpatial
+  static const float kExpValue = 2.4f;
+  return VP8LCombinedShannonEntropy(counts, accumulated) +
+         PredictionCostSpatial(counts, 3, kExpValue);
+}
+
+static float GetPredictionCostCrossColorRed(
+    const uint32_t* argb, int stride, int tile_width, int tile_height,
+    VP8LMultipliers prev_x, VP8LMultipliers prev_y, int green_to_red,
+    const int accumulated_red_histo[256]) {
+  int histo[256] = { 0 };
+  float cur_diff;
+
+  VP8LCollectColorRedTransforms(argb, stride, tile_width, tile_height,
+                                green_to_red, histo);
+
+  cur_diff = PredictionCostCrossColor(accumulated_red_histo, histo);
+  if ((uint8_t)green_to_red == prev_x.green_to_red_) {
+    cur_diff -= 3;  // favor keeping the areas locally similar
+  }
+  if ((uint8_t)green_to_red == prev_y.green_to_red_) {
+    cur_diff -= 3;  // favor keeping the areas locally similar
+  }
+  if (green_to_red == 0) {
+    cur_diff -= 3;
+  }
+  return cur_diff;
+}
+
+static void GetBestGreenToRed(
+    const uint32_t* argb, int stride, int tile_width, int tile_height,
+    VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
+    const int accumulated_red_histo[256], VP8LMultipliers* const best_tx) {
+  const int kMaxIters = 4 + ((7 * quality) >> 8);  // in range [4..6]
+  int green_to_red_best = 0;
+  int iter, offset;
+  float best_diff = GetPredictionCostCrossColorRed(
+      argb, stride, tile_width, tile_height, prev_x, prev_y,
+      green_to_red_best, accumulated_red_histo);
+  for (iter = 0; iter < kMaxIters; ++iter) {
+    // ColorTransformDelta is a 3.5 bit fixed point, so 32 is equal to
+    // one in color computation. Having initial delta here as 1 is sufficient
+    // to explore the range of (-2, 2).
+    const int delta = 32 >> iter;
+    // Try a negative and a positive delta from the best known value.
+    for (offset = -delta; offset <= delta; offset += 2 * delta) {
+      const int green_to_red_cur = offset + green_to_red_best;
+      const float cur_diff = GetPredictionCostCrossColorRed(
+          argb, stride, tile_width, tile_height, prev_x, prev_y,
+          green_to_red_cur, accumulated_red_histo);
+      if (cur_diff < best_diff) {
+        best_diff = cur_diff;
+        green_to_red_best = green_to_red_cur;
+      }
+    }
+  }
+  best_tx->green_to_red_ = (green_to_red_best & 0xff);
+}
+
+static float GetPredictionCostCrossColorBlue(
+    const uint32_t* argb, int stride, int tile_width, int tile_height,
+    VP8LMultipliers prev_x, VP8LMultipliers prev_y,
+    int green_to_blue, int red_to_blue, const int accumulated_blue_histo[256]) {
+  int histo[256] = { 0 };
+  float cur_diff;
+
+  VP8LCollectColorBlueTransforms(argb, stride, tile_width, tile_height,
+                                 green_to_blue, red_to_blue, histo);
+
+  cur_diff = PredictionCostCrossColor(accumulated_blue_histo, histo);
+  if ((uint8_t)green_to_blue == prev_x.green_to_blue_) {
+    cur_diff -= 3;  // favor keeping the areas locally similar
+  }
+  if ((uint8_t)green_to_blue == prev_y.green_to_blue_) {
+    cur_diff -= 3;  // favor keeping the areas locally similar
+  }
+  if ((uint8_t)red_to_blue == prev_x.red_to_blue_) {
+    cur_diff -= 3;  // favor keeping the areas locally similar
+  }
+  if ((uint8_t)red_to_blue == prev_y.red_to_blue_) {
+    cur_diff -= 3;  // favor keeping the areas locally similar
+  }
+  if (green_to_blue == 0) {
+    cur_diff -= 3;
+  }
+  if (red_to_blue == 0) {
+    cur_diff -= 3;
+  }
+  return cur_diff;
+}
+
+#define kGreenRedToBlueNumAxis 8
+#define kGreenRedToBlueMaxIters 7
+static void GetBestGreenRedToBlue(
+    const uint32_t* argb, int stride, int tile_width, int tile_height,
+    VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
+    const int accumulated_blue_histo[256],
+    VP8LMultipliers* const best_tx) {
+  const int8_t offset[kGreenRedToBlueNumAxis][2] =
+      {{0, -1}, {0, 1}, {-1, 0}, {1, 0}, {-1, -1}, {-1, 1}, {1, -1}, {1, 1}};
+  const int8_t delta_lut[kGreenRedToBlueMaxIters] = { 16, 16, 8, 4, 2, 2, 2 };
+  const int iters =
+      (quality < 25) ? 1 : (quality > 50) ? kGreenRedToBlueMaxIters : 4;
+  int green_to_blue_best = 0;
+  int red_to_blue_best = 0;
+  int iter;
+  // Initial value at origin:
+  float best_diff = GetPredictionCostCrossColorBlue(
+      argb, stride, tile_width, tile_height, prev_x, prev_y,
+      green_to_blue_best, red_to_blue_best, accumulated_blue_histo);
+  for (iter = 0; iter < iters; ++iter) {
+    const int delta = delta_lut[iter];
+    int axis;
+    for (axis = 0; axis < kGreenRedToBlueNumAxis; ++axis) {
+      const int green_to_blue_cur =
+          offset[axis][0] * delta + green_to_blue_best;
+      const int red_to_blue_cur = offset[axis][1] * delta + red_to_blue_best;
+      const float cur_diff = GetPredictionCostCrossColorBlue(
+          argb, stride, tile_width, tile_height, prev_x, prev_y,
+          green_to_blue_cur, red_to_blue_cur, accumulated_blue_histo);
+      if (cur_diff < best_diff) {
+        best_diff = cur_diff;
+        green_to_blue_best = green_to_blue_cur;
+        red_to_blue_best = red_to_blue_cur;
+      }
+      if (quality < 25 && iter == 4) {
+        // Only axis aligned diffs for lower quality.
+        break;  // next iter.
+      }
+    }
+    if (delta == 2 && green_to_blue_best == 0 && red_to_blue_best == 0) {
+      // Further iterations would not help.
+      break;  // out of iter-loop.
+    }
+  }
+  best_tx->green_to_blue_ = green_to_blue_best & 0xff;
+  best_tx->red_to_blue_ = red_to_blue_best & 0xff;
+}
+#undef kGreenRedToBlueMaxIters
+#undef kGreenRedToBlueNumAxis
+
+static VP8LMultipliers GetBestColorTransformForTile(
+    int tile_x, int tile_y, int bits,
+    VP8LMultipliers prev_x,
+    VP8LMultipliers prev_y,
+    int quality, int xsize, int ysize,
+    const int accumulated_red_histo[256],
+    const int accumulated_blue_histo[256],
+    const uint32_t* const argb) {
+  const int max_tile_size = 1 << bits;
+  const int tile_y_offset = tile_y * max_tile_size;
+  const int tile_x_offset = tile_x * max_tile_size;
+  const int all_x_max = GetMin(tile_x_offset + max_tile_size, xsize);
+  const int all_y_max = GetMin(tile_y_offset + max_tile_size, ysize);
+  const int tile_width = all_x_max - tile_x_offset;
+  const int tile_height = all_y_max - tile_y_offset;
+  const uint32_t* const tile_argb = argb + tile_y_offset * xsize
+                                  + tile_x_offset;
+  VP8LMultipliers best_tx;
+  MultipliersClear(&best_tx);
+
+  GetBestGreenToRed(tile_argb, xsize, tile_width, tile_height,
+                    prev_x, prev_y, quality, accumulated_red_histo, &best_tx);
+  GetBestGreenRedToBlue(tile_argb, xsize, tile_width, tile_height,
+                        prev_x, prev_y, quality, accumulated_blue_histo,
+                        &best_tx);
+  return best_tx;
+}
+
+static void CopyTileWithColorTransform(int xsize, int ysize,
+                                       int tile_x, int tile_y,
+                                       int max_tile_size,
+                                       VP8LMultipliers color_transform,
+                                       uint32_t* argb) {
+  const int xscan = GetMin(max_tile_size, xsize - tile_x);
+  int yscan = GetMin(max_tile_size, ysize - tile_y);
+  argb += tile_y * xsize + tile_x;
+  while (yscan-- > 0) {
+    VP8LTransformColor(&color_transform, argb, xscan);
+    argb += xsize;
+  }
+}
+
+int VP8LColorSpaceTransform(int width, int height, int bits, int quality,
+                            uint32_t* const argb, uint32_t* image,
+                            const WebPPicture* const pic, int percent_range,
+                            int* const percent) {
+  const int max_tile_size = 1 << bits;
+  const int tile_xsize = VP8LSubSampleSize(width, bits);
+  const int tile_ysize = VP8LSubSampleSize(height, bits);
+  int percent_start = *percent;
+  int accumulated_red_histo[256] = { 0 };
+  int accumulated_blue_histo[256] = { 0 };
+  int tile_x, tile_y;
+  VP8LMultipliers prev_x, prev_y;
+  MultipliersClear(&prev_y);
+  MultipliersClear(&prev_x);
+  for (tile_y = 0; tile_y < tile_ysize; ++tile_y) {
+    for (tile_x = 0; tile_x < tile_xsize; ++tile_x) {
+      int y;
+      const int tile_x_offset = tile_x * max_tile_size;
+      const int tile_y_offset = tile_y * max_tile_size;
+      const int all_x_max = GetMin(tile_x_offset + max_tile_size, width);
+      const int all_y_max = GetMin(tile_y_offset + max_tile_size, height);
+      const int offset = tile_y * tile_xsize + tile_x;
+      if (tile_y != 0) {
+        ColorCodeToMultipliers(image[offset - tile_xsize], &prev_y);
+      }
+      prev_x = GetBestColorTransformForTile(tile_x, tile_y, bits,
+                                            prev_x, prev_y,
+                                            quality, width, height,
+                                            accumulated_red_histo,
+                                            accumulated_blue_histo,
+                                            argb);
+      image[offset] = MultipliersToColorCode(&prev_x);
+      CopyTileWithColorTransform(width, height, tile_x_offset, tile_y_offset,
+                                 max_tile_size, prev_x, argb);
+
+      // Gather accumulated histogram data.
+      for (y = tile_y_offset; y < all_y_max; ++y) {
+        int ix = y * width + tile_x_offset;
+        const int ix_end = ix + all_x_max - tile_x_offset;
+        for (; ix < ix_end; ++ix) {
+          const uint32_t pix = argb[ix];
+          if (ix >= 2 &&
+              pix == argb[ix - 2] &&
+              pix == argb[ix - 1]) {
+            continue;  // repeated pixels are handled by backward references
+          }
+          if (ix >= width + 2 &&
+              argb[ix - 2] == argb[ix - width - 2] &&
+              argb[ix - 1] == argb[ix - width - 1] &&
+              pix == argb[ix - width]) {
+            continue;  // repeated pixels are handled by backward references
+          }
+          ++accumulated_red_histo[(pix >> 16) & 0xff];
+          ++accumulated_blue_histo[(pix >> 0) & 0xff];
+        }
+      }
+    }
+    if (!WebPReportProgress(
+            pic, percent_start + percent_range * tile_y / tile_ysize,
+            percent)) {
+      return 0;
+    }
+  }
+  return 1;
+}
diff --git a/media/libwebp/src/enc/quant_enc.c b/media/libwebp/src/enc/quant_enc.c
new file mode 100644
index 0000000000..6d8202d277
--- /dev/null
+++ b/media/libwebp/src/enc/quant_enc.c
@@ -0,0 +1,1398 @@
+// 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 "src/dsp/quant.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/enc/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 0       // I16 mode (special case)
+#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 uint16_t VP8I16ModeOffsets[4] = { I16DC16, I16TM16, I16VE16, I16HE16 };
+const uint16_t VP8UVModeOffsets[4] = { C8DC8, C8TM8, C8VE8, C8HE8 };
+
+// Must be indexed using {B_DC_PRED -> B_HU_PRED} as index
+const uint16_t 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 uint16_t 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 uint16_t 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* WEBP_RESTRICT const dst,
+                      const VP8ModeScore* WEBP_RESTRICT 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* WEBP_RESTRICT const dst,
+                     const VP8ModeScore* WEBP_RESTRICT 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;
+}
+
+// Coefficient type.
+enum { TYPE_I16_AC = 0, TYPE_I16_DC = 1, TYPE_CHROMA_A = 2, TYPE_I4_AC = 3 };
+
+static int TrellisQuantizeBlock(const VP8Encoder* WEBP_RESTRICT const enc,
+                                int16_t in[16], int16_t out[16],
+                                int ctx0, int coeff_type,
+                                const VP8Matrix* WEBP_RESTRICT 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 == TYPE_I16_AC) ? 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);
+      const 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;
+      int best_prev;
+      score_t cost, score;
+
+      ss_cur[m].costs = costs[n + 1][ctx];
+      if (level < 0 || level > thresh_level) {
+        ss_cur[m].score = MAX_COST;
+        // 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.
+      // The base_score is added to all scores so it is only added for the final
+      // value after the loop.
+      cost = VP8LevelCost(ss_prev[-MIN_DELTA].costs, level);
+      best_cur_score =
+          ss_prev[-MIN_DELTA].score + RDScoreTrellis(lambda, cost, 0);
+      best_prev = -MIN_DELTA;
+      for (p = -MIN_DELTA + 1; 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.
+        cost = VP8LevelCost(ss_prev[p].costs, level);
+        // Examine node assuming it's a non-terminal one.
+        score = ss_prev[p].score + RDScoreTrellis(lambda, cost, 0);
+        if (score < best_cur_score) {
+          best_cur_score = score;
+          best_prev = p;
+        }
+      }
+      best_cur_score += base_score;
+      // 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 && best_cur_score < best_score) {
+        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);
+        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
+  // Beware! We must preserve in[0]/out[0] value for TYPE_I16_AC case.
+  if (coeff_type == TYPE_I16_AC) {
+    memset(in + 1, 0, 15 * sizeof(*in));
+    memset(out + 1, 0, 15 * sizeof(*out));
+  } else {
+    memset(in, 0, 16 * sizeof(*in));
+    memset(out, 0, 16 * 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* WEBP_RESTRICT const it,
+                              VP8ModeScore* WEBP_RESTRICT const rd,
+                              uint8_t* WEBP_RESTRICT 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, TYPE_I16_AC, &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* WEBP_RESTRICT const it,
+                             int16_t levels[16],
+                             const uint8_t* WEBP_RESTRICT const src,
+                             uint8_t* WEBP_RESTRICT 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, TYPE_I4_AC, &dqm->y1_,
+                              dqm->lambda_trellis_i4_);
+  } else {
+    nz = VP8EncQuantizeBlock(tmp, levels, &dqm->y1_);
+  }
+  VP8ITransform(ref, tmp, yuv_out, 0);
+  return nz;
+}
+
+//------------------------------------------------------------------------------
+// DC-error diffusion
+
+// Diffusion weights. We under-correct a bit (15/16th of the error is actually
+// diffused) to avoid 'rainbow' chessboard pattern of blocks at q~=0.
+#define C1 7    // fraction of error sent to the 4x4 block below
+#define C2 8    // fraction of error sent to the 4x4 block on the right
+#define DSHIFT 4
+#define DSCALE 1   // storage descaling, needed to make the error fit int8_t
+
+// Quantize as usual, but also compute and return the quantization error.
+// Error is already divided by DSHIFT.
+static int QuantizeSingle(int16_t* WEBP_RESTRICT const v,
+                          const VP8Matrix* WEBP_RESTRICT const mtx) {
+  int V = *v;
+  const int sign = (V < 0);
+  if (sign) V = -V;
+  if (V > (int)mtx->zthresh_[0]) {
+    const int qV = QUANTDIV(V, mtx->iq_[0], mtx->bias_[0]) * mtx->q_[0];
+    const int err = (V - qV);
+    *v = sign ? -qV : qV;
+    return (sign ? -err : err) >> DSCALE;
+  }
+  *v = 0;
+  return (sign ? -V : V) >> DSCALE;
+}
+
+static void CorrectDCValues(const VP8EncIterator* WEBP_RESTRICT const it,
+                            const VP8Matrix* WEBP_RESTRICT const mtx,
+                            int16_t tmp[][16],
+                            VP8ModeScore* WEBP_RESTRICT const rd) {
+  //         | top[0] | top[1]
+  // --------+--------+---------
+  // left[0] | tmp[0]   tmp[1]  <->   err0 err1
+  // left[1] | tmp[2]   tmp[3]        err2 err3
+  //
+  // Final errors {err1,err2,err3} are preserved and later restored
+  // as top[]/left[] on the next block.
+  int ch;
+  for (ch = 0; ch <= 1; ++ch) {
+    const int8_t* const top = it->top_derr_[it->x_][ch];
+    const int8_t* const left = it->left_derr_[ch];
+    int16_t (* const c)[16] = &tmp[ch * 4];
+    int err0, err1, err2, err3;
+    c[0][0] += (C1 * top[0] + C2 * left[0]) >> (DSHIFT - DSCALE);
+    err0 = QuantizeSingle(&c[0][0], mtx);
+    c[1][0] += (C1 * top[1] + C2 * err0) >> (DSHIFT - DSCALE);
+    err1 = QuantizeSingle(&c[1][0], mtx);
+    c[2][0] += (C1 * err0 + C2 * left[1]) >> (DSHIFT - DSCALE);
+    err2 = QuantizeSingle(&c[2][0], mtx);
+    c[3][0] += (C1 * err1 + C2 * err2) >> (DSHIFT - DSCALE);
+    err3 = QuantizeSingle(&c[3][0], mtx);
+    // error 'err' is bounded by mtx->q_[0] which is 132 at max. Hence
+    // err >> DSCALE will fit in an int8_t type if DSCALE>=1.
+    assert(abs(err1) <= 127 && abs(err2) <= 127 && abs(err3) <= 127);
+    rd->derr[ch][0] = (int8_t)err1;
+    rd->derr[ch][1] = (int8_t)err2;
+    rd->derr[ch][2] = (int8_t)err3;
+  }
+}
+
+static void StoreDiffusionErrors(VP8EncIterator* WEBP_RESTRICT const it,
+                                 const VP8ModeScore* WEBP_RESTRICT const rd) {
+  int ch;
+  for (ch = 0; ch <= 1; ++ch) {
+    int8_t* const top = it->top_derr_[it->x_][ch];
+    int8_t* const left = it->left_derr_[ch];
+    left[0] = rd->derr[ch][0];            // restore err1
+    left[1] = 3 * rd->derr[ch][2] >> 2;   //     ... 3/4th of err3
+    top[0]  = rd->derr[ch][1];            //     ... err2
+    top[1]  = rd->derr[ch][2] - left[1];  //     ... 1/4th of err3.
+  }
+}
+
+#undef C1
+#undef C2
+#undef DSHIFT
+#undef DSCALE
+
+//------------------------------------------------------------------------------
+
+static int ReconstructUV(VP8EncIterator* WEBP_RESTRICT const it,
+                         VP8ModeScore* WEBP_RESTRICT const rd,
+                         uint8_t* WEBP_RESTRICT 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 (it->top_derr_ != NULL) CorrectDCValues(it, &dqm->uv_, tmp, rd);
+
+  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, TYPE_CHROMA_A, &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 void PickBestIntra16(VP8EncIterator* WEBP_RESTRICT const it,
+                            VP8ModeScore* WEBP_RESTRICT 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;
+  int is_flat = IsFlatSource16(it->yuv_in_ + Y_OFF_ENC);
+
+  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 (is_flat) {
+      // refine the first impression (which was in pixel space)
+      is_flat = IsFlat(rd_cur->y_ac_levels[0], kNumBlocks, FLATNESS_LIMIT_I16);
+      if (is_flat) {
+        // Block is very flat. We put emphasis on the distortion being very low!
+        rd_cur->D *= 2;
+        rd_cur->SD *= 2;
+      }
+    }
+
+    // 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* WEBP_RESTRICT 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* WEBP_RESTRICT const it,
+                          VP8ModeScore* WEBP_RESTRICT 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, to avoid flat area to be mispredicted
+      // by a complex mode.
+      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* WEBP_RESTRICT const it,
+                       VP8ModeScore* WEBP_RESTRICT 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));
+      if (it->top_derr_ != NULL) {
+        memcpy(rd->derr, rd_uv.derr, sizeof(rd_uv.derr));
+      }
+      SwapPtr(&dst, &tmp_dst);
+    }
+  }
+  VP8SetIntraUVMode(it, rd->mode_uv);
+  AddScore(rd, &rd_best);
+  if (dst != dst0) {   // copy 16x8 block if needed
+    VP8Copy16x8(dst, dst0);
+  }
+  if (it->top_derr_ != NULL) {  // store diffusion errors for next block
+    StoreDiffusionErrors(it, rd);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Final reconstruction and quantization.
+
+static void SimpleQuantize(VP8EncIterator* WEBP_RESTRICT const it,
+                           VP8ModeScore* WEBP_RESTRICT 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* WEBP_RESTRICT const it,
+                                  int try_both_modes, int refine_uv_mode,
+                                  VP8ModeScore* WEBP_RESTRICT 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 = (score_t)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;
+      }
+    }
+    if (it->x_ == 0 || it->y_ == 0) {
+      // avoid starting a checkerboard resonance from the border. See bug #432.
+      if (IsFlatSource16(src)) {
+        best_mode = (it->x_ == 0) ? 0 : 2;
+        try_both_modes = 0;  // stick to i16
+      }
+    }
+    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* WEBP_RESTRICT const it,
+                VP8ModeScore* WEBP_RESTRICT 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;
+}
diff --git a/media/libwebp/src/enc/syntax_enc.c b/media/libwebp/src/enc/syntax_enc.c
new file mode 100644
index 0000000000..e18cf650ca
--- /dev/null
+++ b/media/libwebp/src/enc/syntax_enc.c
@@ -0,0 +1,388 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Header syntax writing
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+
+#include "src/utils/utils.h"
+#include "src/webp/format_constants.h"  // RIFF constants
+#include "src/webp/mux_types.h"         // ALPHA_FLAG
+#include "src/enc/vp8i_enc.h"
+
+//------------------------------------------------------------------------------
+// Helper functions
+
+static int IsVP8XNeeded(const VP8Encoder* const enc) {
+  return !!enc->has_alpha_;  // Currently the only case when VP8X is needed.
+                             // This could change in the future.
+}
+
+static int PutPaddingByte(const WebPPicture* const pic) {
+  const uint8_t pad_byte[1] = { 0 };
+  return !!pic->writer(pad_byte, 1, pic);
+}
+
+//------------------------------------------------------------------------------
+// Writers for header's various pieces (in order of appearance)
+
+static WebPEncodingError PutRIFFHeader(const VP8Encoder* const enc,
+                                       size_t riff_size) {
+  const WebPPicture* const pic = enc->pic_;
+  uint8_t riff[RIFF_HEADER_SIZE] = {
+    'R', 'I', 'F', 'F', 0, 0, 0, 0, 'W', 'E', 'B', 'P'
+  };
+  assert(riff_size == (uint32_t)riff_size);
+  PutLE32(riff + TAG_SIZE, (uint32_t)riff_size);
+  if (!pic->writer(riff, sizeof(riff), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static WebPEncodingError PutVP8XHeader(const VP8Encoder* const enc) {
+  const WebPPicture* const pic = enc->pic_;
+  uint8_t vp8x[CHUNK_HEADER_SIZE + VP8X_CHUNK_SIZE] = {
+    'V', 'P', '8', 'X'
+  };
+  uint32_t flags = 0;
+
+  assert(IsVP8XNeeded(enc));
+  assert(pic->width >= 1 && pic->height >= 1);
+  assert(pic->width <= MAX_CANVAS_SIZE && pic->height <= MAX_CANVAS_SIZE);
+
+  if (enc->has_alpha_) {
+    flags |= ALPHA_FLAG;
+  }
+
+  PutLE32(vp8x + TAG_SIZE,              VP8X_CHUNK_SIZE);
+  PutLE32(vp8x + CHUNK_HEADER_SIZE,     flags);
+  PutLE24(vp8x + CHUNK_HEADER_SIZE + 4, pic->width - 1);
+  PutLE24(vp8x + CHUNK_HEADER_SIZE + 7, pic->height - 1);
+  if (!pic->writer(vp8x, sizeof(vp8x), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static WebPEncodingError PutAlphaChunk(const VP8Encoder* const enc) {
+  const WebPPicture* const pic = enc->pic_;
+  uint8_t alpha_chunk_hdr[CHUNK_HEADER_SIZE] = {
+    'A', 'L', 'P', 'H'
+  };
+
+  assert(enc->has_alpha_);
+
+  // Alpha chunk header.
+  PutLE32(alpha_chunk_hdr + TAG_SIZE, enc->alpha_data_size_);
+  if (!pic->writer(alpha_chunk_hdr, sizeof(alpha_chunk_hdr), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+
+  // Alpha chunk data.
+  if (!pic->writer(enc->alpha_data_, enc->alpha_data_size_, pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+
+  // Padding.
