Adding libwebp: Adding clean copy of libwebp 0.4.0

This commit imports libwebp 0.4.0, including AUTHORS, COPYING, ChangeLog,
NEWS, PATENTS, README and src directories. In src, only includes header
and source files.

The patches required to build it in Qt will follow in separate
commit(s).

Task-number: QTBUG-14205
Change-Id: I747918fe2a07716f66cffb795129f7aa5e067759
Reviewed-by: Lars Knoll <lars.knoll@digia.com>

22 files changed, 11177 insertions, 0 deletions

diff --git a/src/3rdparty/libwebp/src/enc/alpha.c b/src/3rdparty/libwebp/src/enc/alpha.c
new file mode 100644
index 0000000..21d4b5c
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/alpha.c
@@ -0,0 +1,410 @@
+// 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 "./vp8enci.h"
+#include "../utils/filters.h"
+#include "../utils/quant_levels.h"
+#include "../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
+//          free(*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 "../enc/vp8li.h"
+
+static int EncodeLossless(const uint8_t* const data, int width, int height,
+                          int effort_level,  // in [0..6] range
+                          VP8BitWriter* const bw,
+                          WebPAuxStats* const stats) {
+  int ok = 0;
+  WebPConfig config;
+  WebPPicture picture;
+  VP8LBitWriter tmp_bw;
+
+  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.
+  {
+    int i, j;
+    uint32_t* dst = picture.argb;
+    const uint8_t* src = data;
+    for (j = 0; j < picture.height; ++j) {
+      for (i = 0; i < picture.width; ++i) {
+        dst[i] = src[i] << 8;  // we leave A/R/B channels zero'd.
+      }
+      src += width;
+      dst += picture.argb_stride;
+    }
+  }
+
+  WebPConfigInit(&config);
+  config.lossless = 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'.
+  config.quality = 8.f * effort_level;
+  assert(config.quality >= 0 && config.quality <= 100.f);
+
+  ok = VP8LBitWriterInit(&tmp_bw, (width * height) >> 3);
+  ok = ok && (VP8LEncodeStream(&config, &picture, &tmp_bw) == VP8_ENC_OK);
+  WebPPictureFree(&picture);
+  if (ok) {
+    const uint8_t* const buffer = VP8LBitWriterFinish(&tmp_bw);
+    const size_t buffer_size = VP8LBitWriterNumBytes(&tmp_bw);
+    VP8BitWriterAppend(bw, buffer, buffer_size);
+  }
+  VP8LBitWriterDestroy(&tmp_bw);
+  return ok && !bw->error_;
+}
+
+// -----------------------------------------------------------------------------
+
+// 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;
+  size_t expected_size;
+  const size_t data_size = width * height;
+
+  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);
+  // TODO(skal): have a common function and #define's to validate alpha params.
+
+  expected_size =
+      (method == ALPHA_NO_COMPRESSION) ? (ALPHA_HEADER_LEN + data_size)
+                                       : (data_size >> 5);
+  header = method | (filter << 2);
+  if (reduce_levels) header |= ALPHA_PREPROCESSED_LEVELS << 4;
+
+  VP8BitWriterInit(&result->bw, expected_size);
+  VP8BitWriterAppend(&result->bw, &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 = VP8BitWriterAppend(&result->bw, alpha_src, width * height);
+    ok = ok && !result->bw.error_;
+  } else {
+    ok = EncodeLossless(alpha_src, width, height, effort_level,
+                        &result->bw, &result->stats);
+    VP8BitWriterFinish(&result->bw);
+  }
+  result->score = VP8BitWriterSize(&result->bw);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+
+// TODO(skal): move to dsp/ ?
+static void CopyPlane(const uint8_t* src, int src_stride,
+                      uint8_t* dst, int dst_stride, int width, int height) {
+  while (height-- > 0) {
+    memcpy(dst, src, width);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+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 :
+             EstimateBestFilter(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*)malloc(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);
+        }
+      }
+    }
+    free(filtered_alpha);
+  } else {
+    ok = EncodeAlphaInternal(alpha, width, height, method, WEBP_FILTER_NONE,
+                             reduce_levels, effort_level, NULL, &best);
+  }
+  if (ok) {
+    if (stats != NULL) *stats = best.stats;
+    *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 sanity 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*)malloc(data_size);
+  if (quant_alpha == NULL) {
+    return 0;
+  }
+
+  // Extract alpha data (width x height) from raw_data (stride x height).
+  CopyPlane(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) {
+    ok = ApplyFiltersAndEncode(quant_alpha, width, height, data_size, method,
+                               filter, reduce_levels, effort_level, output,
+                               output_size, pic->stats);
+    if (pic->stats != NULL) {  // need stats?
+      pic->stats->coded_size += (int)(*output_size);
+      enc->sse_[3] = sse;
+    }
+  }
+
+  free(quant_alpha);
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+// Main calls
+
+static int CompressAlphaJob(VP8Encoder* const enc, void* dummy) {
+  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) {  // Sanity check.
+    free(alpha_data);
+    return 0;
+  }
+  enc->alpha_data_size_ = (uint32_t)alpha_size;
+  enc->alpha_data_ = alpha_data;
+  (void)dummy;
+  return 1;
+}
+
+void VP8EncInitAlpha(VP8Encoder* const enc) {
+  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_;
+    WebPWorkerInit(worker);
+    worker->data1 = enc;
+    worker->data2 = NULL;
+    worker->hook = (WebPWorkerHook)CompressAlphaJob;
+  }
+}
+
+int VP8EncStartAlpha(VP8Encoder* const enc) {
+  if (enc->has_alpha_) {
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      if (!WebPWorkerReset(worker)) {    // Makes sure worker is good to go.
+        return 0;
+      }
+      WebPWorkerLaunch(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 (!WebPWorkerSync(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_;
+    ok = WebPWorkerSync(worker);  // finish anything left in flight
+    WebPWorkerEnd(worker);  // still need to end the worker, even if !ok
+  }
+  free(enc->alpha_data_);
+  enc->alpha_data_ = NULL;
+  enc->alpha_data_size_ = 0;
+  enc->has_alpha_ = 0;
+  return ok;
+}
+
diff --git a/src/3rdparty/libwebp/src/enc/analysis.c b/src/3rdparty/libwebp/src/enc/analysis.c
new file mode 100644
index 0000000..7d4cfdc
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/analysis.c
@@ -0,0 +1,497 @@
+// 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 "./vp8enci.h"
+#include "./cost.h"
+#include "../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((uint64_t)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];
+    }
+  }
+  free(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) {
+  int max_value = 0, last_non_zero = 1;
+  int k;
+  int alpha;
+  for (k = 0; k <= MAX_COEFF_THRESH; ++k) {
+    const int value = histo->distribution[k];
+    if (value > 0) {
+      if (value > max_value) max_value = value;
+      last_non_zero = k;
+    }
+  }
+  // '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.
+  alpha = (max_value > 1) ? ALPHA_SCALE * last_non_zero / max_value : 0;
+  return alpha;
+}
+
+static void MergeHistograms(const VP8Histogram* const in,
+                            VP8Histogram* const out) {
+  int i;
+  for (i = 0; i <= MAX_COEFF_THRESH; ++i) {
+    out->distribution[i] += in->distribution[i];
+  }
+}
+
+//------------------------------------------------------------------------------
+// Simplified k-Means, to assign Nb segments based on alpha-histogram
+
+static void AssignSegments(VP8Encoder* const enc,
+                           const int alphas[MAX_ALPHA + 1]) {
+  const int nb = enc->segment_hdr_.num_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);
+
+  // 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. For high-quality settings
+// (method >= FAST_ANALYSIS_METHOD) we don't need to test all the possible modes
+// during the analysis phase.
+#define FAST_ANALYSIS_METHOD 4  // method above which we do partial analysis
+#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 =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA16_MODE
+                                                  : NUM_PRED_MODES;
+  int mode;
+  int best_alpha = DEFAULT_ALPHA;
+  int best_mode = 0;
+
+  VP8MakeLuma16Preds(it);
+  for (mode = 0; mode < max_mode; ++mode) {
+    VP8Histogram histo = { { 0 } };
+    int alpha;
+
+    VP8CollectHistogram(it->yuv_in_ + Y_OFF,
+                        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 MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
+                                   int best_alpha) {
+  uint8_t modes[16];
+  const int max_mode =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA4_MODE
+                                                  : NUM_BMODES;
+  int i4_alpha;
+  VP8Histogram total_histo = { { 0 } };
+  int cur_histo = 0;
+
+  VP8IteratorStartI4(it);
+  do {
+    int mode;
+    int best_mode_alpha = DEFAULT_ALPHA;
+    VP8Histogram histos[2];
+    const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
+
+    VP8MakeIntra4Preds(it);
+    for (mode = 0; mode < max_mode; ++mode) {
+      int alpha;
+
+      memset(&histos[cur_histo], 0, sizeof(histos[cur_histo]));
+      VP8CollectHistogram(src, it->yuv_p_ + VP8I4ModeOffsets[mode],
+                          0, 1, &histos[cur_histo]);
+      alpha = GetAlpha(&histos[cur_histo]);
+      if (IS_BETTER_ALPHA(alpha, best_mode_alpha)) {
+        best_mode_alpha = alpha;
+        modes[it->i4_] = mode;
+        cur_histo ^= 1;   // keep track of best histo so far.
+      }
+    }
+    // accumulate best histogram
+    MergeHistograms(&histos[cur_histo ^ 1], &total_histo);
+    // Note: we reuse the original samples for predictors
+  } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF));
+
+  i4_alpha = GetAlpha(&total_histo);
+  if (IS_BETTER_ALPHA(i4_alpha, best_alpha)) {
+    VP8SetIntra4Mode(it, modes);
+    best_alpha = i4_alpha;
+  }
+  return best_alpha;
+}
+
+static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
+  int best_alpha = DEFAULT_ALPHA;
+  int best_mode = 0;
+  const int max_mode =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_UV_MODE
+                                                  : NUM_PRED_MODES;
+  int mode;
+  VP8MakeChroma8Preds(it);
+  for (mode = 0; mode < max_mode; ++mode) {
+    VP8Histogram histo = { { 0 } };
+    int alpha;
+    VP8CollectHistogram(it->yuv_in_ + U_OFF,
+                        it->yuv_p_ + VP8UVModeOffsets[mode],
+                        16, 16 + 4 + 4, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
+      best_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.
+
+  best_alpha = MBAnalyzeBestIntra16Mode(it);
+  if (enc->method_ >= 5) {
+    // We go and make a fast decision for intra4/intra16.
+    // It's usually not a good and definitive pick, but helps seeding the stats
+    // about level bit-cost.
+    // TODO(skal): improve criterion.
+    best_alpha = MBAnalyzeBestIntra4Mode(it, best_alpha);
+  }
+  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(SegmentJob* const job, VP8EncIterator* const it) {
+  int ok = 1;
+  if (!VP8IteratorIsDone(it)) {
+    uint8_t tmp[32 + ALIGN_CST];
+    uint8_t* const scratch = (uint8_t*)DO_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 TODOs
+static void InitSegmentJob(VP8Encoder* const enc, SegmentJob* const job,
+                           int start_row, int end_row) {
+  WebPWorkerInit(&job->worker);
+  job->worker.data1 = job;
+  job->worker.data2 = &job->it;
+  job->worker.hook = (WebPWorkerHook)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_ == 0);  // for method 0, 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
+    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 &= WebPWorkerReset(&side_job.worker);
+      // launch the two jobs in parallel
+      if (ok) {
+        WebPWorkerLaunch(&side_job.worker);
+        WebPWorkerExecute(&main_job.worker);
+        ok &= WebPWorkerSync(&side_job.worker);
+        ok &= WebPWorkerSync(&main_job.worker);
+      }
+      WebPWorkerEnd(&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);
+      WebPWorkerExecute(&main_job.worker);
+      ok &= WebPWorkerSync(&main_job.worker);
+    }
+    WebPWorkerEnd(&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/src/3rdparty/libwebp/src/enc/backward_references.c b/src/3rdparty/libwebp/src/enc/backward_references.c
new file mode 100644
index 0000000..77b4be7
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/backward_references.c
@@ -0,0 +1,894 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+//
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./backward_references.h"
+#include "./histogram.h"
+#include "../dsp/lossless.h"
+#include "../utils/color_cache.h"
+#include "../utils/utils.h"
+
+#define VALUES_IN_BYTE 256
+
+#define HASH_BITS 18
+#define HASH_SIZE (1 << HASH_BITS)
+#define HASH_MULTIPLIER (0xc6a4a7935bd1e995ULL)
+
+// 1M window (4M bytes) minus 120 special codes for short distances.
+#define WINDOW_SIZE ((1 << 20) - 120)
+
+// Bounds for the match length.
+#define MIN_LENGTH 2
+#define MAX_LENGTH 4096
+
+typedef struct {
+  // Stores the most recently added position with the given hash value.
+  int32_t hash_to_first_index_[HASH_SIZE];
+  // chain_[pos] stores the previous position with the same hash value
+  // for every pixel in the image.
+  int32_t* chain_;
+} HashChain;
+
+// -----------------------------------------------------------------------------
+
+static const uint8_t plane_to_code_lut[128] = {
+ 96,   73,  55,  39,  23,  13,   5,  1,  255, 255, 255, 255, 255, 255, 255, 255,
+ 101,  78,  58,  42,  26,  16,   8,  2,    0,   3,  9,   17,  27,  43,  59,  79,
+ 102,  86,  62,  46,  32,  20,  10,  6,    4,   7,  11,  21,  33,  47,  63,  87,
+ 105,  90,  70,  52,  37,  28,  18,  14,  12,  15,  19,  29,  38,  53,  71,  91,
+ 110,  99,  82,  66,  48,  35,  30,  24,  22,  25,  31,  36,  49,  67,  83, 100,
+ 115, 108,  94,  76,  64,  50,  44,  40,  34,  41,  45,  51,  65,  77,  95, 109,
+ 118, 113, 103,  92,  80,  68,  60,  56,  54,  57,  61,  69,  81,  93, 104, 114,
+ 119, 116, 111, 106,  97,  88,  84,  74,  72,  75,  85,  89,  98, 107, 112, 117
+};
+
+static int DistanceToPlaneCode(int xsize, int dist) {
+  const int yoffset = dist / xsize;
+  const int xoffset = dist - yoffset * xsize;
+  if (xoffset <= 8 && yoffset < 8) {
+    return plane_to_code_lut[yoffset * 16 + 8 - xoffset] + 1;
+  } else if (xoffset > xsize - 8 && yoffset < 7) {
+    return plane_to_code_lut[(yoffset + 1) * 16 + 8 + (xsize - xoffset)] + 1;
+  }
+  return dist + 120;
+}
+
+static WEBP_INLINE int FindMatchLength(const uint32_t* const array1,
+                                       const uint32_t* const array2,
+                                       const int max_limit) {
+  int match_len = 0;
+  while (match_len < max_limit && array1[match_len] == array2[match_len]) {
+    ++match_len;
+  }
+  return match_len;
+}
+
+// -----------------------------------------------------------------------------
+//  VP8LBackwardRefs
+
+void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs) {
+  if (refs != NULL) {
+    refs->refs = NULL;
+    refs->size = 0;
+    refs->max_size = 0;
+  }
+}
+
+void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs) {
+  if (refs != NULL) {
+    free(refs->refs);
+    VP8LInitBackwardRefs(refs);
+  }
+}
+
+int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size) {
+  assert(refs != NULL);
+  refs->size = 0;
+  refs->max_size = 0;
+  refs->refs = (PixOrCopy*)WebPSafeMalloc((uint64_t)max_size,
+                                          sizeof(*refs->refs));
+  if (refs->refs == NULL) return 0;
+  refs->max_size = max_size;
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+// Hash chains
+
+static WEBP_INLINE uint64_t GetPixPairHash64(const uint32_t* const argb) {
+  uint64_t key = ((uint64_t)(argb[1]) << 32) | argb[0];
+  key = (key * HASH_MULTIPLIER) >> (64 - HASH_BITS);
+  return key;
+}
+
+static int HashChainInit(HashChain* const p, int size) {
+  int i;
+  p->chain_ = (int*)WebPSafeMalloc((uint64_t)size, sizeof(*p->chain_));
+  if (p->chain_ == NULL) {
+    return 0;
+  }
+  for (i = 0; i < size; ++i) {
+    p->chain_[i] = -1;
+  }
+  for (i = 0; i < HASH_SIZE; ++i) {
+    p->hash_to_first_index_[i] = -1;
+  }
+  return 1;
+}
+
+static void HashChainDelete(HashChain* const p) {
+  if (p != NULL) {
+    free(p->chain_);
+    free(p);
+  }
+}
+
+// Insertion of two pixels at a time.
+static void HashChainInsert(HashChain* const p,
+                            const uint32_t* const argb, int pos) {
+  const uint64_t hash_code = GetPixPairHash64(argb);
+  p->chain_[pos] = p->hash_to_first_index_[hash_code];
+  p->hash_to_first_index_[hash_code] = pos;
+}
+
+static void GetParamsForHashChainFindCopy(int quality, int xsize,
+                                          int cache_bits, int* window_size,
+                                          int* iter_pos, int* iter_limit) {
+  const int iter_mult = (quality < 27) ? 1 : 1 + ((quality - 27) >> 4);
+  const int iter_neg = -iter_mult * (quality >> 1);
+  // Limit the backward-ref window size for lower qualities.
+  const int max_window_size = (quality > 50) ? WINDOW_SIZE
+                            : (quality > 25) ? (xsize << 8)
+                            : (xsize << 4);
+  assert(xsize > 0);
+  *window_size = (max_window_size > WINDOW_SIZE) ? WINDOW_SIZE
+               : max_window_size;
+  *iter_pos = 8 + (quality >> 3);
+  // For lower entropy images, the rigorous search loop in HashChainFindCopy
+  // can be relaxed.
+  *iter_limit = (cache_bits > 0) ? iter_neg : iter_neg / 2;
+}
+
+static int HashChainFindCopy(const HashChain* const p,
+                             int base_position, int xsize_signed,
+                             const uint32_t* const argb, int max_len,
+                             int window_size, int iter_pos, int iter_limit,
+                             int* const distance_ptr,
+                             int* const length_ptr) {
+  const uint32_t* const argb_start = argb + base_position;
+  uint64_t best_val = 0;
+  uint32_t best_length = 1;
+  uint32_t best_distance = 0;
+  const uint32_t xsize = (uint32_t)xsize_signed;
+  const int min_pos =
+      (base_position > window_size) ? base_position - window_size : 0;
+  int pos;
+  assert(xsize > 0);
+  if (max_len > MAX_LENGTH) {
+    max_len = MAX_LENGTH;
+  }
+  for (pos = p->hash_to_first_index_[GetPixPairHash64(argb_start)];
+       pos >= min_pos;
+       pos = p->chain_[pos]) {
+    uint64_t val;
+    uint32_t curr_length;
+    uint32_t distance;
+    const uint64_t* const ptr1 =
+        (const uint64_t*)(argb + pos + best_length - 1);
+    const uint64_t* const ptr2 =
+        (const uint64_t*)(argb_start + best_length - 1);
+
+    if (iter_pos < 0) {
+      if (iter_pos < iter_limit || best_val >= 0xff0000) {
+        break;
+      }
+    }
+    --iter_pos;
+
+    // Before 'expensive' linear match, check if the two arrays match at the
+    // current best length index and also for the succeeding elements.
+    if (*ptr1 != *ptr2) continue;
+
+    curr_length = FindMatchLength(argb + pos, argb_start, max_len);
+    if (curr_length < best_length) continue;
+
+    distance = (uint32_t)(base_position - pos);
+    val = curr_length << 16;
+    // Favoring 2d locality here gives savings for certain images.
+    if (distance < 9 * xsize) {
+      const uint32_t y = distance / xsize;
+      uint32_t x = distance % xsize;
+      if (x > (xsize >> 1)) {
+        x = xsize - x;
+      }
+      if (x <= 7) {
+        val += 9 * 9 + 9 * 9;
+        val -= y * y + x * x;
+      }
+    }
+    if (best_val < val) {
+      best_val = val;
+      best_length = curr_length;
+      best_distance = distance;
+      if (curr_length >= (uint32_t)max_len) {
+        break;
+      }
+      if ((best_distance == 1 || distance == xsize) &&
+          best_length >= 128) {
+        break;
+      }
+    }
+  }
+  *distance_ptr = (int)best_distance;
+  *length_ptr = best_length;
+  return (best_length >= MIN_LENGTH);
+}
+
+static WEBP_INLINE void PushBackCopy(VP8LBackwardRefs* const refs, int length) {
+  int size = refs->size;
+  while (length >= MAX_LENGTH) {
+    refs->refs[size++] = PixOrCopyCreateCopy(1, MAX_LENGTH);
+    length -= MAX_LENGTH;
+  }
+  if (length > 0) {
+    refs->refs[size++] = PixOrCopyCreateCopy(1, length);
+  }
+  refs->size = size;
+}
+
+static void BackwardReferencesRle(int xsize, int ysize,
+                                  const uint32_t* const argb,
+                                  VP8LBackwardRefs* const refs) {
+  const int pix_count = xsize * ysize;
+  int match_len = 0;
+  int i;
+  refs->size = 0;
+  PushBackCopy(refs, match_len);    // i=0 case
+  refs->refs[refs->size++] = PixOrCopyCreateLiteral(argb[0]);
+  for (i = 1; i < pix_count; ++i) {
+    if (argb[i] == argb[i - 1]) {
+      ++match_len;
+    } else {
+      PushBackCopy(refs, match_len);
+      match_len = 0;
+      refs->refs[refs->size++] = PixOrCopyCreateLiteral(argb[i]);
+    }
+  }
+  PushBackCopy(refs, match_len);
+}
+
+static int BackwardReferencesHashChain(int xsize, int ysize,
+                                       const uint32_t* const argb,
+                                       int cache_bits, int quality,
+                                       VP8LBackwardRefs* const refs) {
+  int i;
+  int ok = 0;
+  int cc_init = 0;
+  const int use_color_cache = (cache_bits > 0);
+  const int pix_count = xsize * ysize;
+  HashChain* const hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+  VP8LColorCache hashers;
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
+
+  if (hash_chain == NULL) return 0;
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+
+  if (!HashChainInit(hash_chain, pix_count)) goto Error;
+
+  refs->size = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
+                                &window_size, &iter_pos, &iter_limit);
+  for (i = 0; i < pix_count; ) {
+    // Alternative#1: Code the pixels starting at 'i' using backward reference.
+    int offset = 0;
+    int len = 0;
+    if (i < pix_count - 1) {  // FindCopy(i,..) reads pixels at [i] and [i + 1].
+      int max_len = pix_count - i;
+      HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
+                        window_size, iter_pos, iter_limit,
+                        &offset, &len);
+    }
+    if (len >= MIN_LENGTH) {
+      // Alternative#2: Insert the pixel at 'i' as literal, and code the
+      // pixels starting at 'i + 1' using backward reference.
+      int offset2 = 0;
+      int len2 = 0;
+      int k;
+      HashChainInsert(hash_chain, &argb[i], i);
+      if (i < pix_count - 2) {  // FindCopy(i+1,..) reads [i + 1] and [i + 2].
+        int max_len = pix_count - (i + 1);
+        HashChainFindCopy(hash_chain, i + 1, xsize, argb, max_len,
+                          window_size, iter_pos, iter_limit,
+                          &offset2, &len2);
+        if (len2 > len + 1) {
+          const uint32_t pixel = argb[i];
+          // Alternative#2 is a better match. So push pixel at 'i' as literal.
+          if (use_color_cache && VP8LColorCacheContains(&hashers, pixel)) {
+            const int ix = VP8LColorCacheGetIndex(&hashers, pixel);
+            refs->refs[refs->size] = PixOrCopyCreateCacheIdx(ix);
+          } else {
+            if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
+            refs->refs[refs->size] = PixOrCopyCreateLiteral(pixel);
+          }
+          ++refs->size;
+          i++;  // Backward reference to be done for next pixel.
+          len = len2;
+          offset = offset2;
+        }
+      }
+      if (len >= MAX_LENGTH) {
+        len = MAX_LENGTH - 1;
+      }
+      refs->refs[refs->size++] = PixOrCopyCreateCopy(offset, len);
+      if (use_color_cache) {
+        for (k = 0; k < len; ++k) {
+          VP8LColorCacheInsert(&hashers, argb[i + k]);
+        }
+      }
+      // Add to the hash_chain (but cannot add the last pixel).