+  if ((enc->alpha_data_size_ & 1) && !PutPaddingByte(pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static WebPEncodingError PutVP8Header(const WebPPicture* const pic,
+                                      size_t vp8_size) {
+  uint8_t vp8_chunk_hdr[CHUNK_HEADER_SIZE] = {
+    'V', 'P', '8', ' '
+  };
+  assert(vp8_size == (uint32_t)vp8_size);
+  PutLE32(vp8_chunk_hdr + TAG_SIZE, (uint32_t)vp8_size);
+  if (!pic->writer(vp8_chunk_hdr, sizeof(vp8_chunk_hdr), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static WebPEncodingError PutVP8FrameHeader(const WebPPicture* const pic,
+                                           int profile, size_t size0) {
+  uint8_t vp8_frm_hdr[VP8_FRAME_HEADER_SIZE];
+  uint32_t bits;
+
+  if (size0 >= VP8_MAX_PARTITION0_SIZE) {  // partition #0 is too big to fit
+    return VP8_ENC_ERROR_PARTITION0_OVERFLOW;
+  }
+
+  // Paragraph 9.1.
+  bits = 0                         // keyframe (1b)
+       | (profile << 1)            // profile (3b)
+       | (1 << 4)                  // visible (1b)
+       | ((uint32_t)size0 << 5);   // partition length (19b)
+  vp8_frm_hdr[0] = (bits >>  0) & 0xff;
+  vp8_frm_hdr[1] = (bits >>  8) & 0xff;
+  vp8_frm_hdr[2] = (bits >> 16) & 0xff;
+  // signature
+  vp8_frm_hdr[3] = (VP8_SIGNATURE >> 16) & 0xff;
+  vp8_frm_hdr[4] = (VP8_SIGNATURE >>  8) & 0xff;
+  vp8_frm_hdr[5] = (VP8_SIGNATURE >>  0) & 0xff;
+  // dimensions
+  vp8_frm_hdr[6] = pic->width & 0xff;
+  vp8_frm_hdr[7] = pic->width >> 8;
+  vp8_frm_hdr[8] = pic->height & 0xff;
+  vp8_frm_hdr[9] = pic->height >> 8;
+
+  if (!pic->writer(vp8_frm_hdr, sizeof(vp8_frm_hdr), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+// WebP Headers.
+static int PutWebPHeaders(const VP8Encoder* const enc, size_t size0,
+                          size_t vp8_size, size_t riff_size) {
+  WebPPicture* const pic = enc->pic_;
+  WebPEncodingError err = VP8_ENC_OK;
+
+  // RIFF header.
+  err = PutRIFFHeader(enc, riff_size);
+  if (err != VP8_ENC_OK) goto Error;
+
+  // VP8X.
+  if (IsVP8XNeeded(enc)) {
+    err = PutVP8XHeader(enc);
+    if (err != VP8_ENC_OK) goto Error;
+  }
+
+  // Alpha.
+  if (enc->has_alpha_) {
+    err = PutAlphaChunk(enc);
+    if (err != VP8_ENC_OK) goto Error;
+  }
+
+  // VP8 header.
+  err = PutVP8Header(pic, vp8_size);
+  if (err != VP8_ENC_OK) goto Error;
+
+  // VP8 frame header.
+  err = PutVP8FrameHeader(pic, enc->profile_, size0);
+  if (err != VP8_ENC_OK) goto Error;
+
+  // All OK.
+  return 1;
+
+  // Error.
+ Error:
+  return WebPEncodingSetError(pic, err);
+}
+
+// Segmentation header
+static void PutSegmentHeader(VP8BitWriter* const bw,
+                             const VP8Encoder* const enc) {
+  const VP8EncSegmentHeader* const hdr = &enc->segment_hdr_;
+  const VP8EncProba* const proba = &enc->proba_;
+  if (VP8PutBitUniform(bw, (hdr->num_segments_ > 1))) {
+    // We always 'update' the quant and filter strength values
+    const int update_data = 1;
+    int s;
+    VP8PutBitUniform(bw, hdr->update_map_);
+    if (VP8PutBitUniform(bw, update_data)) {
+      // we always use absolute values, not relative ones
+      VP8PutBitUniform(bw, 1);   // (segment_feature_mode = 1. Paragraph 9.3.)
+      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+        VP8PutSignedBits(bw, enc->dqm_[s].quant_, 7);
+      }
+      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+        VP8PutSignedBits(bw, enc->dqm_[s].fstrength_, 6);
+      }
+    }
+    if (hdr->update_map_) {
+      for (s = 0; s < 3; ++s) {
+        if (VP8PutBitUniform(bw, (proba->segments_[s] != 255u))) {
+          VP8PutBits(bw, proba->segments_[s], 8);
+        }
+      }
+    }
+  }
+}
+
+// Filtering parameters header
+static void PutFilterHeader(VP8BitWriter* const bw,
+                            const VP8EncFilterHeader* const hdr) {
+  const int use_lf_delta = (hdr->i4x4_lf_delta_ != 0);
+  VP8PutBitUniform(bw, hdr->simple_);
+  VP8PutBits(bw, hdr->level_, 6);
+  VP8PutBits(bw, hdr->sharpness_, 3);
+  if (VP8PutBitUniform(bw, use_lf_delta)) {
+    // '0' is the default value for i4x4_lf_delta_ at frame #0.
+    const int need_update = (hdr->i4x4_lf_delta_ != 0);
+    if (VP8PutBitUniform(bw, need_update)) {
+      // we don't use ref_lf_delta => emit four 0 bits
+      VP8PutBits(bw, 0, 4);
+      // we use mode_lf_delta for i4x4
+      VP8PutSignedBits(bw, hdr->i4x4_lf_delta_, 6);
+      VP8PutBits(bw, 0, 3);    // all others unused
+    }
+  }
+}
+
+// Nominal quantization parameters
+static void PutQuant(VP8BitWriter* const bw,
+                     const VP8Encoder* const enc) {
+  VP8PutBits(bw, enc->base_quant_, 7);
+  VP8PutSignedBits(bw, enc->dq_y1_dc_, 4);
+  VP8PutSignedBits(bw, enc->dq_y2_dc_, 4);
+  VP8PutSignedBits(bw, enc->dq_y2_ac_, 4);
+  VP8PutSignedBits(bw, enc->dq_uv_dc_, 4);
+  VP8PutSignedBits(bw, enc->dq_uv_ac_, 4);
+}
+
+// Partition sizes
+static int EmitPartitionsSize(const VP8Encoder* const enc,
+                              WebPPicture* const pic) {
+  uint8_t buf[3 * (MAX_NUM_PARTITIONS - 1)];
+  int p;
+  for (p = 0; p < enc->num_parts_ - 1; ++p) {
+    const size_t part_size = VP8BitWriterSize(enc->parts_ + p);
+    if (part_size >= VP8_MAX_PARTITION_SIZE) {
+      return WebPEncodingSetError(pic, VP8_ENC_ERROR_PARTITION_OVERFLOW);
+    }
+    buf[3 * p + 0] = (part_size >>  0) & 0xff;
+    buf[3 * p + 1] = (part_size >>  8) & 0xff;
+    buf[3 * p + 2] = (part_size >> 16) & 0xff;
+  }
+  return p ? pic->writer(buf, 3 * p, pic) : 1;
+}
+
+//------------------------------------------------------------------------------
+
+static int GeneratePartition0(VP8Encoder* const enc) {
+  VP8BitWriter* const bw = &enc->bw_;
+  const int mb_size = enc->mb_w_ * enc->mb_h_;
+  uint64_t pos1, pos2, pos3;
+
+  pos1 = VP8BitWriterPos(bw);
+  if (!VP8BitWriterInit(bw, mb_size * 7 / 8)) {        // ~7 bits per macroblock
+    return WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  VP8PutBitUniform(bw, 0);   // colorspace
+  VP8PutBitUniform(bw, 0);   // clamp type
+
+  PutSegmentHeader(bw, enc);
+  PutFilterHeader(bw, &enc->filter_hdr_);
+  VP8PutBits(bw, enc->num_parts_ == 8 ? 3 :
+                 enc->num_parts_ == 4 ? 2 :
+                 enc->num_parts_ == 2 ? 1 : 0, 2);
+  PutQuant(bw, enc);
+  VP8PutBitUniform(bw, 0);   // no proba update
+  VP8WriteProbas(bw, &enc->proba_);
+  pos2 = VP8BitWriterPos(bw);
+  VP8CodeIntraModes(enc);
+  VP8BitWriterFinish(bw);
+
+  pos3 = VP8BitWriterPos(bw);
+
+#if !defined(WEBP_DISABLE_STATS)
+  if (enc->pic_->stats) {
+    enc->pic_->stats->header_bytes[0] = (int)((pos2 - pos1 + 7) >> 3);
+    enc->pic_->stats->header_bytes[1] = (int)((pos3 - pos2 + 7) >> 3);
+    enc->pic_->stats->alpha_data_size = (int)enc->alpha_data_size_;
+  }
+#else
+  (void)pos1;
+  (void)pos2;
+  (void)pos3;
+#endif
+  if (bw->error_) {
+    return WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  return 1;
+}
+
+void VP8EncFreeBitWriters(VP8Encoder* const enc) {
+  int p;
+  VP8BitWriterWipeOut(&enc->bw_);
+  for (p = 0; p < enc->num_parts_; ++p) {
+    VP8BitWriterWipeOut(enc->parts_ + p);
+  }
+}
+
+int VP8EncWrite(VP8Encoder* const enc) {
+  WebPPicture* const pic = enc->pic_;
+  VP8BitWriter* const bw = &enc->bw_;
+  const int task_percent = 19;
+  const int percent_per_part = task_percent / enc->num_parts_;
+  const int final_percent = enc->percent_ + task_percent;
+  int ok = 0;
+  size_t vp8_size, pad, riff_size;
+  int p;
+
+  // Partition #0 with header and partition sizes
+  ok = GeneratePartition0(enc);
+  if (!ok) return 0;
+
+  // Compute VP8 size
+  vp8_size = VP8_FRAME_HEADER_SIZE +
+             VP8BitWriterSize(bw) +
+             3 * (enc->num_parts_ - 1);
+  for (p = 0; p < enc->num_parts_; ++p) {
+    vp8_size += VP8BitWriterSize(enc->parts_ + p);
+  }
+  pad = vp8_size & 1;
+  vp8_size += pad;
+
+  // Compute RIFF size
+  // At the minimum it is: "WEBPVP8 nnnn" + VP8 data size.
+  riff_size = TAG_SIZE + CHUNK_HEADER_SIZE + vp8_size;
+  if (IsVP8XNeeded(enc)) {  // Add size for: VP8X header + data.
+    riff_size += CHUNK_HEADER_SIZE + VP8X_CHUNK_SIZE;
+  }
+  if (enc->has_alpha_) {  // Add size for: ALPH header + data.
+    const uint32_t padded_alpha_size = enc->alpha_data_size_ +
+                                       (enc->alpha_data_size_ & 1);
+    riff_size += CHUNK_HEADER_SIZE + padded_alpha_size;
+  }
+  // RIFF size should fit in 32-bits.
+  if (riff_size > 0xfffffffeU) {
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_FILE_TOO_BIG);
+  }
+
+  // Emit headers and partition #0
+  {
+    const uint8_t* const part0 = VP8BitWriterBuf(bw);
+    const size_t size0 = VP8BitWriterSize(bw);
+    ok = ok && PutWebPHeaders(enc, size0, vp8_size, riff_size)
+            && pic->writer(part0, size0, pic)
+            && EmitPartitionsSize(enc, pic);
+    VP8BitWriterWipeOut(bw);    // will free the internal buffer.
+  }
+
+  // Token partitions
+  for (p = 0; p < enc->num_parts_; ++p) {
+    const uint8_t* const buf = VP8BitWriterBuf(enc->parts_ + p);
+    const size_t size = VP8BitWriterSize(enc->parts_ + p);
+    if (size) ok = ok && pic->writer(buf, size, pic);
+    VP8BitWriterWipeOut(enc->parts_ + p);    // will free the internal buffer.
+    ok = ok && WebPReportProgress(pic, enc->percent_ + percent_per_part,
+                                  &enc->percent_);
+  }
+
+  // Padding byte
+  if (ok && pad) {
+    ok = PutPaddingByte(pic);
+  }
+
+  enc->coded_size_ = (int)(CHUNK_HEADER_SIZE + riff_size);
+  ok = ok && WebPReportProgress(pic, final_percent, &enc->percent_);
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/media/libwebp/src/enc/token_enc.c b/media/libwebp/src/enc/token_enc.c
new file mode 100644
index 0000000000..3a2192acac
--- /dev/null
+++ b/media/libwebp/src/enc/token_enc.c
@@ -0,0 +1,262 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Paginated token buffer
+//
+//  A 'token' is a bit value associated with a probability, either fixed
+// or a later-to-be-determined after statistics have been collected.
+// For dynamic probability, we just record the slot id (idx) for the probability
+// value in the final probability array (uint8_t* probas in VP8EmitTokens).
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "src/enc/cost_enc.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/utils/utils.h"
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+// we use pages to reduce the number of memcpy()
+#define MIN_PAGE_SIZE 8192          // minimum number of token per page
+#define FIXED_PROBA_BIT (1u << 14)
+
+typedef uint16_t token_t;  // bit #15: bit value
+                           // bit #14: flags for constant proba or idx
+                           // bits #0..13: slot or constant proba
+struct VP8Tokens {
+  VP8Tokens* next_;        // pointer to next page
+};
+// Token data is located in memory just after the next_ field.
+// This macro is used to return their address and hide the trick.
+#define TOKEN_DATA(p) ((const token_t*)&(p)[1])
+
+//------------------------------------------------------------------------------
+
+void VP8TBufferInit(VP8TBuffer* const b, int page_size) {
+  b->tokens_ = NULL;
+  b->pages_ = NULL;
+  b->last_page_ = &b->pages_;
+  b->left_ = 0;
+  b->page_size_ = (page_size < MIN_PAGE_SIZE) ? MIN_PAGE_SIZE : page_size;
+  b->error_ = 0;
+}
+
+void VP8TBufferClear(VP8TBuffer* const b) {
+  if (b != NULL) {
+    VP8Tokens* p = b->pages_;
+    while (p != NULL) {
+      VP8Tokens* const next = p->next_;
+      WebPSafeFree(p);
+      p = next;
+    }
+    VP8TBufferInit(b, b->page_size_);
+  }
+}
+
+static int TBufferNewPage(VP8TBuffer* const b) {
+  VP8Tokens* page = NULL;
+  if (!b->error_) {
+    const size_t size = sizeof(*page) + b->page_size_ * sizeof(token_t);
+    page = (VP8Tokens*)WebPSafeMalloc(1ULL, size);
+  }
+  if (page == NULL) {
+    b->error_ = 1;
+    return 0;
+  }
+  page->next_ = NULL;
+
+  *b->last_page_ = page;
+  b->last_page_ = &page->next_;
+  b->left_ = b->page_size_;
+  b->tokens_ = (token_t*)TOKEN_DATA(page);
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+#define TOKEN_ID(t, b, ctx) \
+    (NUM_PROBAS * ((ctx) + NUM_CTX * ((b) + NUM_BANDS * (t))))
+
+static WEBP_INLINE uint32_t AddToken(VP8TBuffer* const b, uint32_t bit,
+                                     uint32_t proba_idx,
+                                     proba_t* const stats) {
+  assert(proba_idx < FIXED_PROBA_BIT);
+  assert(bit <= 1);
+  if (b->left_ > 0 || TBufferNewPage(b)) {
+    const int slot = --b->left_;
+    b->tokens_[slot] = (bit << 15) | proba_idx;
+  }
+  VP8RecordStats(bit, stats);
+  return bit;
+}
+
+static WEBP_INLINE void AddConstantToken(VP8TBuffer* const b,
+                                         uint32_t bit, uint32_t proba) {
+  assert(proba < 256);
+  assert(bit <= 1);
+  if (b->left_ > 0 || TBufferNewPage(b)) {
+    const int slot = --b->left_;
+    b->tokens_[slot] = (bit << 15) | FIXED_PROBA_BIT | proba;
+  }
+}
+
+int VP8RecordCoeffTokens(int ctx, const struct VP8Residual* const res,
+                         VP8TBuffer* const tokens) {
+  const int16_t* const coeffs = res->coeffs;
+  const int coeff_type = res->coeff_type;
+  const int last = res->last;
+  int n = res->first;
+  uint32_t base_id = TOKEN_ID(coeff_type, n, ctx);
+  // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  proba_t* s = res->stats[n][ctx];
+  if (!AddToken(tokens, last >= 0, base_id + 0, s + 0)) {
+    return 0;
+  }
+
+  while (n < 16) {
+    const int c = coeffs[n++];
+    const int sign = c < 0;
+    const uint32_t v = sign ? -c : c;
+    if (!AddToken(tokens, v != 0, base_id + 1, s + 1)) {
+      base_id = TOKEN_ID(coeff_type, VP8EncBands[n], 0);  // ctx=0
+      s = res->stats[VP8EncBands[n]][0];
+      continue;
+    }
+    if (!AddToken(tokens, v > 1, base_id + 2, s + 2)) {
+      base_id = TOKEN_ID(coeff_type, VP8EncBands[n], 1);  // ctx=1
+      s = res->stats[VP8EncBands[n]][1];
+    } else {
+      if (!AddToken(tokens, v > 4, base_id + 3, s + 3)) {
+        if (AddToken(tokens, v != 2, base_id + 4, s + 4)) {
+          AddToken(tokens, v == 4, base_id + 5, s + 5);
+        }
+      } else if (!AddToken(tokens, v > 10, base_id + 6, s + 6)) {
+        if (!AddToken(tokens, v > 6, base_id + 7, s + 7)) {
+          AddConstantToken(tokens, v == 6, 159);
+        } else {
+          AddConstantToken(tokens, v >= 9, 165);
+          AddConstantToken(tokens, !(v & 1), 145);
+        }
+      } else {
+        int mask;
+        const uint8_t* tab;
+        uint32_t residue = v - 3;
+        if (residue < (8 << 1)) {          // VP8Cat3  (3b)
+          AddToken(tokens, 0, base_id + 8, s + 8);
+          AddToken(tokens, 0, base_id + 9, s + 9);
+          residue -= (8 << 0);
+          mask = 1 << 2;
+          tab = VP8Cat3;
+        } else if (residue < (8 << 2)) {   // VP8Cat4  (4b)
+          AddToken(tokens, 0, base_id + 8, s + 8);
+          AddToken(tokens, 1, base_id + 9, s + 9);
+          residue -= (8 << 1);
+          mask = 1 << 3;
+          tab = VP8Cat4;
+        } else if (residue < (8 << 3)) {   // VP8Cat5  (5b)
+          AddToken(tokens, 1, base_id + 8, s + 8);
+          AddToken(tokens, 0, base_id + 10, s + 9);
+          residue -= (8 << 2);
+          mask = 1 << 4;
+          tab = VP8Cat5;
+        } else {                         // VP8Cat6 (11b)
+          AddToken(tokens, 1, base_id + 8, s + 8);
+          AddToken(tokens, 1, base_id + 10, s + 9);
+          residue -= (8 << 3);
+          mask = 1 << 10;
+          tab = VP8Cat6;
+        }
+        while (mask) {
+          AddConstantToken(tokens, !!(residue & mask), *tab++);
+          mask >>= 1;
+        }
+      }
+      base_id = TOKEN_ID(coeff_type, VP8EncBands[n], 2);  // ctx=2
+      s = res->stats[VP8EncBands[n]][2];
+    }
+    AddConstantToken(tokens, sign, 128);
+    if (n == 16 || !AddToken(tokens, n <= last, base_id + 0, s + 0)) {
+      return 1;   // EOB
+    }
+  }
+  return 1;
+}
+
+#undef TOKEN_ID
+
+//------------------------------------------------------------------------------
+// Final coding pass, with known probabilities
+
+int VP8EmitTokens(VP8TBuffer* const b, VP8BitWriter* const bw,
+                  const uint8_t* const probas, int final_pass) {
+  const VP8Tokens* p = b->pages_;
+  assert(!b->error_);
+  while (p != NULL) {
+    const VP8Tokens* const next = p->next_;
+    const int N = (next == NULL) ? b->left_ : 0;
+    int n = b->page_size_;
+    const token_t* const tokens = TOKEN_DATA(p);
+    while (n-- > N) {
+      const token_t token = tokens[n];
+      const int bit = (token >> 15) & 1;
+      if (token & FIXED_PROBA_BIT) {
+        VP8PutBit(bw, bit, token & 0xffu);  // constant proba
+      } else {
+        VP8PutBit(bw, bit, probas[token & 0x3fffu]);
+      }
+    }
+    if (final_pass) WebPSafeFree((void*)p);
+    p = next;
+  }
+  if (final_pass) b->pages_ = NULL;
+  return 1;
+}
+
+// Size estimation
+size_t VP8EstimateTokenSize(VP8TBuffer* const b, const uint8_t* const probas) {
+  size_t size = 0;
+  const VP8Tokens* p = b->pages_;
+  assert(!b->error_);
+  while (p != NULL) {
+    const VP8Tokens* const next = p->next_;
+    const int N = (next == NULL) ? b->left_ : 0;
+    int n = b->page_size_;
+    const token_t* const tokens = TOKEN_DATA(p);
+    while (n-- > N) {
+      const token_t token = tokens[n];
+      const int bit = token & (1 << 15);
+      if (token & FIXED_PROBA_BIT) {
+        size += VP8BitCost(bit, token & 0xffu);
+      } else {
+        size += VP8BitCost(bit, probas[token & 0x3fffu]);
+      }
+    }
+    p = next;
+  }
+  return size;
+}
+
+//------------------------------------------------------------------------------
+
+#else     // DISABLE_TOKEN_BUFFER
+
+void VP8TBufferInit(VP8TBuffer* const b, int page_size) {
+  (void)b;
+  (void)page_size;
+}
+void VP8TBufferClear(VP8TBuffer* const b) {
+  (void)b;
+}
+
+#endif    // !DISABLE_TOKEN_BUFFER
+
diff --git a/media/libwebp/src/enc/tree_enc.c b/media/libwebp/src/enc/tree_enc.c
new file mode 100644
index 0000000000..64ed28360b
--- /dev/null
+++ b/media/libwebp/src/enc/tree_enc.c
@@ -0,0 +1,504 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Coding of token probabilities, intra modes and segments.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "src/enc/vp8i_enc.h"
+
+//------------------------------------------------------------------------------
+// Default probabilities
+
+// Paragraph 13.5
+const uint8_t
+  VP8CoeffsProba0[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS] = {
+  { { { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 253, 136, 254, 255, 228, 219, 128, 128, 128, 128, 128 },
+      { 189, 129, 242, 255, 227, 213, 255, 219, 128, 128, 128 },
+      { 106, 126, 227, 252, 214, 209, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 98, 248, 255, 236, 226, 255, 255, 128, 128, 128 },
+      { 181, 133, 238, 254, 221, 234, 255, 154, 128, 128, 128 },
+      { 78, 134, 202, 247, 198, 180, 255, 219, 128, 128, 128 },
+    },
+    { { 1, 185, 249, 255, 243, 255, 128, 128, 128, 128, 128 },
+      { 184, 150, 247, 255, 236, 224, 128, 128, 128, 128, 128 },
+      { 77, 110, 216, 255, 236, 230, 128, 128, 128, 128, 128 },
+    },
+    { { 1, 101, 251, 255, 241, 255, 128, 128, 128, 128, 128 },
+      { 170, 139, 241, 252, 236, 209, 255, 255, 128, 128, 128 },
+      { 37, 116, 196, 243, 228, 255, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 204, 254, 255, 245, 255, 128, 128, 128, 128, 128 },
+      { 207, 160, 250, 255, 238, 128, 128, 128, 128, 128, 128 },
+      { 102, 103, 231, 255, 211, 171, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 152, 252, 255, 240, 255, 128, 128, 128, 128, 128 },
+      { 177, 135, 243, 255, 234, 225, 128, 128, 128, 128, 128 },
+      { 80, 129, 211, 255, 194, 224, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 246, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 255, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 198, 35, 237, 223, 193, 187, 162, 160, 145, 155, 62 },
+      { 131, 45, 198, 221, 172, 176, 220, 157, 252, 221, 1 },
+      { 68, 47, 146, 208, 149, 167, 221, 162, 255, 223, 128 }
+    },
+    { { 1, 149, 241, 255, 221, 224, 255, 255, 128, 128, 128 },
+      { 184, 141, 234, 253, 222, 220, 255, 199, 128, 128, 128 },
+      { 81, 99, 181, 242, 176, 190, 249, 202, 255, 255, 128 }
+    },
+    { { 1, 129, 232, 253, 214, 197, 242, 196, 255, 255, 128 },
+      { 99, 121, 210, 250, 201, 198, 255, 202, 128, 128, 128 },
+      { 23, 91, 163, 242, 170, 187, 247, 210, 255, 255, 128 }
+    },
+    { { 1, 200, 246, 255, 234, 255, 128, 128, 128, 128, 128 },
+      { 109, 178, 241, 255, 231, 245, 255, 255, 128, 128, 128 },
+      { 44, 130, 201, 253, 205, 192, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 132, 239, 251, 219, 209, 255, 165, 128, 128, 128 },
+      { 94, 136, 225, 251, 218, 190, 255, 255, 128, 128, 128 },
+      { 22, 100, 174, 245, 186, 161, 255, 199, 128, 128, 128 }
+    },
+    { { 1, 182, 249, 255, 232, 235, 128, 128, 128, 128, 128 },
+      { 124, 143, 241, 255, 227, 234, 128, 128, 128, 128, 128 },
+      { 35, 77, 181, 251, 193, 211, 255, 205, 128, 128, 128 }
+    },
+    { { 1, 157, 247, 255, 236, 231, 255, 255, 128, 128, 128 },
+      { 121, 141, 235, 255, 225, 227, 255, 255, 128, 128, 128 },
+      { 45, 99, 188, 251, 195, 217, 255, 224, 128, 128, 128 }
+    },
+    { { 1, 1, 251, 255, 213, 255, 128, 128, 128, 128, 128 },
+      { 203, 1, 248, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 137, 1, 177, 255, 224, 255, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 253, 9, 248, 251, 207, 208, 255, 192, 128, 128, 128 },
+      { 175, 13, 224, 243, 193, 185, 249, 198, 255, 255, 128 },
+      { 73, 17, 171, 221, 161, 179, 236, 167, 255, 234, 128 }
+    },
+    { { 1, 95, 247, 253, 212, 183, 255, 255, 128, 128, 128 },
+      { 239, 90, 244, 250, 211, 209, 255, 255, 128, 128, 128 },
+      { 155, 77, 195, 248, 188, 195, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 24, 239, 251, 218, 219, 255, 205, 128, 128, 128 },
+      { 201, 51, 219, 255, 196, 186, 128, 128, 128, 128, 128 },
+      { 69, 46, 190, 239, 201, 218, 255, 228, 128, 128, 128 }
+    },
+    { { 1, 191, 251, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 223, 165, 249, 255, 213, 255, 128, 128, 128, 128, 128 },
+      { 141, 124, 248, 255, 255, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 16, 248, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 190, 36, 230, 255, 236, 255, 128, 128, 128, 128, 128 },
+      { 149, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 226, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 247, 192, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 240, 128, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 134, 252, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 213, 62, 250, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 55, 93, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 202, 24, 213, 235, 186, 191, 220, 160, 240, 175, 255 },
+      { 126, 38, 182, 232, 169, 184, 228, 174, 255, 187, 128 },
+      { 61, 46, 138, 219, 151, 178, 240, 170, 255, 216, 128 }
+    },
+    { { 1, 112, 230, 250, 199, 191, 247, 159, 255, 255, 128 },
+      { 166, 109, 228, 252, 211, 215, 255, 174, 128, 128, 128 },
+      { 39, 77, 162, 232, 172, 180, 245, 178, 255, 255, 128 }
+    },
+    { { 1, 52, 220, 246, 198, 199, 249, 220, 255, 255, 128 },
+      { 124, 74, 191, 243, 183, 193, 250, 221, 255, 255, 128 },
+      { 24, 71, 130, 219, 154, 170, 243, 182, 255, 255, 128 }
+    },
+    { { 1, 182, 225, 249, 219, 240, 255, 224, 128, 128, 128 },
+      { 149, 150, 226, 252, 216, 205, 255, 171, 128, 128, 128 },
+      { 28, 108, 170, 242, 183, 194, 254, 223, 255, 255, 128 }
+    },
+    { { 1, 81, 230, 252, 204, 203, 255, 192, 128, 128, 128 },
+      { 123, 102, 209, 247, 188, 196, 255, 233, 128, 128, 128 },
+      { 20, 95, 153, 243, 164, 173, 255, 203, 128, 128, 128 }
+    },
+    { { 1, 222, 248, 255, 216, 213, 128, 128, 128, 128, 128 },
+      { 168, 175, 246, 252, 235, 205, 255, 255, 128, 128, 128 },
+      { 47, 116, 215, 255, 211, 212, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 121, 236, 253, 212, 214, 255, 255, 128, 128, 128 },
+      { 141, 84, 213, 252, 201, 202, 255, 219, 128, 128, 128 },
+      { 42, 80, 160, 240, 162, 185, 255, 205, 128, 128, 128 }
+    },
+    { { 1, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 244, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 238, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  }
+};
+
+void VP8DefaultProbas(VP8Encoder* const enc) {
+  VP8EncProba* const probas = &enc->proba_;
+  probas->use_skip_proba_ = 0;
+  memset(probas->segments_, 255u, sizeof(probas->segments_));
+  memcpy(probas->coeffs_, VP8CoeffsProba0, sizeof(VP8CoeffsProba0));
+  // Note: we could hard-code the level_costs_ corresponding to VP8CoeffsProba0,
+  // but that's ~11k of static data. Better call VP8CalculateLevelCosts() later.
+  probas->dirty_ = 1;
+}
+
+// Paragraph 11.5.  900bytes.
+static const uint8_t kBModesProba[NUM_BMODES][NUM_BMODES][NUM_BMODES - 1] = {
+  { { 231, 120, 48, 89, 115, 113, 120, 152, 112 },
+    { 152, 179, 64, 126, 170, 118, 46, 70, 95 },
+    { 175, 69, 143, 80, 85, 82, 72, 155, 103 },
+    { 56, 58, 10, 171, 218, 189, 17, 13, 152 },
+    { 114, 26, 17, 163, 44, 195, 21, 10, 173 },
+    { 121, 24, 80, 195, 26, 62, 44, 64, 85 },
+    { 144, 71, 10, 38, 171, 213, 144, 34, 26 },
+    { 170, 46, 55, 19, 136, 160, 33, 206, 71 },
+    { 63, 20, 8, 114, 114, 208, 12, 9, 226 },
+    { 81, 40, 11, 96, 182, 84, 29, 16, 36 } },
+  { { 134, 183, 89, 137, 98, 101, 106, 165, 148 },
+    { 72, 187, 100, 130, 157, 111, 32, 75, 80 },
+    { 66, 102, 167, 99, 74, 62, 40, 234, 128 },
+    { 41, 53, 9, 178, 241, 141, 26, 8, 107 },
+    { 74, 43, 26, 146, 73, 166, 49, 23, 157 },
+    { 65, 38, 105, 160, 51, 52, 31, 115, 128 },
+    { 104, 79, 12, 27, 217, 255, 87, 17, 7 },
+    { 87, 68, 71, 44, 114, 51, 15, 186, 23 },
+    { 47, 41, 14, 110, 182, 183, 21, 17, 194 },
+    { 66, 45, 25, 102, 197, 189, 23, 18, 22 } },
+  { { 88, 88, 147, 150, 42, 46, 45, 196, 205 },
+    { 43, 97, 183, 117, 85, 38, 35, 179, 61 },
+    { 39, 53, 200, 87, 26, 21, 43, 232, 171 },
+    { 56, 34, 51, 104, 114, 102, 29, 93, 77 },
+    { 39, 28, 85, 171, 58, 165, 90, 98, 64 },
+    { 34, 22, 116, 206, 23, 34, 43, 166, 73 },
+    { 107, 54, 32, 26, 51, 1, 81, 43, 31 },
+    { 68, 25, 106, 22, 64, 171, 36, 225, 114 },
+    { 34, 19, 21, 102, 132, 188, 16, 76, 124 },
+    { 62, 18, 78, 95, 85, 57, 50, 48, 51 } },
+  { { 193, 101, 35, 159, 215, 111, 89, 46, 111 },
+    { 60, 148, 31, 172, 219, 228, 21, 18, 111 },
+    { 112, 113, 77, 85, 179, 255, 38, 120, 114 },
+    { 40, 42, 1, 196, 245, 209, 10, 25, 109 },
+    { 88, 43, 29, 140, 166, 213, 37, 43, 154 },
+    { 61, 63, 30, 155, 67, 45, 68, 1, 209 },
+    { 100, 80, 8, 43, 154, 1, 51, 26, 71 },
+    { 142, 78, 78, 16, 255, 128, 34, 197, 171 },
+    { 41, 40, 5, 102, 211, 183, 4, 1, 221 },
+    { 51, 50, 17, 168, 209, 192, 23, 25, 82 } },
+  { { 138, 31, 36, 171, 27, 166, 38, 44, 229 },
+    { 67, 87, 58, 169, 82, 115, 26, 59, 179 },
+    { 63, 59, 90, 180, 59, 166, 93, 73, 154 },
+    { 40, 40, 21, 116, 143, 209, 34, 39, 175 },
+    { 47, 15, 16, 183, 34, 223, 49, 45, 183 },
+    { 46, 17, 33, 183, 6, 98, 15, 32, 183 },
+    { 57, 46, 22, 24, 128, 1, 54, 17, 37 },
+    { 65, 32, 73, 115, 28, 128, 23, 128, 205 },
+    { 40, 3, 9, 115, 51, 192, 18, 6, 223 },
+    { 87, 37, 9, 115, 59, 77, 64, 21, 47 } },
+  { { 104, 55, 44, 218, 9, 54, 53, 130, 226 },
+    { 64, 90, 70, 205, 40, 41, 23, 26, 57 },
+    { 54, 57, 112, 184, 5, 41, 38, 166, 213 },
+    { 30, 34, 26, 133, 152, 116, 10, 32, 134 },
+    { 39, 19, 53, 221, 26, 114, 32, 73, 255 },
+    { 31, 9, 65, 234, 2, 15, 1, 118, 73 },
+    { 75, 32, 12, 51, 192, 255, 160, 43, 51 },
+    { 88, 31, 35, 67, 102, 85, 55, 186, 85 },
+    { 56, 21, 23, 111, 59, 205, 45, 37, 192 },
+    { 55, 38, 70, 124, 73, 102, 1, 34, 98 } },
+  { { 125, 98, 42, 88, 104, 85, 117, 175, 82 },
+    { 95, 84, 53, 89, 128, 100, 113, 101, 45 },
+    { 75, 79, 123, 47, 51, 128, 81, 171, 1 },
+    { 57, 17, 5, 71, 102, 57, 53, 41, 49 },
+    { 38, 33, 13, 121, 57, 73, 26, 1, 85 },
+    { 41, 10, 67, 138, 77, 110, 90, 47, 114 },
+    { 115, 21, 2, 10, 102, 255, 166, 23, 6 },
+    { 101, 29, 16, 10, 85, 128, 101, 196, 26 },
+    { 57, 18, 10, 102, 102, 213, 34, 20, 43 },
+    { 117, 20, 15, 36, 163, 128, 68, 1, 26 } },
+  { { 102, 61, 71, 37, 34, 53, 31, 243, 192 },
+    { 69, 60, 71, 38, 73, 119, 28, 222, 37 },
+    { 68, 45, 128, 34, 1, 47, 11, 245, 171 },
+    { 62, 17, 19, 70, 146, 85, 55, 62, 70 },
+    { 37, 43, 37, 154, 100, 163, 85, 160, 1 },
+    { 63, 9, 92, 136, 28, 64, 32, 201, 85 },
+    { 75, 15, 9, 9, 64, 255, 184, 119, 16 },
+    { 86, 6, 28, 5, 64, 255, 25, 248, 1 },
+    { 56, 8, 17, 132, 137, 255, 55, 116, 128 },
+    { 58, 15, 20, 82, 135, 57, 26, 121, 40 } },
+  { { 164, 50, 31, 137, 154, 133, 25, 35, 218 },
+    { 51, 103, 44, 131, 131, 123, 31, 6, 158 },
+    { 86, 40, 64, 135, 148, 224, 45, 183, 128 },
+    { 22, 26, 17, 131, 240, 154, 14, 1, 209 },
+    { 45, 16, 21, 91, 64, 222, 7, 1, 197 },
+    { 56, 21, 39, 155, 60, 138, 23, 102, 213 },
+    { 83, 12, 13, 54, 192, 255, 68, 47, 28 },
+    { 85, 26, 85, 85, 128, 128, 32, 146, 171 },
+    { 18, 11, 7, 63, 144, 171, 4, 4, 246 },
+    { 35, 27, 10, 146, 174, 171, 12, 26, 128 } },
+  { { 190, 80, 35, 99, 180, 80, 126, 54, 45 },
+    { 85, 126, 47, 87, 176, 51, 41, 20, 32 },
+    { 101, 75, 128, 139, 118, 146, 116, 128, 85 },
+    { 56, 41, 15, 176, 236, 85, 37, 9, 62 },
+    { 71, 30, 17, 119, 118, 255, 17, 18, 138 },
+    { 101, 38, 60, 138, 55, 70, 43, 26, 142 },
+    { 146, 36, 19, 30, 171, 255, 97, 27, 20 },
+    { 138, 45, 61, 62, 219, 1, 81, 188, 64 },
+    { 32, 41, 20, 117, 151, 142, 20, 21, 163 },
+    { 112, 19, 12, 61, 195, 128, 48, 4, 24 } }
+};
+
+static int PutI4Mode(VP8BitWriter* const bw, int mode,
+                     const uint8_t* const prob) {
+  if (VP8PutBit(bw, mode != B_DC_PRED, prob[0])) {
+    if (VP8PutBit(bw, mode != B_TM_PRED, prob[1])) {
+      if (VP8PutBit(bw, mode != B_VE_PRED, prob[2])) {
+        if (!VP8PutBit(bw, mode >= B_LD_PRED, prob[3])) {
+          if (VP8PutBit(bw, mode != B_HE_PRED, prob[4])) {
+            VP8PutBit(bw, mode != B_RD_PRED, prob[5]);
+          }
+        } else {
+          if (VP8PutBit(bw, mode != B_LD_PRED, prob[6])) {
+            if (VP8PutBit(bw, mode != B_VL_PRED, prob[7])) {
+              VP8PutBit(bw, mode != B_HD_PRED, prob[8]);
+            }
+          }
+        }
+      }
+    }
+  }
+  return mode;
+}
+
+static void PutI16Mode(VP8BitWriter* const bw, int mode) {
+  if (VP8PutBit(bw, (mode == TM_PRED || mode == H_PRED), 156)) {
+    VP8PutBit(bw, mode == TM_PRED, 128);    // TM or HE
+  } else {
+    VP8PutBit(bw, mode == V_PRED, 163);     // VE or DC
+  }
+}
+
+static void PutUVMode(VP8BitWriter* const bw, int uv_mode) {
+  if (VP8PutBit(bw, uv_mode != DC_PRED, 142)) {
+    if (VP8PutBit(bw, uv_mode != V_PRED, 114)) {
+      VP8PutBit(bw, uv_mode != H_PRED, 183);    // else: TM_PRED
+    }
+  }
+}
+
+static void PutSegment(VP8BitWriter* const bw, int s, const uint8_t* p) {
+  if (VP8PutBit(bw, s >= 2, p[0])) p += 1;
+  VP8PutBit(bw, s & 1, p[1]);
+}
+
+void VP8CodeIntraModes(VP8Encoder* const enc) {
+  VP8BitWriter* const bw = &enc->bw_;
+  VP8EncIterator it;
+  VP8IteratorInit(enc, &it);
+  do {
+    const VP8MBInfo* const mb = it.mb_;
+    const uint8_t* preds = it.preds_;
+    if (enc->segment_hdr_.update_map_) {
+      PutSegment(bw, mb->segment_, enc->proba_.segments_);
+    }
+    if (enc->proba_.use_skip_proba_) {
+      VP8PutBit(bw, mb->skip_, enc->proba_.skip_proba_);
+    }
+    if (VP8PutBit(bw, (mb->type_ != 0), 145)) {  // i16x16
+      PutI16Mode(bw, preds[0]);
+    } else {
+      const int preds_w = enc->preds_w_;
+      const uint8_t* top_pred = preds - preds_w;
+      int x, y;
+      for (y = 0; y < 4; ++y) {
+        int left = preds[-1];
+        for (x = 0; x < 4; ++x) {
+          const uint8_t* const probas = kBModesProba[top_pred[x]][left];
+          left = PutI4Mode(bw, preds[x], probas);
+        }
+        top_pred = preds;
+        preds += preds_w;
+      }
+    }
+    PutUVMode(bw, mb->uv_mode_);
+  } while (VP8IteratorNext(&it));
+}
+
+//------------------------------------------------------------------------------
+// Paragraph 13
+
+const uint8_t
+    VP8CoeffsUpdateProba[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS] = {
+  { { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 176, 246, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 223, 241, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 244, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 234, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 246, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 239, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 253, 255, 254, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 254, 255, 254, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 217, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 225, 252, 241, 253, 255, 255, 254, 255, 255, 255, 255 },
+      { 234, 250, 241, 250, 253, 255, 253, 254, 255, 255, 255 }
+    },
+    { { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 223, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 238, 253, 254, 254, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 247, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 186, 251, 250, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 234, 251, 244, 254, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 251, 243, 253, 254, 255, 254, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 236, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 253, 253, 254, 254, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 248, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 254, 252, 254, 255, 255, 255, 255, 255, 255, 255 },
+      { 248, 254, 249, 253, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 246, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 254, 251, 254, 254, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 248, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 254, 254, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 245, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 251, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 252, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  }
+};
+
+void VP8WriteProbas(VP8BitWriter* const bw, const VP8EncProba* const probas) {
+  int t, b, c, p;
+  for (t = 0; t < NUM_TYPES; ++t) {
+    for (b = 0; b < NUM_BANDS; ++b) {
+      for (c = 0; c < NUM_CTX; ++c) {
+        for (p = 0; p < NUM_PROBAS; ++p) {
+          const uint8_t p0 = probas->coeffs_[t][b][c][p];
+          const int update = (p0 != VP8CoeffsProba0[t][b][c][p]);
+          if (VP8PutBit(bw, update, VP8CoeffsUpdateProba[t][b][c][p])) {
+            VP8PutBits(bw, p0, 8);
+          }
+        }
+      }
+    }
+  }
+  if (VP8PutBitUniform(bw, probas->use_skip_proba_)) {
+    VP8PutBits(bw, probas->skip_proba_, 8);
+  }
+}
+
diff --git a/media/libwebp/src/enc/vp8i_enc.h b/media/libwebp/src/enc/vp8i_enc.h
new file mode 100644
index 0000000000..71f76702ae
--- /dev/null
+++ b/media/libwebp/src/enc/vp8i_enc.h
@@ -0,0 +1,523 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+//   WebP encoder: internal header.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_ENC_VP8I_ENC_H_
+#define WEBP_ENC_VP8I_ENC_H_
+
+#include <string.h>     // for memcpy()
+#include "src/dec/common_dec.h"
+#include "src/dsp/dsp.h"
+#include "src/utils/bit_writer_utils.h"
+#include "src/utils/thread_utils.h"
+#include "src/utils/utils.h"
+#include "src/webp/encode.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// Various defines and enums
+
+// version numbers
+#define ENC_MAJ_VERSION 1
+#define ENC_MIN_VERSION 2
+#define ENC_REV_VERSION 4
+
+enum { MAX_LF_LEVELS = 64,       // Maximum loop filter level
+       MAX_VARIABLE_LEVEL = 67,  // last (inclusive) level with variable cost
+       MAX_LEVEL = 2047          // max level (note: max codable is 2047 + 67)
+     };
+
+typedef enum {   // Rate-distortion optimization levels
+  RD_OPT_NONE        = 0,  // no rd-opt
+  RD_OPT_BASIC       = 1,  // basic scoring (no trellis)
+  RD_OPT_TRELLIS     = 2,  // perform trellis-quant on the final decision only
+  RD_OPT_TRELLIS_ALL = 3   // trellis-quant for every scoring (much slower)
+} VP8RDLevel;
+
+// YUV-cache parameters. Cache is 32-bytes wide (= one cacheline).
+// The original or reconstructed samples can be accessed using VP8Scan[].
+// The predicted blocks can be accessed using offsets to yuv_p_ and
+// the arrays VP8*ModeOffsets[].
+// * YUV Samples area (yuv_in_/yuv_out_/yuv_out2_)
+//   (see VP8Scan[] for accessing the blocks, along with
+//   Y_OFF_ENC/U_OFF_ENC/V_OFF_ENC):
+//             +----+----+
+//  Y_OFF_ENC  |YYYY|UUVV|
+//  U_OFF_ENC  |YYYY|UUVV|
+//  V_OFF_ENC  |YYYY|....| <- 25% wasted U/V area
+//             |YYYY|....|
+//             +----+----+
+// * Prediction area ('yuv_p_', size = PRED_SIZE_ENC)
+//   Intra16 predictions (16x16 block each, two per row):
+//         |I16DC16|I16TM16|
+//         |I16VE16|I16HE16|
+//   Chroma U/V predictions (16x8 block each, two per row):
+//         |C8DC8|C8TM8|
+//         |C8VE8|C8HE8|
+//   Intra 4x4 predictions (4x4 block each)
+//         |I4DC4 I4TM4 I4VE4 I4HE4|I4RD4 I4VR4 I4LD4 I4VL4|
+//         |I4HD4 I4HU4 I4TMP .....|.......................| <- ~31% wasted
+#define YUV_SIZE_ENC (BPS * 16)
+#define PRED_SIZE_ENC (32 * BPS + 16 * BPS + 8 * BPS)   // I16+Chroma+I4 preds
+#define Y_OFF_ENC    (0)
+#define U_OFF_ENC    (16)
+#define V_OFF_ENC    (16 + 8)
+
+extern const uint16_t VP8Scan[16];
+extern const uint16_t VP8UVModeOffsets[4];
+extern const uint16_t VP8I16ModeOffsets[4];
+extern const uint16_t VP8I4ModeOffsets[NUM_BMODES];
+
+// Layout of prediction blocks
+// intra 16x16
+#define I16DC16 (0 * 16 * BPS)
+#define I16TM16 (I16DC16 + 16)
+#define I16VE16 (1 * 16 * BPS)
+#define I16HE16 (I16VE16 + 16)
+// chroma 8x8, two U/V blocks side by side (hence: 16x8 each)
+#define C8DC8 (2 * 16 * BPS)
+#define C8TM8 (C8DC8 + 1 * 16)
+#define C8VE8 (2 * 16 * BPS + 8 * BPS)
+#define C8HE8 (C8VE8 + 1 * 16)
+// intra 4x4
+#define I4DC4 (3 * 16 * BPS +  0)
+#define I4TM4 (I4DC4 +  4)
+#define I4VE4 (I4DC4 +  8)
+#define I4HE4 (I4DC4 + 12)
+#define I4RD4 (I4DC4 + 16)
+#define I4VR4 (I4DC4 + 20)
+#define I4LD4 (I4DC4 + 24)
+#define I4VL4 (I4DC4 + 28)
+#define I4HD4 (3 * 16 * BPS + 4 * BPS)
+#define I4HU4 (I4HD4 + 4)
+#define I4TMP (I4HD4 + 8)
+
+typedef int64_t score_t;     // type used for scores, rate, distortion
+// Note that MAX_COST is not the maximum allowed by sizeof(score_t),
+// in order to allow overflowing computations.
+#define MAX_COST ((score_t)0x7fffffffffffffLL)
+
+#define QFIX 17
+#define BIAS(b)  ((b) << (QFIX - 8))
+// Fun fact: this is the _only_ line where we're actually being lossy and
+// discarding bits.
+static WEBP_INLINE int QUANTDIV(uint32_t n, uint32_t iQ, uint32_t B) {
+  return (int)((n * iQ + B) >> QFIX);
+}
+
+// Uncomment the following to remove token-buffer code:
+// #define DISABLE_TOKEN_BUFFER
+
+// quality below which error-diffusion is enabled
+#define ERROR_DIFFUSION_QUALITY 98
+
+//------------------------------------------------------------------------------
+// Headers
+
+typedef uint32_t proba_t;   // 16b + 16b
+typedef uint8_t ProbaArray[NUM_CTX][NUM_PROBAS];
+typedef proba_t StatsArray[NUM_CTX][NUM_PROBAS];
+typedef uint16_t CostArray[NUM_CTX][MAX_VARIABLE_LEVEL + 1];
+typedef const uint16_t* (*CostArrayPtr)[NUM_CTX];   // for easy casting
+typedef const uint16_t* CostArrayMap[16][NUM_CTX];
+typedef double LFStats[NUM_MB_SEGMENTS][MAX_LF_LEVELS];  // filter stats
+
+typedef struct VP8Encoder VP8Encoder;
+
+// segment features
+typedef struct {
+  int num_segments_;      // Actual number of segments. 1 segment only = unused.
+  int update_map_;        // whether to update the segment map or not.
+                          // must be 0 if there's only 1 segment.
+  int size_;              // bit-cost for transmitting the segment map
+} VP8EncSegmentHeader;
+
+// Struct collecting all frame-persistent probabilities.
+typedef struct {
+  uint8_t segments_[3];     // probabilities for segment tree
+  uint8_t skip_proba_;      // final probability of being skipped.
+  ProbaArray coeffs_[NUM_TYPES][NUM_BANDS];      // 1056 bytes
+  StatsArray stats_[NUM_TYPES][NUM_BANDS];       // 4224 bytes
+  CostArray level_cost_[NUM_TYPES][NUM_BANDS];   // 13056 bytes
+  CostArrayMap remapped_costs_[NUM_TYPES];       // 1536 bytes
+  int dirty_;               // if true, need to call VP8CalculateLevelCosts()
+  int use_skip_proba_;      // Note: we always use skip_proba for now.
+  int nb_skip_;             // number of skipped blocks
+} VP8EncProba;
+
+// Filter parameters. Not actually used in the code (we don't perform
+// the in-loop filtering), but filled from user's config
+typedef struct {
+  int simple_;             // filtering type: 0=complex, 1=simple
+  int level_;              // base filter level [0..63]
+  int sharpness_;          // [0..7]
+  int i4x4_lf_delta_;      // delta filter level for i4x4 relative to i16x16
+} VP8EncFilterHeader;
+
+//------------------------------------------------------------------------------
+// Informations about the macroblocks.
+
+typedef struct {
+  // block type
+  unsigned int type_:2;     // 0=i4x4, 1=i16x16
+  unsigned int uv_mode_:2;
+  unsigned int skip_:1;
+  unsigned int segment_:2;
+  uint8_t alpha_;      // quantization-susceptibility
+} VP8MBInfo;
+
+typedef struct VP8Matrix {
+  uint16_t q_[16];        // quantizer steps
+  uint16_t iq_[16];       // reciprocals, fixed point.
+  uint32_t bias_[16];     // rounding bias
+  uint32_t zthresh_[16];  // value below which a coefficient is zeroed
+  uint16_t sharpen_[16];  // frequency boosters for slight sharpening
+} VP8Matrix;
+
+typedef struct {
+  VP8Matrix y1_, y2_, uv_;  // quantization matrices
+  int alpha_;      // quant-susceptibility, range [-127,127]. Zero is neutral.