+      {
+        const int last = (len < pix_count - 1 - i) ? len : pix_count - 1 - i;
+        for (k = 1; k < last; ++k) {
+          HashChainInsert(hash_chain, &argb[i + k], i + k);
+        }
+      }
+      i += len;
+    } else {
+      const uint32_t pixel = argb[i];
+      if (use_color_cache && VP8LColorCacheContains(&hashers, pixel)) {
+        // push pixel as a PixOrCopyCreateCacheIdx pixel
+        const int ix = VP8LColorCacheGetIndex(&hashers, pixel);
+        refs->refs[refs->size] = PixOrCopyCreateCacheIdx(ix);
+      } else {
+        if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
+        refs->refs[refs->size] = PixOrCopyCreateLiteral(pixel);
+      }
+      ++refs->size;
+      if (i + 1 < pix_count) {
+        HashChainInsert(hash_chain, &argb[i], i);
+      }
+      ++i;
+    }
+  }
+  ok = 1;
+Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  HashChainDelete(hash_chain);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+
+typedef struct {
+  double alpha_[VALUES_IN_BYTE];
+  double red_[VALUES_IN_BYTE];
+  double literal_[PIX_OR_COPY_CODES_MAX];
+  double blue_[VALUES_IN_BYTE];
+  double distance_[NUM_DISTANCE_CODES];
+} CostModel;
+
+static int BackwardReferencesTraceBackwards(
+    int xsize, int ysize, int recursive_cost_model,
+    const uint32_t* const argb, int quality, int cache_bits,
+    VP8LBackwardRefs* const refs);
+
+static void ConvertPopulationCountTableToBitEstimates(
+    int num_symbols, const int population_counts[], double output[]) {
+  int sum = 0;
+  int nonzeros = 0;
+  int i;
+  for (i = 0; i < num_symbols; ++i) {
+    sum += population_counts[i];
+    if (population_counts[i] > 0) {
+      ++nonzeros;
+    }
+  }
+  if (nonzeros <= 1) {
+    memset(output, 0, num_symbols * sizeof(*output));
+  } else {
+    const double logsum = VP8LFastLog2(sum);
+    for (i = 0; i < num_symbols; ++i) {
+      output[i] = logsum - VP8LFastLog2(population_counts[i]);
+    }
+  }
+}
+
+static int CostModelBuild(CostModel* const m, int xsize, int ysize,
+                          int recursion_level, const uint32_t* const argb,
+                          int quality, int cache_bits) {
+  int ok = 0;
+  VP8LHistogram histo;
+  VP8LBackwardRefs refs;
+
+  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize)) goto Error;
+
+  if (recursion_level > 0) {
+    if (!BackwardReferencesTraceBackwards(xsize, ysize, recursion_level - 1,
+                                          argb, quality, cache_bits, &refs)) {
+      goto Error;
+    }
+  } else {
+    if (!BackwardReferencesHashChain(xsize, ysize, argb, cache_bits, quality,
+                                     &refs)) {
+      goto Error;
+    }
+  }
+  VP8LHistogramCreate(&histo, &refs, cache_bits);
+  ConvertPopulationCountTableToBitEstimates(
+      VP8LHistogramNumCodes(&histo), 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:
+  VP8LClearBackwardRefs(&refs);
+  return ok;
+}
+
+static WEBP_INLINE double GetLiteralCost(const CostModel* const m, uint32_t v) {
+  return m->alpha_[v >> 24] +
+         m->red_[(v >> 16) & 0xff] +
+         m->literal_[(v >> 8) & 0xff] +
+         m->blue_[v & 0xff];
+}
+
+static WEBP_INLINE double GetCacheCost(const CostModel* const m, uint32_t idx) {
+  const int literal_idx = VALUES_IN_BYTE + NUM_LENGTH_CODES + idx;
+  return m->literal_[literal_idx];
+}
+
+static WEBP_INLINE double GetLengthCost(const CostModel* const m,
+                                        uint32_t length) {
+  int code, extra_bits;
+  VP8LPrefixEncodeBits(length, &code, &extra_bits);
+  return m->literal_[VALUES_IN_BYTE + code] + extra_bits;
+}
+
+static WEBP_INLINE double GetDistanceCost(const CostModel* const m,
+                                          uint32_t distance) {
+  int code, extra_bits;
+  VP8LPrefixEncodeBits(distance, &code, &extra_bits);
+  return m->distance_[code] + extra_bits;
+}
+
+static int BackwardReferencesHashChainDistanceOnly(
+    int xsize, int ysize, int recursive_cost_model, const uint32_t* const argb,
+    int quality, int cache_bits, uint32_t* const dist_array) {
+  int i;
+  int ok = 0;
+  int cc_init = 0;
+  const int pix_count = xsize * ysize;
+  const int use_color_cache = (cache_bits > 0);
+  float* const cost =
+      (float*)WebPSafeMalloc((uint64_t)pix_count, sizeof(*cost));
+  CostModel* cost_model = (CostModel*)malloc(sizeof(*cost_model));
+  HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+  VP8LColorCache hashers;
+  const double mul0 = (recursive_cost_model != 0) ? 1.0 : 0.68;
+  const double mul1 = (recursive_cost_model != 0) ? 1.0 : 0.82;
+  const int min_distance_code = 2;  // TODO(vikasa): tune as function of quality
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
+
+  if (cost == NULL || cost_model == NULL || hash_chain == NULL) goto Error;
+
+  if (!HashChainInit(hash_chain, pix_count)) goto Error;
+
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+
+  if (!CostModelBuild(cost_model, xsize, ysize, recursive_cost_model, argb,
+                      quality, cache_bits)) {
+    goto Error;
+  }
+
+  for (i = 0; i < pix_count; ++i) cost[i] = 1e38f;
+
+  // We loop one pixel at a time, but store all currently best points to
+  // non-processed locations from this point.
+  dist_array[0] = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
+                                &window_size, &iter_pos, &iter_limit);
+  for (i = 0; i < pix_count; ++i) {
+    double prev_cost = 0.0;
+    int shortmax;
+    if (i > 0) {
+      prev_cost = cost[i - 1];
+    }
+    for (shortmax = 0; shortmax < 2; ++shortmax) {
+      int offset = 0;
+      int len = 0;
+      if (i < pix_count - 1) {  // FindCopy reads pixels at [i] and [i + 1].
+        int max_len = shortmax ? 2 : pix_count - i;
+        HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
+                          window_size, iter_pos, iter_limit,
+                          &offset, &len);
+      }
+      if (len >= MIN_LENGTH) {
+        const int code = DistanceToPlaneCode(xsize, offset);
+        const double distance_cost =
+            prev_cost + GetDistanceCost(cost_model, code);
+        int k;
+        for (k = 1; k < len; ++k) {
+          const double cost_val = distance_cost + GetLengthCost(cost_model, k);
+          if (cost[i + k] > cost_val) {
+            cost[i + k] = (float)cost_val;
+            dist_array[i + k] = k + 1;
+          }
+        }
+        // This if is for speedup only. It roughly doubles the speed, and
+        // makes compression worse by .1 %.
+        if (len >= 128 && code <= min_distance_code) {
+          // Long copy for short distances, let's skip the middle
+          // lookups for better copies.
+          // 1) insert the hashes.
+          if (use_color_cache) {
+            for (k = 0; k < len; ++k) {
+              VP8LColorCacheInsert(&hashers, argb[i + k]);
+            }
+          }
+          // 2) Add to the hash_chain (but cannot add the last pixel)
+          {
+            const int last = (len + i < pix_count - 1) ? len + i
+                                                       : pix_count - 1;
+            for (k = i; k < last; ++k) {
+              HashChainInsert(hash_chain, &argb[k], k);
+            }
+          }
+          // 3) jump.
+          i += len - 1;  // for loop does ++i, thus -1 here.
+          goto next_symbol;
+        }
+      }
+    }
+    if (i < pix_count - 1) {
+      HashChainInsert(hash_chain, &argb[i], i);
+    }
+    {
+      // inserting a literal pixel
+      double cost_val = prev_cost;
+      if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
+        const int ix = VP8LColorCacheGetIndex(&hashers, argb[i]);
+        cost_val += GetCacheCost(cost_model, ix) * mul0;
+      } else {
+        if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
+        cost_val += GetLiteralCost(cost_model, argb[i]) * mul1;
+      }
+      if (cost[i] > cost_val) {
+        cost[i] = (float)cost_val;
+        dist_array[i] = 1;  // only one is inserted.
+      }
+    }
+ next_symbol: ;
+  }
+  // Last pixel still to do, it can only be a single step if not reached
+  // through cheaper means already.
+  ok = 1;
+Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  HashChainDelete(hash_chain);
+  free(cost_model);
+  free(cost);
+  return ok;
+}
+
+// We pack the path at the end of *dist_array and return
+// a pointer to this part of the array. Example:
+// dist_array = [1x2xx3x2] => packed [1x2x1232], chosen_path = [1232]
+static void TraceBackwards(uint32_t* const dist_array,
+                           int dist_array_size,
+                           uint32_t** const chosen_path,
+                           int* const chosen_path_size) {
+  uint32_t* path = dist_array + dist_array_size;
+  uint32_t* cur = dist_array + dist_array_size - 1;
+  while (cur >= dist_array) {
+    const int k = *cur;
+    --path;
+    *path = k;
+    cur -= k;
+  }
+  *chosen_path = path;
+  *chosen_path_size = (int)(dist_array + dist_array_size - path);
+}
+
+static int BackwardReferencesHashChainFollowChosenPath(
+    int xsize, int ysize, const uint32_t* const argb,
+    int quality, int cache_bits,
+    const uint32_t* const chosen_path, int chosen_path_size,
+    VP8LBackwardRefs* const refs) {
+  const int pix_count = xsize * ysize;
+  const int use_color_cache = (cache_bits > 0);
+  int size = 0;
+  int i = 0;
+  int k;
+  int ix;
+  int ok = 0;
+  int cc_init = 0;
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
+  HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+  VP8LColorCache hashers;
+
+  if (hash_chain == NULL || !HashChainInit(hash_chain, pix_count)) {
+    goto Error;
+  }
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+
+  refs->size = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
+                                &window_size, &iter_pos, &iter_limit);
+  for (ix = 0; ix < chosen_path_size; ++ix, ++size) {
+    int offset = 0;
+    int len = 0;
+    int max_len = chosen_path[ix];
+    if (max_len != 1) {
+      HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
+                        window_size, iter_pos, iter_limit,
+                        &offset, &len);
+      assert(len == max_len);
+      refs->refs[size] = PixOrCopyCreateCopy(offset, len);
+      if (use_color_cache) {
+        for (k = 0; k < len; ++k) {
+          VP8LColorCacheInsert(&hashers, argb[i + k]);
+        }
+      }
+      {
+        const int last = (len < pix_count - 1 - i) ? len : pix_count - 1 - i;
+        for (k = 0; k < last; ++k) {
+          HashChainInsert(hash_chain, &argb[i + k], i + k);
+        }
+      }
+      i += len;
+    } else {
+      if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
+        // push pixel as a color cache index
+        const int idx = VP8LColorCacheGetIndex(&hashers, argb[i]);
+        refs->refs[size] = PixOrCopyCreateCacheIdx(idx);
+      } else {
+        if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
+        refs->refs[size] = PixOrCopyCreateLiteral(argb[i]);
+      }
+      if (i + 1 < pix_count) {
+        HashChainInsert(hash_chain, &argb[i], i);
+      }
+      ++i;
+    }
+  }
+  assert(size <= refs->max_size);
+  refs->size = size;
+  ok = 1;
+Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  HashChainDelete(hash_chain);
+  return ok;
+}
+
+// Returns 1 on success.
+static int BackwardReferencesTraceBackwards(int xsize, int ysize,
+                                            int recursive_cost_model,
+                                            const uint32_t* const argb,
+                                            int quality, int cache_bits,
+                                            VP8LBackwardRefs* const refs) {
+  int ok = 0;
+  const int dist_array_size = xsize * ysize;
+  uint32_t* chosen_path = NULL;
+  int chosen_path_size = 0;
+  uint32_t* dist_array =
+      (uint32_t*)WebPSafeMalloc((uint64_t)dist_array_size, sizeof(*dist_array));
+
+  if (dist_array == NULL) goto Error;
+
+  if (!BackwardReferencesHashChainDistanceOnly(
+      xsize, ysize, recursive_cost_model, argb, quality, cache_bits,
+      dist_array)) {
+    goto Error;
+  }
+  TraceBackwards(dist_array, dist_array_size, &chosen_path, &chosen_path_size);
+  if (!BackwardReferencesHashChainFollowChosenPath(
+      xsize, ysize, argb, quality, cache_bits, chosen_path, chosen_path_size,
+      refs)) {
+    goto Error;
+  }
+  ok = 1;
+ Error:
+  free(dist_array);
+  return ok;
+}
+
+static void BackwardReferences2DLocality(int xsize,
+                                         VP8LBackwardRefs* const refs) {
+  int i;
+  for (i = 0; i < refs->size; ++i) {
+    if (PixOrCopyIsCopy(&refs->refs[i])) {
+      const int dist = refs->refs[i].argb_or_distance;
+      const int transformed_dist = DistanceToPlaneCode(xsize, dist);
+      refs->refs[i].argb_or_distance = transformed_dist;
+    }
+  }
+}
+
+int VP8LGetBackwardReferences(int width, int height,
+                              const uint32_t* const argb,
+                              int quality, int cache_bits, int use_2d_locality,
+                              VP8LBackwardRefs* const best) {
+  int ok = 0;
+  int lz77_is_useful;
+  VP8LBackwardRefs refs_rle, refs_lz77;
+  const int num_pix = width * height;
+
+  VP8LBackwardRefsAlloc(&refs_rle, num_pix);
+  VP8LBackwardRefsAlloc(&refs_lz77, num_pix);
+  VP8LInitBackwardRefs(best);
+  if (refs_rle.refs == NULL || refs_lz77.refs == NULL) {
+ Error1:
+    VP8LClearBackwardRefs(&refs_rle);
+    VP8LClearBackwardRefs(&refs_lz77);
+    goto End;
+  }
+
+  if (!BackwardReferencesHashChain(width, height, argb, cache_bits, quality,
+                                   &refs_lz77)) {
+    goto End;
+  }
+  // Backward Reference using RLE only.
+  BackwardReferencesRle(width, height, argb, &refs_rle);
+
+  {
+    double bit_cost_lz77, bit_cost_rle;
+    VP8LHistogram* const histo = (VP8LHistogram*)malloc(sizeof(*histo));
+    if (histo == NULL) goto Error1;
+    // Evaluate lz77 coding
+    VP8LHistogramCreate(histo, &refs_lz77, cache_bits);
+    bit_cost_lz77 = VP8LHistogramEstimateBits(histo);
+    // Evaluate RLE coding
+    VP8LHistogramCreate(histo, &refs_rle, cache_bits);
+    bit_cost_rle = VP8LHistogramEstimateBits(histo);
+    // Decide if LZ77 is useful.
+    lz77_is_useful = (bit_cost_lz77 < bit_cost_rle);
+    free(histo);
+  }
+
+  // Choose appropriate backward reference.
+  if (lz77_is_useful) {
+    // TraceBackwards is costly. Don't execute it at lower quality.
+    const int try_lz77_trace_backwards = (quality >= 25);
+    *best = refs_lz77;   // default guess: lz77 is better
+    VP8LClearBackwardRefs(&refs_rle);
+    if (try_lz77_trace_backwards) {
+      // Set recursion level for large images using a color cache.
+      const int recursion_level =
+          (num_pix < 320 * 200) && (cache_bits > 0) ? 1 : 0;
+      VP8LBackwardRefs refs_trace;
+      if (!VP8LBackwardRefsAlloc(&refs_trace, num_pix)) {
+        goto End;
+      }
+      if (BackwardReferencesTraceBackwards(width, height, recursion_level, argb,
+                                           quality, cache_bits, &refs_trace)) {
+        VP8LClearBackwardRefs(&refs_lz77);
+        *best = refs_trace;
+      }
+    }
+  } else {
+    VP8LClearBackwardRefs(&refs_lz77);
+    *best = refs_rle;
+  }
+
+  if (use_2d_locality) BackwardReferences2DLocality(width, best);
+
+  ok = 1;
+
+ End:
+  if (!ok) {
+    VP8LClearBackwardRefs(best);
+  }
+  return ok;
+}
+
+// Returns 1 on success.
+static int ComputeCacheHistogram(const uint32_t* const argb,
+                                 int xsize, int ysize,
+                                 const VP8LBackwardRefs* const refs,
+                                 int cache_bits,
+                                 VP8LHistogram* const histo) {
+  int pixel_index = 0;
+  int i;
+  uint32_t k;
+  VP8LColorCache hashers;
+  const int use_color_cache = (cache_bits > 0);
+  int cc_init = 0;
+
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) return 0;
+  }
+
+  for (i = 0; i < refs->size; ++i) {
+    const PixOrCopy* const v = &refs->refs[i];
+    if (PixOrCopyIsLiteral(v)) {
+      if (use_color_cache &&
+          VP8LColorCacheContains(&hashers, argb[pixel_index])) {
+        // push pixel as a cache index
+        const int ix = VP8LColorCacheGetIndex(&hashers, argb[pixel_index]);
+        const PixOrCopy token = PixOrCopyCreateCacheIdx(ix);
+        VP8LHistogramAddSinglePixOrCopy(histo, &token);
+      } else {
+        VP8LHistogramAddSinglePixOrCopy(histo, v);
+      }
+    } else {
+      VP8LHistogramAddSinglePixOrCopy(histo, v);
+    }
+    if (use_color_cache) {
+      for (k = 0; k < PixOrCopyLength(v); ++k) {
+        VP8LColorCacheInsert(&hashers, argb[pixel_index + k]);
+      }
+    }
+    pixel_index += PixOrCopyLength(v);
+  }
+  assert(pixel_index == xsize * ysize);
+  (void)xsize;  // xsize is not used in non-debug compilations otherwise.
+  (void)ysize;  // ysize is not used in non-debug compilations otherwise.
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  return 1;
+}
+
+// Returns how many bits are to be used for a color cache.
+int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
+                                      int xsize, int ysize,
+                                      int* const best_cache_bits) {
+  int ok = 0;
+  int cache_bits;
+  double lowest_entropy = 1e99;
+  VP8LBackwardRefs refs;
+  static const double kSmallPenaltyForLargeCache = 4.0;
+  static const int quality = 30;
+  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize) ||
+      !BackwardReferencesHashChain(xsize, ysize, argb, 0, quality, &refs)) {
+    goto Error;
+  }
+  for (cache_bits = 0; cache_bits <= MAX_COLOR_CACHE_BITS; ++cache_bits) {
+    double cur_entropy;
+    VP8LHistogram histo;
+    VP8LHistogramInit(&histo, cache_bits);
+    ComputeCacheHistogram(argb, xsize, ysize, &refs, cache_bits, &histo);
+    cur_entropy = VP8LHistogramEstimateBits(&histo) +
+        kSmallPenaltyForLargeCache * cache_bits;
+    if (cache_bits == 0 || cur_entropy < lowest_entropy) {
+      *best_cache_bits = cache_bits;
+      lowest_entropy = cur_entropy;
+    }
+  }
+  ok = 1;
+ Error:
+  VP8LClearBackwardRefs(&refs);
+  return ok;
+}
diff --git a/src/3rdparty/libwebp/src/enc/backward_references.h b/src/3rdparty/libwebp/src/enc/backward_references.h
new file mode 100644
index 0000000..e1c75f0
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/backward_references.h
@@ -0,0 +1,152 @@
+// 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_H_
+#define WEBP_ENC_BACKWARD_REFERENCES_H_
+
+#include <assert.h>
+#include <stdlib.h>
+#include "../webp/types.h"
+#include "../webp/format_constants.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The spec allows 11, we use 9 bits to reduce memory consumption in encoding.
+// Having 9 instead of 11 only removes about 0.25 % of compression density.
+#define MAX_COLOR_CACHE_BITS 9
+
+// Max ever number of codes we'll use:
+#define PIX_OR_COPY_CODES_MAX \
+    (NUM_LITERAL_CODES + NUM_LENGTH_CODES + (1 << MAX_COLOR_CACHE_BITS))
+
+// -----------------------------------------------------------------------------
+// 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 PixOrCopyArgb(const PixOrCopy* const p) {
+  assert(p->mode == kLiteral);
+  return p->argb_or_distance;
+}
+
+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;
+}
+
+// -----------------------------------------------------------------------------
+// VP8LBackwardRefs
+
+typedef struct {
+  PixOrCopy* refs;
+  int size;      // currently used
+  int max_size;  // maximum capacity
+} VP8LBackwardRefs;
+
+// Initialize the object. Must be called first. 'refs' can be NULL.
+void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs);
+
+// Release memory and re-initialize the object. 'refs' can be NULL.
+void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs);
+
+// Allocate 'max_size' references. Returns false in case of memory error.
+int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size);
+
+// -----------------------------------------------------------------------------
+// Main entry points
+
+// Evaluates best possible backward references for specified quality.
+// Further optimize for 2D locality if use_2d_locality flag is set.
+int VP8LGetBackwardReferences(int width, int height,
+                              const uint32_t* const argb,
+                              int quality, int cache_bits, int use_2d_locality,
+                              VP8LBackwardRefs* const best);
+
+// Produce an estimate for a good color cache size for the image.
+int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
+                                      int xsize, int ysize,
+                                      int* const best_cache_bits);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // WEBP_ENC_BACKWARD_REFERENCES_H_
diff --git a/src/3rdparty/libwebp/src/enc/config.c b/src/3rdparty/libwebp/src/enc/config.c
new file mode 100644
index 0000000..af7f0b0
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/config.c
@@ -0,0 +1,140 @@
+// 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)
+
+#include "../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->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->image_hint = WEBP_HINT_DEFAULT;
+  config->emulate_jpeg_size = 0;
+  config->thread_level = 0;
+  config->low_memory = 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->show_compressed < 0 || config->show_compressed > 1)
+    return 0;
+  if (config->preprocessing < 0 || config->preprocessing > 3)
+    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->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;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/src/3rdparty/libwebp/src/enc/cost.c b/src/3rdparty/libwebp/src/enc/cost.c
new file mode 100644
index 0000000..09699f8
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/cost.c
@@ -0,0 +1,489 @@
+// 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 "./cost.h"
+
+//------------------------------------------------------------------------------
+// Boolean-cost cost table
+
+const uint16_t VP8EntropyCost[256] = {
+  1792, 1792, 1792, 1536, 1536, 1408, 1366, 1280, 1280, 1216,
+  1178, 1152, 1110, 1076, 1061, 1024, 1024,  992,  968,  951,
+   939,  911,  896,  878,  871,  854,  838,  820,  811,  794,
+   786,  768,  768,  752,  740,  732,  720,  709,  704,  690,
+   683,  672,  666,  655,  647,  640,  631,  622,  615,  607,
+   598,  592,  586,  576,  572,  564,  559,  555,  547,  541,
+   534,  528,  522,  512,  512,  504,  500,  494,  488,  483,
+   477,  473,  467,  461,  458,  452,  448,  443,  438,  434,
+   427,  424,  419,  415,  410,  406,  403,  399,  394,  390,
+   384,  384,  377,  374,  370,  366,  362,  359,  355,  351,
+   347,  342,  342,  336,  333,  330,  326,  323,  320,  316,
+   312,  308,  305,  302,  299,  296,  293,  288,  287,  283,
+   280,  277,  274,  272,  268,  266,  262,  256,  256,  256,
+   251,  248,  245,  242,  240,  237,  234,  232,  228,  226,
+   223,  221,  218,  216,  214,  211,  208,  205,  203,  201,
+   198,  196,  192,  191,  188,  187,  183,  181,  179,  176,
+   175,  171,  171,  168,  165,  163,  160,  159,  156,  154,
+   152,  150,  148,  146,  144,  142,  139,  138,  135,  133,
+   131,  128,  128,  125,  123,  121,  119,  117,  115,  113,
+   111,  110,  107,  105,  103,  102,  100,   98,   96,   94,
+    92,   91,   89,   86,   86,   83,   82,   80,   77,   76,
+    74,   73,   71,   69,   67,   66,   64,   63,   61,   59,
+    57,   55,   54,   52,   51,   49,   47,   46,   44,   43,
+    41,   40,   38,   36,   35,   33,   32,   30,   29,   27,
+    25,   24,   22,   21,   19,   18,   16,   15,   13,   12,
+    10,    9,    7,    6,    4,    3
+};
+
+//------------------------------------------------------------------------------
+// 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}
+};
+
+// fixed costs for coding levels, deduce from the coding tree.
+// This is only the part that doesn't depend on the probability state.
+const uint16_t VP8LevelFixedCosts[MAX_LEVEL + 1] = {
+     0,  256,  256,  256,  256,  432,  618,  630,
+   731,  640,  640,  828,  901,  948, 1021, 1101,
+  1174, 1221, 1294, 1042, 1085, 1115, 1158, 1202,
+  1245, 1275, 1318, 1337, 1380, 1410, 1453, 1497,
+  1540, 1570, 1613, 1280, 1295, 1317, 1332, 1358,
+  1373, 1395, 1410, 1454, 1469, 1491, 1506, 1532,
+  1547, 1569, 1584, 1601, 1616, 1638, 1653, 1679,
+  1694, 1716, 1731, 1775, 1790, 1812, 1827, 1853,
+  1868, 1890, 1905, 1727, 1733, 1742, 1748, 1759,
+  1765, 1774, 1780, 1800, 1806, 1815, 1821, 1832,
+  1838, 1847, 1853, 1878, 1884, 1893, 1899, 1910,
+  1916, 1925, 1931, 1951, 1957, 1966, 1972, 1983,
+  1989, 1998, 2004, 2027, 2033, 2042, 2048, 2059,
+  2065, 2074, 2080, 2100, 2106, 2115, 2121, 2132,
+  2138, 2147, 2153, 2178, 2184, 2193, 2199, 2210,
+  2216, 2225, 2231, 2251, 2257, 2266, 2272, 2283,
+  2289, 2298, 2304, 2168, 2174, 2183, 2189, 2200,
+  2206, 2215, 2221, 2241, 2247, 2256, 2262, 2273,
+  2279, 2288, 2294, 2319, 2325, 2334, 2340, 2351,
+  2357, 2366, 2372, 2392, 2398, 2407, 2413, 2424,
+  2430, 2439, 2445, 2468, 2474, 2483, 2489, 2500,
+  2506, 2515, 2521, 2541, 2547, 2556, 2562, 2573,
+  2579, 2588, 2594, 2619, 2625, 2634, 2640, 2651,
+  2657, 2666, 2672, 2692, 2698, 2707, 2713, 2724,
+  2730, 2739, 2745, 2540, 2546, 2555, 2561, 2572,
+  2578, 2587, 2593, 2613, 2619, 2628, 2634, 2645,
+  2651, 2660, 2666, 2691, 2697, 2706, 2712, 2723,
+  2729, 2738, 2744, 2764, 2770, 2779, 2785, 2796,
+  2802, 2811, 2817, 2840, 2846, 2855, 2861, 2872,
+  2878, 2887, 2893, 2913, 2919, 2928, 2934, 2945,
+  2951, 2960, 2966, 2991, 2997, 3006, 3012, 3023,
+  3029, 3038, 3044, 3064, 3070, 3079, 3085, 3096,
+  3102, 3111, 3117, 2981, 2987, 2996, 3002, 3013,
+  3019, 3028, 3034, 3054, 3060, 3069, 3075, 3086,
+  3092, 3101, 3107, 3132, 3138, 3147, 3153, 3164,
+  3170, 3179, 3185, 3205, 3211, 3220, 3226, 3237,
+  3243, 3252, 3258, 3281, 3287, 3296, 3302, 3313,
+  3319, 3328, 3334, 3354, 3360, 3369, 3375, 3386,
+  3392, 3401, 3407, 3432, 3438, 3447, 3453, 3464,
+  3470, 3479, 3485, 3505, 3511, 3520, 3526, 3537,
+  3543, 3552, 3558, 2816, 2822, 2831, 2837, 2848,
+  2854, 2863, 2869, 2889, 2895, 2904, 2910, 2921,
+  2927, 2936, 2942, 2967, 2973, 2982, 2988, 2999,
+  3005, 3014, 3020, 3040, 3046, 3055, 3061, 3072,
+  3078, 3087, 3093, 3116, 3122, 3131, 3137, 3148,
+  3154, 3163, 3169, 3189, 3195, 3204, 3210, 3221,
+  3227, 3236, 3242, 3267, 3273, 3282, 3288, 3299,
+  3305, 3314, 3320, 3340, 3346, 3355, 3361, 3372,
+  3378, 3387, 3393, 3257, 3263, 3272, 3278, 3289,
+  3295, 3304, 3310, 3330, 3336, 3345, 3351, 3362,
+  3368, 3377, 3383, 3408, 3414, 3423, 3429, 3440,
+  3446, 3455, 3461, 3481, 3487, 3496, 3502, 3513,
+  3519, 3528, 3534, 3557, 3563, 3572, 3578, 3589,
+  3595, 3604, 3610, 3630, 3636, 3645, 3651, 3662,
+  3668, 3677, 3683, 3708, 3714, 3723, 3729, 3740,
+  3746, 3755, 3761, 3781, 3787, 3796, 3802, 3813,
+  3819, 3828, 3834, 3629, 3635, 3644, 3650, 3661,
+  3667, 3676, 3682, 3702, 3708, 3717, 3723, 3734,
+  3740, 3749, 3755, 3780, 3786, 3795, 3801, 3812,
+  3818, 3827, 3833, 3853, 3859, 3868, 3874, 3885,
+  3891, 3900, 3906, 3929, 3935, 3944, 3950, 3961,
+  3967, 3976, 3982, 4002, 4008, 4017, 4023, 4034,
+  4040, 4049, 4055, 4080, 4086, 4095, 4101, 4112,
+  4118, 4127, 4133, 4153, 4159, 4168, 4174, 4185,
+  4191, 4200, 4206, 4070, 4076, 4085, 4091, 4102,
+  4108, 4117, 4123, 4143, 4149, 4158, 4164, 4175,
+  4181, 4190, 4196, 4221, 4227, 4236, 4242, 4253,
+  4259, 4268, 4274, 4294, 4300, 4309, 4315, 4326,
+  4332, 4341, 4347, 4370, 4376, 4385, 4391, 4402,
+  4408, 4417, 4423, 4443, 4449, 4458, 4464, 4475,
+  4481, 4490, 4496, 4521, 4527, 4536, 4542, 4553,
+  4559, 4568, 4574, 4594, 4600, 4609, 4615, 4626,
+  4632, 4641, 4647, 3515, 3521, 3530, 3536, 3547,
+  3553, 3562, 3568, 3588, 3594, 3603, 3609, 3620,
+  3626, 3635, 3641, 3666, 3672, 3681, 3687, 3698,
+  3704, 3713, 3719, 3739, 3745, 3754, 3760, 3771,
+  3777, 3786, 3792, 3815, 3821, 3830, 3836, 3847,
+  3853, 3862, 3868, 3888, 3894, 3903, 3909, 3920,
+  3926, 3935, 3941, 3966, 3972, 3981, 3987, 3998,
+  4004, 4013, 4019, 4039, 4045, 4054, 4060, 4071,
+  4077, 4086, 4092, 3956, 3962, 3971, 3977, 3988,
+  3994, 4003, 4009, 4029, 4035, 4044, 4050, 4061,
+  4067, 4076, 4082, 4107, 4113, 4122, 4128, 4139,
+  4145, 4154, 4160, 4180, 4186, 4195, 4201, 4212,
+  4218, 4227, 4233, 4256, 4262, 4271, 4277, 4288,
+  4294, 4303, 4309, 4329, 4335, 4344, 4350, 4361,
+  4367, 4376, 4382, 4407, 4413, 4422, 4428, 4439,
+  4445, 4454, 4460, 4480, 4486, 4495, 4501, 4512,
+  4518, 4527, 4533, 4328, 4334, 4343, 4349, 4360,
+  4366, 4375, 4381, 4401, 4407, 4416, 4422, 4433,
+  4439, 4448, 4454, 4479, 4485, 4494, 4500, 4511,
+  4517, 4526, 4532, 4552, 4558, 4567, 4573, 4584,
+  4590, 4599, 4605, 4628, 4634, 4643, 4649, 4660,
+  4666, 4675, 4681, 4701, 4707, 4716, 4722, 4733,
+  4739, 4748, 4754, 4779, 4785, 4794, 4800, 4811,
+  4817, 4826, 4832, 4852, 4858, 4867, 4873, 4884,
+  4890, 4899, 4905, 4769, 4775, 4784, 4790, 4801,
+  4807, 4816, 4822, 4842, 4848, 4857, 4863, 4874,
+  4880, 4889, 4895, 4920, 4926, 4935, 4941, 4952,
+  4958, 4967, 4973, 4993, 4999, 5008, 5014, 5025,
+  5031, 5040, 5046, 5069, 5075, 5084, 5090, 5101,
+  5107, 5116, 5122, 5142, 5148, 5157, 5163, 5174,
+  5180, 5189, 5195, 5220, 5226, 5235, 5241, 5252,
+  5258, 5267, 5273, 5293, 5299, 5308, 5314, 5325,
+  5331, 5340, 5346, 4604, 4610, 4619, 4625, 4636,
+  4642, 4651, 4657, 4677, 4683, 4692, 4698, 4709,
+  4715, 4724, 4730, 4755, 4761, 4770, 4776, 4787,
+  4793, 4802, 4808, 4828, 4834, 4843, 4849, 4860,
+  4866, 4875, 4881, 4904, 4910, 4919, 4925, 4936,
+  4942, 4951, 4957, 4977, 4983, 4992, 4998, 5009,
+  5015, 5024, 5030, 5055, 5061, 5070, 5076, 5087,
+  5093, 5102, 5108, 5128, 5134, 5143, 5149, 5160,
+  5166, 5175, 5181, 5045, 5051, 5060, 5066, 5077,
+  5083, 5092, 5098, 5118, 5124, 5133, 5139, 5150,
+  5156, 5165, 5171, 5196, 5202, 5211, 5217, 5228,
+  5234, 5243, 5249, 5269, 5275, 5284, 5290, 5301,
+  5307, 5316, 5322, 5345, 5351, 5360, 5366, 5377,
+  5383, 5392, 5398, 5418, 5424, 5433, 5439, 5450,
+  5456, 5465, 5471, 5496, 5502, 5511, 5517, 5528,
+  5534, 5543, 5549, 5569, 5575, 5584, 5590, 5601,
+  5607, 5616, 5622, 5417, 5423, 5432, 5438, 5449,
+  5455, 5464, 5470, 5490, 5496, 5505, 5511, 5522,
+  5528, 5537, 5543, 5568, 5574, 5583, 5589, 5600,
+  5606, 5615, 5621, 5641, 5647, 5656, 5662, 5673,
+  5679, 5688, 5694, 5717, 5723, 5732, 5738, 5749,
+  5755, 5764, 5770, 5790, 5796, 5805, 5811, 5822,
+  5828, 5837, 5843, 5868, 5874, 5883, 5889, 5900,
+  5906, 5915, 5921, 5941, 5947, 5956, 5962, 5973,
+  5979, 5988, 5994, 5858, 5864, 5873, 5879, 5890,
+  5896, 5905, 5911, 5931, 5937, 5946, 5952, 5963,
+  5969, 5978, 5984, 6009, 6015, 6024, 6030, 6041,
+  6047, 6056, 6062, 6082, 6088, 6097, 6103, 6114,
+  6120, 6129, 6135, 6158, 6164, 6173, 6179, 6190,
+  6196, 6205, 6211, 6231, 6237, 6246, 6252, 6263,
+  6269, 6278, 6284, 6309, 6315, 6324, 6330, 6341,
+  6347, 6356, 6362, 6382, 6388, 6397, 6403, 6414,
+  6420, 6429, 6435, 3515, 3521, 3530, 3536, 3547,
+  3553, 3562, 3568, 3588, 3594, 3603, 3609, 3620,
+  3626, 3635, 3641, 3666, 3672, 3681, 3687, 3698,
+  3704, 3713, 3719, 3739, 3745, 3754, 3760, 3771,
+  3777, 3786, 3792, 3815, 3821, 3830, 3836, 3847,
+  3853, 3862, 3868, 3888, 3894, 3903, 3909, 3920,
+  3926, 3935, 3941, 3966, 3972, 3981, 3987, 3998,
+  4004, 4013, 4019, 4039, 4045, 4054, 4060, 4071,
+  4077, 4086, 4092, 3956, 3962, 3971, 3977, 3988,
+  3994, 4003, 4009, 4029, 4035, 4044, 4050, 4061,
+  4067, 4076, 4082, 4107, 4113, 4122, 4128, 4139,
+  4145, 4154, 4160, 4180, 4186, 4195, 4201, 4212,
+  4218, 4227, 4233, 4256, 4262, 4271, 4277, 4288,
+  4294, 4303, 4309, 4329, 4335, 4344, 4350, 4361,
+  4367, 4376, 4382, 4407, 4413, 4422, 4428, 4439,
+  4445, 4454, 4460, 4480, 4486, 4495, 4501, 4512,
+  4518, 4527, 4533, 4328, 4334, 4343, 4349, 4360,
+  4366, 4375, 4381, 4401, 4407, 4416, 4422, 4433,
+  4439, 4448, 4454, 4479, 4485, 4494, 4500, 4511,
+  4517, 4526, 4532, 4552, 4558, 4567, 4573, 4584,
+  4590, 4599, 4605, 4628, 4634, 4643, 4649, 4660,
+  4666, 4675, 4681, 4701, 4707, 4716, 4722, 4733,
+  4739, 4748, 4754, 4779, 4785, 4794, 4800, 4811,
+  4817, 4826, 4832, 4852, 4858, 4867, 4873, 4884,
+  4890, 4899, 4905, 4769, 4775, 4784, 4790, 4801,
+  4807, 4816, 4822, 4842, 4848, 4857, 4863, 4874,
+  4880, 4889, 4895, 4920, 4926, 4935, 4941, 4952,
+  4958, 4967, 4973, 4993, 4999, 5008, 5014, 5025,
+  5031, 5040, 5046, 5069, 5075, 5084, 5090, 5101,
+  5107, 5116, 5122, 5142, 5148, 5157, 5163, 5174,
+  5180, 5189, 5195, 5220, 5226, 5235, 5241, 5252,
+  5258, 5267, 5273, 5293, 5299, 5308, 5314, 5325,
+  5331, 5340, 5346, 4604, 4610, 4619, 4625, 4636,
+  4642, 4651, 4657, 4677, 4683, 4692, 4698, 4709,
+  4715, 4724, 4730, 4755, 4761, 4770, 4776, 4787,
+  4793, 4802, 4808, 4828, 4834, 4843, 4849, 4860,
+  4866, 4875, 4881, 4904, 4910, 4919, 4925, 4936,
+  4942, 4951, 4957, 4977, 4983, 4992, 4998, 5009,
+  5015, 5024, 5030, 5055, 5061, 5070, 5076, 5087,
+  5093, 5102, 5108, 5128, 5134, 5143, 5149, 5160,
+  5166, 5175, 5181, 5045, 5051, 5060, 5066, 5077,
+  5083, 5092, 5098, 5118, 5124, 5133, 5139, 5150,
+  5156, 5165, 5171, 5196, 5202, 5211, 5217, 5228,
+  5234, 5243, 5249, 5269, 5275, 5284, 5290, 5301,
+  5307, 5316, 5322, 5345, 5351, 5360, 5366, 5377,
+  5383, 5392, 5398, 5418, 5424, 5433, 5439, 5450,
+  5456, 5465, 5471, 5496, 5502, 5511, 5517, 5528,
+  5534, 5543, 5549, 5569, 5575, 5584, 5590, 5601,
+  5607, 5616, 5622, 5417, 5423, 5432, 5438, 5449,
+  5455, 5464, 5470, 5490, 5496, 5505, 5511, 5522,
+  5528, 5537, 5543, 5568, 5574, 5583, 5589, 5600,
+  5606, 5615, 5621, 5641, 5647, 5656, 5662, 5673,
+  5679, 5688, 5694, 5717, 5723, 5732, 5738, 5749,
+  5755, 5764, 5770, 5790, 5796, 5805, 5811, 5822,
+  5828, 5837, 5843, 5868, 5874, 5883, 5889, 5900,
+  5906, 5915, 5921, 5941, 5947, 5956, 5962, 5973,
+  5979, 5988, 5994, 5858, 5864, 5873, 5879, 5890,
+  5896, 5905, 5911, 5931, 5937, 5946, 5952, 5963,
+  5969, 5978, 5984, 6009, 6015, 6024, 6030, 6041,
+  6047, 6056, 6062, 6082, 6088, 6097, 6103, 6114,
+  6120, 6129, 6135, 6158, 6164, 6173, 6179, 6190,
+  6196, 6205, 6211, 6231, 6237, 6246, 6252, 6263,
+  6269, 6278, 6284, 6309, 6315, 6324, 6330, 6341,
+  6347, 6356, 6362, 6382, 6388, 6397, 6403, 6414,
+  6420, 6429, 6435, 5303, 5309, 5318, 5324, 5335,
+  5341, 5350, 5356, 5376, 5382, 5391, 5397, 5408,
+  5414, 5423, 5429, 5454, 5460, 5469, 5475, 5486,
+  5492, 5501, 5507, 5527, 5533, 5542, 5548, 5559,
+  5565, 5574, 5580, 5603, 5609, 5618, 5624, 5635,
+  5641, 5650, 5656, 5676, 5682, 5691, 5697, 5708,
+  5714, 5723, 5729, 5754, 5760, 5769, 5775, 5786,
+  5792, 5801, 5807, 5827, 5833, 5842, 5848, 5859,
+  5865, 5874, 5880, 5744, 5750, 5759, 5765, 5776,
+  5782, 5791, 5797, 5817, 5823, 5832, 5838, 5849,
+  5855, 5864, 5870, 5895, 5901, 5910, 5916, 5927,
+  5933, 5942, 5948, 5968, 5974, 5983, 5989, 6000,
+  6006, 6015, 6021, 6044, 6050, 6059, 6065, 6076,
+  6082, 6091, 6097, 6117, 6123, 6132, 6138, 6149,
+  6155, 6164, 6170, 6195, 6201, 6210, 6216, 6227,
+  6233, 6242, 6248, 6268, 6274, 6283, 6289, 6300,
+  6306, 6315, 6321, 6116, 6122, 6131, 6137, 6148,
+  6154, 6163, 6169, 6189, 6195, 6204, 6210, 6221,
+  6227, 6236, 6242, 6267, 6273, 6282, 6288, 6299,
+  6305, 6314, 6320, 6340, 6346, 6355, 6361, 6372,
+  6378, 6387, 6393, 6416, 6422, 6431, 6437, 6448,
+  6454, 6463, 6469, 6489, 6495, 6504, 6510, 6521,
+  6527, 6536, 6542, 6567, 6573, 6582, 6588, 6599,
+  6605, 6614, 6620, 6640, 6646, 6655, 6661, 6672,
+  6678, 6687, 6693, 6557, 6563, 6572, 6578, 6589,
+  6595, 6604, 6610, 6630, 6636, 6645, 6651, 6662,
+  6668, 6677, 6683, 6708, 6714, 6723, 6729, 6740,
+  6746, 6755, 6761, 6781, 6787, 6796, 6802, 6813,
+  6819, 6828, 6834, 6857, 6863, 6872, 6878, 6889,
+  6895, 6904, 6910, 6930, 6936, 6945, 6951, 6962,
+  6968, 6977, 6983, 7008, 7014, 7023, 7029, 7040,
+  7046, 7055, 7061, 7081, 7087, 7096, 7102, 7113,
+  7119, 7128, 7134, 6392, 6398, 6407, 6413, 6424,
+  6430, 6439, 6445, 6465, 6471, 6480, 6486, 6497,
+  6503, 6512, 6518, 6543, 6549, 6558, 6564, 6575,
+  6581, 6590, 6596, 6616, 6622, 6631, 6637, 6648,
+  6654, 6663, 6669, 6692, 6698, 6707, 6713, 6724,
+  6730, 6739, 6745, 6765, 6771, 6780, 6786, 6797,
+  6803, 6812, 6818, 6843, 6849, 6858, 6864, 6875,
+  6881, 6890, 6896, 6916, 6922, 6931, 6937, 6948,
+  6954, 6963, 6969, 6833, 6839, 6848, 6854, 6865,
+  6871, 6880, 6886, 6906, 6912, 6921, 6927, 6938,
+  6944, 6953, 6959, 6984, 6990, 6999, 7005, 7016,
+  7022, 7031, 7037, 7057, 7063, 7072, 7078, 7089,
+  7095, 7104, 7110, 7133, 7139, 7148, 7154, 7165,
+  7171, 7180, 7186, 7206, 7212, 7221, 7227, 7238,
+  7244, 7253, 7259, 7284, 7290, 7299, 7305, 7316,
+  7322, 7331, 7337, 7357, 7363, 7372, 7378, 7389,
+  7395, 7404, 7410, 7205, 7211, 7220, 7226, 7237,
+  7243, 7252, 7258, 7278, 7284, 7293, 7299, 7310,
+  7316, 7325, 7331, 7356, 7362, 7371, 7377, 7388,
+  7394, 7403, 7409, 7429, 7435, 7444, 7450, 7461,
+  7467, 7476, 7482, 7505, 7511, 7520, 7526, 7537,
+  7543, 7552, 7558, 7578, 7584, 7593, 7599, 7610,
+  7616, 7625, 7631, 7656, 7662, 7671, 7677, 7688,
+  7694, 7703, 7709, 7729, 7735, 7744, 7750, 7761
+};
+
+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(VP8Proba* const proba) {
+  int ctype, band, ctx;
+
+  if (!proba->dirty_) return;  // nothing to do.
+
+  for (ctype = 0; ctype < NUM_TYPES; ++ctype) {
+    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 cost_base = VP8BitCost(1, p[1]);
+        int v;
+        table[0] = VP8BitCost(0, p[1]);
+        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.