+                   // Lower values indicate a lower risk of blurriness.
+  int beta_;       // filter-susceptibility, range [0,255].
+  int quant_;      // final segment quantizer.
+  int fstrength_;  // final in-loop filtering strength
+  int max_edge_;   // max edge delta (for filtering strength)
+  int min_disto_;  // minimum distortion required to trigger filtering record
+  // reactivities
+  int lambda_i16_, lambda_i4_, lambda_uv_;
+  int lambda_mode_, lambda_trellis_, tlambda_;
+  int lambda_trellis_i16_, lambda_trellis_i4_, lambda_trellis_uv_;
+
+  // lambda values for distortion-based evaluation
+  score_t i4_penalty_;   // penalty for using Intra4
+} VP8SegmentInfo;
+
+typedef int8_t DError[2 /* u/v */][2 /* top or left */];
+
+// Handy transient struct to accumulate score and info during RD-optimization
+// and mode evaluation.
+typedef struct {
+  score_t D, SD;              // Distortion, spectral distortion
+  score_t H, R, score;        // header bits, rate, score.
+  int16_t y_dc_levels[16];    // Quantized levels for luma-DC, luma-AC, chroma.
+  int16_t y_ac_levels[16][16];
+  int16_t uv_levels[4 + 4][16];
+  int mode_i16;               // mode number for intra16 prediction
+  uint8_t modes_i4[16];       // mode numbers for intra4 predictions
+  int mode_uv;                // mode number of chroma prediction
+  uint32_t nz;                // non-zero blocks
+  int8_t derr[2][3];          // DC diffusion errors for U/V for blocks #1/2/3
+} VP8ModeScore;
+
+// Iterator structure to iterate through macroblocks, pointing to the
+// right neighbouring data (samples, predictions, contexts, ...)
+typedef struct {
+  int x_, y_;                      // current macroblock
+  uint8_t*      yuv_in_;           // input samples
+  uint8_t*      yuv_out_;          // output samples
+  uint8_t*      yuv_out2_;         // secondary buffer swapped with yuv_out_.
+  uint8_t*      yuv_p_;            // scratch buffer for prediction
+  VP8Encoder*   enc_;              // back-pointer
+  VP8MBInfo*    mb_;               // current macroblock
+  VP8BitWriter* bw_;               // current bit-writer
+  uint8_t*      preds_;            // intra mode predictors (4x4 blocks)
+  uint32_t*     nz_;               // non-zero pattern
+  uint8_t       i4_boundary_[37];  // 32+5 boundary samples needed by intra4x4
+  uint8_t*      i4_top_;           // pointer to the current top boundary sample
+  int           i4_;               // current intra4x4 mode being tested
+  int           top_nz_[9];        // top-non-zero context.
+  int           left_nz_[9];       // left-non-zero. left_nz[8] is independent.
+  uint64_t      bit_count_[4][3];  // bit counters for coded levels.
+  uint64_t      luma_bits_;        // macroblock bit-cost for luma
+  uint64_t      uv_bits_;          // macroblock bit-cost for chroma
+  LFStats*      lf_stats_;         // filter stats (borrowed from enc_)
+  int           do_trellis_;       // if true, perform extra level optimisation
+  int           count_down_;       // number of mb still to be processed
+  int           count_down0_;      // starting counter value (for progress)
+  int           percent0_;         // saved initial progress percent
+
+  DError        left_derr_;        // left error diffusion (u/v)
+  DError*       top_derr_;         // top diffusion error - NULL if disabled
+
+  uint8_t* y_left_;    // left luma samples (addressable from index -1 to 15).
+  uint8_t* u_left_;    // left u samples (addressable from index -1 to 7)
+  uint8_t* v_left_;    // left v samples (addressable from index -1 to 7)
+
+  uint8_t* y_top_;     // top luma samples at position 'x_'
+  uint8_t* uv_top_;    // top u/v samples at position 'x_', packed as 16 bytes
+
+  // memory for storing y/u/v_left_
+  uint8_t yuv_left_mem_[17 + 16 + 16 + 8 + WEBP_ALIGN_CST];
+  // memory for yuv_*
+  uint8_t yuv_mem_[3 * YUV_SIZE_ENC + PRED_SIZE_ENC + WEBP_ALIGN_CST];
+} VP8EncIterator;
+
+  // in iterator.c
+// must be called first
+void VP8IteratorInit(VP8Encoder* const enc, VP8EncIterator* const it);
+// restart a scan
+void VP8IteratorReset(VP8EncIterator* const it);
+// reset iterator position to row 'y'
+void VP8IteratorSetRow(VP8EncIterator* const it, int y);
+// set count down (=number of iterations to go)
+void VP8IteratorSetCountDown(VP8EncIterator* const it, int count_down);
+// return true if iteration is finished
+int VP8IteratorIsDone(const VP8EncIterator* const it);
+// Import uncompressed samples from source.
+// If tmp_32 is not NULL, import boundary samples too.
+// tmp_32 is a 32-bytes scratch buffer that must be aligned in memory.
+void VP8IteratorImport(VP8EncIterator* const it, uint8_t* const tmp_32);
+// export decimated samples
+void VP8IteratorExport(const VP8EncIterator* const it);
+// go to next macroblock. Returns false if not finished.
+int VP8IteratorNext(VP8EncIterator* const it);
+// save the yuv_out_ boundary values to top_/left_ arrays for next iterations.
+void VP8IteratorSaveBoundary(VP8EncIterator* const it);
+// Report progression based on macroblock rows. Return 0 for user-abort request.
+int VP8IteratorProgress(const VP8EncIterator* const it, int delta);
+// Intra4x4 iterations
+void VP8IteratorStartI4(VP8EncIterator* const it);
+// returns true if not done.
+int VP8IteratorRotateI4(VP8EncIterator* const it,
+                        const uint8_t* const yuv_out);
+
+// Non-zero context setup/teardown
+void VP8IteratorNzToBytes(VP8EncIterator* const it);
+void VP8IteratorBytesToNz(VP8EncIterator* const it);
+
+// Helper functions to set mode properties
+void VP8SetIntra16Mode(const VP8EncIterator* const it, int mode);
+void VP8SetIntra4Mode(const VP8EncIterator* const it, const uint8_t* modes);
+void VP8SetIntraUVMode(const VP8EncIterator* const it, int mode);
+void VP8SetSkip(const VP8EncIterator* const it, int skip);
+void VP8SetSegment(const VP8EncIterator* const it, int segment);
+
+//------------------------------------------------------------------------------
+// Paginated token buffer
+
+typedef struct VP8Tokens VP8Tokens;  // struct details in token.c
+
+typedef struct {
+#if !defined(DISABLE_TOKEN_BUFFER)
+  VP8Tokens* pages_;        // first page
+  VP8Tokens** last_page_;   // last page
+  uint16_t* tokens_;        // set to (*last_page_)->tokens_
+  int left_;                // how many free tokens left before the page is full
+  int page_size_;           // number of tokens per page
+#endif
+  int error_;         // true in case of malloc error
+} VP8TBuffer;
+
+// initialize an empty buffer
+void VP8TBufferInit(VP8TBuffer* const b, int page_size);
+void VP8TBufferClear(VP8TBuffer* const b);   // de-allocate pages memory
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+// Finalizes bitstream when probabilities are known.
+// Deletes the allocated token memory if final_pass is true.
+int VP8EmitTokens(VP8TBuffer* const b, VP8BitWriter* const bw,
+                  const uint8_t* const probas, int final_pass);
+
+// record the coding of coefficients without knowing the probabilities yet
+int VP8RecordCoeffTokens(int ctx, const struct VP8Residual* const res,
+                         VP8TBuffer* const tokens);
+
+// Estimate the final coded size given a set of 'probas'.
+size_t VP8EstimateTokenSize(VP8TBuffer* const b, const uint8_t* const probas);
+
+#endif  // !DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+// VP8Encoder
+
+struct VP8Encoder {
+  const WebPConfig* config_;    // user configuration and parameters
+  WebPPicture* pic_;            // input / output picture
+
+  // headers
+  VP8EncFilterHeader   filter_hdr_;     // filtering information
+  VP8EncSegmentHeader  segment_hdr_;    // segment information
+
+  int profile_;                      // VP8's profile, deduced from Config.
+
+  // dimension, in macroblock units.
+  int mb_w_, mb_h_;
+  int preds_w_;   // stride of the *preds_ prediction plane (=4*mb_w + 1)
+
+  // number of partitions (1, 2, 4 or 8 = MAX_NUM_PARTITIONS)
+  int num_parts_;
+
+  // per-partition boolean decoders.
+  VP8BitWriter bw_;                         // part0
+  VP8BitWriter parts_[MAX_NUM_PARTITIONS];  // token partitions
+  VP8TBuffer tokens_;                       // token buffer
+
+  int percent_;                             // for progress
+
+  // transparency blob
+  int has_alpha_;
+  uint8_t* alpha_data_;       // non-NULL if transparency is present
+  uint32_t alpha_data_size_;
+  WebPWorker alpha_worker_;
+
+  // quantization info (one set of DC/AC dequant factor per segment)
+  VP8SegmentInfo dqm_[NUM_MB_SEGMENTS];
+  int base_quant_;                 // nominal quantizer value. Only used
+                                   // for relative coding of segments' quant.
+  int alpha_;                      // global susceptibility (<=> complexity)
+  int uv_alpha_;                   // U/V quantization susceptibility
+  // global offset of quantizers, shared by all segments
+  int dq_y1_dc_;
+  int dq_y2_dc_, dq_y2_ac_;
+  int dq_uv_dc_, dq_uv_ac_;
+
+  // probabilities and statistics
+  VP8EncProba proba_;
+  uint64_t    sse_[4];      // sum of Y/U/V/A squared errors for all macroblocks
+  uint64_t    sse_count_;   // pixel count for the sse_[] stats
+  int         coded_size_;
+  int         residual_bytes_[3][4];
+  int         block_count_[3];
+
+  // quality/speed settings
+  int method_;               // 0=fastest, 6=best/slowest.
+  VP8RDLevel rd_opt_level_;  // Deduced from method_.
+  int max_i4_header_bits_;   // partition #0 safeness factor
+  int mb_header_limit_;      // rough limit for header bits per MB
+  int thread_level_;         // derived from config->thread_level
+  int do_search_;            // derived from config->target_XXX
+  int use_tokens_;           // if true, use token buffer
+
+  // Memory
+  VP8MBInfo* mb_info_;   // contextual macroblock infos (mb_w_ + 1)
+  uint8_t*   preds_;     // predictions modes: (4*mb_w+1) * (4*mb_h+1)
+  uint32_t*  nz_;        // non-zero bit context: mb_w+1
+  uint8_t*   y_top_;     // top luma samples.
+  uint8_t*   uv_top_;    // top u/v samples.
+                         // U and V are packed into 16 bytes (8 U + 8 V)
+  LFStats*   lf_stats_;  // autofilter stats (if NULL, autofilter is off)
+  DError*    top_derr_;  // diffusion error (NULL if disabled)
+};
+
+//------------------------------------------------------------------------------
+// internal functions. Not public.
+
+  // in tree.c
+extern const uint8_t VP8CoeffsProba0[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS];
+extern const uint8_t
+    VP8CoeffsUpdateProba[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS];
+// Reset the token probabilities to their initial (default) values
+void VP8DefaultProbas(VP8Encoder* const enc);
+// Write the token probabilities
+void VP8WriteProbas(VP8BitWriter* const bw, const VP8EncProba* const probas);
+// Writes the partition #0 modes (that is: all intra modes)
+void VP8CodeIntraModes(VP8Encoder* const enc);
+
+  // in syntax.c
+// Generates the final bitstream by coding the partition0 and headers,
+// and appending an assembly of all the pre-coded token partitions.
+// Return true if everything is ok.
+int VP8EncWrite(VP8Encoder* const enc);
+// Release memory allocated for bit-writing in VP8EncLoop & seq.
+void VP8EncFreeBitWriters(VP8Encoder* const enc);
+
+  // in frame.c
+extern const uint8_t VP8Cat3[];
+extern const uint8_t VP8Cat4[];
+extern const uint8_t VP8Cat5[];
+extern const uint8_t VP8Cat6[];
+
+// Form all the four Intra16x16 predictions in the yuv_p_ cache
+void VP8MakeLuma16Preds(const VP8EncIterator* const it);
+// Form all the four Chroma8x8 predictions in the yuv_p_ cache
+void VP8MakeChroma8Preds(const VP8EncIterator* const it);
+// Form all the ten Intra4x4 predictions in the yuv_p_ cache
+// for the 4x4 block it->i4_
+void VP8MakeIntra4Preds(const VP8EncIterator* const it);
+// Rate calculation
+int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd);
+int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]);
+int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd);
+// Main coding calls
+int VP8EncLoop(VP8Encoder* const enc);
+int VP8EncTokenLoop(VP8Encoder* const enc);
+
+  // in webpenc.c
+// Assign an error code to a picture. Return false for convenience.
+int WebPEncodingSetError(const WebPPicture* const pic, WebPEncodingError error);
+int WebPReportProgress(const WebPPicture* const pic,
+                       int percent, int* const percent_store);
+
+  // in analysis.c
+// Main analysis loop. Decides the segmentations and complexity.
+// Assigns a first guess for Intra16 and uvmode_ prediction modes.
+int VP8EncAnalyze(VP8Encoder* const enc);
+
+  // in quant.c
+// Sets up segment's quantization values, base_quant_ and filter strengths.
+void VP8SetSegmentParams(VP8Encoder* const enc, float quality);
+// Pick best modes and fills the levels. Returns true if skipped.
+int VP8Decimate(VP8EncIterator* WEBP_RESTRICT const it,
+                VP8ModeScore* WEBP_RESTRICT const rd,
+                VP8RDLevel rd_opt);
+
+  // in alpha.c
+void VP8EncInitAlpha(VP8Encoder* const enc);    // initialize alpha compression
+int VP8EncStartAlpha(VP8Encoder* const enc);    // start alpha coding process
+int VP8EncFinishAlpha(VP8Encoder* const enc);   // finalize compressed data
+int VP8EncDeleteAlpha(VP8Encoder* const enc);   // delete compressed data
+
+// autofilter
+void VP8InitFilter(VP8EncIterator* const it);
+void VP8StoreFilterStats(VP8EncIterator* const it);
+void VP8AdjustFilterStrength(VP8EncIterator* const it);
+
+// returns the approximate filtering strength needed to smooth a edge
+// step of 'delta', given a sharpness parameter 'sharpness'.
+int VP8FilterStrengthFromDelta(int sharpness, int delta);
+
+  // misc utils for picture_*.c:
+
+// Returns true if 'picture' is non-NULL and dimensions/colorspace are within
+// their valid ranges. If returning false, the 'error_code' in 'picture' is
+// updated.
+int WebPValidatePicture(const WebPPicture* const picture);
+
+// Remove reference to the ARGB/YUVA buffer (doesn't free anything).
+void WebPPictureResetBuffers(WebPPicture* const picture);
+
+// Allocates ARGB buffer according to set width/height (previous one is
+// always free'd). Preserves the YUV(A) buffer. Returns false in case of error
+// (invalid param, out-of-memory).
+int WebPPictureAllocARGB(WebPPicture* const picture);
+
+// Allocates YUVA buffer according to set width/height (previous one is always
+// free'd). Uses picture->csp to determine whether an alpha buffer is needed.
+// Preserves the ARGB buffer.
+// Returns false in case of error (invalid param, out-of-memory).
+int WebPPictureAllocYUVA(WebPPicture* const picture);
+
+// Replace samples that are fully transparent by 'color' to help compressibility
+// (no guarantee, though). Assumes pic->use_argb is true.
+void WebPReplaceTransparentPixels(WebPPicture* const pic, uint32_t color);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // WEBP_ENC_VP8I_ENC_H_
diff --git a/media/libwebp/src/enc/vp8l_enc.c b/media/libwebp/src/enc/vp8l_enc.c
new file mode 100644
index 0000000000..2b345df610
--- /dev/null
+++ b/media/libwebp/src/enc/vp8l_enc.c
@@ -0,0 +1,2197 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// main entry for the lossless encoder.
+//
+// Author: Vikas Arora (vikaas.arora@gmail.com)
+//
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include "src/dsp/lossless.h"
+#include "src/dsp/lossless_common.h"
+#include "src/enc/backward_references_enc.h"
+#include "src/enc/histogram_enc.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/enc/vp8li_enc.h"
+#include "src/utils/bit_writer_utils.h"
+#include "src/utils/huffman_encode_utils.h"
+#include "src/utils/utils.h"
+#include "src/webp/encode.h"
+#include "src/webp/format_constants.h"
+
+// Maximum number of histogram images (sub-blocks).
+#define MAX_HUFF_IMAGE_SIZE       2600
+
+// Palette reordering for smaller sum of deltas (and for smaller storage).
+
+static int PaletteCompareColorsForQsort(const void* p1, const void* p2) {
+  const uint32_t a = WebPMemToUint32((uint8_t*)p1);
+  const uint32_t b = WebPMemToUint32((uint8_t*)p2);
+  assert(a != b);
+  return (a < b) ? -1 : 1;
+}
+
+static WEBP_INLINE uint32_t PaletteComponentDistance(uint32_t v) {
+  return (v <= 128) ? v : (256 - v);
+}
+
+// Computes a value that is related to the entropy created by the
+// palette entry diff.
+//
+// Note that the last & 0xff is a no-operation in the next statement, but
+// removed by most compilers and is here only for regularity of the code.
+static WEBP_INLINE uint32_t PaletteColorDistance(uint32_t col1, uint32_t col2) {
+  const uint32_t diff = VP8LSubPixels(col1, col2);
+  const int kMoreWeightForRGBThanForAlpha = 9;
+  uint32_t score;
+  score =  PaletteComponentDistance((diff >>  0) & 0xff);
+  score += PaletteComponentDistance((diff >>  8) & 0xff);
+  score += PaletteComponentDistance((diff >> 16) & 0xff);
+  score *= kMoreWeightForRGBThanForAlpha;
+  score += PaletteComponentDistance((diff >> 24) & 0xff);
+  return score;
+}
+
+static WEBP_INLINE void SwapColor(uint32_t* const col1, uint32_t* const col2) {
+  const uint32_t tmp = *col1;
+  *col1 = *col2;
+  *col2 = tmp;
+}
+
+static WEBP_INLINE int SearchColorNoIdx(const uint32_t sorted[], uint32_t color,
+                                        int num_colors) {
+  int low = 0, hi = num_colors;
+  if (sorted[low] == color) return low;  // loop invariant: sorted[low] != color
+  while (1) {
+    const int mid = (low + hi) >> 1;
+    if (sorted[mid] == color) {
+      return mid;
+    } else if (sorted[mid] < color) {
+      low = mid;
+    } else {
+      hi = mid;
+    }
+  }
+  assert(0);
+  return 0;
+}
+
+// The palette has been sorted by alpha. This function checks if the other
+// components of the palette have a monotonic development with regards to
+// position in the palette. If all have monotonic development, there is
+// no benefit to re-organize them greedily. A monotonic development
+// would be spotted in green-only situations (like lossy alpha) or gray-scale
+// images.
+static int PaletteHasNonMonotonousDeltas(const uint32_t* const palette,
+                                         int num_colors) {
+  uint32_t predict = 0x000000;
+  int i;
+  uint8_t sign_found = 0x00;
+  for (i = 0; i < num_colors; ++i) {
+    const uint32_t diff = VP8LSubPixels(palette[i], predict);
+    const uint8_t rd = (diff >> 16) & 0xff;
+    const uint8_t gd = (diff >>  8) & 0xff;
+    const uint8_t bd = (diff >>  0) & 0xff;
+    if (rd != 0x00) {
+      sign_found |= (rd < 0x80) ? 1 : 2;
+    }
+    if (gd != 0x00) {
+      sign_found |= (gd < 0x80) ? 8 : 16;
+    }
+    if (bd != 0x00) {
+      sign_found |= (bd < 0x80) ? 64 : 128;
+    }
+    predict = palette[i];
+  }
+  return (sign_found & (sign_found << 1)) != 0;  // two consequent signs.
+}
+
+static void PaletteSortMinimizeDeltas(const uint32_t* const palette_sorted,
+                                      int num_colors, uint32_t* const palette) {
+  uint32_t predict = 0x00000000;
+  int i, k;
+  memcpy(palette, palette_sorted, num_colors * sizeof(*palette));
+  if (!PaletteHasNonMonotonousDeltas(palette_sorted, num_colors)) return;
+  // Find greedily always the closest color of the predicted color to minimize
+  // deltas in the palette. This reduces storage needs since the
+  // palette is stored with delta encoding.
+  for (i = 0; i < num_colors; ++i) {
+    int best_ix = i;
+    uint32_t best_score = ~0U;
+    for (k = i; k < num_colors; ++k) {
+      const uint32_t cur_score = PaletteColorDistance(palette[k], predict);
+      if (best_score > cur_score) {
+        best_score = cur_score;
+        best_ix = k;
+      }
+    }
+    SwapColor(&palette[best_ix], &palette[i]);
+    predict = palette[i];
+  }
+}
+
+// Sort palette in increasing order and prepare an inverse mapping array.
+static void PrepareMapToPalette(const uint32_t palette[], uint32_t num_colors,
+                                uint32_t sorted[], uint32_t idx_map[]) {
+  uint32_t i;
+  memcpy(sorted, palette, num_colors * sizeof(*sorted));
+  qsort(sorted, num_colors, sizeof(*sorted), PaletteCompareColorsForQsort);
+  for (i = 0; i < num_colors; ++i) {
+    idx_map[SearchColorNoIdx(sorted, palette[i], num_colors)] = i;
+  }
+}
+
+// -----------------------------------------------------------------------------
+// Modified Zeng method from "A Survey on Palette Reordering
+// Methods for Improving the Compression of Color-Indexed Images" by Armando J.
+// Pinho and Antonio J. R. Neves.
+
+// Finds the biggest cooccurrence in the matrix.
+static void CoOccurrenceFindMax(const uint32_t* const cooccurrence,
+                                uint32_t num_colors, uint8_t* const c1,
+                                uint8_t* const c2) {
+  // Find the index that is most frequently located adjacent to other
+  // (different) indexes.
+  uint32_t best_sum = 0u;
+  uint32_t i, j, best_cooccurrence;
+  *c1 = 0u;
+  for (i = 0; i < num_colors; ++i) {
+    uint32_t sum = 0;
+    for (j = 0; j < num_colors; ++j) sum += cooccurrence[i * num_colors + j];
+    if (sum > best_sum) {
+      best_sum = sum;
+      *c1 = i;
+    }
+  }
+  // Find the index that is most frequently found adjacent to *c1.
+  *c2 = 0u;
+  best_cooccurrence = 0u;
+  for (i = 0; i < num_colors; ++i) {
+    if (cooccurrence[*c1 * num_colors + i] > best_cooccurrence) {
+      best_cooccurrence = cooccurrence[*c1 * num_colors + i];
+      *c2 = i;
+    }
+  }
+  assert(*c1 != *c2);
+}
+
+// Builds the cooccurrence matrix
+static int CoOccurrenceBuild(const WebPPicture* const pic,
+                             const uint32_t* const palette, uint32_t num_colors,
+                             uint32_t* cooccurrence) {
+  uint32_t *lines, *line_top, *line_current, *line_tmp;
+  int x, y;
+  const uint32_t* src = pic->argb;
+  uint32_t prev_pix = ~src[0];
+  uint32_t prev_idx = 0u;
+  uint32_t idx_map[MAX_PALETTE_SIZE] = {0};
+  uint32_t palette_sorted[MAX_PALETTE_SIZE];
+  lines = (uint32_t*)WebPSafeMalloc(2 * pic->width, sizeof(*lines));
+  if (lines == NULL) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return 0;
+  }
+  line_top = &lines[0];
+  line_current = &lines[pic->width];
+  PrepareMapToPalette(palette, num_colors, palette_sorted, idx_map);
+  for (y = 0; y < pic->height; ++y) {
+    for (x = 0; x < pic->width; ++x) {
+      const uint32_t pix = src[x];
+      if (pix != prev_pix) {
+        prev_idx = idx_map[SearchColorNoIdx(palette_sorted, pix, num_colors)];
+        prev_pix = pix;
+      }
+      line_current[x] = prev_idx;
+      // 4-connectivity is what works best as mentioned in "On the relation
+      // between Memon's and the modified Zeng's palette reordering methods".
+      if (x > 0 && prev_idx != line_current[x - 1]) {
+        const uint32_t left_idx = line_current[x - 1];
+        ++cooccurrence[prev_idx * num_colors + left_idx];
+        ++cooccurrence[left_idx * num_colors + prev_idx];
+      }
+      if (y > 0 && prev_idx != line_top[x]) {
+        const uint32_t top_idx = line_top[x];
+        ++cooccurrence[prev_idx * num_colors + top_idx];
+        ++cooccurrence[top_idx * num_colors + prev_idx];
+      }
+    }
+    line_tmp = line_top;
+    line_top = line_current;
+    line_current = line_tmp;
+    src += pic->argb_stride;
+  }
+  WebPSafeFree(lines);
+  return 1;
+}
+
+struct Sum {
+  uint8_t index;
+  uint32_t sum;
+};
+
+// Implements the modified Zeng method from "A Survey on Palette Reordering
+// Methods for Improving the Compression of Color-Indexed Images" by Armando J.
+// Pinho and Antonio J. R. Neves.
+static int PaletteSortModifiedZeng(
+    const WebPPicture* const pic, const uint32_t* const palette_sorted,
+    uint32_t num_colors, uint32_t* const palette) {
+  uint32_t i, j, ind;
+  uint8_t remapping[MAX_PALETTE_SIZE];
+  uint32_t* cooccurrence;
+  struct Sum sums[MAX_PALETTE_SIZE];
+  uint32_t first, last;
+  uint32_t num_sums;
+  // TODO(vrabaud) check whether one color images should use palette or not.
+  if (num_colors <= 1) return 1;
+  // Build the co-occurrence matrix.
+  cooccurrence =
+      (uint32_t*)WebPSafeCalloc(num_colors * num_colors, sizeof(*cooccurrence));
+  if (cooccurrence == NULL) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return 0;
+  }
+  if (!CoOccurrenceBuild(pic, palette_sorted, num_colors, cooccurrence)) {
+    return 0;
+  }
+
+  // Initialize the mapping list with the two best indices.
+  CoOccurrenceFindMax(cooccurrence, num_colors, &remapping[0], &remapping[1]);
+
+  // We need to append and prepend to the list of remapping. To this end, we
+  // actually define the next start/end of the list as indices in a vector (with
+  // a wrap around when the end is reached).
+  first = 0;
+  last = 1;
+  num_sums = num_colors - 2;  // -2 because we know the first two values
+  if (num_sums > 0) {
+    // Initialize the sums with the first two remappings and find the best one
+    struct Sum* best_sum = &sums[0];
+    best_sum->index = 0u;
+    best_sum->sum = 0u;
+    for (i = 0, j = 0; i < num_colors; ++i) {
+      if (i == remapping[0] || i == remapping[1]) continue;
+      sums[j].index = i;
+      sums[j].sum = cooccurrence[i * num_colors + remapping[0]] +
+                    cooccurrence[i * num_colors + remapping[1]];
+      if (sums[j].sum > best_sum->sum) best_sum = &sums[j];
+      ++j;
+    }
+
+    while (num_sums > 0) {
+      const uint8_t best_index = best_sum->index;
+      // Compute delta to know if we need to prepend or append the best index.
+      int32_t delta = 0;
+      const int32_t n = num_colors - num_sums;
+      for (ind = first, j = 0; (ind + j) % num_colors != last + 1; ++j) {
+        const uint16_t l_j = remapping[(ind + j) % num_colors];
+        delta += (n - 1 - 2 * (int32_t)j) *
+                 (int32_t)cooccurrence[best_index * num_colors + l_j];
+      }
+      if (delta > 0) {
+        first = (first == 0) ? num_colors - 1 : first - 1;
+        remapping[first] = best_index;
+      } else {
+        ++last;
+        remapping[last] = best_index;
+      }
+      // Remove best_sum from sums.
+      *best_sum = sums[num_sums - 1];
+      --num_sums;
+      // Update all the sums and find the best one.
+      best_sum = &sums[0];
+      for (i = 0; i < num_sums; ++i) {
+        sums[i].sum += cooccurrence[best_index * num_colors + sums[i].index];
+        if (sums[i].sum > best_sum->sum) best_sum = &sums[i];
+      }
+    }
+  }
+  assert((last + 1) % num_colors == first);
+  WebPSafeFree(cooccurrence);
+
+  // Re-map the palette.
+  for (i = 0; i < num_colors; ++i) {
+    palette[i] = palette_sorted[remapping[(first + i) % num_colors]];
+  }
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+// Palette
+
+// These five modes are evaluated and their respective entropy is computed.
+typedef enum {
+  kDirect = 0,
+  kSpatial = 1,
+  kSubGreen = 2,
+  kSpatialSubGreen = 3,
+  kPalette = 4,
+  kPaletteAndSpatial = 5,
+  kNumEntropyIx = 6
+} EntropyIx;
+
+typedef enum {
+  kSortedDefault = 0,
+  kMinimizeDelta = 1,
+  kModifiedZeng = 2,
+  kUnusedPalette = 3,
+} PaletteSorting;
+
+typedef enum {
+  kHistoAlpha = 0,
+  kHistoAlphaPred,
+  kHistoGreen,
+  kHistoGreenPred,
+  kHistoRed,
+  kHistoRedPred,
+  kHistoBlue,
+  kHistoBluePred,
+  kHistoRedSubGreen,
+  kHistoRedPredSubGreen,
+  kHistoBlueSubGreen,
+  kHistoBluePredSubGreen,
+  kHistoPalette,
+  kHistoTotal  // Must be last.
+} HistoIx;
+
+static void AddSingleSubGreen(int p, uint32_t* const r, uint32_t* const b) {
+  const int green = p >> 8;  // The upper bits are masked away later.
+  ++r[((p >> 16) - green) & 0xff];
+  ++b[((p >>  0) - green) & 0xff];
+}
+
+static void AddSingle(uint32_t p,
+                      uint32_t* const a, uint32_t* const r,
+                      uint32_t* const g, uint32_t* const b) {
+  ++a[(p >> 24) & 0xff];
+  ++r[(p >> 16) & 0xff];
+  ++g[(p >>  8) & 0xff];
+  ++b[(p >>  0) & 0xff];
+}
+
+static WEBP_INLINE uint32_t HashPix(uint32_t pix) {
+  // Note that masking with 0xffffffffu is for preventing an
+  // 'unsigned int overflow' warning. Doesn't impact the compiled code.
+  return ((((uint64_t)pix + (pix >> 19)) * 0x39c5fba7ull) & 0xffffffffu) >> 24;
+}
+
+static int AnalyzeEntropy(const uint32_t* argb,
+                          int width, int height, int argb_stride,
+                          int use_palette,
+                          int palette_size, int transform_bits,
+                          EntropyIx* const min_entropy_ix,
+                          int* const red_and_blue_always_zero) {
+  // Allocate histogram set with cache_bits = 0.
+  uint32_t* histo;
+
+  if (use_palette && palette_size <= 16) {
+    // In the case of small palettes, we pack 2, 4 or 8 pixels together. In
+    // practice, small palettes are better than any other transform.
+    *min_entropy_ix = kPalette;
+    *red_and_blue_always_zero = 1;
+    return 1;
+  }
+  histo = (uint32_t*)WebPSafeCalloc(kHistoTotal, sizeof(*histo) * 256);
+  if (histo != NULL) {
+    int i, x, y;
+    const uint32_t* prev_row = NULL;
+    const uint32_t* curr_row = argb;
+    uint32_t pix_prev = argb[0];  // Skip the first pixel.
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const uint32_t pix = curr_row[x];
+        const uint32_t pix_diff = VP8LSubPixels(pix, pix_prev);
+        pix_prev = pix;
+        if ((pix_diff == 0) || (prev_row != NULL && pix == prev_row[x])) {
+          continue;
+        }
+        AddSingle(pix,
+                  &histo[kHistoAlpha * 256],
+                  &histo[kHistoRed * 256],
+                  &histo[kHistoGreen * 256],
+                  &histo[kHistoBlue * 256]);
+        AddSingle(pix_diff,
+                  &histo[kHistoAlphaPred * 256],
+                  &histo[kHistoRedPred * 256],
+                  &histo[kHistoGreenPred * 256],
+                  &histo[kHistoBluePred * 256]);
+        AddSingleSubGreen(pix,
+                          &histo[kHistoRedSubGreen * 256],
+                          &histo[kHistoBlueSubGreen * 256]);
+        AddSingleSubGreen(pix_diff,
+                          &histo[kHistoRedPredSubGreen * 256],
+                          &histo[kHistoBluePredSubGreen * 256]);
+        {
+          // Approximate the palette by the entropy of the multiplicative hash.
+          const uint32_t hash = HashPix(pix);
+          ++histo[kHistoPalette * 256 + hash];
+        }
+      }
+      prev_row = curr_row;
+      curr_row += argb_stride;
+    }
+    {
+      float entropy_comp[kHistoTotal];
+      float entropy[kNumEntropyIx];
+      int k;
+      int last_mode_to_analyze = use_palette ? kPalette : kSpatialSubGreen;
+      int j;
+      // Let's add one zero to the predicted histograms. The zeros are removed
+      // too efficiently by the pix_diff == 0 comparison, at least one of the
+      // zeros is likely to exist.
+      ++histo[kHistoRedPredSubGreen * 256];
+      ++histo[kHistoBluePredSubGreen * 256];
+      ++histo[kHistoRedPred * 256];
+      ++histo[kHistoGreenPred * 256];
+      ++histo[kHistoBluePred * 256];
+      ++histo[kHistoAlphaPred * 256];
+
+      for (j = 0; j < kHistoTotal; ++j) {
+        entropy_comp[j] = VP8LBitsEntropy(&histo[j * 256], 256);
+      }
+      entropy[kDirect] = entropy_comp[kHistoAlpha] +
+          entropy_comp[kHistoRed] +
+          entropy_comp[kHistoGreen] +
+          entropy_comp[kHistoBlue];
+      entropy[kSpatial] = entropy_comp[kHistoAlphaPred] +
+          entropy_comp[kHistoRedPred] +
+          entropy_comp[kHistoGreenPred] +
+          entropy_comp[kHistoBluePred];
+      entropy[kSubGreen] = entropy_comp[kHistoAlpha] +
+          entropy_comp[kHistoRedSubGreen] +
+          entropy_comp[kHistoGreen] +
+          entropy_comp[kHistoBlueSubGreen];
+      entropy[kSpatialSubGreen] = entropy_comp[kHistoAlphaPred] +
+          entropy_comp[kHistoRedPredSubGreen] +
+          entropy_comp[kHistoGreenPred] +
+          entropy_comp[kHistoBluePredSubGreen];
+      entropy[kPalette] = entropy_comp[kHistoPalette];
+
+      // When including transforms, there is an overhead in bits from
+      // storing them. This overhead is small but matters for small images.
+      // For spatial, there are 14 transformations.
+      entropy[kSpatial] += VP8LSubSampleSize(width, transform_bits) *
+                           VP8LSubSampleSize(height, transform_bits) *
+                           VP8LFastLog2(14);
+      // For color transforms: 24 as only 3 channels are considered in a
+      // ColorTransformElement.
+      entropy[kSpatialSubGreen] += VP8LSubSampleSize(width, transform_bits) *
+                                   VP8LSubSampleSize(height, transform_bits) *
+                                   VP8LFastLog2(24);
+      // For palettes, add the cost of storing the palette.
+      // We empirically estimate the cost of a compressed entry as 8 bits.
+      // The palette is differential-coded when compressed hence a much
+      // lower cost than sizeof(uint32_t)*8.
+      entropy[kPalette] += palette_size * 8;
+
+      *min_entropy_ix = kDirect;
+      for (k = kDirect + 1; k <= last_mode_to_analyze; ++k) {
+        if (entropy[*min_entropy_ix] > entropy[k]) {
+          *min_entropy_ix = (EntropyIx)k;
+        }
+      }
+      assert((int)*min_entropy_ix <= last_mode_to_analyze);
+      *red_and_blue_always_zero = 1;
+      // Let's check if the histogram of the chosen entropy mode has
+      // non-zero red and blue values. If all are zero, we can later skip
+      // the cross color optimization.
+      {
+        static const uint8_t kHistoPairs[5][2] = {
+          { kHistoRed, kHistoBlue },
+          { kHistoRedPred, kHistoBluePred },
+          { kHistoRedSubGreen, kHistoBlueSubGreen },
+          { kHistoRedPredSubGreen, kHistoBluePredSubGreen },
+          { kHistoRed, kHistoBlue }
+        };
+        const uint32_t* const red_histo =
+            &histo[256 * kHistoPairs[*min_entropy_ix][0]];
+        const uint32_t* const blue_histo =
+            &histo[256 * kHistoPairs[*min_entropy_ix][1]];
+        for (i = 1; i < 256; ++i) {
+          if ((red_histo[i] | blue_histo[i]) != 0) {
+            *red_and_blue_always_zero = 0;
+            break;
+          }
+        }
+      }
+    }
+    WebPSafeFree(histo);
+    return 1;
+  } else {
+    return 0;
+  }
+}
+
+static int GetHistoBits(int method, int use_palette, int width, int height) {
+  // Make tile size a function of encoding method (Range: 0 to 6).
+  int histo_bits = (use_palette ? 9 : 7) - method;
+  while (1) {
+    const int huff_image_size = VP8LSubSampleSize(width, histo_bits) *
+                                VP8LSubSampleSize(height, histo_bits);
+    if (huff_image_size <= MAX_HUFF_IMAGE_SIZE) break;
+    ++histo_bits;
+  }
+  return (histo_bits < MIN_HUFFMAN_BITS) ? MIN_HUFFMAN_BITS :
+         (histo_bits > MAX_HUFFMAN_BITS) ? MAX_HUFFMAN_BITS : histo_bits;
+}
+
+static int GetTransformBits(int method, int histo_bits) {
+  const int max_transform_bits = (method < 4) ? 6 : (method > 4) ? 4 : 5;
+  const int res =
+      (histo_bits > max_transform_bits) ? max_transform_bits : histo_bits;
+  assert(res <= MAX_TRANSFORM_BITS);
+  return res;
+}
+
+// Set of parameters to be used in each iteration of the cruncher.
+#define CRUNCH_SUBCONFIGS_MAX 2
+typedef struct {
+  int lz77_;
+  int do_no_cache_;
+} CrunchSubConfig;
+typedef struct {
+  int entropy_idx_;
+  PaletteSorting palette_sorting_type_;
+  CrunchSubConfig sub_configs_[CRUNCH_SUBCONFIGS_MAX];
+  int sub_configs_size_;
+} CrunchConfig;
+
+// +2 because we add a palette sorting configuration for kPalette and
+// kPaletteAndSpatial.
+#define CRUNCH_CONFIGS_MAX (kNumEntropyIx + 2)
+
+static int EncoderAnalyze(VP8LEncoder* const enc,
+                          CrunchConfig crunch_configs[CRUNCH_CONFIGS_MAX],
+                          int* const crunch_configs_size,
+                          int* const red_and_blue_always_zero) {
+  const WebPPicture* const pic = enc->pic_;
+  const int width = pic->width;
+  const int height = pic->height;
+  const WebPConfig* const config = enc->config_;
+  const int method = config->method;
+  const int low_effort = (config->method == 0);
+  int i;
+  int use_palette;
+  int n_lz77s;
+  // If set to 0, analyze the cache with the computed cache value. If 1, also
+  // analyze with no-cache.
+  int do_no_cache = 0;
+  assert(pic != NULL && pic->argb != NULL);
+
+  // Check whether a palette is possible.
+  enc->palette_size_ = WebPGetColorPalette(pic, enc->palette_sorted_);
+  use_palette = (enc->palette_size_ <= MAX_PALETTE_SIZE);
+  if (!use_palette) {
+    enc->palette_size_ = 0;
+  } else {
+    qsort(enc->palette_sorted_, enc->palette_size_,
+          sizeof(*enc->palette_sorted_), PaletteCompareColorsForQsort);
+  }
+
+  // Empirical bit sizes.
+  enc->histo_bits_ = GetHistoBits(method, use_palette,
+                                  pic->width, pic->height);
+  enc->transform_bits_ = GetTransformBits(method, enc->histo_bits_);
+
+  if (low_effort) {
+    // AnalyzeEntropy is somewhat slow.
+    crunch_configs[0].entropy_idx_ = use_palette ? kPalette : kSpatialSubGreen;
+    crunch_configs[0].palette_sorting_type_ =
+        use_palette ? kSortedDefault : kUnusedPalette;
+    n_lz77s = 1;
+    *crunch_configs_size = 1;
+  } else {
+    EntropyIx min_entropy_ix;
+    // Try out multiple LZ77 on images with few colors.
+    n_lz77s = (enc->palette_size_ > 0 && enc->palette_size_ <= 16) ? 2 : 1;
+    if (!AnalyzeEntropy(pic->argb, width, height, pic->argb_stride, use_palette,
+                        enc->palette_size_, enc->transform_bits_,
+                        &min_entropy_ix, red_and_blue_always_zero)) {
+      return 0;
+    }
+    if (method == 6 && config->quality == 100) {
+      do_no_cache = 1;
+      // Go brute force on all transforms.
+      *crunch_configs_size = 0;
+      for (i = 0; i < kNumEntropyIx; ++i) {
+        // We can only apply kPalette or kPaletteAndSpatial if we can indeed use
+        // a palette.
+        if ((i != kPalette && i != kPaletteAndSpatial) || use_palette) {
+          assert(*crunch_configs_size < CRUNCH_CONFIGS_MAX);
+          crunch_configs[(*crunch_configs_size)].entropy_idx_ = i;
+          if (use_palette && (i == kPalette || i == kPaletteAndSpatial)) {
+            crunch_configs[(*crunch_configs_size)].palette_sorting_type_ =
+                kMinimizeDelta;
+            ++*crunch_configs_size;
+            // Also add modified Zeng's method.
+            crunch_configs[(*crunch_configs_size)].entropy_idx_ = i;
+            crunch_configs[(*crunch_configs_size)].palette_sorting_type_ =
+                kModifiedZeng;
+          } else {
+            crunch_configs[(*crunch_configs_size)].palette_sorting_type_ =
+                kUnusedPalette;
+          }
+          ++*crunch_configs_size;
+        }
+      }
+    } else {
+      // Only choose the guessed best transform.
+      *crunch_configs_size = 1;
+      crunch_configs[0].entropy_idx_ = min_entropy_ix;
+      crunch_configs[0].palette_sorting_type_ =
+          use_palette ? kMinimizeDelta : kUnusedPalette;
+      if (config->quality >= 75 && method == 5) {
+        // Test with and without color cache.
+        do_no_cache = 1;
+        // If we have a palette, also check in combination with spatial.
+        if (min_entropy_ix == kPalette) {
+          *crunch_configs_size = 2;
+          crunch_configs[1].entropy_idx_ = kPaletteAndSpatial;
+          crunch_configs[1].palette_sorting_type_ = kMinimizeDelta;
+        }
+      }
+    }
+  }
+  // Fill in the different LZ77s.
+  assert(n_lz77s <= CRUNCH_SUBCONFIGS_MAX);
+  for (i = 0; i < *crunch_configs_size; ++i) {
+    int j;
+    for (j = 0; j < n_lz77s; ++j) {
+      assert(j < CRUNCH_SUBCONFIGS_MAX);
+      crunch_configs[i].sub_configs_[j].lz77_ =
+          (j == 0) ? kLZ77Standard | kLZ77RLE : kLZ77Box;
+      crunch_configs[i].sub_configs_[j].do_no_cache_ = do_no_cache;
+    }
+    crunch_configs[i].sub_configs_size_ = n_lz77s;
+  }
+  return 1;
+}
+
+static int EncoderInit(VP8LEncoder* const enc) {
+  const WebPPicture* const pic = enc->pic_;
+  const int width = pic->width;
+  const int height = pic->height;
+  const int pix_cnt = width * height;
+  // we round the block size up, so we're guaranteed to have
+  // at most MAX_REFS_BLOCK_PER_IMAGE blocks used:
+  const int refs_block_size = (pix_cnt - 1) / MAX_REFS_BLOCK_PER_IMAGE + 1;
+  int i;
+  if (!VP8LHashChainInit(&enc->hash_chain_, pix_cnt)) return 0;
+
+  for (i = 0; i < 4; ++i) VP8LBackwardRefsInit(&enc->refs_[i], refs_block_size);
+
+  return 1;
+}
+
+// Returns false in case of memory error.
+static int GetHuffBitLengthsAndCodes(
+    const VP8LHistogramSet* const histogram_image,
+    HuffmanTreeCode* const huffman_codes) {
+  int i, k;
+  int ok = 0;
+  uint64_t total_length_size = 0;
+  uint8_t* mem_buf = NULL;
+  const int histogram_image_size = histogram_image->size;
+  int max_num_symbols = 0;
+  uint8_t* buf_rle = NULL;
+  HuffmanTree* huff_tree = NULL;
+
+  // Iterate over all histograms and get the aggregate number of codes used.
+  for (i = 0; i < histogram_image_size; ++i) {
+    const VP8LHistogram* const histo = histogram_image->histograms[i];
+    HuffmanTreeCode* const codes = &huffman_codes[5 * i];
+    assert(histo != NULL);
+    for (k = 0; k < 5; ++k) {
+      const int num_symbols =
+          (k == 0) ? VP8LHistogramNumCodes(histo->palette_code_bits_) :
+          (k == 4) ? NUM_DISTANCE_CODES : 256;
+      codes[k].num_symbols = num_symbols;
+      total_length_size += num_symbols;
+    }
+  }
+
+  // Allocate and Set Huffman codes.
+  {
+    uint16_t* codes;
+    uint8_t* lengths;
+    mem_buf = (uint8_t*)WebPSafeCalloc(total_length_size,
+                                       sizeof(*lengths) + sizeof(*codes));
+    if (mem_buf == NULL) goto End;
+
+    codes = (uint16_t*)mem_buf;
+    lengths = (uint8_t*)&codes[total_length_size];
+    for (i = 0; i < 5 * histogram_image_size; ++i) {
+      const int bit_length = huffman_codes[i].num_symbols;
+      huffman_codes[i].codes = codes;
+      huffman_codes[i].code_lengths = lengths;
+      codes += bit_length;
+      lengths += bit_length;
+      if (max_num_symbols < bit_length) {
+        max_num_symbols = bit_length;
+      }
+    }
+  }
+
+  buf_rle = (uint8_t*)WebPSafeMalloc(1ULL, max_num_symbols);
+  huff_tree = (HuffmanTree*)WebPSafeMalloc(3ULL * max_num_symbols,
+                                           sizeof(*huff_tree));
+  if (buf_rle == NULL || huff_tree == NULL) goto End;
+
+  // Create Huffman trees.
+  for (i = 0; i < histogram_image_size; ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[5 * i];
+    VP8LHistogram* const histo = histogram_image->histograms[i];
+    VP8LCreateHuffmanTree(histo->literal_, 15, buf_rle, huff_tree, codes + 0);
+    VP8LCreateHuffmanTree(histo->red_, 15, buf_rle, huff_tree, codes + 1);
+    VP8LCreateHuffmanTree(histo->blue_, 15, buf_rle, huff_tree, codes + 2);
+    VP8LCreateHuffmanTree(histo->alpha_, 15, buf_rle, huff_tree, codes + 3);
+    VP8LCreateHuffmanTree(histo->distance_, 15, buf_rle, huff_tree, codes + 4);
+  }
+  ok = 1;
+ End:
+  WebPSafeFree(huff_tree);
+  WebPSafeFree(buf_rle);
+  if (!ok) {
+    WebPSafeFree(mem_buf);
+    memset(huffman_codes, 0, 5 * histogram_image_size * sizeof(*huffman_codes));
+  }
+  return ok;
+}
+
+static void StoreHuffmanTreeOfHuffmanTreeToBitMask(
+    VP8LBitWriter* const bw, const uint8_t* code_length_bitdepth) {
+  // RFC 1951 will calm you down if you are worried about this funny sequence.
+  // This sequence is tuned from that, but more weighted for lower symbol count,
+  // and more spiking histograms.
+  static const uint8_t kStorageOrder[CODE_LENGTH_CODES] = {
+    17, 18, 0, 1, 2, 3, 4, 5, 16, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
+  };
+  int i;
+  // Throw away trailing zeros:
+  int codes_to_store = CODE_LENGTH_CODES;
+  for (; codes_to_store > 4; --codes_to_store) {
+    if (code_length_bitdepth[kStorageOrder[codes_to_store - 1]] != 0) {
+      break;
+    }
+  }
+  VP8LPutBits(bw, codes_to_store - 4, 4);
+  for (i = 0; i < codes_to_store; ++i) {
+    VP8LPutBits(bw, code_length_bitdepth[kStorageOrder[i]], 3);
+  }
+}
+
+static void ClearHuffmanTreeIfOnlyOneSymbol(
+    HuffmanTreeCode* const huffman_code) {
+  int k;
+  int count = 0;
+  for (k = 0; k < huffman_code->num_symbols; ++k) {
+    if (huffman_code->code_lengths[k] != 0) {
+      ++count;
+      if (count > 1) return;
+    }
+  }
+  for (k = 0; k < huffman_code->num_symbols; ++k) {
+    huffman_code->code_lengths[k] = 0;
+    huffman_code->codes[k] = 0;
+  }
+}
+
+static void StoreHuffmanTreeToBitMask(
+    VP8LBitWriter* const bw,
+    const HuffmanTreeToken* const tokens, const int num_tokens,
+    const HuffmanTreeCode* const huffman_code) {
+  int i;
+  for (i = 0; i < num_tokens; ++i) {
+    const int ix = tokens[i].code;
+    const int extra_bits = tokens[i].extra_bits;
+    VP8LPutBits(bw, huffman_code->codes[ix], huffman_code->code_lengths[ix]);
+    switch (ix) {
+      case 16:
+        VP8LPutBits(bw, extra_bits, 2);
+        break;
+      case 17:
+        VP8LPutBits(bw, extra_bits, 3);
+        break;
+      case 18:
+        VP8LPutBits(bw, extra_bits, 7);
+        break;
+    }
+  }
+}
+
+// 'huff_tree' and 'tokens' are pre-alloacted buffers.