+      }
+    }
+  }
+  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 } }
+};
+
+//------------------------------------------------------------------------------
+
diff --git a/src/3rdparty/libwebp/src/enc/cost.h b/src/3rdparty/libwebp/src/enc/cost.h
new file mode 100644
index 0000000..3cbad1a
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/cost.h
@@ -0,0 +1,51 @@
+// 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_H_
+#define WEBP_ENC_COST_H_
+
+#include "./vp8enci.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// approximate cost per level:
+extern const uint16_t VP8LevelFixedCosts[MAX_LEVEL + 1];
+extern const uint16_t VP8EntropyCost[256];        // 8bit fixed-point log(p)
+
+// 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(VP8Proba* 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_H_ */
diff --git a/src/3rdparty/libwebp/src/enc/filter.c b/src/3rdparty/libwebp/src/enc/filter.c
new file mode 100644
index 0000000..dd27804
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/filter.c
@@ -0,0 +1,471 @@
+// 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 "./vp8enci.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];
+}
+
+// -----------------------------------------------------------------------------
+// NOTE: clip1, tables and InitTables are repeated entries of dsp.c
+static uint8_t abs0[255 + 255 + 1];     // abs(i)
+static uint8_t abs1[255 + 255 + 1];     // abs(i)>>1
+static int8_t sclip1[1020 + 1020 + 1];  // clips [-1020, 1020] to [-128, 127]
+static int8_t sclip2[112 + 112 + 1];    // clips [-112, 112] to [-16, 15]
+static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
+
+static int tables_ok = 0;
+
+static void InitTables(void) {
+  if (!tables_ok) {
+    int i;
+    for (i = -255; i <= 255; ++i) {
+      abs0[255 + i] = (i < 0) ? -i : i;
+      abs1[255 + i] = abs0[255 + i] >> 1;
+    }
+    for (i = -1020; i <= 1020; ++i) {
+      sclip1[1020 + i] = (i < -128) ? -128 : (i > 127) ? 127 : i;
+    }
+    for (i = -112; i <= 112; ++i) {
+      sclip2[112 + i] = (i < -16) ? -16 : (i > 15) ? 15 : i;
+    }
+    for (i = -255; i <= 255 + 255; ++i) {
+      clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i;
+    }
+    tables_ok = 1;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Edge filtering functions
+
+// 4 pixels in, 2 pixels out
+static WEBP_INLINE void do_filter2(uint8_t* p, int step) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  const int a = 3 * (q0 - p0) + sclip1[1020 + p1 - q1];
+  const int a1 = sclip2[112 + ((a + 4) >> 3)];
+  const int a2 = sclip2[112 + ((a + 3) >> 3)];
+  p[-step] = clip1[255 + p0 + a2];
+  p[    0] = clip1[255 + q0 - a1];
+}
+
+// 4 pixels in, 4 pixels out
+static WEBP_INLINE void do_filter4(uint8_t* p, int step) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  const int a = 3 * (q0 - p0);
+  const int a1 = sclip2[112 + ((a + 4) >> 3)];
+  const int a2 = sclip2[112 + ((a + 3) >> 3)];
+  const int a3 = (a1 + 1) >> 1;
+  p[-2*step] = clip1[255 + p1 + a3];
+  p[-  step] = clip1[255 + p0 + a2];
+  p[      0] = clip1[255 + q0 - a1];
+  p[   step] = clip1[255 + q1 - a3];
+}
+
+// high edge-variance
+static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return (abs0[255 + p1 - p0] > thresh) || (abs0[255 + q1 - q0] > thresh);
+}
+
+static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int thresh) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return (2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) <= thresh;
+}
+
+static WEBP_INLINE int needs_filter2(const uint8_t* p,
+                                     int step, int t, int it) {
+  const int p3 = p[-4*step], p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step];
+  const int q0 = p[0], q1 = p[step], q2 = p[2*step], q3 = p[3*step];
+  if ((2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) > t)
+    return 0;
+  return abs0[255 + p3 - p2] <= it && abs0[255 + p2 - p1] <= it &&
+         abs0[255 + p1 - p0] <= it && abs0[255 + q3 - q2] <= it &&
+         abs0[255 + q2 - q1] <= it && abs0[255 + q1 - q0] <= it;
+}
+
+//------------------------------------------------------------------------------
+// Simple In-loop filtering (Paragraph 15.2)
+
+static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    if (needs_filter(p + i, stride, thresh)) {
+      do_filter2(p + i, stride);
+    }
+  }
+}
+
+static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    if (needs_filter(p + i * stride, 1, thresh)) {
+      do_filter2(p + i * stride, 1);
+    }
+  }
+}
+
+static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    SimpleVFilter16(p, stride, thresh);
+  }
+}
+
+static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    SimpleHFilter16(p, stride, thresh);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Complex In-loop filtering (Paragraph 15.3)
+
+static WEBP_INLINE void FilterLoop24(uint8_t* p,
+                                     int hstride, int vstride, int size,
+                                     int thresh, int ithresh, int hev_thresh) {
+  while (size-- > 0) {
+    if (needs_filter2(p, hstride, thresh, ithresh)) {
+      if (hev(p, hstride, hev_thresh)) {
+        do_filter2(p, hstride);
+      } else {
+        do_filter4(p, hstride);
+      }
+    }
+    p += vstride;
+  }
+}
+
+// on three inner edges
+static void VFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh);
+  }
+}
+
+static void HFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh);
+  }
+}
+
+static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
+  FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
+}
+
+static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
+  FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
+}
+
+//------------------------------------------------------------------------------
+
+void (*VP8EncVFilter16i)(uint8_t*, int, int, int, int) = VFilter16i;
+void (*VP8EncHFilter16i)(uint8_t*, int, int, int, int) = HFilter16i;
+void (*VP8EncVFilter8i)(uint8_t*, uint8_t*, int, int, int, int) = VFilter8i;
+void (*VP8EncHFilter8i)(uint8_t*, uint8_t*, int, int, int, int) = HFilter8i;
+
+void (*VP8EncSimpleVFilter16i)(uint8_t*, int, int) = SimpleVFilter16i;
+void (*VP8EncSimpleHFilter16i)(uint8_t*, int, int) = SimpleHFilter16i;
+
+//------------------------------------------------------------------------------
+// Paragraph 15.4: compute the inner-edge filtering strength
+
+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;
+  uint8_t* const u_dst = it->yuv_out2_ + U_OFF;
+  uint8_t* const v_dst = it->yuv_out2_ + V_OFF;
+
+  // copy current block to yuv_out2_
+  memcpy(y_dst, it->yuv_out_, YUV_SIZE * sizeof(uint8_t));
+
+  if (enc->filter_hdr_.simple_ == 1) {   // simple
+    VP8EncSimpleHFilter16i(y_dst, BPS, limit);
+    VP8EncSimpleVFilter16i(y_dst, BPS, limit);
+  } else {    // complex
+    const int hev_thresh = (level >= 40) ? 2 : (level >= 15) ? 1 : 0;
+    VP8EncHFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8EncHFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+    VP8EncVFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8EncVFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+  }
+}
+
+//------------------------------------------------------------------------------
+// SSIM metric
+
+enum { KERNEL = 3 };
+static const double kMinValue = 1.e-10;  // minimal threshold
+
+void VP8SSIMAddStats(const DistoStats* const src, DistoStats* const dst) {
+  dst->w   += src->w;
+  dst->xm  += src->xm;
+  dst->ym  += src->ym;
+  dst->xxm += src->xxm;
+  dst->xym += src->xym;
+  dst->yym += src->yym;
+}
+
+static void VP8SSIMAccumulate(const uint8_t* src1, int stride1,
+                              const uint8_t* src2, int stride2,
+                              int xo, int yo, int W, int H,
+                              DistoStats* const stats) {
+  const int ymin = (yo - KERNEL < 0) ? 0 : yo - KERNEL;
+  const int ymax = (yo + KERNEL > H - 1) ? H - 1 : yo + KERNEL;
+  const int xmin = (xo - KERNEL < 0) ? 0 : xo - KERNEL;
+  const int xmax = (xo + KERNEL > W - 1) ? W - 1 : xo + KERNEL;
+  int x, y;
+  src1 += ymin * stride1;
+  src2 += ymin * stride2;
+  for (y = ymin; y <= ymax; ++y, src1 += stride1, src2 += stride2) {
+    for (x = xmin; x <= xmax; ++x) {
+      const int s1 = src1[x];
+      const int s2 = src2[x];
+      stats->w   += 1;
+      stats->xm  += s1;
+      stats->ym  += s2;
+      stats->xxm += s1 * s1;
+      stats->xym += s1 * s2;
+      stats->yym += s2 * s2;
+    }
+  }
+}
+
+double VP8SSIMGet(const DistoStats* const stats) {
+  const double xmxm = stats->xm * stats->xm;
+  const double ymym = stats->ym * stats->ym;
+  const double xmym = stats->xm * stats->ym;
+  const double w2 = stats->w * stats->w;
+  double sxx = stats->xxm * stats->w - xmxm;
+  double syy = stats->yym * stats->w - ymym;
+  double sxy = stats->xym * stats->w - xmym;
+  double C1, C2;
+  double fnum;
+  double fden;
+  // small errors are possible, due to rounding. Clamp to zero.
+  if (sxx < 0.) sxx = 0.;
+  if (syy < 0.) syy = 0.;
+  C1 = 6.5025 * w2;
+  C2 = 58.5225 * w2;
+  fnum = (2 * xmym + C1) * (2 * sxy + C2);
+  fden = (xmxm + ymym + C1) * (sxx + syy + C2);
+  return (fden != 0.) ? fnum / fden : kMinValue;
+}
+
+double VP8SSIMGetSquaredError(const DistoStats* const s) {
+  if (s->w > 0.) {
+    const double iw2 = 1. / (s->w * s->w);
+    const double sxx = s->xxm * s->w - s->xm * s->xm;
+    const double syy = s->yym * s->w - s->ym * s->ym;
+    const double sxy = s->xym * s->w - s->xm * s->ym;
+    const double SSE = iw2 * (sxx + syy - 2. * sxy);
+    if (SSE > kMinValue) return SSE;
+  }
+  return kMinValue;
+}
+
+void VP8SSIMAccumulatePlane(const uint8_t* src1, int stride1,
+                            const uint8_t* src2, int stride2,
+                            int W, int H, DistoStats* const stats) {
+  int x, y;
+  for (y = 0; y < H; ++y) {
+    for (x = 0; x < W; ++x) {
+      VP8SSIMAccumulate(src1, stride1, src2, stride2, x, y, W, H, stats);
+    }
+  }
+}
+
+static double GetMBSSIM(const uint8_t* yuv1, const uint8_t* yuv2) {
+  int x, y;
+  DistoStats s = { .0, .0, .0, .0, .0, .0 };
+
+  // compute SSIM in a 10 x 10 window
+  for (x = 3; x < 13; x++) {
+    for (y = 3; y < 13; y++) {
+      VP8SSIMAccumulate(yuv1 + Y_OFF, BPS, yuv2 + Y_OFF, BPS, x, y, 16, 16, &s);
+    }
+  }
+  for (x = 1; x < 7; x++) {
+    for (y = 1; y < 7; y++) {
+      VP8SSIMAccumulate(yuv1 + U_OFF, BPS, yuv2 + U_OFF, BPS, x, y, 8, 8, &s);
+      VP8SSIMAccumulate(yuv1 + V_OFF, BPS, yuv2 + V_OFF, BPS, x, y, 8, 8, &s);
+    }
+  }
+  return VP8SSIMGet(&s);
+}
+
+//------------------------------------------------------------------------------
+// Exposed APIs: Encoder should call the following 3 functions to adjust
+// loop filter strength
+
+void VP8InitFilter(VP8EncIterator* const it) {
+  if (it->lf_stats_ != NULL) {
+    int s, i;
+    InitTables();
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      for (i = 0; i < MAX_LF_LEVELS; i++) {
+        (*it->lf_stats_)[s][i] = 0;
+      }
+    }
+  }
+}
+
+void VP8StoreFilterStats(VP8EncIterator* const it) {
+  int d;
+  VP8Encoder* const enc = it->enc_;
+  const int s = it->mb_->segment_;
+  const int level0 = enc->dqm_[s].fstrength_;  // TODO: ref_lf_delta[]
+
+  // 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_);
+  }
+}
+
+void VP8AdjustFilterStrength(VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  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;
+    }
+  } else 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/src/3rdparty/libwebp/src/enc/frame.c b/src/3rdparty/libwebp/src/enc/frame.c
new file mode 100644
index 0000000..2582244
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/frame.c
@@ -0,0 +1,1068 @@
+// 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 <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+
+#include "./vp8enci.h"
+#include "./cost.h"
+#include "../webp/format_constants.h"  // RIFF constants
+
+#define SEGMENT_VISU 0
+#define DEBUG_SEARCH 0    // useful to track search convergence
+
+// On-the-fly info about the current set of residuals. Handy to avoid
+// passing zillions of params.
+typedef struct {
+  int first;
+  int last;
+  const int16_t* coeffs;
+
+  int coeff_type;
+  ProbaArray* prob;
+  StatsArray* stats;
+  CostArray*  cost;
+} VP8Residual;
+
+//------------------------------------------------------------------------------
+// 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)
+
+typedef struct {  // struct for organizing convergence in either size or PSNR
+  int is_first;
+  float dq;
+  float q, last_q;
+  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->q = s->last_q = enc->config_->quality;
+  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 Clamp(float v, float min, float max) {
+  return (v < min) ? min : (v > max) ? max : v;
+}
+
+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, 0.f, 100.f);
+  return s->q;
+}
+
+//------------------------------------------------------------------------------
+// Tables for level coding
+
+const uint8_t VP8EncBands[16 + 1] = {
+  0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
+  0  // sentinel
+};
+
+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) {
+  VP8Proba* 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) {
+  VP8Proba* 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;
+}
+
+//------------------------------------------------------------------------------
+// Recording of token probabilities.
+
+static void ResetTokenStats(VP8Encoder* const enc) {
+  VP8Proba* const proba = &enc->proba_;
+  memset(proba->stats_, 0, sizeof(proba->stats_));
+}
+
+// Record proba context used
+static int Record(int bit, proba_t* const stats) {
+  proba_t p = *stats;
+  if (p >= 0xffff0000u) {               // an overflow is inbound.
+    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;
+}
+
+// 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.
+static int RecordCoeffs(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) {
+    Record(0, s + 0);
+    return 0;
+  }
+  while (n <= res->last) {
+    int v;
+    Record(1, s + 0);  // order of record doesn't matter
+    while ((v = res->coeffs[n++]) == 0) {
+      Record(0, s + 1);
+      s = res->stats[VP8EncBands[n]][0];
+    }
+    Record(1, s + 1);
+    if (!Record(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 (!Record(v > 4, s + 3)) {
+        if (Record(v != 2, s + 4))
+          Record(v == 4, s + 5);
+      } else if (!Record(v > 10, s + 6)) {
+        Record(v > 6, s + 7);
+      } else if (!Record((v >= 3 + (8 << 2)), s + 8)) {
+        Record((v >= 3 + (8 << 1)), s + 9);
+      } else {
+        Record((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) Record(!!(bits & mask), s + 3 + i);
+        }
+      }
+#endif
+      s = res->stats[VP8EncBands[n]][2];
+    }
+  }
+  if (n < 16) Record(0, s + 0);
+  return 1;
+}
+
+// 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 int FinalizeTokenProbas(VP8Proba* 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 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 (enc->pic_->stats != NULL) {
+    for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
+      enc->pic_->stats->segment_size[n] = p[n];
+    }
+  }
+  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);
+    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;
+  }
+}
+
+//------------------------------------------------------------------------------
+// helper functions for residuals struct VP8Residual.
+
+static void InitResidual(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->cost  = enc->proba_.level_cost_[coeff_type];
+  res->first = first;
+}
+
+static void SetResidualCoeffs(const int16_t* const coeffs,
+                              VP8Residual* const res) {
+  int n;
+  res->last = -1;
+  for (n = 15; n >= res->first; --n) {
+    if (coeffs[n]) {
+      res->last = n;
+      break;
+    }
+  }
+  res->coeffs = coeffs;
+}
+
+//------------------------------------------------------------------------------
+// Mode costs
+
+static int GetResidualCost(int ctx0, const VP8Residual* const res) {
+  int n = res->first;
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  int p0 = res->prob[n][ctx0][0];
+  const uint16_t* t = res->cost[n][ctx0];
+  int cost;
+
+  if (res->last < 0) {
+    return VP8BitCost(0, p0);
+  }
+  cost = VP8BitCost(1, p0);
+  for (; n < res->last; ++n) {
+    const int v = abs(res->coeffs[n]);
+    const int b = VP8EncBands[n + 1];
+    const int ctx = (v >= 2) ? 2 : v;
+    cost += VP8LevelCost(t, v);
+    t = res->cost[b][ctx];
+    // the masking trick is faster than "if (v) cost += ..." with clang
+    cost += (v ? ~0U : 0) & VP8BitCost(1, res->prob[b][ctx][0]);
+  }
+  // Last coefficient is always non-zero
+  {
+    const int v = abs(res->coeffs[n]);
+    assert(v != 0);
+    cost += VP8LevelCost(t, v);
+    if (n < 15) {
+      const int b = VP8EncBands[n + 1];
+      const int ctx = (v == 1) ? 1 : 2;
+      const int last_p0 = res->prob[b][ctx][0];
+      cost += VP8BitCost(0, last_p0);
+    }
+  }
+  return cost;
+}
+
+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;
+
+  InitResidual(0, 3, enc, &res);
+  ctx = it->top_nz_[x] + it->left_nz_[y];
+  SetResidualCoeffs(levels, &res);
+  R += GetResidualCost(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
+  InitResidual(0, 1, enc, &res);
+  SetResidualCoeffs(rd->y_dc_levels, &res);
+  R += GetResidualCost(it->top_nz_[8] + it->left_nz_[8], &res);
+
+  // AC
+  InitResidual(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];
+      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      R += GetResidualCost(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
+
+  InitResidual(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];
+        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        R += GetResidualCost(ctx, &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = (res.last >= 0);
+      }
+    }
+  }
+  return R;
+}
+
+//------------------------------------------------------------------------------
+// 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) {
+    InitResidual(0, 1, enc, &res);
+    SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+      PutCoeffs(bw, it->top_nz_[8] + it->left_nz_[8], &res);
+    InitResidual(1, 0, enc, &res);
+  } else {
+    InitResidual(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];
+      SetResidualCoeffs(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
+  InitResidual(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];
+        SetResidualCoeffs(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
+    InitResidual(0, 1, enc, &res);
+    SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+      RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res);
+    InitResidual(1, 0, enc, &res);
+  } else {
+    InitResidual(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];
+      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] = RecordCoeffs(ctx, &res);
+    }
+  }
+
+  // U/V
+  InitResidual(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];
+        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
+            RecordCoeffs(ctx, &res);
+      }
+    }
+  }
+
+  VP8IteratorBytesToNz(it);
+}
+
+//------------------------------------------------------------------------------
+// Token buffer
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+static void 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];
+    InitResidual(0, 1, enc, &res);
+    SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+        VP8RecordCoeffTokens(ctx, 1,
+                             res.first, res.last, res.coeffs, tokens);
+    RecordCoeffs(ctx, &res);
+    InitResidual(1, 0, enc, &res);
+  } else {
+    InitResidual(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];
+      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] =
+          VP8RecordCoeffTokens(ctx, res.coeff_type,
+                               res.first, res.last, res.coeffs, tokens);
+      RecordCoeffs(ctx, &res);
+    }
+  }
+
+  // U/V
+  InitResidual(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];
+        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
+            VP8RecordCoeffTokens(ctx, 2,
+                                 res.first, res.last, res.coeffs, tokens);
+        RecordCoeffs(ctx, &res);
+      }
+    }
+  }
+  VP8IteratorBytesToNz(it);
+}
+
+#endif    // !DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+// ExtraInfo map / Debug function
+
+#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, out + Y_OFF);
+  enc->sse_[1] += VP8SSE8x8(in + U_OFF, out + U_OFF);
+  enc->sse_[2] += VP8SSE8x8(in + V_OFF, out + V_OFF);
+  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, mb->segment_ * 64, 16);
+  SetBlock(it->yuv_out_ + U_OFF, it->preds_[0] * 64, 8);
+  SetBlock(it->yuv_out_ + V_OFF, mb->uv_mode_ * 64, 8);
+#endif
+}
+
+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 int 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
+  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 (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);
+        }
+      }
+    }
+    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);
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (enc->use_layer_) {
+      VP8EncCodeLayerBlock(&it);
+    }
+#endif
+    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;
+  const int do_search = enc->do_search_;
+  VP8EncIterator it;
+  VP8Proba* 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;
+    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);
+      RecordTokens(&it, &info, &enc->tokens_);
+      size_p0 += info.H;
+      distortion += info.D;
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+      if (enc->use_layer_) {
+        VP8EncCodeLayerBlock(&it);
+      }
+#endif
+      if (is_last_pass) {
+        StoreSideInfo(&it);
+        VP8StoreFilterStats(&it);
+        VP8IteratorExport(&it);
+        ok = VP8IteratorProgress(&it, 20);
+      }
+      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\n",
+           num_pass_left, stats.last_value, stats.value,
+           stats.last_q, stats.q, stats.dq);
+#endif
+    if (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;   // 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_ + 20, &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/src/3rdparty/libwebp/src/enc/histogram.c b/src/3rdparty/libwebp/src/enc/histogram.c
new file mode 100644
index 0000000..abd253b
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/histogram.c
@@ -0,0 +1,512 @@
+// 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 "config.h"
+#endif
+
+#include <math.h>
+#include <stdio.h>
+
+#include "./backward_references.h"
+#include "./histogram.h"
+#include "../dsp/lossless.h"
+#include "../utils/utils.h"
+
+static void HistogramClear(VP8LHistogram* const p) {
+  memset(p->literal_, 0, sizeof(p->literal_));
+  memset(p->red_, 0, sizeof(p->red_));
+  memset(p->blue_, 0, sizeof(p->blue_));
+  memset(p->alpha_, 0, sizeof(p->alpha_));
+  memset(p->distance_, 0, sizeof(p->distance_));
+  p->bit_cost_ = 0;
+}
+
+void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
+                            VP8LHistogram* const histo) {
+  int i;
+  for (i = 0; i < refs->size; ++i) {
+    VP8LHistogramAddSinglePixOrCopy(histo, &refs->refs[i]);
+  }
+}
+
+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) {
+  p->palette_code_bits_ = palette_code_bits;
+  HistogramClear(p);
+}
+
+VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
+  int i;
+  VP8LHistogramSet* set;
+  VP8LHistogram* bulk;
+  const uint64_t total_size = sizeof(*set)
+                            + (uint64_t)size * sizeof(*set->histograms)
+                            + (uint64_t)size * sizeof(**set->histograms);
+  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;
+  memory += size * sizeof(*set->histograms);
+  bulk = (VP8LHistogram*)memory;
+  set->max_size = size;
+  set->size = size;
+  for (i = 0; i < size; ++i) {
+    set->histograms[i] = bulk + i;
+    VP8LHistogramInit(set->histograms[i], cache_bits);
+  }
+  return set;
+}
+
+// -----------------------------------------------------------------------------
+
+void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
+                                     const PixOrCopy* const v) {
+  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)) {
+    int literal_ix = 256 + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
+    ++histo->literal_[literal_ix];
+  } else {
+    int code, extra_bits;
+    VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
+    ++histo->literal_[256 + code];
+    VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
+    ++histo->distance_[code];
+  }
+}
+
+static double BitsEntropy(const int* const array, int n) {
+  double retval = 0.;
+  int sum = 0;
+  int nonzeros = 0;
+  int max_val = 0;
+  int i;
+  double mix;
+  for (i = 0; i < n; ++i) {
+    if (array[i] != 0) {
+      sum += array[i];
+      ++nonzeros;
+      retval -= VP8LFastSLog2(array[i]);
+      if (max_val < array[i]) {
+        max_val = array[i];
+      }
+    }
+  }
+  retval += VP8LFastSLog2(sum);
+
+  if (nonzeros < 5) {
+    if (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 (nonzeros == 2) {
+      return 0.99 * sum + 0.01 * retval;
+    }
+    // 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 (nonzeros == 3) {
+      mix = 0.95;
+    } else {
+      mix = 0.7;  // nonzeros == 4.
+    }
+  } else {
+    mix = 0.627;
+  }
+
+  {
+    double min_limit = 2 * sum - max_val;
+    min_limit = mix * min_limit + (1.0 - mix) * retval;
+    return (retval < min_limit) ? min_limit : retval;
+  }
+}
+
+// Returns the cost encode the rle-encoded entropy code.
+// The constants in this function are experimental.
+static double HuffmanCost(const int* const population, int length) {
+  // Small bias because Huffman code length is typically not stored in
+  // full length.
+  static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3;
+  static const double kSmallBias = 9.1;
+  double retval = kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
+  int streak = 0;
+  int i = 0;
+  for (; i < length - 1; ++i) {
+    ++streak;
+    if (population[i] == population[i + 1]) {
+      continue;
+    }
+ last_streak_hack:
+    // population[i] points now to the symbol in the streak of same values.
+    if (streak > 3) {
+      if (population[i] == 0) {
+        retval += 1.5625 + 0.234375 * streak;
+      } else {
+        retval += 2.578125 + 0.703125 * streak;
+      }
+    } else {
+      if (population[i] == 0) {
+        retval += 1.796875 * streak;
+      } else {
+        retval += 3.28125 * streak;
+      }
+    }
+    streak = 0;
+  }
+  if (i == length - 1) {
+    ++streak;
+    goto last_streak_hack;
+  }
+  return retval;
+}
+
+static double PopulationCost(const int* const population, int length) {
+  return BitsEntropy(population, length) + HuffmanCost(population, length);
+}
+
+static double ExtraCost(const int* const population, int length) {
+  int i;
+  double cost = 0.;
+  for (i = 2; i < length - 2; ++i) cost += (i >> 1) * population[i + 2];
+  return cost;
+}
+
+// Estimates the Entropy + Huffman + other block overhead size cost.
+double VP8LHistogramEstimateBits(const VP8LHistogram* const p) {
+  return PopulationCost(p->literal_, VP8LHistogramNumCodes(p))
+       + PopulationCost(p->red_, 256)
+       + PopulationCost(p->blue_, 256)
+       + PopulationCost(p->alpha_, 256)
+       + PopulationCost(p->distance_, NUM_DISTANCE_CODES)
+       + ExtraCost(p->literal_ + 256, NUM_LENGTH_CODES)
+       + ExtraCost(p->distance_, NUM_DISTANCE_CODES);
+}
+
+double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p) {
+  return BitsEntropy(p->literal_, VP8LHistogramNumCodes(p))
+       + BitsEntropy(p->red_, 256)
+       + BitsEntropy(p->blue_, 256)
+       + BitsEntropy(p->alpha_, 256)
+       + BitsEntropy(p->distance_, NUM_DISTANCE_CODES)
+       + ExtraCost(p->literal_ + 256, NUM_LENGTH_CODES)
+       + ExtraCost(p->distance_, NUM_DISTANCE_CODES);
+}
+
+// -----------------------------------------------------------------------------
+// Various histogram combine/cost-eval functions
+
+// Adds 'in' histogram to 'out'
+static void HistogramAdd(const VP8LHistogram* const in,
+                         VP8LHistogram* const out) {
+  int i;
+  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
+    out->literal_[i] += in->literal_[i];
+  }
+  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
+    out->distance_[i] += in->distance_[i];
+  }
+  for (i = 0; i < 256; ++i) {
+    out->red_[i] += in->red_[i];
+    out->blue_[i] += in->blue_[i];
+    out->alpha_[i] += in->alpha_[i];
+  }
+}
+
+// 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 double HistogramAddEval(const VP8LHistogram* const a,
+                               const VP8LHistogram* const b,
+                               VP8LHistogram* const out,
+                               double cost_threshold) {
+  double cost = 0;
+  const double sum_cost = a->bit_cost_ + b->bit_cost_;
+  int i;
+
+  cost_threshold += sum_cost;
+
+  // palette_code_bits_ is part of the cost evaluation for literal_.
+  // TODO(skal): remove/simplify this palette_code_bits_?
+  out->palette_code_bits_ =
+      (a->palette_code_bits_ > b->palette_code_bits_) ? a->palette_code_bits_ :
+                                                        b->palette_code_bits_;
+  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
+    out->literal_[i] = a->literal_[i] + b->literal_[i];
+  }
+  cost += PopulationCost(out->literal_, VP8LHistogramNumCodes(out));
+  cost += ExtraCost(out->literal_ + 256, NUM_LENGTH_CODES);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) out->red_[i] = a->red_[i] + b->red_[i];
+  cost += PopulationCost(out->red_, 256);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) out->blue_[i] = a->blue_[i] + b->blue_[i];
+  cost += PopulationCost(out->blue_, 256);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
+    out->distance_[i] = a->distance_[i] + b->distance_[i];
+  }
+  cost += PopulationCost(out->distance_, NUM_DISTANCE_CODES);
+  cost += ExtraCost(out->distance_, NUM_DISTANCE_CODES);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) out->alpha_[i] = a->alpha_[i] + b->alpha_[i];
+  cost += PopulationCost(out->alpha_, 256);
+
+  out->bit_cost_ = cost;
+  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 double HistogramAddThresh(const VP8LHistogram* const a,
+                                 const VP8LHistogram* const b,
+                                 double cost_threshold) {
+  int tmp[PIX_OR_COPY_CODES_MAX];  // <= max storage we'll need
+  int i;
+  double cost = -a->bit_cost_;
+
+  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
+    tmp[i] = a->literal_[i] + b->literal_[i];
+  }
+  // note that the tests are ordered so that the usually largest
+  // cost shares come first.
+  cost += PopulationCost(tmp, VP8LHistogramNumCodes(a));
+  cost += ExtraCost(tmp + 256, NUM_LENGTH_CODES);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) tmp[i] = a->red_[i] + b->red_[i];
+  cost += PopulationCost(tmp, 256);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) tmp[i] = a->blue_[i] + b->blue_[i];
+  cost += PopulationCost(tmp, 256);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
+    tmp[i] = a->distance_[i] + b->distance_[i];
+  }
+  cost += PopulationCost(tmp, NUM_DISTANCE_CODES);
+  cost += ExtraCost(tmp, NUM_DISTANCE_CODES);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) tmp[i] = a->alpha_[i] + b->alpha_[i];
+  cost += PopulationCost(tmp, 256);
+
+  return cost;
+}
+
+// -----------------------------------------------------------------------------
+
+static void HistogramBuildImage(int xsize, int histo_bits,
+                                const VP8LBackwardRefs* const backward_refs,
+                                VP8LHistogramSet* const image) {
+  int i;
+  int x = 0, y = 0;
+  const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
+  VP8LHistogram** const histograms = image->histograms;
+  assert(histo_bits > 0);
+  for (i = 0; i < backward_refs->size; ++i) {
+    const PixOrCopy* const v = &backward_refs->refs[i];
+    const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
+    VP8LHistogramAddSinglePixOrCopy(histograms[ix], v);
+    x += PixOrCopyLength(v);
+    while (x >= xsize) {
+      x -= xsize;
+      ++y;
+    }
+  }
+}
+
+static uint32_t MyRand(uint32_t *seed) {
+  *seed *= 16807U;
+  if (*seed == 0) {
+    *seed = 1;
+  }
+  return *seed;
+}
+
+static int HistogramCombine(const VP8LHistogramSet* const in,
+                            VP8LHistogramSet* const out, int iter_mult,
+                            int num_pairs, int num_tries_no_success) {
+  int ok = 0;
+  int i, iter;
+  uint32_t seed = 0;
+  int tries_with_no_success = 0;
+  int out_size = in->size;
+  const int outer_iters = in->size * iter_mult;
+  const int min_cluster_size = 2;
+  VP8LHistogram* const histos = (VP8LHistogram*)malloc(2 * sizeof(*histos));
+  VP8LHistogram* cur_combo = histos + 0;    // trial merged histogram
+  VP8LHistogram* best_combo = histos + 1;   // best merged histogram so far
+  if (histos == NULL) goto End;
+
+  // Copy histograms from in[] to out[].
+  assert(in->size <= out->size);
+  for (i = 0; i < in->size; ++i) {
+    in->histograms[i]->bit_cost_ = VP8LHistogramEstimateBits(in->histograms[i]);
+    *out->histograms[i] = *in->histograms[i];
+  }
+
+  // Collapse similar histograms in 'out'.
+  for (iter = 0; iter < outer_iters && out_size >= min_cluster_size; ++iter) {
+    double best_cost_diff = 0.;
+    int best_idx1 = -1, best_idx2 = 1;
+    int j;
+    const int num_tries = (num_pairs < out_size) ? num_pairs : out_size;
+    seed += iter;
+    for (j = 0; j < num_tries; ++j) {
+      double curr_cost_diff;
+      // Choose two histograms at random and try to combine them.
+      const uint32_t idx1 = MyRand(&seed) % out_size;
+      const uint32_t tmp = (j & 7) + 1;
+      const uint32_t diff = (tmp < 3) ? tmp : MyRand(&seed) % (out_size - 1);
+      const uint32_t idx2 = (idx1 + diff + 1) % out_size;
+      if (idx1 == idx2) {
+        continue;
+      }
+      // Calculate cost reduction on combining.
+      curr_cost_diff = HistogramAddEval(out->histograms[idx1],
+                                        out->histograms[idx2],
+                                        cur_combo, best_cost_diff);
+      if (curr_cost_diff < best_cost_diff) {    // found a better pair?
+        {     // swap cur/best combo histograms
+          VP8LHistogram* const tmp_histo = cur_combo;
+          cur_combo = best_combo;
+          best_combo = tmp_histo;
+        }
+        best_cost_diff = curr_cost_diff;
+        best_idx1 = idx1;
+        best_idx2 = idx2;
+      }
+    }
+
+    if (best_idx1 >= 0) {
+      *out->histograms[best_idx1] = *best_combo;
+      // swap best_idx2 slot with last one (which is now unused)
+      --out_size;
+      if (best_idx2 != out_size) {
+        out->histograms[best_idx2] = out->histograms[out_size];
+        out->histograms[out_size] = NULL;   // just for sanity check.
+      }
+      tries_with_no_success = 0;
+    }
+    if (++tries_with_no_success >= num_tries_no_success) {
+      break;
+    }
+  }
+  out->size = out_size;
+  ok = 1;
+
+ End:
+  free(histos);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+// Histogram refinement
+
+// What is the bit cost of moving square_histogram from cur_symbol to candidate.
+static double HistogramDistance(const VP8LHistogram* const square_histogram,
+                                const VP8LHistogram* const candidate,
+                                double cost_threshold) {
+  return HistogramAddThresh(candidate, square_histogram, cost_threshold);
+}
+
+// Find the best 'out' histogram for each of the 'in' histograms.
+// Note: we assume that out[]->bit_cost_ is already up-to-date.
+static void HistogramRemap(const VP8LHistogramSet* const in,
+                           const VP8LHistogramSet* const out,
+                           uint16_t* const symbols) {
+  int i;
+  for (i = 0; i < in->size; ++i) {
+    int best_out = 0;
+    double best_bits =
+        HistogramDistance(in->histograms[i], out->histograms[0], 1.e38);
+    int k;
+    for (k = 1; k < out->size; ++k) {
+      const double cur_bits =
+          HistogramDistance(in->histograms[i], out->histograms[k], best_bits);
+      if (cur_bits < best_bits) {
+        best_bits = cur_bits;
+        best_out = k;
+      }
+    }
+    symbols[i] = best_out;
+  }
+
+  // Recompute each out based on raw and symbols.
+  for (i = 0; i < out->size; ++i) {
+    HistogramClear(out->histograms[i]);
+  }
+  for (i = 0; i < in->size; ++i) {
+    HistogramAdd(in->histograms[i], out->histograms[symbols[i]]);
+  }
+}
+
+int VP8LGetHistoImageSymbols(int xsize, int ysize,
+                             const VP8LBackwardRefs* const refs,
+                             int quality, int histo_bits, int cache_bits,
+                             VP8LHistogramSet* const image_in,
+                             uint16_t* const histogram_symbols) {
+  int ok = 0;
+  const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
+  const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
+  const int histo_image_raw_size = histo_xsize * histo_ysize;
+
+  // Heuristic params for HistogramCombine().
+  const int num_tries_no_success = 8 + (quality >> 1);
+  const int iter_mult = (quality < 27) ? 1 : 1 + ((quality - 27) >> 4);
+  const int num_pairs = (quality < 25) ? 10 : (5 * quality) >> 3;
+
+  VP8LHistogramSet* const image_out =
+      VP8LAllocateHistogramSet(histo_image_raw_size, cache_bits);
+  if (image_out == NULL) return 0;
+
+  // Build histogram image.
+  HistogramBuildImage(xsize, histo_bits, refs, image_out);
+  // Collapse similar histograms.
+  if (!HistogramCombine(image_out, image_in, iter_mult, num_pairs,
+                        num_tries_no_success)) {
+    goto Error;
+  }
+  // Find the optimal map from original histograms to the final ones.
+  HistogramRemap(image_out, image_in, histogram_symbols);
+  ok = 1;
+
+Error:
+  free(image_out);
+  return ok;
+}
diff --git a/src/3rdparty/libwebp/src/enc/histogram.h b/src/3rdparty/libwebp/src/enc/histogram.h
new file mode 100644
index 0000000..4d346a8
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/histogram.h
@@ -0,0 +1,101 @@
+// 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_H_
+#define WEBP_ENC_HISTOGRAM_H_
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "./backward_references.h"
+#include "../webp/format_constants.h"
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// A simple container for histograms of data.
+typedef struct {
+  // literal_ contains green literal, palette-code and
+  // copy-length-prefix histogram
+  int literal_[PIX_OR_COPY_CODES_MAX];
+  int red_[256];
+  int blue_[256];
+  int alpha_[256];
+  // Backward reference prefix-code histogram.
+  int distance_[NUM_DISTANCE_CODES];
+  int palette_code_bits_;
+  double bit_cost_;   // cached value of VP8LHistogramEstimateBits(this)
+} VP8LHistogram;
+
+// Collection of histograms with fixed capacity, allocated as one
+// big memory chunk. Can be destroyed by simply calling 'free()'.
+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);
+
+// Set the palette_code_bits and reset the stats.
+void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits);
+
+// Collect all the references into a histogram (without reset)
+void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
+                            VP8LHistogram* 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);
+
+// Accumulate a token 'v' into a histogram.
+void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
+                                     const PixOrCopy* const v);
+
+// Estimate how many bits the combined entropy of literals and distance
+// approximately maps to.
+double VP8LHistogramEstimateBits(const VP8LHistogram* const p);
+
+// This function estimates the cost in bits excluding the bits needed to
+// represent the entropy code itself.
+double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p);
+
+static WEBP_INLINE int VP8LHistogramNumCodes(const VP8LHistogram* const p) {
+  return 256 + NUM_LENGTH_CODES +
+      ((p->palette_code_bits_ > 0) ? (1 << p->palette_code_bits_) : 0);
+}
+
+// Builds the histogram image.
+int VP8LGetHistoImageSymbols(int xsize, int ysize,
+                             const VP8LBackwardRefs* const refs,
+                             int quality, int histogram_bits, int cache_bits,
+                             VP8LHistogramSet* const image_in,
+                             uint16_t* const histogram_symbols);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // WEBP_ENC_HISTOGRAM_H_
diff --git a/src/3rdparty/libwebp/src/enc/iterator.c b/src/3rdparty/libwebp/src/enc/iterator.c
new file mode 100644
index 0000000..e42ad00
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/iterator.c
@@ -0,0 +1,456 @@
+// 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 "./vp8enci.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;
+}
+
+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_));
+}
+
+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);
+  InitLeft(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->y_stride_  = enc->pic_->y_stride;
+  it->uv_stride_ = enc->pic_->uv_stride;
+  it->yuv_in_   = (uint8_t*)DO_ALIGN(it->yuv_mem_);
+  it->yuv_out_  = it->yuv_in_ + YUV_SIZE;
+  it->yuv_out2_ = it->yuv_out_ + YUV_SIZE;
+  it->yuv_p_    = it->yuv_out2_ + YUV_SIZE;
+  it->lf_stats_ = enc->lf_stats_;
+  it->percent0_ = enc->percent_;
+  it->y_left_ = (uint8_t*)DO_ALIGN(it->yuv_left_mem_ + 1);
+  it->u_left_ = it->y_left_ + 16 + 16;
+  it->v_left_ = it->u_left_ + 16;
+  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* 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, w, h, 16);
+  ImportBlock(usrc, pic->uv_stride, it->yuv_in_ + U_OFF, uv_w, uv_h, 8);
+  ImportBlock(vsrc, pic->uv_stride, it->yuv_in_ + V_OFF, 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;
+    const uint8_t* const usrc = it->yuv_out_ + U_OFF;
+    const uint8_t* const vsrc = it->yuv_out_ + V_OFF;
+    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;
+  const uint8_t* const uvsrc = it->yuv_out_ + U_OFF;
+  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) {
+  it->preds_ += 4;
+  it->mb_ += 1;
+  it->nz_ += 1;
+  it->y_top_ += 16;
+  it->uv_top_ += 16;
+  it->x_ += 1;
+  if (it->x_ == it->enc_->mb_w_) {
+    VP8IteratorSetRow(it, ++it->y_);
+  }
+  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/src/3rdparty/libwebp/src/enc/layer.c b/src/3rdparty/libwebp/src/enc/layer.c
new file mode 100644
index 0000000..2402362
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/layer.c
@@ -0,0 +1,44 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// Enhancement layer (for YUV444/422)
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+
+#include "./vp8enci.h"
+
+//------------------------------------------------------------------------------
+
+void VP8EncInitLayer(VP8Encoder* const enc) {
+  enc->use_layer_ = (enc->pic_->u0 != NULL);
+  enc->layer_data_size_ = 0;
+  enc->layer_data_ = NULL;
+  if (enc->use_layer_) {
+    VP8BitWriterInit(&enc->layer_bw_, enc->mb_w_ * enc->mb_h_ * 3);
+  }
+}
+
+void VP8EncCodeLayerBlock(VP8EncIterator* it) {
+  (void)it;   // remove a warning
+}
+
+int VP8EncFinishLayer(VP8Encoder* const enc) {
+  if (enc->use_layer_) {
+    enc->layer_data_ = VP8BitWriterFinish(&enc->layer_bw_);
+    enc->layer_data_size_ = VP8BitWriterSize(&enc->layer_bw_);
+  }
+  return 1;
+}
+
+void VP8EncDeleteLayer(VP8Encoder* enc) {
+  free(enc->layer_data_);
+}
+
diff --git a/src/3rdparty/libwebp/src/enc/picture.c b/src/3rdparty/libwebp/src/enc/picture.c
new file mode 100644
index 0000000..011690d
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/picture.c
@@ -0,0 +1,1324 @@
+// 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 utils: colorspace conversion, crop, ...
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <math.h>
+
+#include "./vp8enci.h"
+#include "../utils/alpha_processing.h"
+#include "../utils/random.h"
+#include "../utils/rescaler.h"
+#include "../utils/utils.h"
+#include "../dsp/dsp.h"
+#include "../dsp/yuv.h"
+
+// Uncomment to disable gamma-compression during RGB->U/V averaging
+#define USE_GAMMA_COMPRESSION
+
+#define HALVE(x) (((x) + 1) >> 1)
+#define IS_YUV_CSP(csp, YUV_CSP) (((csp) & WEBP_CSP_UV_MASK) == (YUV_CSP))
+
+static const union {
+  uint32_t argb;
+  uint8_t  bytes[4];
+} test_endian = { 0xff000000u };
+#define ALPHA_IS_LAST (test_endian.bytes[3] == 0xff)
+
+static WEBP_INLINE uint32_t MakeARGB32(int r, int g, int b) {
+  return (0xff000000u | (r << 16) | (g << 8) | b);
+}
+
+//------------------------------------------------------------------------------
+// WebPPicture
+//------------------------------------------------------------------------------
+
+int WebPPictureAlloc(WebPPicture* picture) {
+  if (picture != NULL) {
+    const WebPEncCSP uv_csp = picture->colorspace & WEBP_CSP_UV_MASK;
+    const int has_alpha = picture->colorspace & WEBP_CSP_ALPHA_BIT;
+    const int width = picture->width;
+    const int height = picture->height;
+
+    if (!picture->use_argb) {
+      const int y_stride = width;
+      const int uv_width = HALVE(width);
+      const int uv_height = HALVE(height);
+      const int uv_stride = uv_width;
+      int uv0_stride = 0;
+      int a_width, a_stride;
+      uint64_t y_size, uv_size, uv0_size, a_size, total_size;
+      uint8_t* mem;
+
+      // U/V
+      switch (uv_csp) {
+        case WEBP_YUV420:
+          break;
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+        case WEBP_YUV400:    // for now, we'll just reset the U/V samples
+          break;
+        case WEBP_YUV422:
+          uv0_stride = uv_width;
+          break;
+        case WEBP_YUV444:
+          uv0_stride = width;
+          break;
+#endif
+        default:
+          return 0;
+      }
+      uv0_size = height * uv0_stride;
+
+      // 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 + 2 * uv0_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 0;
+      }
+      // Clear previous buffer and allocate a new one.
+      WebPPictureFree(picture);   // erase previous buffer
+      mem = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*mem));
+      if (mem == NULL) return 0;
+
+      // 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;
+      picture->uv0_stride = uv0_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) {
+        picture->a = mem;
+        mem += a_size;
+      }
+      if (uv0_size) {
+        picture->u0 = mem;
+        mem += uv0_size;
+        picture->v0 = mem;
+        mem += uv0_size;
+      }
+      (void)mem;  // makes the static analyzer happy
+    } else {
+      void* memory;
+      const uint64_t argb_size = (uint64_t)width * height;
+      if (width <= 0 || height <= 0) {
+        return 0;
+      }
+      // Clear previous buffer and allocate a new one.
+      WebPPictureFree(picture);   // erase previous buffer
+      memory = WebPSafeMalloc(argb_size, sizeof(*picture->argb));
+      if (memory == NULL) return 0;
+
+      // TODO(skal): align plane to cache line?
+      picture->memory_argb_ = memory;
+      picture->argb = (uint32_t*)memory;
+      picture->argb_stride = width;
+    }
+  }
+  return 1;
+}
+
+// Remove reference to the ARGB buffer (doesn't free anything).
+static void PictureResetARGB(WebPPicture* const picture) {
+  picture->memory_argb_ = NULL;
+  picture->argb = NULL;
+  picture->argb_stride = 0;
+}
+
+// Remove reference to the YUVA buffer (doesn't free anything).
+static void PictureResetYUVA(WebPPicture* const picture) {
+  picture->memory_ = NULL;
+  picture->y = picture->u = picture->v = picture->a = NULL;
+  picture->u0 = picture->v0 = NULL;
+  picture->y_stride = picture->uv_stride = 0;
+  picture->a_stride = 0;
+  picture->uv0_stride = 0;
+}
+
+// Grab the 'specs' (writer, *opaque, width, height...) from 'src' and copy them
+// into 'dst'. Mark 'dst' as not owning any memory.
+static void WebPPictureGrabSpecs(const WebPPicture* const src,
+                                 WebPPicture* const dst) {
+  assert(src != NULL && dst != NULL);
+  *dst = *src;
+  PictureResetYUVA(dst);
+  PictureResetARGB(dst);
+}
+
+// Allocate a new argb buffer, discarding any existing one and preserving
+// the other YUV(A) buffer.
+static int PictureAllocARGB(WebPPicture* const picture) {
+  WebPPicture tmp;
+  free(picture->memory_argb_);
+  PictureResetARGB(picture);
+  picture->use_argb = 1;
+  WebPPictureGrabSpecs(picture, &tmp);
+  if (!WebPPictureAlloc(&tmp)) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  picture->memory_argb_ = tmp.memory_argb_;
+  picture->argb = tmp.argb;
+  picture->argb_stride = tmp.argb_stride;
+  return 1;
+}
+
+// Release memory owned by 'picture' (both YUV and ARGB buffers).
+void WebPPictureFree(WebPPicture* picture) {
+  if (picture != NULL) {
+    free(picture->memory_);
+    free(picture->memory_argb_);
+    PictureResetYUVA(picture);
+    PictureResetARGB(picture);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Picture copying
+
+// Not worth moving to dsp/enc.c (only used here).
+static void CopyPlane(const uint8_t* src, int src_stride,
+                      uint8_t* dst, int dst_stride, int width, int height) {
+  while (height-- > 0) {
+    memcpy(dst, src, width);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+// 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) {
+    const int is_yuv422 = IS_YUV_CSP(pic->colorspace, WEBP_YUV422);
+    if (IS_YUV_CSP(pic->colorspace, WEBP_YUV420) || is_yuv422) {
+      *left &= ~1;
+      if (!is_yuv422) *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;
+}
+
+int WebPPictureCopy(const WebPPicture* src, WebPPicture* dst) {
+  if (src == NULL || dst == NULL) return 0;
+  if (src == dst) return 1;
+
+  WebPPictureGrabSpecs(src, dst);
+  if (!WebPPictureAlloc(dst)) return 0;
+
+  if (!src->use_argb) {
+    CopyPlane(src->y, src->y_stride,
+              dst->y, dst->y_stride, dst->width, dst->height);
+    CopyPlane(src->u, src->uv_stride,
+              dst->u, dst->uv_stride, HALVE(dst->width), HALVE(dst->height));
+    CopyPlane(src->v, src->uv_stride,
+              dst->v, dst->uv_stride, HALVE(dst->width), HALVE(dst->height));
+    if (dst->a != NULL)  {
+      CopyPlane(src->a, src->a_stride,
+                dst->a, dst->a_stride, dst->width, dst->height);
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (dst->u0 != NULL)  {
+      int uv0_width = src->width;
+      if (IS_YUV_CSP(dst->colorspace, WEBP_YUV422)) {
+        uv0_width = HALVE(uv0_width);
+      }
+      CopyPlane(src->u0, src->uv0_stride,
+                dst->u0, dst->uv0_stride, uv0_width, dst->height);
+      CopyPlane(src->v0, src->uv0_stride,
+                dst->v0, dst->uv0_stride, uv0_width, dst->height);
+    }
+#endif
+  } else {
+    CopyPlane((const uint8_t*)src->argb, 4 * src->argb_stride,
+              (uint8_t*)dst->argb, 4 * dst->argb_stride,
+              4 * dst->width, dst->height);
+  }
+  return 1;
+}
+
+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_'.
+    WebPPictureGrabSpecs(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;
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (src->u0 != NULL) {
+      const int left_pos =
+          IS_YUV_CSP(dst->colorspace, WEBP_YUV422) ? (left >> 1) : left;
+      dst->u0 = src->u0 + top * src->uv0_stride + left_pos;
+      dst->v0 = src->v0 + top * src->uv0_stride + left_pos;
+      dst->uv0_stride = src->uv0_stride;
+    }
+#endif
+  } else {
+    dst->argb = src->argb + top * src->argb_stride + left;
+    dst->argb_stride = src->argb_stride;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// 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;
+
+  WebPPictureGrabSpecs(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;
+    CopyPlane(pic->y + y_offset, pic->y_stride,
+              tmp.y, tmp.y_stride, width, height);
+    CopyPlane(pic->u + uv_offset, pic->uv_stride,
+              tmp.u, tmp.uv_stride, HALVE(width), HALVE(height));
+    CopyPlane(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;
+      CopyPlane(pic->a + a_offset, pic->a_stride,
+                tmp.a, tmp.a_stride, width, height);
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (tmp.u0 != NULL) {
+      int w = width;
+      int left_pos = left;
+      if (IS_YUV_CSP(tmp.colorspace, WEBP_YUV422)) {
+        w = HALVE(w);
+        left_pos = HALVE(left_pos);
+      }
+      CopyPlane(pic->u0 + top * pic->uv0_stride + left_pos, pic->uv0_stride,
+                tmp.u0, tmp.uv0_stride, w, height);
+      CopyPlane(pic->v0 + top * pic->uv0_stride + left_pos, pic->uv0_stride,
+                tmp.v0, tmp.uv0_stride, w, height);
+    }
+#endif
+  } else {
+    const uint8_t* const src =
+        (const uint8_t*)(pic->argb + top * pic->argb_stride + left);
+    CopyPlane(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 void 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,
+                         int32_t* const work,
+                         int num_channels) {
+  WebPRescaler rescaler;
+  int y = 0;
+  WebPRescalerInit(&rescaler, src_width, src_height,
+                   dst, dst_width, dst_height, dst_stride,
+                   num_channels,
+                   src_width, dst_width,
+                   src_height, dst_height,
+                   work);
+  memset(work, 0, 2 * dst_width * num_channels * sizeof(*work));
+  while (y < src_height) {
+    y += WebPRescalerImport(&rescaler, src_height - y,
+                            src + y * src_stride, src_stride);
+    WebPRescalerExport(&rescaler);
+  }
+}
+
+static void AlphaMultiplyARGB(WebPPicture* const pic, int inverse) {
+  uint32_t* ptr = pic->argb;
+  int y;
+  for (y = 0; y < pic->height; ++y) {
+    WebPMultARGBRow(ptr, pic->width, inverse);
+    ptr += pic->argb_stride;
+  }
+}
+
+static void AlphaMultiplyY(WebPPicture* const pic, int inverse) {
+  const uint8_t* ptr_a = pic->a;
+  if (ptr_a != NULL) {
+    uint8_t* ptr_y = pic->y;
+    int y;
+    for (y = 0; y < pic->height; ++y) {
+      WebPMultRow(ptr_y, ptr_a, pic->width, inverse);
+      ptr_y += pic->y_stride;
+      ptr_a += pic->a_stride;
+    }
+  }
+}
+
+int WebPPictureRescale(WebPPicture* pic, int width, int height) {
+  WebPPicture tmp;
+  int prev_width, prev_height;
+  int32_t* work;
+
+  if (pic == NULL) return 0;
+  prev_width = pic->width;
+  prev_height = pic->height;
+  // if width is unspecified, scale original proportionally to height ratio.