+static void StoreFullHuffmanCode(VP8LBitWriter* const bw,
+                                 HuffmanTree* const huff_tree,
+                                 HuffmanTreeToken* const tokens,
+                                 const HuffmanTreeCode* const tree) {
+  uint8_t code_length_bitdepth[CODE_LENGTH_CODES] = { 0 };
+  uint16_t code_length_bitdepth_symbols[CODE_LENGTH_CODES] = { 0 };
+  const int max_tokens = tree->num_symbols;
+  int num_tokens;
+  HuffmanTreeCode huffman_code;
+  huffman_code.num_symbols = CODE_LENGTH_CODES;
+  huffman_code.code_lengths = code_length_bitdepth;
+  huffman_code.codes = code_length_bitdepth_symbols;
+
+  VP8LPutBits(bw, 0, 1);
+  num_tokens = VP8LCreateCompressedHuffmanTree(tree, tokens, max_tokens);
+  {
+    uint32_t histogram[CODE_LENGTH_CODES] = { 0 };
+    uint8_t buf_rle[CODE_LENGTH_CODES] = { 0 };
+    int i;
+    for (i = 0; i < num_tokens; ++i) {
+      ++histogram[tokens[i].code];
+    }
+
+    VP8LCreateHuffmanTree(histogram, 7, buf_rle, huff_tree, &huffman_code);
+  }
+
+  StoreHuffmanTreeOfHuffmanTreeToBitMask(bw, code_length_bitdepth);
+  ClearHuffmanTreeIfOnlyOneSymbol(&huffman_code);
+  {
+    int trailing_zero_bits = 0;
+    int trimmed_length = num_tokens;
+    int write_trimmed_length;
+    int length;
+    int i = num_tokens;
+    while (i-- > 0) {
+      const int ix = tokens[i].code;
+      if (ix == 0 || ix == 17 || ix == 18) {
+        --trimmed_length;   // discount trailing zeros
+        trailing_zero_bits += code_length_bitdepth[ix];
+        if (ix == 17) {
+          trailing_zero_bits += 3;
+        } else if (ix == 18) {
+          trailing_zero_bits += 7;
+        }
+      } else {
+        break;
+      }
+    }
+    write_trimmed_length = (trimmed_length > 1 && trailing_zero_bits > 12);
+    length = write_trimmed_length ? trimmed_length : num_tokens;
+    VP8LPutBits(bw, write_trimmed_length, 1);
+    if (write_trimmed_length) {
+      if (trimmed_length == 2) {
+        VP8LPutBits(bw, 0, 3 + 2);     // nbitpairs=1, trimmed_length=2
+      } else {
+        const int nbits = BitsLog2Floor(trimmed_length - 2);
+        const int nbitpairs = nbits / 2 + 1;
+        assert(trimmed_length > 2);
+        assert(nbitpairs - 1 < 8);
+        VP8LPutBits(bw, nbitpairs - 1, 3);
+        VP8LPutBits(bw, trimmed_length - 2, nbitpairs * 2);
+      }
+    }
+    StoreHuffmanTreeToBitMask(bw, tokens, length, &huffman_code);
+  }
+}
+
+// 'huff_tree' and 'tokens' are pre-alloacted buffers.
+static void StoreHuffmanCode(VP8LBitWriter* const bw,
+                             HuffmanTree* const huff_tree,
+                             HuffmanTreeToken* const tokens,
+                             const HuffmanTreeCode* const huffman_code) {
+  int i;
+  int count = 0;
+  int symbols[2] = { 0, 0 };
+  const int kMaxBits = 8;
+  const int kMaxSymbol = 1 << kMaxBits;
+
+  // Check whether it's a small tree.
+  for (i = 0; i < huffman_code->num_symbols && count < 3; ++i) {
+    if (huffman_code->code_lengths[i] != 0) {
+      if (count < 2) symbols[count] = i;
+      ++count;
+    }
+  }
+
+  if (count == 0) {   // emit minimal tree for empty cases
+    // bits: small tree marker: 1, count-1: 0, large 8-bit code: 0, code: 0
+    VP8LPutBits(bw, 0x01, 4);
+  } else if (count <= 2 && symbols[0] < kMaxSymbol && symbols[1] < kMaxSymbol) {
+    VP8LPutBits(bw, 1, 1);  // Small tree marker to encode 1 or 2 symbols.
+    VP8LPutBits(bw, count - 1, 1);
+    if (symbols[0] <= 1) {
+      VP8LPutBits(bw, 0, 1);  // Code bit for small (1 bit) symbol value.
+      VP8LPutBits(bw, symbols[0], 1);
+    } else {
+      VP8LPutBits(bw, 1, 1);
+      VP8LPutBits(bw, symbols[0], 8);
+    }
+    if (count == 2) {
+      VP8LPutBits(bw, symbols[1], 8);
+    }
+  } else {
+    StoreFullHuffmanCode(bw, huff_tree, tokens, huffman_code);
+  }
+}
+
+static WEBP_INLINE void WriteHuffmanCode(VP8LBitWriter* const bw,
+                             const HuffmanTreeCode* const code,
+                             int code_index) {
+  const int depth = code->code_lengths[code_index];
+  const int symbol = code->codes[code_index];
+  VP8LPutBits(bw, symbol, depth);
+}
+
+static WEBP_INLINE void WriteHuffmanCodeWithExtraBits(
+    VP8LBitWriter* const bw,
+    const HuffmanTreeCode* const code,
+    int code_index,
+    int bits,
+    int n_bits) {
+  const int depth = code->code_lengths[code_index];
+  const int symbol = code->codes[code_index];
+  VP8LPutBits(bw, (bits << depth) | symbol, depth + n_bits);
+}
+
+static int StoreImageToBitMask(
+    VP8LBitWriter* const bw, int width, int histo_bits,
+    const VP8LBackwardRefs* const refs,
+    const uint16_t* histogram_symbols,
+    const HuffmanTreeCode* const huffman_codes, const WebPPicture* const pic) {
+  const int histo_xsize = histo_bits ? VP8LSubSampleSize(width, histo_bits) : 1;
+  const int tile_mask = (histo_bits == 0) ? 0 : -(1 << histo_bits);
+  // x and y trace the position in the image.
+  int x = 0;
+  int y = 0;
+  int tile_x = x & tile_mask;
+  int tile_y = y & tile_mask;
+  int histogram_ix = histogram_symbols[0];
+  const HuffmanTreeCode* codes = huffman_codes + 5 * histogram_ix;
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  while (VP8LRefsCursorOk(&c)) {
+    const PixOrCopy* const v = c.cur_pos;
+    if ((tile_x != (x & tile_mask)) || (tile_y != (y & tile_mask))) {
+      tile_x = x & tile_mask;
+      tile_y = y & tile_mask;
+      histogram_ix = histogram_symbols[(y >> histo_bits) * histo_xsize +
+                                       (x >> histo_bits)];
+      codes = huffman_codes + 5 * histogram_ix;
+    }
+    if (PixOrCopyIsLiteral(v)) {
+      static const uint8_t order[] = { 1, 2, 0, 3 };
+      int k;
+      for (k = 0; k < 4; ++k) {
+        const int code = PixOrCopyLiteral(v, order[k]);
+        WriteHuffmanCode(bw, codes + k, code);
+      }
+    } else if (PixOrCopyIsCacheIdx(v)) {
+      const int code = PixOrCopyCacheIdx(v);
+      const int literal_ix = 256 + NUM_LENGTH_CODES + code;
+      WriteHuffmanCode(bw, codes, literal_ix);
+    } else {
+      int bits, n_bits;
+      int code;
+
+      const int distance = PixOrCopyDistance(v);
+      VP8LPrefixEncode(v->len, &code, &n_bits, &bits);
+      WriteHuffmanCodeWithExtraBits(bw, codes, 256 + code, bits, n_bits);
+
+      // Don't write the distance with the extra bits code since
+      // the distance can be up to 18 bits of extra bits, and the prefix
+      // 15 bits, totaling to 33, and our PutBits only supports up to 32 bits.
+      VP8LPrefixEncode(distance, &code, &n_bits, &bits);
+      WriteHuffmanCode(bw, codes + 4, code);
+      VP8LPutBits(bw, bits, n_bits);
+    }
+    x += PixOrCopyLength(v);
+    while (x >= width) {
+      x -= width;
+      ++y;
+    }
+    VP8LRefsCursorNext(&c);
+  }
+  if (bw->error_) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return 0;
+  }
+  return 1;
+}
+
+// Special case of EncodeImageInternal() for cache-bits=0, histo_bits=31.
+// pic and percent are for progress.
+static int EncodeImageNoHuffman(VP8LBitWriter* const bw,
+                                const uint32_t* const argb,
+                                VP8LHashChain* const hash_chain,
+                                VP8LBackwardRefs* const refs_array, int width,
+                                int height, int quality, int low_effort,
+                                const WebPPicture* const pic, int percent_range,
+                                int* const percent) {
+  int i;
+  int max_tokens = 0;
+  VP8LBackwardRefs* refs;
+  HuffmanTreeToken* tokens = NULL;
+  HuffmanTreeCode huffman_codes[5] = {{0, NULL, NULL}};
+  const uint16_t histogram_symbols[1] = {0};  // only one tree, one symbol
+  int cache_bits = 0;
+  VP8LHistogramSet* histogram_image = NULL;
+  HuffmanTree* const huff_tree = (HuffmanTree*)WebPSafeMalloc(
+      3ULL * CODE_LENGTH_CODES, sizeof(*huff_tree));
+  if (huff_tree == NULL) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  // Calculate backward references from ARGB image.
+  if (!VP8LHashChainFill(hash_chain, quality, argb, width, height, low_effort,
+                         pic, percent_range / 2, percent)) {
+    goto Error;
+  }
+  if (!VP8LGetBackwardReferences(width, height, argb, quality, /*low_effort=*/0,
+                                 kLZ77Standard | kLZ77RLE, cache_bits,
+                                 /*do_no_cache=*/0, hash_chain, refs_array,
+                                 &cache_bits, pic,
+                                 percent_range - percent_range / 2, percent)) {
+    goto Error;
+  }
+  refs = &refs_array[0];
+  histogram_image = VP8LAllocateHistogramSet(1, cache_bits);
+  if (histogram_image == NULL) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+  VP8LHistogramSetClear(histogram_image);
+
+  // Build histogram image and symbols from backward references.
+  VP8LHistogramStoreRefs(refs, histogram_image->histograms[0]);
+
+  // Create Huffman bit lengths and codes for each histogram image.
+  assert(histogram_image->size == 1);
+  if (!GetHuffBitLengthsAndCodes(histogram_image, huffman_codes)) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  // No color cache, no Huffman image.
+  VP8LPutBits(bw, 0, 1);
+
+  // Find maximum number of symbols for the huffman tree-set.
+  for (i = 0; i < 5; ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[i];
+    if (max_tokens < codes->num_symbols) {
+      max_tokens = codes->num_symbols;
+    }
+  }
+
+  tokens = (HuffmanTreeToken*)WebPSafeMalloc(max_tokens, sizeof(*tokens));
+  if (tokens == NULL) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  // Store Huffman codes.
+  for (i = 0; i < 5; ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[i];
+    StoreHuffmanCode(bw, huff_tree, tokens, codes);
+    ClearHuffmanTreeIfOnlyOneSymbol(codes);
+  }
+
+  // Store actual literals.
+  if (!StoreImageToBitMask(bw, width, 0, refs, histogram_symbols, huffman_codes,
+                           pic)) {
+    goto Error;
+  }
+
+ Error:
+  WebPSafeFree(tokens);
+  WebPSafeFree(huff_tree);
+  VP8LFreeHistogramSet(histogram_image);
+  WebPSafeFree(huffman_codes[0].codes);
+  return (pic->error_code == VP8_ENC_OK);
+}
+
+// pic and percent are for progress.
+static int EncodeImageInternal(
+    VP8LBitWriter* const bw, const uint32_t* const argb,
+    VP8LHashChain* const hash_chain, VP8LBackwardRefs refs_array[4], int width,
+    int height, int quality, int low_effort, int use_cache,
+    const CrunchConfig* const config, int* cache_bits, int histogram_bits,
+    size_t init_byte_position, int* const hdr_size, int* const data_size,
+    const WebPPicture* const pic, int percent_range, int* const percent) {
+  const uint32_t histogram_image_xysize =
+      VP8LSubSampleSize(width, histogram_bits) *
+      VP8LSubSampleSize(height, histogram_bits);
+  int remaining_percent = percent_range;
+  int percent_start = *percent;
+  VP8LHistogramSet* histogram_image = NULL;
+  VP8LHistogram* tmp_histo = NULL;
+  int histogram_image_size = 0;
+  size_t bit_array_size = 0;
+  HuffmanTree* const huff_tree = (HuffmanTree*)WebPSafeMalloc(
+      3ULL * CODE_LENGTH_CODES, sizeof(*huff_tree));
+  HuffmanTreeToken* tokens = NULL;
+  HuffmanTreeCode* huffman_codes = NULL;
+  uint16_t* const histogram_symbols = (uint16_t*)WebPSafeMalloc(
+      histogram_image_xysize, sizeof(*histogram_symbols));
+  int sub_configs_idx;
+  int cache_bits_init, write_histogram_image;
+  VP8LBitWriter bw_init = *bw, bw_best;
+  int hdr_size_tmp;
+  VP8LHashChain hash_chain_histogram;  // histogram image hash chain
+  size_t bw_size_best = ~(size_t)0;
+  assert(histogram_bits >= MIN_HUFFMAN_BITS);
+  assert(histogram_bits <= MAX_HUFFMAN_BITS);
+  assert(hdr_size != NULL);
+  assert(data_size != NULL);
+
+  memset(&hash_chain_histogram, 0, sizeof(hash_chain_histogram));
+  if (!VP8LBitWriterInit(&bw_best, 0)) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  // Make sure we can allocate the different objects.
+  if (huff_tree == NULL || histogram_symbols == NULL ||
+      !VP8LHashChainInit(&hash_chain_histogram, histogram_image_xysize)) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  percent_range = remaining_percent / 5;
+  if (!VP8LHashChainFill(hash_chain, quality, argb, width, height,
+                         low_effort, pic, percent_range, percent)) {
+    goto Error;
+  }
+  percent_start += percent_range;
+  remaining_percent -= percent_range;
+
+  if (use_cache) {
+    // If the value is different from zero, it has been set during the
+    // palette analysis.
+    cache_bits_init = (*cache_bits == 0) ? MAX_COLOR_CACHE_BITS : *cache_bits;
+  } else {
+    cache_bits_init = 0;
+  }
+  // If several iterations will happen, clone into bw_best.
+  if ((config->sub_configs_size_ > 1 || config->sub_configs_[0].do_no_cache_) &&
+      !VP8LBitWriterClone(bw, &bw_best)) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  for (sub_configs_idx = 0; sub_configs_idx < config->sub_configs_size_;
+       ++sub_configs_idx) {
+    const CrunchSubConfig* const sub_config =
+        &config->sub_configs_[sub_configs_idx];
+    int cache_bits_best, i_cache;
+    int i_remaining_percent = remaining_percent / config->sub_configs_size_;
+    int i_percent_range = i_remaining_percent / 4;
+    i_remaining_percent -= i_percent_range;
+
+    if (!VP8LGetBackwardReferences(
+            width, height, argb, quality, low_effort, sub_config->lz77_,
+            cache_bits_init, sub_config->do_no_cache_, hash_chain,
+            &refs_array[0], &cache_bits_best, pic, i_percent_range, percent)) {
+      goto Error;
+    }
+
+    for (i_cache = 0; i_cache < (sub_config->do_no_cache_ ? 2 : 1); ++i_cache) {
+      const int cache_bits_tmp = (i_cache == 0) ? cache_bits_best : 0;
+      // Speed-up: no need to study the no-cache case if it was already studied
+      // in i_cache == 0.
+      if (i_cache == 1 && cache_bits_best == 0) break;
+
+      // Reset the bit writer for this iteration.
+      VP8LBitWriterReset(&bw_init, bw);
+
+      // Build histogram image and symbols from backward references.
+      histogram_image =
+          VP8LAllocateHistogramSet(histogram_image_xysize, cache_bits_tmp);
+      tmp_histo = VP8LAllocateHistogram(cache_bits_tmp);
+      if (histogram_image == NULL || tmp_histo == NULL) {
+        WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+        goto Error;
+      }
+
+      i_percent_range = i_remaining_percent / 3;
+      i_remaining_percent -= i_percent_range;
+      if (!VP8LGetHistoImageSymbols(
+              width, height, &refs_array[i_cache], quality, low_effort,
+              histogram_bits, cache_bits_tmp, histogram_image, tmp_histo,
+              histogram_symbols, pic, i_percent_range, percent)) {
+        goto Error;
+      }
+      // Create Huffman bit lengths and codes for each histogram image.
+      histogram_image_size = histogram_image->size;
+      bit_array_size = 5 * histogram_image_size;
+      huffman_codes = (HuffmanTreeCode*)WebPSafeCalloc(bit_array_size,
+                                                       sizeof(*huffman_codes));
+      // Note: some histogram_image entries may point to tmp_histos[], so the
+      // latter need to outlive the following call to
+      // GetHuffBitLengthsAndCodes().
+      if (huffman_codes == NULL ||
+          !GetHuffBitLengthsAndCodes(histogram_image, huffman_codes)) {
+        WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+        goto Error;
+      }
+      // Free combined histograms.
+      VP8LFreeHistogramSet(histogram_image);
+      histogram_image = NULL;
+
+      // Free scratch histograms.
+      VP8LFreeHistogram(tmp_histo);
+      tmp_histo = NULL;
+
+      // Color Cache parameters.
+      if (cache_bits_tmp > 0) {
+        VP8LPutBits(bw, 1, 1);
+        VP8LPutBits(bw, cache_bits_tmp, 4);
+      } else {
+        VP8LPutBits(bw, 0, 1);
+      }
+
+      // Huffman image + meta huffman.
+      write_histogram_image = (histogram_image_size > 1);
+      VP8LPutBits(bw, write_histogram_image, 1);
+      if (write_histogram_image) {
+        uint32_t* const histogram_argb = (uint32_t*)WebPSafeMalloc(
+            histogram_image_xysize, sizeof(*histogram_argb));
+        int max_index = 0;
+        uint32_t i;
+        if (histogram_argb == NULL) {
+          WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+          goto Error;
+        }
+        for (i = 0; i < histogram_image_xysize; ++i) {
+          const int symbol_index = histogram_symbols[i] & 0xffff;
+          histogram_argb[i] = (symbol_index << 8);
+          if (symbol_index >= max_index) {
+            max_index = symbol_index + 1;
+          }
+        }
+        histogram_image_size = max_index;
+
+        VP8LPutBits(bw, histogram_bits - 2, 3);
+        i_percent_range = i_remaining_percent / 2;
+        i_remaining_percent -= i_percent_range;
+        if (!EncodeImageNoHuffman(
+                bw, histogram_argb, &hash_chain_histogram, &refs_array[2],
+                VP8LSubSampleSize(width, histogram_bits),
+                VP8LSubSampleSize(height, histogram_bits), quality, low_effort,
+                pic, i_percent_range, percent)) {
+          WebPSafeFree(histogram_argb);
+          goto Error;
+        }
+        WebPSafeFree(histogram_argb);
+      }
+
+      // Store Huffman codes.
+      {
+        int i;
+        int max_tokens = 0;
+        // Find maximum number of symbols for the huffman tree-set.
+        for (i = 0; i < 5 * histogram_image_size; ++i) {
+          HuffmanTreeCode* const codes = &huffman_codes[i];
+          if (max_tokens < codes->num_symbols) {
+            max_tokens = codes->num_symbols;
+          }
+        }
+        tokens = (HuffmanTreeToken*)WebPSafeMalloc(max_tokens, sizeof(*tokens));
+        if (tokens == NULL) goto Error;
+        for (i = 0; i < 5 * histogram_image_size; ++i) {
+          HuffmanTreeCode* const codes = &huffman_codes[i];
+          StoreHuffmanCode(bw, huff_tree, tokens, codes);
+          ClearHuffmanTreeIfOnlyOneSymbol(codes);
+        }
+      }
+      // Store actual literals.
+      hdr_size_tmp = (int)(VP8LBitWriterNumBytes(bw) - init_byte_position);
+      if (!StoreImageToBitMask(bw, width, histogram_bits, &refs_array[i_cache],
+                               histogram_symbols, huffman_codes, pic)) {
+        goto Error;
+      }
+      // Keep track of the smallest image so far.
+      if (VP8LBitWriterNumBytes(bw) < bw_size_best) {
+        bw_size_best = VP8LBitWriterNumBytes(bw);
+        *cache_bits = cache_bits_tmp;
+        *hdr_size = hdr_size_tmp;
+        *data_size =
+            (int)(VP8LBitWriterNumBytes(bw) - init_byte_position - *hdr_size);
+        VP8LBitWriterSwap(bw, &bw_best);
+      }
+      WebPSafeFree(tokens);
+      tokens = NULL;
+      if (huffman_codes != NULL) {
+        WebPSafeFree(huffman_codes->codes);
+        WebPSafeFree(huffman_codes);
+        huffman_codes = NULL;
+      }
+    }
+  }
+  VP8LBitWriterSwap(bw, &bw_best);
+
+  if (!WebPReportProgress(pic, percent_start + remaining_percent, percent)) {
+    goto Error;
+  }
+
+ Error:
+  WebPSafeFree(tokens);
+  WebPSafeFree(huff_tree);
+  VP8LFreeHistogramSet(histogram_image);
+  VP8LFreeHistogram(tmp_histo);
+  VP8LHashChainClear(&hash_chain_histogram);
+  if (huffman_codes != NULL) {
+    WebPSafeFree(huffman_codes->codes);
+    WebPSafeFree(huffman_codes);
+  }
+  WebPSafeFree(histogram_symbols);
+  VP8LBitWriterWipeOut(&bw_best);
+  return (pic->error_code == VP8_ENC_OK);
+}
+
+// -----------------------------------------------------------------------------
+// Transforms
+
+static void ApplySubtractGreen(VP8LEncoder* const enc, int width, int height,
+                               VP8LBitWriter* const bw) {
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, SUBTRACT_GREEN, 2);
+  VP8LSubtractGreenFromBlueAndRed(enc->argb_, width * height);
+}
+
+static int ApplyPredictFilter(const VP8LEncoder* const enc, int width,
+                              int height, int quality, int low_effort,
+                              int used_subtract_green, VP8LBitWriter* const bw,
+                              int percent_range, int* const percent) {
+  const int pred_bits = enc->transform_bits_;
+  const int transform_width = VP8LSubSampleSize(width, pred_bits);
+  const int transform_height = VP8LSubSampleSize(height, pred_bits);
+  // we disable near-lossless quantization if palette is used.
+  const int near_lossless_strength =
+      enc->use_palette_ ? 100 : enc->config_->near_lossless;
+
+  if (!VP8LResidualImage(
+          width, height, pred_bits, low_effort, enc->argb_, enc->argb_scratch_,
+          enc->transform_data_, near_lossless_strength, enc->config_->exact,
+          used_subtract_green, enc->pic_, percent_range / 2, percent)) {
+    return 0;
+  }
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, PREDICTOR_TRANSFORM, 2);
+  assert(pred_bits >= 2);
+  VP8LPutBits(bw, pred_bits - 2, 3);
+  return EncodeImageNoHuffman(
+      bw, enc->transform_data_, (VP8LHashChain*)&enc->hash_chain_,
+      (VP8LBackwardRefs*)&enc->refs_[0], transform_width, transform_height,
+      quality, low_effort, enc->pic_, percent_range - percent_range / 2,
+      percent);
+}
+
+static int ApplyCrossColorFilter(const VP8LEncoder* const enc, int width,
+                                 int height, int quality, int low_effort,
+                                 VP8LBitWriter* const bw, int percent_range,
+                                 int* const percent) {
+  const int ccolor_transform_bits = enc->transform_bits_;
+  const int transform_width = VP8LSubSampleSize(width, ccolor_transform_bits);
+  const int transform_height = VP8LSubSampleSize(height, ccolor_transform_bits);
+
+  if (!VP8LColorSpaceTransform(width, height, ccolor_transform_bits, quality,
+                               enc->argb_, enc->transform_data_, enc->pic_,
+                               percent_range / 2, percent)) {
+    return 0;
+  }
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, CROSS_COLOR_TRANSFORM, 2);
+  assert(ccolor_transform_bits >= 2);
+  VP8LPutBits(bw, ccolor_transform_bits - 2, 3);
+  return EncodeImageNoHuffman(
+      bw, enc->transform_data_, (VP8LHashChain*)&enc->hash_chain_,
+      (VP8LBackwardRefs*)&enc->refs_[0], transform_width, transform_height,
+      quality, low_effort, enc->pic_, percent_range - percent_range / 2,
+      percent);
+}
+
+// -----------------------------------------------------------------------------
+
+static int WriteRiffHeader(const WebPPicture* const pic, size_t riff_size,
+                           size_t vp8l_size) {
+  uint8_t riff[RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE + VP8L_SIGNATURE_SIZE] = {
+    'R', 'I', 'F', 'F', 0, 0, 0, 0, 'W', 'E', 'B', 'P',
+    'V', 'P', '8', 'L', 0, 0, 0, 0, VP8L_MAGIC_BYTE,
+  };
+  PutLE32(riff + TAG_SIZE, (uint32_t)riff_size);
+  PutLE32(riff + RIFF_HEADER_SIZE + TAG_SIZE, (uint32_t)vp8l_size);
+  return pic->writer(riff, sizeof(riff), pic);
+}
+
+static int WriteImageSize(const WebPPicture* const pic,
+                          VP8LBitWriter* const bw) {
+  const int width = pic->width - 1;
+  const int height = pic->height - 1;
+  assert(width < WEBP_MAX_DIMENSION && height < WEBP_MAX_DIMENSION);
+
+  VP8LPutBits(bw, width, VP8L_IMAGE_SIZE_BITS);
+  VP8LPutBits(bw, height, VP8L_IMAGE_SIZE_BITS);
+  return !bw->error_;
+}
+
+static int WriteRealAlphaAndVersion(VP8LBitWriter* const bw, int has_alpha) {
+  VP8LPutBits(bw, has_alpha, 1);
+  VP8LPutBits(bw, VP8L_VERSION, VP8L_VERSION_BITS);
+  return !bw->error_;
+}
+
+static int WriteImage(const WebPPicture* const pic, VP8LBitWriter* const bw,
+                      size_t* const coded_size) {
+  const uint8_t* const webpll_data = VP8LBitWriterFinish(bw);
+  const size_t webpll_size = VP8LBitWriterNumBytes(bw);
+  const size_t vp8l_size = VP8L_SIGNATURE_SIZE + webpll_size;
+  const size_t pad = vp8l_size & 1;
+  const size_t riff_size = TAG_SIZE + CHUNK_HEADER_SIZE + vp8l_size + pad;
+
+  if (!WriteRiffHeader(pic, riff_size, vp8l_size) ||
+      !pic->writer(webpll_data, webpll_size, pic)) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_BAD_WRITE);
+    return 0;
+  }
+
+  if (pad) {
+    const uint8_t pad_byte[1] = { 0 };
+    if (!pic->writer(pad_byte, 1, pic)) {
+      WebPEncodingSetError(pic, VP8_ENC_ERROR_BAD_WRITE);
+      return 0;
+    }
+  }
+  *coded_size = CHUNK_HEADER_SIZE + riff_size;
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+
+static void ClearTransformBuffer(VP8LEncoder* const enc) {
+  WebPSafeFree(enc->transform_mem_);
+  enc->transform_mem_ = NULL;
+  enc->transform_mem_size_ = 0;
+}
+
+// Allocates the memory for argb (W x H) buffer, 2 rows of context for
+// prediction and transform data.
+// Flags influencing the memory allocated:
+//  enc->transform_bits_
+//  enc->use_predict_, enc->use_cross_color_
+static int AllocateTransformBuffer(VP8LEncoder* const enc, int width,
+                                   int height) {
+  const uint64_t image_size = width * height;
+  // VP8LResidualImage needs room for 2 scanlines of uint32 pixels with an extra
+  // pixel in each, plus 2 regular scanlines of bytes.
+  // TODO(skal): Clean up by using arithmetic in bytes instead of words.
+  const uint64_t argb_scratch_size =
+      enc->use_predict_ ? (width + 1) * 2 + (width * 2 + sizeof(uint32_t) - 1) /
+                                                sizeof(uint32_t)
+                        : 0;
+  const uint64_t transform_data_size =
+      (enc->use_predict_ || enc->use_cross_color_)
+          ? VP8LSubSampleSize(width, enc->transform_bits_) *
+                VP8LSubSampleSize(height, enc->transform_bits_)
+          : 0;
+  const uint64_t max_alignment_in_words =
+      (WEBP_ALIGN_CST + sizeof(uint32_t) - 1) / sizeof(uint32_t);
+  const uint64_t mem_size = image_size + max_alignment_in_words +
+                            argb_scratch_size + max_alignment_in_words +
+                            transform_data_size;
+  uint32_t* mem = enc->transform_mem_;
+  if (mem == NULL || mem_size > enc->transform_mem_size_) {
+    ClearTransformBuffer(enc);
+    mem = (uint32_t*)WebPSafeMalloc(mem_size, sizeof(*mem));
+    if (mem == NULL) {
+      WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+      return 0;
+    }
+    enc->transform_mem_ = mem;
+    enc->transform_mem_size_ = (size_t)mem_size;
+    enc->argb_content_ = kEncoderNone;
+  }
+  enc->argb_ = mem;
+  mem = (uint32_t*)WEBP_ALIGN(mem + image_size);
+  enc->argb_scratch_ = mem;
+  mem = (uint32_t*)WEBP_ALIGN(mem + argb_scratch_size);
+  enc->transform_data_ = mem;
+
+  enc->current_width_ = width;
+  return 1;
+}
+
+static int MakeInputImageCopy(VP8LEncoder* const enc) {
+  const WebPPicture* const picture = enc->pic_;
+  const int width = picture->width;
+  const int height = picture->height;
+
+  if (!AllocateTransformBuffer(enc, width, height)) return 0;
+  if (enc->argb_content_ == kEncoderARGB) return 1;
+
+  {
+    uint32_t* dst = enc->argb_;
+    const uint32_t* src = picture->argb;
+    int y;
+    for (y = 0; y < height; ++y) {
+      memcpy(dst, src, width * sizeof(*dst));
+      dst += width;
+      src += picture->argb_stride;
+    }
+  }
+  enc->argb_content_ = kEncoderARGB;
+  assert(enc->current_width_ == width);
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+
+#define APPLY_PALETTE_GREEDY_MAX 4
+
+static WEBP_INLINE uint32_t SearchColorGreedy(const uint32_t palette[],
+                                              int palette_size,
+                                              uint32_t color) {
+  (void)palette_size;
+  assert(palette_size < APPLY_PALETTE_GREEDY_MAX);
+  assert(3 == APPLY_PALETTE_GREEDY_MAX - 1);
+  if (color == palette[0]) return 0;
+  if (color == palette[1]) return 1;
+  if (color == palette[2]) return 2;
+  return 3;
+}
+
+static WEBP_INLINE uint32_t ApplyPaletteHash0(uint32_t color) {
+  // Focus on the green color.
+  return (color >> 8) & 0xff;
+}
+
+#define PALETTE_INV_SIZE_BITS 11
+#define PALETTE_INV_SIZE (1 << PALETTE_INV_SIZE_BITS)
+
+static WEBP_INLINE uint32_t ApplyPaletteHash1(uint32_t color) {
+  // Forget about alpha.
+  return ((uint32_t)((color & 0x00ffffffu) * 4222244071ull)) >>
+         (32 - PALETTE_INV_SIZE_BITS);
+}
+
+static WEBP_INLINE uint32_t ApplyPaletteHash2(uint32_t color) {
+  // Forget about alpha.
+  return ((uint32_t)((color & 0x00ffffffu) * ((1ull << 31) - 1))) >>
+         (32 - PALETTE_INV_SIZE_BITS);
+}
+
+// Use 1 pixel cache for ARGB pixels.
+#define APPLY_PALETTE_FOR(COLOR_INDEX) do {         \
+  uint32_t prev_pix = palette[0];                   \
+  uint32_t prev_idx = 0;                            \
+  for (y = 0; y < height; ++y) {                    \
+    for (x = 0; x < width; ++x) {                   \
+      const uint32_t pix = src[x];                  \
+      if (pix != prev_pix) {                        \
+        prev_idx = COLOR_INDEX;                     \
+        prev_pix = pix;                             \
+      }                                             \
+      tmp_row[x] = prev_idx;                        \
+    }                                               \
+    VP8LBundleColorMap(tmp_row, width, xbits, dst); \
+    src += src_stride;                              \
+    dst += dst_stride;                              \
+  }                                                 \
+} while (0)
+
+// Remap argb values in src[] to packed palettes entries in dst[]
+// using 'row' as a temporary buffer of size 'width'.
+// We assume that all src[] values have a corresponding entry in the palette.
+// Note: src[] can be the same as dst[]
+static int ApplyPalette(const uint32_t* src, uint32_t src_stride, uint32_t* dst,
+                        uint32_t dst_stride, const uint32_t* palette,
+                        int palette_size, int width, int height, int xbits,
+                        const WebPPicture* const pic) {
+  // TODO(skal): this tmp buffer is not needed if VP8LBundleColorMap() can be
+  // made to work in-place.
+  uint8_t* const tmp_row = (uint8_t*)WebPSafeMalloc(width, sizeof(*tmp_row));
+  int x, y;
+
+  if (tmp_row == NULL) {
+    WebPEncodingSetError(pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return 0;
+  }
+
+  if (palette_size < APPLY_PALETTE_GREEDY_MAX) {
+    APPLY_PALETTE_FOR(SearchColorGreedy(palette, palette_size, pix));
+  } else {
+    int i, j;
+    uint16_t buffer[PALETTE_INV_SIZE];
+    uint32_t (*const hash_functions[])(uint32_t) = {
+        ApplyPaletteHash0, ApplyPaletteHash1, ApplyPaletteHash2
+    };
+
+    // Try to find a perfect hash function able to go from a color to an index
+    // within 1 << PALETTE_INV_SIZE_BITS in order to build a hash map to go
+    // from color to index in palette.
+    for (i = 0; i < 3; ++i) {
+      int use_LUT = 1;
+      // Set each element in buffer to max uint16_t.
+      memset(buffer, 0xff, sizeof(buffer));
+      for (j = 0; j < palette_size; ++j) {
+        const uint32_t ind = hash_functions[i](palette[j]);
+        if (buffer[ind] != 0xffffu) {
+          use_LUT = 0;
+          break;
+        } else {
+          buffer[ind] = j;
+        }
+      }
+      if (use_LUT) break;
+    }
+
+    if (i == 0) {
+      APPLY_PALETTE_FOR(buffer[ApplyPaletteHash0(pix)]);
+    } else if (i == 1) {
+      APPLY_PALETTE_FOR(buffer[ApplyPaletteHash1(pix)]);
+    } else if (i == 2) {
+      APPLY_PALETTE_FOR(buffer[ApplyPaletteHash2(pix)]);
+    } else {
+      uint32_t idx_map[MAX_PALETTE_SIZE];
+      uint32_t palette_sorted[MAX_PALETTE_SIZE];
+      PrepareMapToPalette(palette, palette_size, palette_sorted, idx_map);
+      APPLY_PALETTE_FOR(
+          idx_map[SearchColorNoIdx(palette_sorted, pix, palette_size)]);
+    }
+  }
+  WebPSafeFree(tmp_row);
+  return 1;
+}
+#undef APPLY_PALETTE_FOR
+#undef PALETTE_INV_SIZE_BITS
+#undef PALETTE_INV_SIZE
+#undef APPLY_PALETTE_GREEDY_MAX
+
+// Note: Expects "enc->palette_" to be set properly.
+static int MapImageFromPalette(VP8LEncoder* const enc, int in_place) {
+  const WebPPicture* const pic = enc->pic_;
+  const int width = pic->width;
+  const int height = pic->height;
+  const uint32_t* const palette = enc->palette_;
+  const uint32_t* src = in_place ? enc->argb_ : pic->argb;
+  const int src_stride = in_place ? enc->current_width_ : pic->argb_stride;
+  const int palette_size = enc->palette_size_;
+  int xbits;
+
+  // Replace each input pixel by corresponding palette index.
+  // This is done line by line.
+  if (palette_size <= 4) {
+    xbits = (palette_size <= 2) ? 3 : 2;
+  } else {
+    xbits = (palette_size <= 16) ? 1 : 0;
+  }
+
+  if (!AllocateTransformBuffer(enc, VP8LSubSampleSize(width, xbits), height)) {
+    return 0;
+  }
+  if (!ApplyPalette(src, src_stride,
+                     enc->argb_, enc->current_width_,
+                     palette, palette_size, width, height, xbits, pic)) {
+    return 0;
+  }
+  enc->argb_content_ = kEncoderPalette;
+  return 1;
+}
+
+// Save palette_[] to bitstream.
+static WebPEncodingError EncodePalette(VP8LBitWriter* const bw, int low_effort,
+                                       VP8LEncoder* const enc,
+                                       int percent_range, int* const percent) {
+  int i;
+  uint32_t tmp_palette[MAX_PALETTE_SIZE];
+  const int palette_size = enc->palette_size_;
+  const uint32_t* const palette = enc->palette_;
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, COLOR_INDEXING_TRANSFORM, 2);
+  assert(palette_size >= 1 && palette_size <= MAX_PALETTE_SIZE);
+  VP8LPutBits(bw, palette_size - 1, 8);
+  for (i = palette_size - 1; i >= 1; --i) {
+    tmp_palette[i] = VP8LSubPixels(palette[i], palette[i - 1]);
+  }
+  tmp_palette[0] = palette[0];
+  return EncodeImageNoHuffman(bw, tmp_palette, &enc->hash_chain_,
+                              &enc->refs_[0], palette_size, 1, /*quality=*/20,
+                              low_effort, enc->pic_, percent_range, percent);
+}
+
+// -----------------------------------------------------------------------------
+// VP8LEncoder
+
+static VP8LEncoder* VP8LEncoderNew(const WebPConfig* const config,
+                                   const WebPPicture* const picture) {
+  VP8LEncoder* const enc = (VP8LEncoder*)WebPSafeCalloc(1ULL, sizeof(*enc));
+  if (enc == NULL) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return NULL;
+  }
+  enc->config_ = config;
+  enc->pic_ = picture;
+  enc->argb_content_ = kEncoderNone;
+
+  VP8LEncDspInit();
+
+  return enc;
+}
+
+static void VP8LEncoderDelete(VP8LEncoder* enc) {
+  if (enc != NULL) {
+    int i;
+    VP8LHashChainClear(&enc->hash_chain_);
+    for (i = 0; i < 4; ++i) VP8LBackwardRefsClear(&enc->refs_[i]);
+    ClearTransformBuffer(enc);
+    WebPSafeFree(enc);
+  }
+}
+
+// -----------------------------------------------------------------------------
+// Main call
+
+typedef struct {
+  const WebPConfig* config_;
+  const WebPPicture* picture_;
+  VP8LBitWriter* bw_;
+  VP8LEncoder* enc_;
+  int use_cache_;
+  CrunchConfig crunch_configs_[CRUNCH_CONFIGS_MAX];
+  int num_crunch_configs_;
+  int red_and_blue_always_zero_;
+  WebPAuxStats* stats_;
+} StreamEncodeContext;
+
+static int EncodeStreamHook(void* input, void* data2) {
+  StreamEncodeContext* const params = (StreamEncodeContext*)input;
+  const WebPConfig* const config = params->config_;
+  const WebPPicture* const picture = params->picture_;
+  VP8LBitWriter* const bw = params->bw_;
+  VP8LEncoder* const enc = params->enc_;
+  const int use_cache = params->use_cache_;
+  const CrunchConfig* const crunch_configs = params->crunch_configs_;
+  const int num_crunch_configs = params->num_crunch_configs_;
+  const int red_and_blue_always_zero = params->red_and_blue_always_zero_;
+#if !defined(WEBP_DISABLE_STATS)
+  WebPAuxStats* const stats = params->stats_;
+#endif
+  const int quality = (int)config->quality;
+  const int low_effort = (config->method == 0);
+#if (WEBP_NEAR_LOSSLESS == 1)
+  const int width = picture->width;
+#endif
+  const int height = picture->height;
+  const size_t byte_position = VP8LBitWriterNumBytes(bw);
+  int percent = 2;  // for WebPProgressHook
+#if (WEBP_NEAR_LOSSLESS == 1)
+  int use_near_lossless = 0;
+#endif
+  int hdr_size = 0;
+  int data_size = 0;
+  int use_delta_palette = 0;
+  int idx;
+  size_t best_size = ~(size_t)0;
+  VP8LBitWriter bw_init = *bw, bw_best;
+  (void)data2;
+
+  if (!VP8LBitWriterInit(&bw_best, 0) ||
+      (num_crunch_configs > 1 && !VP8LBitWriterClone(bw, &bw_best))) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  for (idx = 0; idx < num_crunch_configs; ++idx) {
+    const int entropy_idx = crunch_configs[idx].entropy_idx_;
+    int remaining_percent = 97 / num_crunch_configs, percent_range;
+    enc->use_palette_ =
+        (entropy_idx == kPalette) || (entropy_idx == kPaletteAndSpatial);
+    enc->use_subtract_green_ =
+        (entropy_idx == kSubGreen) || (entropy_idx == kSpatialSubGreen);
+    enc->use_predict_ = (entropy_idx == kSpatial) ||
+                        (entropy_idx == kSpatialSubGreen) ||
+                        (entropy_idx == kPaletteAndSpatial);
+    // When using a palette, R/B==0, hence no need to test for cross-color.
+    if (low_effort || enc->use_palette_) {
+      enc->use_cross_color_ = 0;
+    } else {
+      enc->use_cross_color_ = red_and_blue_always_zero ? 0 : enc->use_predict_;
+    }
+    // Reset any parameter in the encoder that is set in the previous iteration.
+    enc->cache_bits_ = 0;
+    VP8LBackwardRefsClear(&enc->refs_[0]);
+    VP8LBackwardRefsClear(&enc->refs_[1]);
+
+#if (WEBP_NEAR_LOSSLESS == 1)
+    // Apply near-lossless preprocessing.
+    use_near_lossless = (config->near_lossless < 100) && !enc->use_palette_ &&
+                        !enc->use_predict_;
+    if (use_near_lossless) {
+      if (!AllocateTransformBuffer(enc, width, height)) goto Error;
+      if ((enc->argb_content_ != kEncoderNearLossless) &&
+          !VP8ApplyNearLossless(picture, config->near_lossless, enc->argb_)) {
+        WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+        goto Error;
+      }
+      enc->argb_content_ = kEncoderNearLossless;
+    } else {
+      enc->argb_content_ = kEncoderNone;
+    }
+#else
+    enc->argb_content_ = kEncoderNone;
+#endif
+
+    // Encode palette
+    if (enc->use_palette_) {
+      if (crunch_configs[idx].palette_sorting_type_ == kSortedDefault) {
+        // Nothing to do, we have already sorted the palette.
+        memcpy(enc->palette_, enc->palette_sorted_,
+               enc->palette_size_ * sizeof(*enc->palette_));
+      } else if (crunch_configs[idx].palette_sorting_type_ == kMinimizeDelta) {
+        PaletteSortMinimizeDeltas(enc->palette_sorted_, enc->palette_size_,
+                                  enc->palette_);
+      } else {
+        assert(crunch_configs[idx].palette_sorting_type_ == kModifiedZeng);
+        if (!PaletteSortModifiedZeng(enc->pic_, enc->palette_sorted_,
+                                      enc->palette_size_, enc->palette_)) {
+          goto Error;
+        }
+      }
+      percent_range = remaining_percent / 4;
+      if (!EncodePalette(bw, low_effort, enc, percent_range, &percent)) {
+        goto Error;
+      }
+      remaining_percent -= percent_range;
+      if (!MapImageFromPalette(enc, use_delta_palette)) goto Error;
+      // If using a color cache, do not have it bigger than the number of
+      // colors.
+      if (use_cache && enc->palette_size_ < (1 << MAX_COLOR_CACHE_BITS)) {
+        enc->cache_bits_ = BitsLog2Floor(enc->palette_size_) + 1;
+      }
+    }
+    if (!use_delta_palette) {
+      // In case image is not packed.
+      if (enc->argb_content_ != kEncoderNearLossless &&
+          enc->argb_content_ != kEncoderPalette) {
+        if (!MakeInputImageCopy(enc)) goto Error;
+      }
+
+      // -----------------------------------------------------------------------
+      // Apply transforms and write transform data.
+
+      if (enc->use_subtract_green_) {
+        ApplySubtractGreen(enc, enc->current_width_, height, bw);
+      }
+
+      if (enc->use_predict_) {
+        percent_range = remaining_percent / 3;
+        if (!ApplyPredictFilter(enc, enc->current_width_, height, quality,
+                                low_effort, enc->use_subtract_green_, bw,
+                                percent_range, &percent)) {
+          goto Error;
+        }
+        remaining_percent -= percent_range;
+      }
+
+      if (enc->use_cross_color_) {
+        percent_range = remaining_percent / 2;
+        if (!ApplyCrossColorFilter(enc, enc->current_width_, height, quality,
+                                   low_effort, bw, percent_range, &percent)) {
+          goto Error;
+        }
+        remaining_percent -= percent_range;
+      }
+    }
+
+    VP8LPutBits(bw, !TRANSFORM_PRESENT, 1);  // No more transforms.
+
+    // -------------------------------------------------------------------------
+    // Encode and write the transformed image.
+    if (!EncodeImageInternal(
+            bw, enc->argb_, &enc->hash_chain_, enc->refs_, enc->current_width_,
+            height, quality, low_effort, use_cache, &crunch_configs[idx],
+            &enc->cache_bits_, enc->histo_bits_, byte_position, &hdr_size,
+            &data_size, picture, remaining_percent, &percent)) {
+      goto Error;
+    }
+
+    // If we are better than what we already have.
+    if (VP8LBitWriterNumBytes(bw) < best_size) {
+      best_size = VP8LBitWriterNumBytes(bw);
+      // Store the BitWriter.
+      VP8LBitWriterSwap(bw, &bw_best);
+#if !defined(WEBP_DISABLE_STATS)
+      // Update the stats.
+      if (stats != NULL) {
+        stats->lossless_features = 0;
+        if (enc->use_predict_) stats->lossless_features |= 1;
+        if (enc->use_cross_color_) stats->lossless_features |= 2;
+        if (enc->use_subtract_green_) stats->lossless_features |= 4;
+        if (enc->use_palette_) stats->lossless_features |= 8;
+        stats->histogram_bits = enc->histo_bits_;
+        stats->transform_bits = enc->transform_bits_;
+        stats->cache_bits = enc->cache_bits_;
+        stats->palette_size = enc->palette_size_;
+        stats->lossless_size = (int)(best_size - byte_position);
+        stats->lossless_hdr_size = hdr_size;
+        stats->lossless_data_size = data_size;
+      }
+#endif
+    }
+    // Reset the bit writer for the following iteration if any.
+    if (num_crunch_configs > 1) VP8LBitWriterReset(&bw_init, bw);
+  }
+  VP8LBitWriterSwap(&bw_best, bw);
+
+ Error:
+  VP8LBitWriterWipeOut(&bw_best);
+  // The hook should return false in case of error.
+  return (params->picture_->error_code == VP8_ENC_OK);
+}
+
+int VP8LEncodeStream(const WebPConfig* const config,
+                     const WebPPicture* const picture,
+                     VP8LBitWriter* const bw_main, int use_cache) {
+  VP8LEncoder* const enc_main = VP8LEncoderNew(config, picture);
+  VP8LEncoder* enc_side = NULL;
+  CrunchConfig crunch_configs[CRUNCH_CONFIGS_MAX];
+  int num_crunch_configs_main, num_crunch_configs_side = 0;
+  int idx;
+  int red_and_blue_always_zero = 0;
+  WebPWorker worker_main, worker_side;
+  StreamEncodeContext params_main, params_side;
+  // The main thread uses picture->stats, the side thread uses stats_side.
+  WebPAuxStats stats_side;
+  VP8LBitWriter bw_side;
+  WebPPicture picture_side;
+  const WebPWorkerInterface* const worker_interface = WebPGetWorkerInterface();
+  int ok_main;
+
+  if (enc_main == NULL || !VP8LBitWriterInit(&bw_side, 0)) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    VP8LEncoderDelete(enc_main);
+    return 0;
+  }
+
+  // Avoid "garbage value" error from Clang's static analysis tool.
+  WebPPictureInit(&picture_side);
+
+  // Analyze image (entropy, num_palettes etc)
+  if (!EncoderAnalyze(enc_main, crunch_configs, &num_crunch_configs_main,
+                      &red_and_blue_always_zero) ||
+      !EncoderInit(enc_main)) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  // Split the configs between the main and side threads (if any).
+  if (config->thread_level > 0) {
+    num_crunch_configs_side = num_crunch_configs_main / 2;
+    for (idx = 0; idx < num_crunch_configs_side; ++idx) {
+      params_side.crunch_configs_[idx] =
+          crunch_configs[num_crunch_configs_main - num_crunch_configs_side +
+                         idx];
+    }
+    params_side.num_crunch_configs_ = num_crunch_configs_side;
+  }
+  num_crunch_configs_main -= num_crunch_configs_side;
+  for (idx = 0; idx < num_crunch_configs_main; ++idx) {
+    params_main.crunch_configs_[idx] = crunch_configs[idx];
+  }
+  params_main.num_crunch_configs_ = num_crunch_configs_main;
+
+  // Fill in the parameters for the thread workers.
+  {
+    const int params_size = (num_crunch_configs_side > 0) ? 2 : 1;
+    for (idx = 0; idx < params_size; ++idx) {
+      // Create the parameters for each worker.
+      WebPWorker* const worker = (idx == 0) ? &worker_main : &worker_side;
+      StreamEncodeContext* const param =
+          (idx == 0) ? &params_main : &params_side;
+      param->config_ = config;
+      param->use_cache_ = use_cache;
+      param->red_and_blue_always_zero_ = red_and_blue_always_zero;
+      if (idx == 0) {
+        param->picture_ = picture;
+        param->stats_ = picture->stats;
+        param->bw_ = bw_main;
+        param->enc_ = enc_main;
+      } else {
+        // Create a side picture (error_code is not thread-safe).
+        if (!WebPPictureView(picture, /*left=*/0, /*top=*/0, picture->width,
+                             picture->height, &picture_side)) {
+          assert(0);
+        }
+        picture_side.progress_hook = NULL;  // Progress hook is not thread-safe.
+        param->picture_ = &picture_side;  // No need to free a view afterwards.
+        param->stats_ = (picture->stats == NULL) ? NULL : &stats_side;
+        // Create a side bit writer.
+        if (!VP8LBitWriterClone(bw_main, &bw_side)) {
+          WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+          goto Error;
+        }
+        param->bw_ = &bw_side;
+        // Create a side encoder.
+        enc_side = VP8LEncoderNew(config, &picture_side);
+        if (enc_side == NULL || !EncoderInit(enc_side)) {
+          WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+          goto Error;
+        }
+        // Copy the values that were computed for the main encoder.
+        enc_side->histo_bits_ = enc_main->histo_bits_;
+        enc_side->transform_bits_ = enc_main->transform_bits_;
+        enc_side->palette_size_ = enc_main->palette_size_;
+        memcpy(enc_side->palette_, enc_main->palette_,
+               sizeof(enc_main->palette_));
+        memcpy(enc_side->palette_sorted_, enc_main->palette_sorted_,
+               sizeof(enc_main->palette_sorted_));
+        param->enc_ = enc_side;
+      }
+      // Create the workers.
+      worker_interface->Init(worker);
+      worker->data1 = param;
+      worker->data2 = NULL;
+      worker->hook = EncodeStreamHook;
+    }
+  }
+
+  // Start the second thread if needed.