+  if (width == 0) {
+    width = (prev_width * height + prev_height / 2) / prev_height;
+  }
+  // if height is unspecified, scale original proportionally to width ratio.
+  if (height == 0) {
+    height = (prev_height * width + prev_width / 2) / prev_width;
+  }
+  // Check if the overall dimensions still make sense.
+  if (width <= 0 || height <= 0) return 0;
+
+  WebPPictureGrabSpecs(pic, &tmp);
+  tmp.width = width;
+  tmp.height = height;
+  if (!WebPPictureAlloc(&tmp)) return 0;
+
+  if (!pic->use_argb) {
+    work = (int32_t*)WebPSafeMalloc(2ULL * width, sizeof(*work));
+    if (work == NULL) {
+      WebPPictureFree(&tmp);
+      return 0;
+    }
+    // If present, we need to rescale alpha first (for AlphaMultiplyY).
+    if (pic->a != NULL) {
+      RescalePlane(pic->a, prev_width, prev_height, pic->a_stride,
+                   tmp.a, width, height, tmp.a_stride, work, 1);
+    }
+
+    // We take transparency into account on the luma plane only. That's not
+    // totally exact blending, but still is a good approximation.
+    AlphaMultiplyY(pic, 0);
+    RescalePlane(pic->y, prev_width, prev_height, pic->y_stride,
+                 tmp.y, width, height, tmp.y_stride, work, 1);
+    AlphaMultiplyY(&tmp, 1);
+
+    RescalePlane(pic->u,
+                 HALVE(prev_width), HALVE(prev_height), pic->uv_stride,
+                 tmp.u,
+                 HALVE(width), HALVE(height), tmp.uv_stride, work, 1);
+    RescalePlane(pic->v,
+                 HALVE(prev_width), HALVE(prev_height), pic->uv_stride,
+                 tmp.v,
+                 HALVE(width), HALVE(height), tmp.uv_stride, work, 1);
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (tmp.u0 != NULL) {
+      const int s = IS_YUV_CSP(tmp.colorspace, WEBP_YUV422) ? 2 : 1;
+      RescalePlane(
+          pic->u0, (prev_width + s / 2) / s, prev_height, pic->uv0_stride,
+          tmp.u0, (width + s / 2) / s, height, tmp.uv0_stride, work, 1);
+      RescalePlane(
+          pic->v0, (prev_width + s / 2) / s, prev_height, pic->uv0_stride,
+          tmp.v0, (width + s / 2) / s, height, tmp.uv0_stride, work, 1);
+    }
+#endif
+  } else {
+    work = (int32_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).
+    AlphaMultiplyARGB(pic, 0);
+    RescalePlane((const uint8_t*)pic->argb, prev_width, prev_height,
+                 pic->argb_stride * 4,
+                 (uint8_t*)tmp.argb, width, height,
+                 tmp.argb_stride * 4,
+                 work, 4);
+    AlphaMultiplyARGB(&tmp, 1);
+  }
+  WebPPictureFree(pic);
+  free(work);
+  *pic = tmp;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// 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);
+    }
+    free(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;
+}
+
+//------------------------------------------------------------------------------
+// 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;
+  while (height-- > 0) {
+    int x;
+    for (x = 0; x < width * x_step; x += x_step) {
+      if (alpha[x] != 0xff) return 1;  // TODO(skal): check 4/8 bytes at a time.
+    }
+    alpha += y_step;
+  }
+  return 0;
+}
+
+// Checking for the presence of non-opaque alpha.
+int WebPPictureHasTransparency(const WebPPicture* picture) {
+  if (picture == NULL) return 0;
+  if (!picture->use_argb) {
+    return CheckNonOpaque(picture->a, picture->width, picture->height,
+                          1, picture->a_stride);
+  } else {
+    int x, y;
+    const uint32_t* argb = picture->argb;
+    if (argb == NULL) return 0;
+    for (y = 0; y < picture->height; ++y) {
+      for (x = 0; x < picture->width; ++x) {
+        if (argb[x] < 0xff000000u) return 1;   // test any alpha values != 0xff
+      }
+      argb += picture->argb_stride;
+    }
+  }
+  return 0;
+}
+
+//------------------------------------------------------------------------------
+// RGB -> YUV conversion
+
+static int RGBToY(int r, int g, int b, VP8Random* const rg) {
+  return VP8RGBToY(r, g, b, VP8RandomBits(rg, YUV_FIX));
+}
+
+static int RGBToU(int r, int g, int b, VP8Random* const rg) {
+  return VP8RGBToU(r, g, b, VP8RandomBits(rg, YUV_FIX + 2));
+}
+
+static int RGBToV(int r, int g, int b, VP8Random* const rg) {
+  return VP8RGBToV(r, g, b, VP8RandomBits(rg, YUV_FIX + 2));
+}
+
+//------------------------------------------------------------------------------
+
+#if defined(USE_GAMMA_COMPRESSION)
+
+// gamma-compensates loss of resolution during chroma subsampling
+#define kGamma 0.80
+#define kGammaFix 12     // fixed-point precision for linear values
+#define kGammaScale ((1 << kGammaFix) - 1)
+#define kGammaTabFix 7   // fixed-point fractional bits precision
+#define kGammaTabScale (1 << kGammaTabFix)
+#define kGammaTabRounder (kGammaTabScale >> 1)
+#define kGammaTabSize (1 << (kGammaFix - kGammaTabFix))
+
+static int kLinearToGammaTab[kGammaTabSize + 1];
+static uint16_t kGammaToLinearTab[256];
+static int kGammaTablesOk = 0;
+
+static void InitGammaTables(void) {
+  if (!kGammaTablesOk) {
+    int v;
+    const double scale = 1. / kGammaScale;
+    for (v = 0; v <= 255; ++v) {
+      kGammaToLinearTab[v] =
+          (uint16_t)(pow(v / 255., kGamma) * kGammaScale + .5);
+    }
+    for (v = 0; v <= kGammaTabSize; ++v) {
+      const double x = scale * (v << kGammaTabFix);
+      kLinearToGammaTab[v] = (int)(pow(x, 1. / kGamma) * 255. + .5);
+    }
+    kGammaTablesOk = 1;
+  }
+}
+
+static WEBP_INLINE uint32_t GammaToLinear(uint8_t v) {
+  return kGammaToLinearTab[v];
+}
+
+// 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 v = base_value << shift;              // final uplifted value
+  const int tab_pos = v >> (kGammaTabFix + 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
+  return (y + kGammaTabRounder) >> kGammaTabFix;             // 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) {
+  (void)shift;
+  return v;
+}
+
+#endif    // USE_GAMMA_COMPRESSION
+
+//------------------------------------------------------------------------------
+
+#define SUM4(ptr) LinearToGamma(                         \
+    GammaToLinear((ptr)[0]) +                            \
+    GammaToLinear((ptr)[step]) +                         \
+    GammaToLinear((ptr)[rgb_stride]) +                   \
+    GammaToLinear((ptr)[rgb_stride + step]), 0)          \
+
+#define SUM2H(ptr) \
+    LinearToGamma(GammaToLinear((ptr)[0]) + GammaToLinear((ptr)[step]), 1)
+#define SUM2V(ptr) \
+    LinearToGamma(GammaToLinear((ptr)[0]) + GammaToLinear((ptr)[rgb_stride]), 1)
+#define SUM1(ptr)  \
+    LinearToGamma(GammaToLinear((ptr)[0]), 2)
+
+#define RGB_TO_UV(x, y, SUM) {                           \
+  const int src = (2 * (step * (x) + (y) * rgb_stride)); \
+  const int dst = (x) + (y) * picture->uv_stride;        \
+  const int r = SUM(r_ptr + src);                        \
+  const int g = SUM(g_ptr + src);                        \
+  const int b = SUM(b_ptr + src);                        \
+  picture->u[dst] = RGBToU(r, g, b, &rg);                \
+  picture->v[dst] = RGBToV(r, g, b, &rg);                \
+}
+
+#define RGB_TO_UV0(x_in, x_out, y, SUM) {                \
+  const int src = (step * (x_in) + (y) * rgb_stride);    \
+  const int dst = (x_out) + (y) * picture->uv0_stride;   \
+  const int r = SUM(r_ptr + src);                        \
+  const int g = SUM(g_ptr + src);                        \
+  const int b = SUM(b_ptr + src);                        \
+  picture->u0[dst] = RGBToU(r, g, b, &rg);               \
+  picture->v0[dst] = RGBToV(r, g, b, &rg);               \
+}
+
+static void MakeGray(WebPPicture* const picture) {
+  int y;
+  const int uv_width = HALVE(picture->width);
+  const int uv_height = HALVE(picture->height);
+  for (y = 0; y < uv_height; ++y) {
+    memset(picture->u + y * picture->uv_stride, 128, uv_width);
+    memset(picture->v + y * picture->uv_stride, 128, uv_width);
+  }
+}
+
+static int ImportYUVAFromRGBA(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 step,         // bytes per pixel
+                              int rgb_stride,   // bytes per scanline
+                              float dithering,
+                              WebPPicture* const picture) {
+  const WebPEncCSP uv_csp = picture->colorspace & WEBP_CSP_UV_MASK;
+  int x, y;
+  const int width = picture->width;
+  const int height = picture->height;
+  const int has_alpha = CheckNonOpaque(a_ptr, width, height, step, rgb_stride);
+  VP8Random rg;
+
+  picture->colorspace = uv_csp;
+  picture->use_argb = 0;
+  if (has_alpha) {
+    picture->colorspace |= WEBP_CSP_ALPHA_BIT;
+  }
+  if (!WebPPictureAlloc(picture)) return 0;
+
+  VP8InitRandom(&rg, dithering);
+  InitGammaTables();
+
+  // Import luma plane
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      const int offset = step * x + y * rgb_stride;
+      picture->y[x + y * picture->y_stride] =
+          RGBToY(r_ptr[offset], g_ptr[offset], b_ptr[offset], &rg);
+    }
+  }
+
+  // Downsample U/V plane
+  if (uv_csp != WEBP_YUV400) {
+    for (y = 0; y < (height >> 1); ++y) {
+      for (x = 0; x < (width >> 1); ++x) {
+        RGB_TO_UV(x, y, SUM4);
+      }
+      if (width & 1) {
+        RGB_TO_UV(x, y, SUM2V);
+      }
+    }
+    if (height & 1) {
+      for (x = 0; x < (width >> 1); ++x) {
+        RGB_TO_UV(x, y, SUM2H);
+      }
+      if (width & 1) {
+        RGB_TO_UV(x, y, SUM1);
+      }
+    }
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    // Store original U/V samples too
+    if (uv_csp == WEBP_YUV422) {
+      for (y = 0; y < height; ++y) {
+        for (x = 0; x < (width >> 1); ++x) {
+          RGB_TO_UV0(2 * x, x, y, SUM2H);
+        }
+        if (width & 1) {
+          RGB_TO_UV0(2 * x, x, y, SUM1);
+        }
+      }
+    } else if (uv_csp == WEBP_YUV444) {
+      for (y = 0; y < height; ++y) {
+        for (x = 0; x < width; ++x) {
+          RGB_TO_UV0(x, x, y, SUM1);
+        }
+      }
+    }
+#endif
+  } else {
+    MakeGray(picture);
+  }
+
+  if (has_alpha) {
+    assert(step >= 4);
+    assert(picture->a != NULL);
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        picture->a[x + y * picture->a_stride] =
+            a_ptr[step * x + y * rgb_stride];
+      }
+    }
+  }
+  return 1;
+}
+
+static int Import(WebPPicture* const picture,
+                  const uint8_t* const rgb, int rgb_stride,
+                  int step, int swap_rb, int import_alpha) {
+  const uint8_t* const r_ptr = rgb + (swap_rb ? 2 : 0);
+  const uint8_t* const g_ptr = rgb + 1;
+  const uint8_t* const b_ptr = rgb + (swap_rb ? 0 : 2);
+  const uint8_t* const a_ptr = import_alpha ? rgb + 3 : NULL;
+  const int width = picture->width;
+  const int height = picture->height;
+
+  if (!picture->use_argb) {
+    return ImportYUVAFromRGBA(r_ptr, g_ptr, b_ptr, a_ptr, step, rgb_stride,
+                              0.f /* no dithering */, picture);
+  }
+  if (import_alpha) {
+    picture->colorspace |= WEBP_CSP_ALPHA_BIT;
+  } else {
+    picture->colorspace &= ~WEBP_CSP_ALPHA_BIT;
+  }
+  if (!WebPPictureAlloc(picture)) return 0;
+
+  if (!import_alpha) {
+    int x, y;
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const int offset = step * x + y * rgb_stride;
+        const uint32_t argb =
+            MakeARGB32(r_ptr[offset], g_ptr[offset], b_ptr[offset]);
+        picture->argb[x + y * picture->argb_stride] = argb;
+      }
+    }
+  } else {
+    int x, y;
+    assert(step >= 4);
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const int offset = step * x + y * rgb_stride;
+        const uint32_t argb = ((uint32_t)a_ptr[offset] << 24) |
+                              (r_ptr[offset] << 16) |
+                              (g_ptr[offset] <<  8) |
+                              (b_ptr[offset]);
+        picture->argb[x + y * picture->argb_stride] = argb;
+      }
+    }
+  }
+  return 1;
+}
+#undef SUM4
+#undef SUM2V
+#undef SUM2H
+#undef SUM1
+#undef RGB_TO_UV
+
+int WebPPictureImportRGB(WebPPicture* picture,
+                         const uint8_t* rgb, int rgb_stride) {
+  return Import(picture, rgb, rgb_stride, 3, 0, 0);
+}
+
+int WebPPictureImportBGR(WebPPicture* picture,
+                         const uint8_t* rgb, int rgb_stride) {
+  return Import(picture, rgb, rgb_stride, 3, 1, 0);
+}
+
+int WebPPictureImportRGBA(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return Import(picture, rgba, rgba_stride, 4, 0, 1);
+}
+
+int WebPPictureImportBGRA(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return Import(picture, rgba, rgba_stride, 4, 1, 1);
+}
+
+int WebPPictureImportRGBX(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return Import(picture, rgba, rgba_stride, 4, 0, 0);
+}
+
+int WebPPictureImportBGRX(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return Import(picture, rgba, rgba_stride, 4, 1, 0);
+}
+
+//------------------------------------------------------------------------------
+// Automatic YUV <-> ARGB conversions.
+
+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 (!PictureAllocARGB(picture)) return 0;
+
+  // 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_IS_LAST);
+
+    // 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;
+}
+
+int WebPPictureARGBToYUVADithered(WebPPicture* picture, WebPEncCSP colorspace,
+                                  float dithering) {
+  if (picture == NULL) return 0;
+  if (picture->argb == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_NULL_PARAMETER);
+  } else {
+    const uint8_t* const argb = (const uint8_t*)picture->argb;
+    const uint8_t* const r = ALPHA_IS_LAST ? argb + 2 : argb + 1;
+    const uint8_t* const g = ALPHA_IS_LAST ? argb + 1 : argb + 2;
+    const uint8_t* const b = ALPHA_IS_LAST ? argb + 0 : argb + 3;
+    const uint8_t* const a = ALPHA_IS_LAST ? argb + 3 : argb + 0;
+    // We work on a tmp copy of 'picture', because ImportYUVAFromRGBA()
+    // would be calling WebPPictureFree(picture) otherwise.
+    WebPPicture tmp = *picture;
+    PictureResetARGB(&tmp);  // reset ARGB buffer so that it's not free()'d.
+    tmp.use_argb = 0;
+    tmp.colorspace = colorspace & WEBP_CSP_UV_MASK;
+    if (!ImportYUVAFromRGBA(r, g, b, a, 4, 4 * picture->argb_stride, dithering,
+                            &tmp)) {
+      return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    }
+    // Copy back the YUV specs into 'picture'.
+    tmp.argb = picture->argb;
+    tmp.argb_stride = picture->argb_stride;
+    tmp.memory_argb_ = picture->memory_argb_;
+    *picture = tmp;
+  }
+  return 1;
+}
+
+int WebPPictureARGBToYUVA(WebPPicture* picture, WebPEncCSP colorspace) {
+  return WebPPictureARGBToYUVADithered(picture, colorspace, 0.f);
+}
+
+//------------------------------------------------------------------------------
+// Helper: clean up fully transparent area to help compressibility.
+
+#define SIZE 8
+#define SIZE2 (SIZE / 2)
+static int is_transparent_area(const uint8_t* ptr, int stride, int size) {
+  int y, x;
+  for (y = 0; y < size; ++y) {
+    for (x = 0; x < size; ++x) {
+      if (ptr[x]) {
+        return 0;
+      }
+    }
+    ptr += stride;
+  }
+  return 1;
+}
+
+static WEBP_INLINE 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;
+  }
+}
+
+void WebPCleanupTransparentArea(WebPPicture* pic) {
+  int x, y, w, h;
+  const uint8_t* a_ptr;
+  int values[3] = { 0 };
+
+  if (pic == NULL) return;
+
+  a_ptr = pic->a;
+  if (a_ptr == NULL) return;    // nothing to do
+
+  w = pic->width / SIZE;
+  h = pic->height / SIZE;
+  for (y = 0; y < h; ++y) {
+    int need_reset = 1;
+    for (x = 0; x < w; ++x) {
+      const int off_a = (y * pic->a_stride + x) * SIZE;
+      const int off_y = (y * pic->y_stride + x) * SIZE;
+      const int off_uv = (y * pic->uv_stride + x) * SIZE2;
+      if (is_transparent_area(a_ptr + off_a, pic->a_stride, SIZE)) {
+        if (need_reset) {
+          values[0] = pic->y[off_y];
+          values[1] = pic->u[off_uv];
+          values[2] = pic->v[off_uv];
+          need_reset = 0;
+        }
+        flatten(pic->y + off_y, values[0], pic->y_stride, SIZE);
+        flatten(pic->u + off_uv, values[1], pic->uv_stride, SIZE2);
+        flatten(pic->v + off_uv, values[2], pic->uv_stride, SIZE2);
+      } else {
+        need_reset = 1;
+      }
+    }
+    // ignore the left-overs on right/bottom
+  }
+}
+
+#undef SIZE
+#undef SIZE2
+
+//------------------------------------------------------------------------------
+// Blend color and remove transparency info
+
+#define BLEND(V0, V1, ALPHA) \
+    ((((V0) * (255 - (ALPHA)) + (V1) * (ALPHA)) * 0x101) >> 16)
+#define BLEND_10BIT(V0, V1, ALPHA) \
+    ((((V0) * (1020 - (ALPHA)) + (V1) * (ALPHA)) * 0x101) >> 18)
+
+void WebPBlendAlpha(WebPPicture* pic, 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;
+  VP8Random rg;
+  int x, y;
+  if (pic == NULL) return;
+  VP8InitRandom(&rg, 0.f);
+  if (!pic->use_argb) {
+    const int uv_width = (pic->width >> 1);  // omit last pixel during u/v loop
+    const int Y0 = RGBToY(red, green, blue, &rg);
+    // VP8RGBToU/V expects the u/v values summed over four pixels
+    const int U0 = RGBToU(4 * red, 4 * green, 4 * blue, &rg);
+    const int V0 = RGBToV(4 * red, 4 * green, 4 * blue, &rg);
+    const int has_alpha = pic->colorspace & WEBP_CSP_ALPHA_BIT;
+    if (!has_alpha || pic->a == NULL) return;    // nothing to do
+    for (y = 0; y < pic->height; ++y) {
+      // Luma blending
+      uint8_t* const y_ptr = pic->y + y * pic->y_stride;
+      uint8_t* const a_ptr = pic->a + y * pic->a_stride;
+      for (x = 0; x < pic->width; ++x) {
+        const int alpha = a_ptr[x];
+        if (alpha < 0xff) {
+          y_ptr[x] = BLEND(Y0, y_ptr[x], a_ptr[x]);
+        }
+      }
+      // Chroma blending every even line
+      if ((y & 1) == 0) {
+        uint8_t* const u = pic->u + (y >> 1) * pic->uv_stride;
+        uint8_t* const v = pic->v + (y >> 1) * pic->uv_stride;
+        uint8_t* const a_ptr2 =
+            (y + 1 == pic->height) ? a_ptr : a_ptr + pic->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 int alpha =
+              a_ptr[2 * x + 0] + a_ptr[2 * x + 1] +
+              a_ptr2[2 * x + 0] + a_ptr2[2 * x + 1];
+          u[x] = BLEND_10BIT(U0, u[x], alpha);
+          v[x] = BLEND_10BIT(V0, v[x], alpha);
+        }
+        if (pic->width & 1) {   // rightmost pixel
+          const int alpha = 2 * (a_ptr[2 * x + 0] + a_ptr2[2 * x + 0]);
+          u[x] = BLEND_10BIT(U0, u[x], alpha);
+          v[x] = BLEND_10BIT(V0, v[x], alpha);
+        }
+      }
+      memset(a_ptr, 0xff, pic->width);
+    }
+  } else {
+    uint32_t* argb = pic->argb;
+    const uint32_t background = MakeARGB32(red, green, blue);
+    for (y = 0; y < pic->height; ++y) {
+      for (x = 0; x < pic->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 += pic->argb_stride;
+    }
+  }
+}
+
+#undef BLEND
+#undef BLEND_10BIT
+
+//------------------------------------------------------------------------------
+// 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.
+
+// search radius. Shouldn't be too large.
+#define RADIUS 2
+
+static float 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* s = src + j * src_stride;
+        for (i = x_0; i < x_1; ++i) {
+          const double sse = (double)(s[i] - value) * (s[i] - value);
+          if (sse < best_sse) best_sse = sse;
+        }
+      }
+      total_sse += best_sse;
+    }
+  }
+  return (float)total_sse;
+}
+#undef RADIUS
+
+//------------------------------------------------------------------------------
+// Distortion
+
+// Max value returned in case of exact similarity.
+static const double kMinDistortion_dB = 99.;
+static float GetPSNR(const double v) {
+  return (float)((v > 0.) ? -4.3429448 * log(v / (255 * 255.))
+                          : kMinDistortion_dB);
+}
+
+int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
+                          int type, float result[5]) {
+  DistoStats stats[5];
+  int has_alpha;
+  int uv_w, uv_h;
+
+  if (src == NULL || ref == NULL ||
+      src->width != ref->width || src->height != ref->height ||
+      src->y == NULL || ref->y == NULL ||
+      src->u == NULL || ref->u == NULL ||
+      src->v == NULL || ref->v == NULL ||
+      result == NULL) {
+    return 0;
+  }
+  // TODO(skal): provide distortion for ARGB too.