+  if (num_crunch_configs_side != 0) {
+    if (!worker_interface->Reset(&worker_side)) {
+      WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+      goto Error;
+    }
+#if !defined(WEBP_DISABLE_STATS)
+    // This line is here and not in the param initialization above to remove a
+    // Clang static analyzer warning.
+    if (picture->stats != NULL) {
+      memcpy(&stats_side, picture->stats, sizeof(stats_side));
+    }
+#endif
+    worker_interface->Launch(&worker_side);
+  }
+  // Execute the main thread.
+  worker_interface->Execute(&worker_main);
+  ok_main = worker_interface->Sync(&worker_main);
+  worker_interface->End(&worker_main);
+  if (num_crunch_configs_side != 0) {
+    // Wait for the second thread.
+    const int ok_side = worker_interface->Sync(&worker_side);
+    worker_interface->End(&worker_side);
+    if (!ok_main || !ok_side) {
+      if (picture->error_code == VP8_ENC_OK) {
+        assert(picture_side.error_code != VP8_ENC_OK);
+        WebPEncodingSetError(picture, picture_side.error_code);
+      }
+      goto Error;
+    }
+    if (VP8LBitWriterNumBytes(&bw_side) < VP8LBitWriterNumBytes(bw_main)) {
+      VP8LBitWriterSwap(bw_main, &bw_side);
+#if !defined(WEBP_DISABLE_STATS)
+      if (picture->stats != NULL) {
+        memcpy(picture->stats, &stats_side, sizeof(*picture->stats));
+      }
+#endif
+    }
+  }
+
+ Error:
+  VP8LBitWriterWipeOut(&bw_side);
+  VP8LEncoderDelete(enc_main);
+  VP8LEncoderDelete(enc_side);
+  return (picture->error_code == VP8_ENC_OK);
+}
+
+#undef CRUNCH_CONFIGS_MAX
+#undef CRUNCH_SUBCONFIGS_MAX
+
+int VP8LEncodeImage(const WebPConfig* const config,
+                    const WebPPicture* const picture) {
+  int width, height;
+  int has_alpha;
+  size_t coded_size;
+  int percent = 0;
+  int initial_size;
+  VP8LBitWriter bw;
+
+  if (picture == NULL) return 0;
+
+  if (config == NULL || picture->argb == NULL) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_NULL_PARAMETER);
+    return 0;
+  }
+
+  width = picture->width;
+  height = picture->height;
+  // Initialize BitWriter with size corresponding to 16 bpp to photo images and
+  // 8 bpp for graphical images.
+  initial_size = (config->image_hint == WEBP_HINT_GRAPH) ?
+      width * height : width * height * 2;
+  if (!VP8LBitWriterInit(&bw, initial_size)) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  if (!WebPReportProgress(picture, 1, &percent)) {
+ UserAbort:
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_USER_ABORT);
+    goto Error;
+  }
+  // Reset stats (for pure lossless coding)
+  if (picture->stats != NULL) {
+    WebPAuxStats* const stats = picture->stats;
+    memset(stats, 0, sizeof(*stats));
+    stats->PSNR[0] = 99.f;
+    stats->PSNR[1] = 99.f;
+    stats->PSNR[2] = 99.f;
+    stats->PSNR[3] = 99.f;
+    stats->PSNR[4] = 99.f;
+  }
+
+  // Write image size.
+  if (!WriteImageSize(picture, &bw)) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  has_alpha = WebPPictureHasTransparency(picture);
+  // Write the non-trivial Alpha flag and lossless version.
+  if (!WriteRealAlphaAndVersion(&bw, has_alpha)) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    goto Error;
+  }
+
+  if (!WebPReportProgress(picture, 2, &percent)) goto UserAbort;
+
+  // Encode main image stream.
+  if (!VP8LEncodeStream(config, picture, &bw, 1 /*use_cache*/)) goto Error;
+
+  if (!WebPReportProgress(picture, 99, &percent)) goto UserAbort;
+
+  // Finish the RIFF chunk.
+  if (!WriteImage(picture, &bw, &coded_size)) goto Error;
+
+  if (!WebPReportProgress(picture, 100, &percent)) goto UserAbort;
+
+#if !defined(WEBP_DISABLE_STATS)
+  // Save size.
+  if (picture->stats != NULL) {
+    picture->stats->coded_size += (int)coded_size;
+    picture->stats->lossless_size = (int)coded_size;
+  }
+#endif
+
+  if (picture->extra_info != NULL) {
+    const int mb_w = (width + 15) >> 4;
+    const int mb_h = (height + 15) >> 4;
+    memset(picture->extra_info, 0, mb_w * mb_h * sizeof(*picture->extra_info));
+  }
+
+ Error:
+  if (bw.error_) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  VP8LBitWriterWipeOut(&bw);
+  return (picture->error_code == VP8_ENC_OK);
+}
+
+//------------------------------------------------------------------------------
diff --git a/media/libwebp/src/enc/vp8li_enc.h b/media/libwebp/src/enc/vp8li_enc.h
new file mode 100644
index 0000000000..3d35e1612d
--- /dev/null
+++ b/media/libwebp/src/enc/vp8li_enc.h
@@ -0,0 +1,126 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Lossless encoder: internal header.
+//
+// Author: Vikas Arora (vikaas.arora@gmail.com)
+
+#ifndef WEBP_ENC_VP8LI_ENC_H_
+#define WEBP_ENC_VP8LI_ENC_H_
+
+#ifdef HAVE_CONFIG_H
+#include "src/webp/config.h"
+#endif
+// Either WEBP_NEAR_LOSSLESS is defined as 0 in config.h when compiling to
+// disable near-lossless, or it is enabled by default.
+#ifndef WEBP_NEAR_LOSSLESS
+#define WEBP_NEAR_LOSSLESS 1
+#endif
+
+#include "src/enc/backward_references_enc.h"
+#include "src/enc/histogram_enc.h"
+#include "src/utils/bit_writer_utils.h"
+#include "src/webp/encode.h"
+#include "src/webp/format_constants.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// maximum value of transform_bits_ in VP8LEncoder.
+#define MAX_TRANSFORM_BITS 6
+
+typedef enum {
+  kEncoderNone = 0,
+  kEncoderARGB,
+  kEncoderNearLossless,
+  kEncoderPalette
+} VP8LEncoderARGBContent;
+
+typedef struct {
+  const WebPConfig* config_;      // user configuration and parameters
+  const WebPPicture* pic_;        // input picture.
+
+  uint32_t* argb_;                       // Transformed argb image data.
+  VP8LEncoderARGBContent argb_content_;  // Content type of the argb buffer.
+  uint32_t* argb_scratch_;               // Scratch memory for argb rows
+                                         // (used for prediction).
+  uint32_t* transform_data_;             // Scratch memory for transform data.
+  uint32_t* transform_mem_;              // Currently allocated memory.
+  size_t    transform_mem_size_;         // Currently allocated memory size.
+
+  int       current_width_;       // Corresponds to packed image width.
+
+  // Encoding parameters derived from quality parameter.
+  int histo_bits_;
+  int transform_bits_;    // <= MAX_TRANSFORM_BITS.
+  int cache_bits_;        // If equal to 0, don't use color cache.
+
+  // Encoding parameters derived from image characteristics.
+  int use_cross_color_;
+  int use_subtract_green_;
+  int use_predict_;
+  int use_palette_;
+  int palette_size_;
+  uint32_t palette_[MAX_PALETTE_SIZE];
+  // Sorted version of palette_ for cache purposes.
+  uint32_t palette_sorted_[MAX_PALETTE_SIZE];
+
+  // Some 'scratch' (potentially large) objects.
+  struct VP8LBackwardRefs refs_[4];  // Backward Refs array for temporaries.
+  VP8LHashChain hash_chain_;         // HashChain data for constructing
+                                     // backward references.
+} VP8LEncoder;
+
+//------------------------------------------------------------------------------
+// internal functions. Not public.
+
+// Encodes the picture.
+// Returns 0 if config or picture is NULL or picture doesn't have valid argb
+// input.
+int VP8LEncodeImage(const WebPConfig* const config,
+                    const WebPPicture* const picture);
+
+// Encodes the main image stream using the supplied bit writer.
+// If 'use_cache' is false, disables the use of color cache.
+// Returns false in case of error (stored in picture->error_code).
+int VP8LEncodeStream(const WebPConfig* const config,
+                     const WebPPicture* const picture, VP8LBitWriter* const bw,
+                     int use_cache);
+
+#if (WEBP_NEAR_LOSSLESS == 1)
+// in near_lossless.c
+// Near lossless preprocessing in RGB color-space.
+int VP8ApplyNearLossless(const WebPPicture* const picture, int quality,
+                         uint32_t* const argb_dst);
+#endif
+
+//------------------------------------------------------------------------------
+// Image transforms in predictor.c.
+
+// pic and percent are for progress.
+// Returns false in case of error (stored in pic->error_code).
+int VP8LResidualImage(int width, int height, int bits, int low_effort,
+                      uint32_t* const argb, uint32_t* const argb_scratch,
+                      uint32_t* const image, int near_lossless, int exact,
+                      int used_subtract_green, const WebPPicture* const pic,
+                      int percent_range, int* const percent);
+
+int VP8LColorSpaceTransform(int width, int height, int bits, int quality,
+                            uint32_t* const argb, uint32_t* image,
+                            const WebPPicture* const pic, int percent_range,
+                            int* const percent);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // WEBP_ENC_VP8LI_ENC_H_
diff --git a/media/libwebp/src/enc/webp_enc.c b/media/libwebp/src/enc/webp_enc.c
new file mode 100644
index 0000000000..9620e05070
--- /dev/null
+++ b/media/libwebp/src/enc/webp_enc.c
@@ -0,0 +1,408 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// WebP encoder: main entry point
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+
+#include "src/enc/cost_enc.h"
+#include "src/enc/vp8i_enc.h"
+#include "src/enc/vp8li_enc.h"
+#include "src/utils/utils.h"
+
+// #define PRINT_MEMORY_INFO
+
+#ifdef PRINT_MEMORY_INFO
+#include <stdio.h>
+#endif
+
+//------------------------------------------------------------------------------
+
+int WebPGetEncoderVersion(void) {
+  return (ENC_MAJ_VERSION << 16) | (ENC_MIN_VERSION << 8) | ENC_REV_VERSION;
+}
+
+//------------------------------------------------------------------------------
+// VP8Encoder
+//------------------------------------------------------------------------------
+
+static void ResetSegmentHeader(VP8Encoder* const enc) {
+  VP8EncSegmentHeader* const hdr = &enc->segment_hdr_;
+  hdr->num_segments_ = enc->config_->segments;
+  hdr->update_map_  = (hdr->num_segments_ > 1);
+  hdr->size_ = 0;
+}
+
+static void ResetFilterHeader(VP8Encoder* const enc) {
+  VP8EncFilterHeader* const hdr = &enc->filter_hdr_;
+  hdr->simple_ = 1;
+  hdr->level_ = 0;
+  hdr->sharpness_ = 0;
+  hdr->i4x4_lf_delta_ = 0;
+}
+
+static void ResetBoundaryPredictions(VP8Encoder* const enc) {
+  // init boundary values once for all
+  // Note: actually, initializing the preds_[] is only needed for intra4.
+  int i;
+  uint8_t* const top = enc->preds_ - enc->preds_w_;
+  uint8_t* const left = enc->preds_ - 1;
+  for (i = -1; i < 4 * enc->mb_w_; ++i) {
+    top[i] = B_DC_PRED;
+  }
+  for (i = 0; i < 4 * enc->mb_h_; ++i) {
+    left[i * enc->preds_w_] = B_DC_PRED;
+  }
+  enc->nz_[-1] = 0;   // constant
+}
+
+// Mapping from config->method_ to coding tools used.
+//-------------------+---+---+---+---+---+---+---+
+//   Method          | 0 | 1 | 2 | 3 |(4)| 5 | 6 |
+//-------------------+---+---+---+---+---+---+---+
+// fast probe        | x |   |   | x |   |   |   |
+//-------------------+---+---+---+---+---+---+---+
+// dynamic proba     | ~ | x | x | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// fast mode analysis|[x]|[x]|   |   | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// basic rd-opt      |   |   |   | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// disto-refine i4/16| x | x | x |   |   |   |   |
+//-------------------+---+---+---+---+---+---+---+
+// disto-refine uv   |   | x | x |   |   |   |   |
+//-------------------+---+---+---+---+---+---+---+
+// rd-opt i4/16      |   |   | ~ | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// token buffer (opt)|   |   |   | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// Trellis           |   |   |   |   |   | x |Ful|
+//-------------------+---+---+---+---+---+---+---+
+// full-SNS          |   |   |   |   | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+
+static void MapConfigToTools(VP8Encoder* const enc) {
+  const WebPConfig* const config = enc->config_;
+  const int method = config->method;
+  const int limit = 100 - config->partition_limit;
+  enc->method_ = method;
+  enc->rd_opt_level_ = (method >= 6) ? RD_OPT_TRELLIS_ALL
+                     : (method >= 5) ? RD_OPT_TRELLIS
+                     : (method >= 3) ? RD_OPT_BASIC
+                     : RD_OPT_NONE;
+  enc->max_i4_header_bits_ =
+      256 * 16 * 16 *                 // upper bound: up to 16bit per 4x4 block
+      (limit * limit) / (100 * 100);  // ... modulated with a quadratic curve.
+
+  // partition0 = 512k max.
+  enc->mb_header_limit_ =
+      (score_t)256 * 510 * 8 * 1024 / (enc->mb_w_ * enc->mb_h_);
+
+  enc->thread_level_ = config->thread_level;
+
+  enc->do_search_ = (config->target_size > 0 || config->target_PSNR > 0);
+  if (!config->low_memory) {
+#if !defined(DISABLE_TOKEN_BUFFER)
+    enc->use_tokens_ = (enc->rd_opt_level_ >= RD_OPT_BASIC);  // need rd stats
+#endif
+    if (enc->use_tokens_) {
+      enc->num_parts_ = 1;   // doesn't work with multi-partition
+    }
+  }
+}
+
+// Memory scaling with dimensions:
+//  memory (bytes) ~= 2.25 * w + 0.0625 * w * h
+//
+// Typical memory footprint (614x440 picture)
+//              encoder: 22111
+//                 info: 4368
+//                preds: 17741
+//          top samples: 1263
+//             non-zero: 175
+//             lf-stats: 0
+//                total: 45658
+// Transient object sizes:
+//       VP8EncIterator: 3360
+//         VP8ModeScore: 872
+//       VP8SegmentInfo: 732
+//          VP8EncProba: 18352
+//              LFStats: 2048
+// Picture size (yuv): 419328
+
+static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
+                                  WebPPicture* const picture) {
+  VP8Encoder* enc;
+  const int use_filter =
+      (config->filter_strength > 0) || (config->autofilter > 0);
+  const int mb_w = (picture->width + 15) >> 4;
+  const int mb_h = (picture->height + 15) >> 4;
+  const int preds_w = 4 * mb_w + 1;
+  const int preds_h = 4 * mb_h + 1;
+  const size_t preds_size = preds_w * preds_h * sizeof(*enc->preds_);
+  const int top_stride = mb_w * 16;
+  const size_t nz_size = (mb_w + 1) * sizeof(*enc->nz_) + WEBP_ALIGN_CST;
+  const size_t info_size = mb_w * mb_h * sizeof(*enc->mb_info_);
+  const size_t samples_size =
+      2 * top_stride * sizeof(*enc->y_top_)  // top-luma/u/v
+      + WEBP_ALIGN_CST;                      // align all
+  const size_t lf_stats_size =
+      config->autofilter ? sizeof(*enc->lf_stats_) + WEBP_ALIGN_CST : 0;
+  const size_t top_derr_size =
+      (config->quality <= ERROR_DIFFUSION_QUALITY || config->pass > 1) ?
+          mb_w * sizeof(*enc->top_derr_) : 0;
+  uint8_t* mem;
+  const uint64_t size = (uint64_t)sizeof(*enc)   // main struct
+                      + WEBP_ALIGN_CST           // cache alignment
+                      + info_size                // modes info
+                      + preds_size               // prediction modes
+                      + samples_size             // top/left samples
+                      + top_derr_size            // top diffusion error
+                      + nz_size                  // coeff context bits
+                      + lf_stats_size;           // autofilter stats
+
+#ifdef PRINT_MEMORY_INFO
+  printf("===================================\n");
+  printf("Memory used:\n"
+         "             encoder: %ld\n"
+         "                info: %ld\n"
+         "               preds: %ld\n"
+         "         top samples: %ld\n"
+         "       top diffusion: %ld\n"
+         "            non-zero: %ld\n"
+         "            lf-stats: %ld\n"
+         "               total: %ld\n",
+         sizeof(*enc) + WEBP_ALIGN_CST, info_size,
+         preds_size, samples_size, top_derr_size, nz_size, lf_stats_size, size);
+  printf("Transient object sizes:\n"
+         "      VP8EncIterator: %ld\n"
+         "        VP8ModeScore: %ld\n"
+         "      VP8SegmentInfo: %ld\n"
+         "         VP8EncProba: %ld\n"
+         "             LFStats: %ld\n",
+         sizeof(VP8EncIterator), sizeof(VP8ModeScore),
+         sizeof(VP8SegmentInfo), sizeof(VP8EncProba),
+         sizeof(LFStats));
+  printf("Picture size (yuv): %ld\n",
+         mb_w * mb_h * 384 * sizeof(uint8_t));
+  printf("===================================\n");
+#endif
+  mem = (uint8_t*)WebPSafeMalloc(size, sizeof(*mem));
+  if (mem == NULL) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return NULL;
+  }
+  enc = (VP8Encoder*)mem;
+  mem = (uint8_t*)WEBP_ALIGN(mem + sizeof(*enc));
+  memset(enc, 0, sizeof(*enc));
+  enc->num_parts_ = 1 << config->partitions;
+  enc->mb_w_ = mb_w;
+  enc->mb_h_ = mb_h;
+  enc->preds_w_ = preds_w;
+  enc->mb_info_ = (VP8MBInfo*)mem;
+  mem += info_size;
+  enc->preds_ = mem + 1 + enc->preds_w_;
+  mem += preds_size;
+  enc->nz_ = 1 + (uint32_t*)WEBP_ALIGN(mem);
+  mem += nz_size;
+  enc->lf_stats_ = lf_stats_size ? (LFStats*)WEBP_ALIGN(mem) : NULL;
+  mem += lf_stats_size;
+
+  // top samples (all 16-aligned)
+  mem = (uint8_t*)WEBP_ALIGN(mem);
+  enc->y_top_ = mem;
+  enc->uv_top_ = enc->y_top_ + top_stride;
+  mem += 2 * top_stride;
+  enc->top_derr_ = top_derr_size ? (DError*)mem : NULL;
+  mem += top_derr_size;
+  assert(mem <= (uint8_t*)enc + size);
+
+  enc->config_ = config;
+  enc->profile_ = use_filter ? ((config->filter_type == 1) ? 0 : 1) : 2;
+  enc->pic_ = picture;
+  enc->percent_ = 0;
+
+  MapConfigToTools(enc);
+  VP8EncDspInit();
+  VP8DefaultProbas(enc);
+  ResetSegmentHeader(enc);
+  ResetFilterHeader(enc);
+  ResetBoundaryPredictions(enc);
+  VP8EncDspCostInit();
+  VP8EncInitAlpha(enc);
+
+  // lower quality means smaller output -> we modulate a little the page
+  // size based on quality. This is just a crude 1rst-order prediction.
+  {
+    const float scale = 1.f + config->quality * 5.f / 100.f;  // in [1,6]
+    VP8TBufferInit(&enc->tokens_, (int)(mb_w * mb_h * 4 * scale));
+  }
+  return enc;
+}
+
+static int DeleteVP8Encoder(VP8Encoder* enc) {
+  int ok = 1;
+  if (enc != NULL) {
+    ok = VP8EncDeleteAlpha(enc);
+    VP8TBufferClear(&enc->tokens_);
+    WebPSafeFree(enc);
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+
+#if !defined(WEBP_DISABLE_STATS)
+static double GetPSNR(uint64_t err, uint64_t size) {
+  return (err > 0 && size > 0) ? 10. * log10(255. * 255. * size / err) : 99.;
+}
+
+static void FinalizePSNR(const VP8Encoder* const enc) {
+  WebPAuxStats* stats = enc->pic_->stats;
+  const uint64_t size = enc->sse_count_;
+  const uint64_t* const sse = enc->sse_;
+  stats->PSNR[0] = (float)GetPSNR(sse[0], size);
+  stats->PSNR[1] = (float)GetPSNR(sse[1], size / 4);
+  stats->PSNR[2] = (float)GetPSNR(sse[2], size / 4);
+  stats->PSNR[3] = (float)GetPSNR(sse[0] + sse[1] + sse[2], size * 3 / 2);
+  stats->PSNR[4] = (float)GetPSNR(sse[3], size);
+}
+#endif  // !defined(WEBP_DISABLE_STATS)
+
+static void StoreStats(VP8Encoder* const enc) {
+#if !defined(WEBP_DISABLE_STATS)
+  WebPAuxStats* const stats = enc->pic_->stats;
+  if (stats != NULL) {
+    int i, s;
+    for (i = 0; i < NUM_MB_SEGMENTS; ++i) {
+      stats->segment_level[i] = enc->dqm_[i].fstrength_;
+      stats->segment_quant[i] = enc->dqm_[i].quant_;
+      for (s = 0; s <= 2; ++s) {
+        stats->residual_bytes[s][i] = enc->residual_bytes_[s][i];
+      }
+    }
+    FinalizePSNR(enc);
+    stats->coded_size = enc->coded_size_;
+    for (i = 0; i < 3; ++i) {
+      stats->block_count[i] = enc->block_count_[i];
+    }
+  }
+#else  // defined(WEBP_DISABLE_STATS)
+  WebPReportProgress(enc->pic_, 100, &enc->percent_);  // done!
+#endif  // !defined(WEBP_DISABLE_STATS)
+}
+
+int WebPEncodingSetError(const WebPPicture* const pic,
+                         WebPEncodingError error) {
+  assert((int)error < VP8_ENC_ERROR_LAST);
+  assert((int)error >= VP8_ENC_OK);
+  ((WebPPicture*)pic)->error_code = error;
+  return 0;
+}
+
+int WebPReportProgress(const WebPPicture* const pic,
+                       int percent, int* const percent_store) {
+  if (percent_store != NULL && percent != *percent_store) {
+    *percent_store = percent;
+    if (pic->progress_hook && !pic->progress_hook(percent, pic)) {
+      // user abort requested
+      WebPEncodingSetError(pic, VP8_ENC_ERROR_USER_ABORT);
+      return 0;
+    }
+  }
+  return 1;  // ok
+}
+//------------------------------------------------------------------------------
+
+int WebPEncode(const WebPConfig* config, WebPPicture* pic) {
+  int ok = 0;
+  if (pic == NULL) return 0;
+
+  WebPEncodingSetError(pic, VP8_ENC_OK);  // all ok so far
+  if (config == NULL) {  // bad params
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_NULL_PARAMETER);
+  }
+  if (!WebPValidateConfig(config)) {
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_INVALID_CONFIGURATION);
+  }
+  if (!WebPValidatePicture(pic)) return 0;
+  if (pic->width > WEBP_MAX_DIMENSION || pic->height > WEBP_MAX_DIMENSION) {
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_BAD_DIMENSION);
+  }
+
+  if (pic->stats != NULL) memset(pic->stats, 0, sizeof(*pic->stats));
+
+  if (!config->lossless) {
+    VP8Encoder* enc = NULL;
+
+    if (pic->use_argb || pic->y == NULL || pic->u == NULL || pic->v == NULL) {
+      // Make sure we have YUVA samples.
+      if (config->use_sharp_yuv || (config->preprocessing & 4)) {
+        if (!WebPPictureSharpARGBToYUVA(pic)) {
+          return 0;
+        }
+      } else {
+        float dithering = 0.f;
+        if (config->preprocessing & 2) {
+          const float x = config->quality / 100.f;
+          const float x2 = x * x;
+          // slowly decreasing from max dithering at low quality (q->0)
+          // to 0.5 dithering amplitude at high quality (q->100)
+          dithering = 1.0f + (0.5f - 1.0f) * x2 * x2;
+        }
+        if (!WebPPictureARGBToYUVADithered(pic, WEBP_YUV420, dithering)) {
+          return 0;
+        }
+      }
+    }
+
+    if (!config->exact) {
+      WebPCleanupTransparentArea(pic);
+    }
+
+    enc = InitVP8Encoder(config, pic);
+    if (enc == NULL) return 0;  // pic->error is already set.
+    // Note: each of the tasks below account for 20% in the progress report.
+    ok = VP8EncAnalyze(enc);
+
+    // Analysis is done, proceed to actual coding.
+    ok = ok && VP8EncStartAlpha(enc);   // possibly done in parallel
+    if (!enc->use_tokens_) {
+      ok = ok && VP8EncLoop(enc);
+    } else {
+      ok = ok && VP8EncTokenLoop(enc);
+    }
+    ok = ok && VP8EncFinishAlpha(enc);
+
+    ok = ok && VP8EncWrite(enc);
+    StoreStats(enc);
+    if (!ok) {
+      VP8EncFreeBitWriters(enc);
+    }
+    ok &= DeleteVP8Encoder(enc);  // must always be called, even if !ok
+  } else {
+    // Make sure we have ARGB samples.
+    if (pic->argb == NULL && !WebPPictureYUVAToARGB(pic)) {
+      return 0;
+    }
+
+    if (!config->exact) {
+      WebPReplaceTransparentPixels(pic, 0x000000);
+    }
+
+    ok = VP8LEncodeImage(config, pic);  // Sets pic->error in case of problem.
+  }
+
+  return ok;
+}