+  if (src->use_argb == 1 || src->use_argb != ref->use_argb) {
+    return 0;
+  }
+
+  has_alpha = !!(src->colorspace & WEBP_CSP_ALPHA_BIT);
+  if (has_alpha != !!(ref->colorspace & WEBP_CSP_ALPHA_BIT) ||
+      (has_alpha && (src->a == NULL || ref->a == NULL))) {
+    return 0;
+  }
+
+  memset(stats, 0, sizeof(stats));
+
+  uv_w = HALVE(src->width);
+  uv_h = HALVE(src->height);
+  if (type >= 2) {
+    float sse[4];
+    sse[0] = AccumulateLSIM(src->y, src->y_stride,
+                            ref->y, ref->y_stride, src->width, src->height);
+    sse[1] = AccumulateLSIM(src->u, src->uv_stride,
+                            ref->u, ref->uv_stride, uv_w, uv_h);
+    sse[2] = AccumulateLSIM(src->v, src->uv_stride,
+                            ref->v, ref->uv_stride, uv_w, uv_h);
+    sse[3] = has_alpha ? AccumulateLSIM(src->a, src->a_stride,
+                                        ref->a, ref->a_stride,
+                                        src->width, src->height)
+                       : 0.f;
+    result[0] = GetPSNR(sse[0] / (src->width * src->height));
+    result[1] = GetPSNR(sse[1] / (uv_w * uv_h));
+    result[2] = GetPSNR(sse[2] / (uv_w * uv_h));
+    result[3] = GetPSNR(sse[3] / (src->width * src->height));
+    {
+      double total_sse = sse[0] + sse[1] + sse[2];
+      int total_pixels = src->width * src->height + 2 * uv_w * uv_h;
+      if (has_alpha) {
+        total_pixels += src->width * src->height;
+        total_sse += sse[3];
+      }
+      result[4] = GetPSNR(total_sse / total_pixels);
+    }
+  } else {
+    int c;
+    VP8SSIMAccumulatePlane(src->y, src->y_stride,
+                           ref->y, ref->y_stride,
+                           src->width, src->height, &stats[0]);
+    VP8SSIMAccumulatePlane(src->u, src->uv_stride,
+                           ref->u, ref->uv_stride,
+                           uv_w, uv_h, &stats[1]);
+    VP8SSIMAccumulatePlane(src->v, src->uv_stride,
+                           ref->v, ref->uv_stride,
+                           uv_w, uv_h, &stats[2]);
+    if (has_alpha) {
+      VP8SSIMAccumulatePlane(src->a, src->a_stride,
+                             ref->a, ref->a_stride,
+                             src->width, src->height, &stats[3]);
+    }
+    for (c = 0; c <= 4; ++c) {
+      if (type == 1) {
+        const double v = VP8SSIMGet(&stats[c]);
+        result[c] = (float)((v < 1.) ? -10.0 * log10(1. - v)
+                                     : kMinDistortion_dB);
+      } else {
+        const double v = VP8SSIMGetSquaredError(&stats[c]);
+        result[c] = GetPSNR(v);
+      }
+      // Accumulate forward
+      if (c < 4) VP8SSIMAddStats(&stats[c], &stats[4]);
+    }
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// 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 (!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) {
+    free(wrt.mem);
+    *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(WebPEncodeBGR, WebPPictureImportBGR)
+ENCODE_FUNC(WebPEncodeRGBA, WebPPictureImportRGBA)
+ENCODE_FUNC(WebPEncodeBGRA, WebPPictureImportBGRA)
+
+#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(WebPEncodeLosslessBGR, WebPPictureImportBGR)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGBA, WebPPictureImportRGBA)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGRA, WebPPictureImportBGRA)
+
+#undef LOSSLESS_ENCODE_FUNC
+
+//------------------------------------------------------------------------------
+
diff --git a/src/3rdparty/libwebp/src/enc/quant.c b/src/3rdparty/libwebp/src/enc/quant.c
new file mode 100644
index 0000000..e1d202b
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/quant.c
@@ -0,0 +1,1156 @@
+// 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 "./vp8enci.h"
+#include "./cost.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.
+
+#define I4_PENALTY 4000   // Rate-penalty for quick i4/i16 decision
+
+// number of non-zero coeffs below which we consider the block very flat
+// (and apply a penalty to complex predictions)
+#define FLATNESS_LIMIT_I16 10      // I16 mode
+#define FLATNESS_LIMIT_I4  3       // I4 mode
+#define FLATNESS_LIMIT_UV  2       // UV mode
+#define FLATNESS_PENALTY   140     // roughly ~1bit per block
+
+#define MULT_8B(a, b) (((a) * (b) + 128) >> 8)
+
+// #define 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);
+  printf("SOURCE / OUTPUT / ABS DELTA\n");
+  for (j = 0; j < 24; ++j) {
+    if (j == 16) printf("\n");   // newline before the U/V block
+    for (i = 0; i < 16; ++i) printf("%3d ", it->yuv_in_[i + j * BPS]);
+    printf("     ");
+    for (i = 0; i < 16; ++i) printf("%3d ", it->yuv_out_[i + j * BPS]);
+    printf("     ");
+    for (i = 0; i < 16; ++i) {
+      printf("%1d ", abs(it->yuv_out_[i + j * BPS] - it->yuv_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 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 q4, q16, quv;
+    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)];
+
+    q4  = ExpandMatrix(&m->y1_, 0);
+    q16 = ExpandMatrix(&m->y2_, 1);
+    quv = ExpandMatrix(&m->uv_, 2);
+
+    m->lambda_i4_          = (3 * q4 * q4) >> 7;
+    m->lambda_i16_         = (3 * q16 * q16);
+    m->lambda_uv_          = (3 * quv * quv) >> 6;
+    m->lambda_mode_        = (1 * q4 * q4) >> 7;
+    m->lambda_trellis_i4_  = (7 * q4 * q4) >> 3;
+    m->lambda_trellis_i16_ = (q16 * q16) >> 2;
+    m->lambda_trellis_uv_  = (quv *quv) << 1;
+    m->tlambda_            = (tlambda_scale * q4) >> 5;
+
+    m->min_disto_ = 10 * m->y1_.q_[0];   // quantization-aware min disto
+    m->max_edge_  = 0;
+  }
+}
+
+//------------------------------------------------------------------------------
+// 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 };
+  const int num_segments = enc->segment_hdr_.num_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 displeasant).
+  dq_uv_dc = -4 * enc->config_->sns_strength / 100;
+  dq_uv_dc = clip(dq_uv_dc, -15, 15);   // 4bit-signed max allowed
+
+  enc->dq_y1_dc_ = 0;       // TODO(skal): dq-lum
+  enc->dq_y2_dc_ = 0;
+  enc->dq_y2_ac_ = 0;
+  enc->dq_uv_dc_ = dq_uv_dc;
+  enc->dq_uv_ac_ = dq_uv_ac;
+
+  SetupFilterStrength(enc);   // initialize segments' filtering, eventually
+
+  if (num_segments > 1) SimplifySegments(enc);
+
+  SetupMatrices(enc);         // finalize quantization matrices
+}
+
+//------------------------------------------------------------------------------
+// Form the predictions in cache
+
+// Must be ordered using {DC_PRED, TM_PRED, V_PRED, H_PRED} as index
+const int VP8I16ModeOffsets[4] = { I16DC16, I16TM16, I16VE16, I16HE16 };
+const int VP8UVModeOffsets[4] = { C8DC8, C8TM8, C8VE8, C8HE8 };
+
+// Must be indexed using {B_DC_PRED -> B_HU_PRED} as index
+const int VP8I4ModeOffsets[NUM_BMODES] = {
+  I4DC4, I4TM4, I4VE4, I4HE4, I4RD4, I4VR4, I4LD4, I4VL4, I4HD4, I4HU4
+};
+
+void VP8MakeLuma16Preds(const VP8EncIterator* const it) {
+  const uint8_t* const left = it->x_ ? it->y_left_ : NULL;
+  const uint8_t* const top = it->y_ ? it->y_top_ : NULL;
+  VP8EncPredLuma16(it->yuv_p_, left, top);
+}
+
+void VP8MakeChroma8Preds(const VP8EncIterator* const it) {
+  const uint8_t* const left = it->x_ ? it->u_left_ : NULL;
+  const uint8_t* const top = it->y_ ? it->uv_top_ : NULL;
+  VP8EncPredChroma8(it->yuv_p_, left, top);
+}
+
+void VP8MakeIntra4Preds(const VP8EncIterator* const it) {
+  VP8EncPredLuma4(it->yuv_p_, it->i4_top_);
+}
+
+//------------------------------------------------------------------------------
+// Quantize
+
+// Layout:
+// +----+
+// |YYYY| 0
+// |YYYY| 4
+// |YYYY| 8
+// |YYYY| 12
+// +----+
+// |UUVV| 16
+// |UUVV| 20
+// +----+
+
+const int VP8Scan[16 + 4 + 4] = {
+  // 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,
+
+  0 + 0 * BPS,   4 + 0 * BPS, 0 + 4 * BPS,  4 + 4 * BPS,    // U
+  8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
+};
+
+//------------------------------------------------------------------------------
+// Distortion measurement
+
+static const uint16_t kWeightY[16] = {
+  38, 32, 20, 9, 32, 28, 17, 7, 20, 17, 10, 4, 9, 7, 4, 2
+};
+
+static const uint16_t kWeightTrellis[16] = {
+#if USE_TDISTO == 0
+  16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16
+#else
+  30, 27, 19, 11,
+  27, 24, 17, 10,
+  19, 17, 12,  8,
+  11, 10,  8,  6
+#endif
+};
+
+// Init/Copy the common fields in score.
+static void InitScore(VP8ModeScore* const rd) {
+  rd->D  = 0;
+  rd->SD = 0;
+  rd->R  = 0;
+  rd->H  = 0;
+  rd->nz = 0;
+  rd->score = MAX_COST;
+}
+
+static void CopyScore(VP8ModeScore* const dst, const VP8ModeScore* const src) {
+  dst->D  = src->D;
+  dst->SD = src->SD;
+  dst->R  = src->R;
+  dst->H  = src->H;
+  dst->nz = src->nz;      // note that nz is not accumulated, but just copied.
+  dst->score = src->score;
+}
+
+static void AddScore(VP8ModeScore* const dst, const VP8ModeScore* const src) {
+  dst->D  += src->D;
+  dst->SD += src->SD;
+  dst->R  += src->R;
+  dst->H  += src->H;
+  dst->nz |= src->nz;     // here, new nz bits are accumulated.
+  dst->score += src->score;
+}
+
+//------------------------------------------------------------------------------
+// Performs trellis-optimized quantization.
+
+// Trellis
+
+typedef struct {
+  int prev;        // best previous
+  int level;       // level
+  int sign;        // sign of coeff_i
+  score_t cost;    // bit cost
+  score_t error;   // distortion = sum of (|coeff_i| - level_i * Q_i)^2
+  int ctx;         // context (only depends on 'level'. Could be spared.)
+} Node;
+
+// 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) + 1][(l) + MIN_DELTA])
+
+static WEBP_INLINE void SetRDScore(int lambda, VP8ModeScore* const rd) {
+  // TODO: incorporate the "* 256" in the tables?
+  rd->score = (rd->R + rd->H) * lambda + 256 * (rd->D + rd->SD);
+}
+
+static WEBP_INLINE score_t RDScoreTrellis(int lambda, score_t rate,
+                                          score_t distortion) {
+  return rate * lambda + 256 * distortion;
+}
+
+static int TrellisQuantizeBlock(const VP8EncIterator* const it,
+                                int16_t in[16], int16_t out[16],
+                                int ctx0, int coeff_type,
+                                const VP8Matrix* const mtx,
+                                int lambda) {
+  ProbaArray* const last_costs = it->enc_->proba_.coeffs_[coeff_type];
+  CostArray* const costs = it->enc_->proba_.level_cost_[coeff_type];
+  const int first = (coeff_type == 0) ? 1 : 0;
+  Node nodes[17][NUM_NODES];
+  int best_path[3] = {-1, -1, -1};   // store best-last/best-level/best-previous
+  score_t best_score;
+  int best_node;
+  int last = first - 1;
+  int n, m, p, nz;
+
+  {
+    score_t cost;
+    score_t max_error;
+    const int thresh = mtx->q_[1] * mtx->q_[1] / 4;
+    const int last_proba = last_costs[VP8EncBands[first]][ctx0][0];
+
+    // compute maximal distortion.
+    max_error = 0;
+    for (n = first; n < 16; ++n) {
+      const int j  = kZigzag[n];
+      const int err = in[j] * in[j];
+      max_error += kWeightTrellis[j] * err;
+      if (err > thresh) last = n;
+    }
+    // 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, max_error);
+
+    // initialize source node.
+    n = first - 1;
+    for (m = -MIN_DELTA; m <= MAX_DELTA; ++m) {
+      NODE(n, m).cost = 0;
+      NODE(n, m).error = max_error;
+      NODE(n, m).ctx = ctx0;
+    }
+  }
+
+  // traverse trellis.
+  for (n = first; n <= last; ++n) {
+    const int j  = kZigzag[n];
+    const int Q  = mtx->q_[j];
+    const int iQ = mtx->iq_[j];
+    const int 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 int coeff0 = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
+    int level0 = QUANTDIV(coeff0, iQ, B);
+    if (level0 > MAX_LEVEL) level0 = MAX_LEVEL;
+
+    // test all alternate level values around level0.
+    for (m = -MIN_DELTA; m <= MAX_DELTA; ++m) {
+      Node* const cur = &NODE(n, m);
+      int delta_error, new_error;
+      score_t cur_score = MAX_COST;
+      int level = level0 + m;
+      int last_proba;
+
+      cur->sign = sign;
+      cur->level = level;
+      cur->ctx = (level == 0) ? 0 : (level == 1) ? 1 : 2;
+      if (level > MAX_LEVEL || level < 0) {   // node is dead?
+        cur->cost = MAX_COST;
+        continue;
+      }
+      last_proba = last_costs[VP8EncBands[n + 1]][cur->ctx][0];
+
+      // Compute delta_error = how much coding this level will
+      // subtract as distortion to max_error
+      new_error = coeff0 - level * Q;
+      delta_error =
+        kWeightTrellis[j] * (coeff0 * coeff0 - new_error * new_error);
+
+      // Inspect all possible non-dead predecessors. Retain only the best one.
+      for (p = -MIN_DELTA; p <= MAX_DELTA; ++p) {
+        const Node* const prev = &NODE(n - 1, p);
+        const int prev_ctx = prev->ctx;
+        const uint16_t* const tcost = costs[VP8EncBands[n]][prev_ctx];
+        const score_t total_error = prev->error - delta_error;
+        score_t cost, base_cost, score;
+
+        if (prev->cost >= MAX_COST) {   // dead node?
+          continue;
+        }
+
+        // Base cost of both terminal/non-terminal
+        base_cost = prev->cost + VP8LevelCost(tcost, level);
+
+        // Examine node assuming it's a non-terminal one.
+        cost = base_cost;
+        if (level && n < 15) {
+          cost += VP8BitCost(1, last_proba);
+        }
+        score = RDScoreTrellis(lambda, cost, total_error);
+        if (score < cur_score) {
+          cur_score = score;
+          cur->cost  = cost;
+          cur->error = total_error;
+          cur->prev  = p;
+        }
+
+        // Now, record best terminal node (and thus best entry in the graph).
+        if (level) {
+          cost = base_cost;
+          if (n < 15) cost += VP8BitCost(0, last_proba);
+          score = RDScoreTrellis(lambda, cost, total_error);
+          if (score < best_score) {
+            best_score = score;
+            best_path[0] = n;   // best eob position
+            best_path[1] = m;   // best level
+            best_path[2] = p;   // best predecessor
+          }
+        }
+      }
+    }
+  }
+
+  // Fresh start
+  memset(in + first, 0, (16 - first) * sizeof(*in));
+  memset(out + first, 0, (16 - first) * sizeof(*out));
+  if (best_path[0] == -1) {
+    return 0;   // skip!
+  }
+
+  // Unwind the best path.
+  // Note: best-prev on terminal node is not necessarily equal to the
+  // best_prev for non-terminal. So we patch best_path[2] in.
+  n = best_path[0];
+  best_node = best_path[1];
+  NODE(n, best_node).prev = best_path[2];   // force best-prev for terminal
+  nz = 0;
+
+  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 != 0);
+    in[j] = out[n] * mtx->q_[j];
+    best_node = node->prev;
+  }
+  return nz;
+}
+
+#undef NODE
+
+//------------------------------------------------------------------------------
+// Performs: difference, transform, quantize, back-transform, add
+// all at once. Output is the reconstructed block in *yuv_out, and the
+// quantized levels in *levels.
+
+static int ReconstructIntra16(VP8EncIterator* const it,
+                              VP8ModeScore* const rd,
+                              uint8_t* const yuv_out,
+                              int mode) {
+  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;
+  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) {
+    VP8FTransform(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(it, tmp[n], rd->y_ac_levels[n], ctx, 0,
+                                &dqm->y1_, dqm->lambda_trellis_i16_);
+        it->top_nz_[x] = it->left_nz_[y] = non_zero;
+        nz |= non_zero << n;
+      }
+    }
+  } else {
+    for (n = 0; n < 16; ++n) {
+      nz |= VP8EncQuantizeBlock(tmp[n], rd->y_ac_levels[n], 1, &dqm->y1_) << n;
+    }
+  }
+
+  // Transform back
+  VP8ITransformWHT(dc_tmp, tmp[0]);
+  for (n = 0; n < 16; n += 2) {
+    VP8ITransform(ref + VP8Scan[n], tmp[n], yuv_out + VP8Scan[n], 1);
+  }
+
+  return nz;
+}
+
+static int ReconstructIntra4(VP8EncIterator* const it,
+                             int16_t levels[16],
+                             const uint8_t* const src,
+                             uint8_t* const yuv_out,
+                             int mode) {
+  const VP8Encoder* const enc = it->enc_;
+  const uint8_t* const ref = it->yuv_p_ + VP8I4ModeOffsets[mode];
+  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  int nz = 0;
+  int16_t tmp[16];
+
+  VP8FTransform(src, ref, tmp);
+  if (DO_TRELLIS_I4 && it->do_trellis_) {
+    const int x = it->i4_ & 3, y = it->i4_ >> 2;
+    const int ctx = it->top_nz_[x] + it->left_nz_[y];
+    nz = TrellisQuantizeBlock(it, tmp, levels, ctx, 3, &dqm->y1_,
+                              dqm->lambda_trellis_i4_);
+  } else {
+    nz = VP8EncQuantizeBlock(tmp, levels, 0, &dqm->y1_);
+  }
+  VP8ITransform(ref, tmp, yuv_out, 0);
+  return nz;
+}
+
+static int ReconstructUV(VP8EncIterator* const it, VP8ModeScore* const rd,
+                         uint8_t* const yuv_out, int mode) {
+  const VP8Encoder* const enc = it->enc_;
+  const uint8_t* const ref = it->yuv_p_ + VP8UVModeOffsets[mode];
+  const uint8_t* const src = it->yuv_in_ + U_OFF;
+  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) {
+    VP8FTransform(src + VP8Scan[16 + n], ref + VP8Scan[16 + n], tmp[n]);
+  }
+  if (DO_TRELLIS_UV && it->do_trellis_) {
+    int ch, x, y;
+    for (ch = 0, n = 0; ch <= 2; ch += 2) {
+      for (y = 0; y < 2; ++y) {
+        for (x = 0; x < 2; ++x, ++n) {
+          const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+          const int non_zero =
+            TrellisQuantizeBlock(it, tmp[n], rd->uv_levels[n], ctx, 2,
+                                 &dqm->uv_, dqm->lambda_trellis_uv_);
+          it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = non_zero;
+          nz |= non_zero << n;
+        }
+      }
+    }
+  } else {
+    for (n = 0; n < 8; ++n) {
+      nz |= VP8EncQuantizeBlock(tmp[n], rd->uv_levels[n], 0, &dqm->uv_) << n;
+    }
+  }
+
+  for (n = 0; n < 8; n += 2) {
+    VP8ITransform(ref + VP8Scan[16 + n], tmp[n], yuv_out + VP8Scan[16 + 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[4]);
+  const int v2 = abs(DCs[5]);
+  int max_v = (v0 > v1) ? v1 : v0;
+  max_v = (v2 > max_v) ? v2 : max_v;
+  if (max_v > dqm->max_edge_) dqm->max_edge_ = max_v;
+}
+
+static void SwapPtr(uint8_t** a, uint8_t** b) {
+  uint8_t* const tmp = *a;
+  *a = *b;
+  *b = tmp;
+}
+
+static void SwapOut(VP8EncIterator* const it) {
+  SwapPtr(&it->yuv_out_, &it->yuv_out2_);
+}
+
+static score_t IsFlat(const int16_t* levels, int num_blocks, score_t thresh) {
+  score_t score = 0;
+  while (num_blocks-- > 0) {      // TODO(skal): refine positional scoring?
+    int i;
+    for (i = 1; i < 16; ++i) {    // omit DC, we're only interested in AC
+      score += (levels[i] != 0);
+      if (score > thresh) return 0;
+    }
+    levels += 16;
+  }
+  return 1;
+}
+
+static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* const rd) {
+  const int kNumBlocks = 16;
+  VP8Encoder* const enc = it->enc_;
+  VP8SegmentInfo* const dqm = &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;
+  VP8ModeScore rd16;
+  int mode;
+
+  rd->mode_i16 = -1;
+  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
+    uint8_t* const tmp_dst = it->yuv_out2_ + Y_OFF;  // scratch buffer
+    int nz;
+
+    // Reconstruct
+    nz = ReconstructIntra16(it, &rd16, tmp_dst, mode);
+
+    // Measure RD-score
+    rd16.D = VP8SSE16x16(src, tmp_dst);
+    rd16.SD = tlambda ? MULT_8B(tlambda, VP8TDisto16x16(src, tmp_dst, kWeightY))
+            : 0;
+    rd16.H = VP8FixedCostsI16[mode];
+    rd16.R = VP8GetCostLuma16(it, &rd16);
+    if (mode > 0 &&
+        IsFlat(rd16.y_ac_levels[0], kNumBlocks, FLATNESS_LIMIT_I16)) {
+      // penalty to avoid flat area to be mispredicted by complex mode
+      rd16.R += FLATNESS_PENALTY * kNumBlocks;
+    }
+
+    // Since we always examine Intra16 first, we can overwrite *rd directly.
+    SetRDScore(lambda, &rd16);
+    if (mode == 0 || rd16.score < rd->score) {
+      CopyScore(rd, &rd16);
+      rd->mode_i16 = mode;
+      rd->nz = nz;
+      memcpy(rd->y_ac_levels, rd16.y_ac_levels, sizeof(rd16.y_ac_levels));
+      memcpy(rd->y_dc_levels, rd16.y_dc_levels, sizeof(rd16.y_dc_levels));
+      SwapOut(it);
+    }
+  }
+  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 & 0xffff) == 0 && rd->D > dqm->min_disto_) {
+    StoreMaxDelta(dqm, rd->y_dc_levels);
+  }
+}
+
+//------------------------------------------------------------------------------
+
+// return the cost array corresponding to the surrounding prediction modes.
+static const uint16_t* GetCostModeI4(VP8EncIterator* const it,
+                                     const uint8_t modes[16]) {
+  const int preds_w = it->enc_->preds_w_;
+  const int x = (it->i4_ & 3), y = it->i4_ >> 2;
+  const int left = (x == 0) ? it->preds_[y * preds_w - 1] : modes[it->i4_ - 1];
+  const int top = (y == 0) ? it->preds_[-preds_w + x] : modes[it->i4_ - 4];
+  return VP8FixedCostsI4[top][left];
+}
+
+static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
+  const VP8Encoder* const enc = it->enc_;
+  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  const int lambda = dqm->lambda_i4_;
+  const int tlambda = dqm->tlambda_;
+  const uint8_t* const src0 = it->yuv_in_ + Y_OFF;
+  uint8_t* const best_blocks = it->yuv_out2_ + Y_OFF;
+  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];
+      rd_tmp.R = VP8GetCostLuma4(it, tmp_levels);
+      if (mode > 0 && IsFlat(tmp_levels, kNumBlocks, FLATNESS_LIMIT_I4)) {
+        rd_tmp.R += FLATNESS_PENALTY * kNumBlocks;
+      }
+
+      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(tmp_levels));
+      }
+    }
+    SetRDScore(dqm->lambda_mode_, &rd_i4);
+    AddScore(&rd_best, &rd_i4);
+    if (rd_best.score >= rd->score) {
+      return 0;
+    }
+    total_header_bits += (int)rd_i4.H;   // <- equal to mode_costs[best_mode];
+    if (total_header_bits > enc->max_i4_header_bits_) {
+      return 0;
+    }
+    // Copy selected samples if not in the right place already.
+    if (best_block != best_blocks + VP8Scan[it->i4_]) {
+      VP8Copy4x4(best_block, best_blocks + VP8Scan[it->i4_]);
+    }
+    rd->modes_i4[it->i4_] = best_mode;
+    it->top_nz_[it->i4_ & 3] = it->left_nz_[it->i4_ >> 2] = (rd_i4.nz ? 1 : 0);
+  } while (VP8IteratorRotateI4(it, best_blocks));
+
+  // finalize state
+  CopyScore(rd, &rd_best);
+  VP8SetIntra4Mode(it, rd->modes_i4);
+  SwapOut(it);
+  memcpy(rd->y_ac_levels, rd_best.y_ac_levels, sizeof(rd->y_ac_levels));
+  return 1;   // select intra4x4 over intra16x16
+}
+
+//------------------------------------------------------------------------------
+
+static void PickBestUV(VP8EncIterator* const it, VP8ModeScore* const rd) {
+  const int kNumBlocks = 8;
+  const VP8Encoder* const enc = it->enc_;
+  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  const int lambda = dqm->lambda_uv_;
+  const uint8_t* const src = it->yuv_in_ + U_OFF;
+  uint8_t* const tmp_dst = it->yuv_out2_ + U_OFF;  // scratch buffer
+  uint8_t* const dst0 = it->yuv_out_ + U_OFF;
+  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;    // TODO: should we call TDisto? 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));
+      memcpy(dst0, tmp_dst, UV_SIZE);   //  TODO: SwapUVOut() ?
+    }
+  }
+  VP8SetIntraUVMode(it, rd->mode_uv);
+  AddScore(rd, &rd_best);
+}
+
+//------------------------------------------------------------------------------
+// Final reconstruction and quantization.
+
+static void SimpleQuantize(VP8EncIterator* const it, VP8ModeScore* const rd) {
+  const VP8Encoder* const enc = it->enc_;
+  const int is_i16 = (it->mb_->type_ == 1);
+  int nz = 0;
+
+  if (is_i16) {
+    nz = ReconstructIntra16(it, rd, it->yuv_out_ + Y_OFF, 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 + VP8Scan[it->i4_];
+      uint8_t* const dst = it->yuv_out_ + Y_OFF + 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));
+  }
+
+  nz |= ReconstructUV(it, rd, it->yuv_out_ + U_OFF, it->mb_->uv_mode_);
+  rd->nz = nz;
+}
+
+// Refine intra16/intra4 sub-modes based on distortion only (not rate).
+static void DistoRefine(VP8EncIterator* const it, int try_both_i4_i16) {
+  const int is_i16 = (it->mb_->type_ == 1);
+  score_t best_score = MAX_COST;
+
+  if (try_both_i4_i16 || is_i16) {
+    int mode;
+    int best_mode = -1;
+    for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
+      const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
+      const uint8_t* const src = it->yuv_in_ + Y_OFF;
+      const score_t score = VP8SSE16x16(src, ref);
+      if (score < best_score) {
+        best_mode = mode;
+        best_score = score;
+      }
+    }
+    VP8SetIntra16Mode(it, best_mode);
+  }
+  if (try_both_i4_i16 || !is_i16) {
+    uint8_t modes_i4[16];
+    // 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).
+    score_t score_i4 = (score_t)I4_PENALTY;
+
+    VP8IteratorStartI4(it);
+    do {
+      int mode;
+      int best_sub_mode = -1;
+      score_t best_sub_score = MAX_COST;
+      const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
+
+      // TODO(skal): we don't really need the prediction pixels here,
+      // but just the distortion against 'src'.
+      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);
+        if (score < best_sub_score) {
+          best_sub_mode = mode;
+          best_sub_score = score;
+        }
+      }
+      modes_i4[it->i4_] = best_sub_mode;
+      score_i4 += best_sub_score;
+      if (score_i4 >= best_score) break;
+    } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF));
+    if (score_i4 < best_score) {
+      VP8SetIntra4Mode(it, modes_i4);
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+// Entry point
+
+int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd,
+                VP8RDLevel rd_opt) {
+  int is_skipped;
+  const int method = it->enc_->method_;
+
+  InitScore(rd);
+
+  // We can perform predictions for Luma16x16 and Chroma8x8 already.
+  // Luma4x4 predictions needs to be done as-we-go.
+  VP8MakeLuma16Preds(it);
+  VP8MakeChroma8Preds(it);
+
+  if (rd_opt > RD_OPT_NONE) {
+    it->do_trellis_ = (rd_opt >= RD_OPT_TRELLIS_ALL);
+    PickBestIntra16(it, rd);
+    if (method >= 2) {
+      PickBestIntra4(it, rd);
+    }
+    PickBestUV(it, rd);
+    if (rd_opt == RD_OPT_TRELLIS) {   // finish off with trellis-optim now
+      it->do_trellis_ = 1;
+      SimpleQuantize(it, rd);
+    }
+  } else {
+    // For method == 2, pick the best intra4/intra16 based on SSE (~tad slower).
+    // For method <= 1, we refine intra4 or intra16 (but don't re-examine mode).
+    DistoRefine(it, (method >= 2));
+    SimpleQuantize(it, rd);
+  }
+  is_skipped = (rd->nz == 0);
+  VP8SetSkip(it, is_skipped);
+  return is_skipped;
+}
+
diff --git a/src/3rdparty/libwebp/src/enc/syntax.c b/src/3rdparty/libwebp/src/enc/syntax.c
new file mode 100644
index 0000000..08cfe79
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/syntax.c
@@ -0,0 +1,423 @@
+// 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 "../utils/utils.h"
+#include "../webp/format_constants.h"  // RIFF constants
+#include "../webp/mux_types.h"         // ALPHA_FLAG
+#include "./vp8enci.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 VP8SegmentHeader* const hdr = &enc->segment_hdr_;
+  const VP8Proba* 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) {
+        VP8PutSignedValue(bw, enc->dqm_[s].quant_, 7);
+      }
+      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+        VP8PutSignedValue(bw, enc->dqm_[s].fstrength_, 6);
+      }
+    }
+    if (hdr->update_map_) {
+      for (s = 0; s < 3; ++s) {
+        if (VP8PutBitUniform(bw, (proba->segments_[s] != 255u))) {
+          VP8PutValue(bw, proba->segments_[s], 8);
+        }
+      }
+    }
+  }
+}
+
+// Filtering parameters header
+static void PutFilterHeader(VP8BitWriter* const bw,
+                            const VP8FilterHeader* const hdr) {
+  const int use_lf_delta = (hdr->i4x4_lf_delta_ != 0);
+  VP8PutBitUniform(bw, hdr->simple_);
+  VP8PutValue(bw, hdr->level_, 6);
+  VP8PutValue(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
+      VP8PutValue(bw, 0, 4);
+      // we use mode_lf_delta for i4x4
+      VP8PutSignedValue(bw, hdr->i4x4_lf_delta_, 6);
+      VP8PutValue(bw, 0, 3);    // all others unused
+    }
+  }
+}
+
+// Nominal quantization parameters
+static void PutQuant(VP8BitWriter* const bw,
+                     const VP8Encoder* const enc) {
+  VP8PutValue(bw, enc->base_quant_, 7);
+  VP8PutSignedValue(bw, enc->dq_y1_dc_, 4);
+  VP8PutSignedValue(bw, enc->dq_y2_dc_, 4);
+  VP8PutSignedValue(bw, enc->dq_y2_ac_, 4);
+  VP8PutSignedValue(bw, enc->dq_uv_dc_, 4);
+  VP8PutSignedValue(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;
+}
+
+//------------------------------------------------------------------------------
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+
+#define KTRAILER_SIZE 8
+
+static int WriteExtensions(VP8Encoder* const enc) {
+  uint8_t buffer[KTRAILER_SIZE];
+  VP8BitWriter* const bw = &enc->bw_;
+  WebPPicture* const pic = enc->pic_;
+
+  // Layer (bytes 0..3)
+  PutLE24(buffer + 0, enc->layer_data_size_);
+  buffer[3] = enc->pic_->colorspace & WEBP_CSP_UV_MASK;
+  if (enc->layer_data_size_ > 0) {
+    assert(enc->use_layer_);
+    // append layer data to last partition
+    if (!VP8BitWriterAppend(&enc->parts_[enc->num_parts_ - 1],
+                            enc->layer_data_, enc->layer_data_size_)) {
+      return WebPEncodingSetError(pic, VP8_ENC_ERROR_BITSTREAM_OUT_OF_MEMORY);
+    }
+  }
+
+  buffer[KTRAILER_SIZE - 1] = 0x01;  // marker
+  if (!VP8BitWriterAppend(bw, buffer, KTRAILER_SIZE)) {
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_BITSTREAM_OUT_OF_MEMORY);
+  }
+  return 1;
+}
+
+#endif    /* WEBP_EXPERIMENTAL_FEATURES */
+
+//------------------------------------------------------------------------------
+
+static size_t GeneratePartition0(VP8Encoder* const enc) {
+  VP8BitWriter* const bw = &enc->bw_;
+  const int mb_size = enc->mb_w_ * enc->mb_h_;
+  uint64_t pos1, pos2, pos3;
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  const int need_extensions = enc->use_layer_;
+#endif
+
+  pos1 = VP8BitWriterPos(bw);
+  VP8BitWriterInit(bw, mb_size * 7 / 8);        // ~7 bits per macroblock
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  VP8PutBitUniform(bw, need_extensions);   // extensions
+#else
+  VP8PutBitUniform(bw, 0);   // colorspace
+#endif
+  VP8PutBitUniform(bw, 0);   // clamp type
+
+  PutSegmentHeader(bw, enc);
+  PutFilterHeader(bw, &enc->filter_hdr_);
+  VP8PutValue(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);
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  if (need_extensions && !WriteExtensions(enc)) {
+    return 0;
+  }
+#endif
+
+  pos3 = VP8BitWriterPos(bw);
+
+  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_;
+    enc->pic_->stats->layer_data_size = (int)enc->layer_data_size_;
+  }
+  return !bw->error_;
+}
+
+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);
+
+  // 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;
+  }
+  // Sanity check.
+  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/src/3rdparty/libwebp/src/enc/token.c b/src/3rdparty/libwebp/src/enc/token.c
new file mode 100644
index 0000000..e696642
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/token.c
@@ -0,0 +1,273 @@
+// 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 "./cost.h"
+#include "./vp8enci.h"
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+// we use pages to reduce the number of memcpy()
+#define MAX_NUM_TOKEN 8192          // max number of token per page
+#define FIXED_PROBA_BIT (1u << 14)
+
+struct VP8Tokens {
+  uint16_t tokens_[MAX_NUM_TOKEN];  // bit#15: bit
+                                    // bit #14: constant proba or idx
+                                    // bits 0..13: slot or constant proba
+  VP8Tokens* next_;
+};
+
+//------------------------------------------------------------------------------
+
+void VP8TBufferInit(VP8TBuffer* const b) {
+  b->tokens_ = NULL;
+  b->pages_ = NULL;
+  b->last_page_ = &b->pages_;
+  b->left_ = 0;
+  b->error_ = 0;
+}
+
+void VP8TBufferClear(VP8TBuffer* const b) {
+  if (b != NULL) {
+    const VP8Tokens* p = b->pages_;
+    while (p != NULL) {
+      const VP8Tokens* const next = p->next_;
+      free((void*)p);
+      p = next;
+    }
+    VP8TBufferInit(b);
+  }
+}
+
+static int TBufferNewPage(VP8TBuffer* const b) {
+  VP8Tokens* const page = b->error_ ? NULL : (VP8Tokens*)malloc(sizeof(*page));
+  if (page == NULL) {
+    b->error_ = 1;
+    return 0;
+  }
+  *b->last_page_ = page;
+  b->last_page_ = &page->next_;
+  b->left_ = MAX_NUM_TOKEN;
+  b->tokens_ = page->tokens_;
+  page->next_ = NULL;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+#define TOKEN_ID(t, b, ctx, p) \
+    ((p) + NUM_PROBAS * ((ctx) + NUM_CTX * ((b) + NUM_BANDS * (t))))
+
+static WEBP_INLINE int AddToken(VP8TBuffer* const b,
+                                int bit, uint32_t proba_idx) {
+  assert(proba_idx < FIXED_PROBA_BIT);
+  assert(bit == 0 || bit == 1);
+  if (b->left_ > 0 || TBufferNewPage(b)) {
+    const int slot = --b->left_;
+    b->tokens_[slot] = (bit << 15) | proba_idx;
+  }
+  return bit;
+}
+
+static WEBP_INLINE void AddConstantToken(VP8TBuffer* const b,
+                                         int bit, int proba) {
+  assert(proba < 256);
+  assert(bit == 0 || 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, int coeff_type, int first, int last,
+                         const int16_t* const coeffs,
+                         VP8TBuffer* const tokens) {
+  int n = first;
+  uint32_t base_id = TOKEN_ID(coeff_type, n, ctx, 0);
+  if (!AddToken(tokens, last >= 0, base_id + 0)) {
+    return 0;
+  }
+
+  while (n < 16) {
+    const int c = coeffs[n++];
+    const int sign = c < 0;
+    int v = sign ? -c : c;
+    if (!AddToken(tokens, v != 0, base_id + 1)) {
+      ctx = 0;
+      base_id = TOKEN_ID(coeff_type, VP8EncBands[n], ctx, 0);
+      continue;
+    }
+    if (!AddToken(tokens, v > 1, base_id + 2)) {
+      ctx = 1;
+    } else {
+      if (!AddToken(tokens, v > 4, base_id + 3)) {
+        if (AddToken(tokens, v != 2, base_id + 4))
+          AddToken(tokens, v == 4, base_id + 5);
+      } else if (!AddToken(tokens, v > 10, base_id + 6)) {
+        if (!AddToken(tokens, v > 6, base_id + 7)) {
+          AddConstantToken(tokens, v == 6, 159);
+        } else {
+          AddConstantToken(tokens, v >= 9, 165);
+          AddConstantToken(tokens, !(v & 1), 145);
+        }
+      } else {
+        int mask;
+        const uint8_t* tab;
+        if (v < 3 + (8 << 1)) {          // VP8Cat3  (3b)
+          AddToken(tokens, 0, base_id + 8);
+          AddToken(tokens, 0, base_id + 9);
+          v -= 3 + (8 << 0);
+          mask = 1 << 2;
+          tab = VP8Cat3;
+        } else if (v < 3 + (8 << 2)) {   // VP8Cat4  (4b)
+          AddToken(tokens, 0, base_id + 8);
+          AddToken(tokens, 1, base_id + 9);
+          v -= 3 + (8 << 1);
+          mask = 1 << 3;
+          tab = VP8Cat4;
+        } else if (v < 3 + (8 << 3)) {   // VP8Cat5  (5b)
+          AddToken(tokens, 1, base_id + 8);
+          AddToken(tokens, 0, base_id + 10);
+          v -= 3 + (8 << 2);
+          mask = 1 << 4;
+          tab = VP8Cat5;
+        } else {                         // VP8Cat6 (11b)
+          AddToken(tokens, 1, base_id + 8);
+          AddToken(tokens, 1, base_id + 10);
+          v -= 3 + (8 << 3);
+          mask = 1 << 10;
+          tab = VP8Cat6;
+        }
+        while (mask) {
+          AddConstantToken(tokens, !!(v & mask), *tab++);
+          mask >>= 1;
+        }
+      }
+      ctx = 2;
+    }
+    AddConstantToken(tokens, sign, 128);
+    base_id = TOKEN_ID(coeff_type, VP8EncBands[n], ctx, 0);
+    if (n == 16 || !AddToken(tokens, n <= last, base_id + 0)) {
+      return 1;   // EOB
+    }
+  }
+  return 1;
+}
+
+#undef TOKEN_ID
+
+//------------------------------------------------------------------------------
+// This function works, but isn't currently used. Saved for later.
+
+#if 0
+
+static void Record(int bit, proba_t* const stats) {
+  proba_t p = *stats;
+  if (p >= 0xffff0000u) {               // an overflow is inbound.
+    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;
+}
+
+void VP8TokenToStats(const VP8TBuffer* const b, proba_t* const stats) {
+  const VP8Tokens* p = b->pages_;
+  while (p != NULL) {
+    const int N = (p->next_ == NULL) ? b->left_ : 0;
+    int n = MAX_NUM_TOKEN;
+    while (n-- > N) {
+      const uint16_t token = p->tokens_[n];
+      if (!(token & FIXED_PROBA_BIT)) {
+        Record((token >> 15) & 1, stats + (token & 0x3fffu));
+      }
+    }
+    p = p->next_;
+  }
+}
+
+#endif   // 0
+
+//------------------------------------------------------------------------------
+// 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_;
+  (void)final_pass;
+  if (b->error_) return 0;
+  while (p != NULL) {
+    const VP8Tokens* const next = p->next_;
+    const int N = (next == NULL) ? b->left_ : 0;
+    int n = MAX_NUM_TOKEN;
+    while (n-- > N) {
+      const uint16_t token = p->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) free((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_;
+  if (b->error_) return 0;
+  while (p != NULL) {
+    const VP8Tokens* const next = p->next_;
+    const int N = (next == NULL) ? b->left_ : 0;
+    int n = MAX_NUM_TOKEN;
+    while (n-- > N) {
+      const uint16_t token = p->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) {
+  (void)b;
+}
+void VP8TBufferClear(VP8TBuffer* const b) {
+  (void)b;
+}
+
+#endif    // !DISABLE_TOKEN_BUFFER
+
diff --git a/src/3rdparty/libwebp/src/enc/tree.c b/src/3rdparty/libwebp/src/enc/tree.c
new file mode 100644
index 0000000..e5d05e5
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/tree.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 "./vp8enci.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) {
+  VP8Proba* 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 VP8Proba* 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])) {
+            VP8PutValue(bw, p0, 8);
+          }
+        }
+      }
+    }
+  }
+  if (VP8PutBitUniform(bw, probas->use_skip_proba_)) {
+    VP8PutValue(bw, probas->skip_proba_, 8);
+  }
+}
+
diff --git a/src/3rdparty/libwebp/src/enc/vp8enci.h b/src/3rdparty/libwebp/src/enc/vp8enci.h
new file mode 100644
index 0000000..71adf6c
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/vp8enci.h
@@ -0,0 +1,570 @@
+// 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_VP8ENCI_H_
+#define WEBP_ENC_VP8ENCI_H_
+
+#include <string.h>     // for memcpy()
+#include "../webp/encode.h"
+#include "../dsp/dsp.h"
+#include "../utils/bit_writer.h"
+#include "../utils/thread.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// Various defines and enums
+
+// version numbers
+#define ENC_MAJ_VERSION 0
+#define ENC_MIN_VERSION 4
+#define ENC_REV_VERSION 0
+
+// intra prediction modes
+enum { B_DC_PRED = 0,   // 4x4 modes
+       B_TM_PRED = 1,
+       B_VE_PRED = 2,
+       B_HE_PRED = 3,
+       B_RD_PRED = 4,
+       B_VR_PRED = 5,
+       B_LD_PRED = 6,
+       B_VL_PRED = 7,
+       B_HD_PRED = 8,
+       B_HU_PRED = 9,
+       NUM_BMODES = B_HU_PRED + 1 - B_DC_PRED,  // = 10
+
+       // Luma16 or UV modes
+       DC_PRED = B_DC_PRED, V_PRED = B_VE_PRED,
+       H_PRED = B_HE_PRED, TM_PRED = B_TM_PRED,
+       NUM_PRED_MODES = 4
+     };
+
+enum { NUM_MB_SEGMENTS = 4,
+       MAX_NUM_PARTITIONS = 8,
+       NUM_TYPES = 4,   // 0: i16-AC,  1: i16-DC,  2:chroma-AC,  3:i4-AC
+       NUM_BANDS = 8,
+       NUM_CTX = 3,
+       NUM_PROBAS = 11,
+       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 16-pixels wide.
+// 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. See VP8Scan[] for accessing the blocks.
+//  Y_OFF  |YYYY| <- original samples  ('yuv_in_')
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+//  U_OFF  |UUVV| V_OFF  (=U_OFF + 8)
+//         |UUVV|
+//         +----+
+//  Y_OFF  |YYYY| <- compressed/decoded samples  ('yuv_out_')
+//         |YYYY|    There are two buffers like this ('yuv_out_'/'yuv_out2_')
+//         |YYYY|
+//         |YYYY|
+//  U_OFF  |UUVV| V_OFF
+//         |UUVV|
+//          x2 (for yuv_out2_)
+//         +----+     Prediction area ('yuv_p_', size = PRED_SIZE)
+// I16DC16 |YYYY|  Intra16 predictions (16x16 block each)
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+// I16TM16 |YYYY|
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+// I16VE16 |YYYY|
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+// I16HE16 |YYYY|
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+//         +----+  Chroma U/V predictions (16x8 block each)
+// C8DC8   |UUVV|
+//         |UUVV|
+// C8TM8   |UUVV|
+//         |UUVV|
+// C8VE8   |UUVV|
+//         |UUVV|
+// C8HE8   |UUVV|
+//         |UUVV|
+//         +----+  Intra 4x4 predictions (4x4 block each)
+//         |YYYY| I4DC4 I4TM4 I4VE4 I4HE4
+//         |YYYY| I4RD4 I4VR4 I4LD4 I4VL4
+//         |YY..| I4HD4 I4HU4 I4TMP
+//         +----+
+#define BPS       16   // this is the common stride
+#define Y_SIZE   (BPS * 16)
+#define UV_SIZE  (BPS * 8)
+#define YUV_SIZE (Y_SIZE + UV_SIZE)
+#define PRED_SIZE (6 * 16 * BPS + 12 * BPS)
+#define Y_OFF    (0)
+#define U_OFF    (Y_SIZE)
+#define V_OFF    (U_OFF + 8)
+#define ALIGN_CST 15
+#define DO_ALIGN(PTR) ((uintptr_t)((PTR) + ALIGN_CST) & ~ALIGN_CST)
+
+extern const int VP8Scan[16 + 4 + 4];           // in quant.c
+extern const int VP8UVModeOffsets[4];           // in analyze.c
+extern const int VP8I16ModeOffsets[4];
+extern const int VP8I4ModeOffsets[NUM_BMODES];
+
+// Layout of prediction blocks
+// intra 16x16
+#define I16DC16 (0 * 16 * BPS)
+#define I16TM16 (1 * 16 * BPS)
+#define I16VE16 (2 * 16 * BPS)
+#define I16HE16 (3 * 16 * BPS)
+// chroma 8x8, two U/V blocks side by side (hence: 16x8 each)
+#define C8DC8 (4 * 16 * BPS)
+#define C8TM8 (4 * 16 * BPS + 8 * BPS)
+#define C8VE8 (5 * 16 * BPS)
+#define C8HE8 (5 * 16 * BPS + 8 * BPS)
+// intra 4x4
+#define I4DC4 (6 * 16 * BPS +  0)
+#define I4TM4 (6 * 16 * BPS +  4)
+#define I4VE4 (6 * 16 * BPS +  8)
+#define I4HE4 (6 * 16 * BPS + 12)
+#define I4RD4 (6 * 16 * BPS + 4 * BPS +  0)
+#define I4VR4 (6 * 16 * BPS + 4 * BPS +  4)
+#define I4LD4 (6 * 16 * BPS + 4 * BPS +  8)
+#define I4VL4 (6 * 16 * BPS + 4 * BPS + 12)
+#define I4HD4 (6 * 16 * BPS + 8 * BPS +  0)
+#define I4HU4 (6 * 16 * BPS + 8 * BPS +  4)
+#define I4TMP (6 * 16 * BPS + 8 * BPS +  8)
+
+typedef int64_t score_t;     // type used for scores, rate, distortion
+#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(int n, int iQ, int B) {
+  return (n * iQ + B) >> QFIX;
+}
+
+// size of histogram used by CollectHistogram.
+#define MAX_COEFF_THRESH   31
+typedef struct VP8Histogram VP8Histogram;
+struct VP8Histogram {
+  // TODO(skal): we only need to store the max_value and last_non_zero actually.
+  int distribution[MAX_COEFF_THRESH + 1];
+};
+
+// Uncomment the following to remove token-buffer code:
+// #define DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+// 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 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
+} VP8SegmentHeader;
+
+// 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];      // 924 bytes
+  StatsArray stats_[NUM_TYPES][NUM_BANDS];       // 4224 bytes
+  CostArray level_cost_[NUM_TYPES][NUM_BANDS];   // 11.4k
+  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
+} VP8Proba;
+
+// 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
+} VP8FilterHeader;
+
+//------------------------------------------------------------------------------
+// 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.
+  uint16_t bias_[16];     // rounding bias
+  uint16_t zthresh_[16];  // value under 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_;
+} VP8SegmentInfo;
+
+// 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
+} VP8ModeScore;
+
+// Iterator structure to iterate through macroblocks, pointing to the
+// right neighbouring data (samples, predictions, contexts, ...)
+typedef struct {
+  int x_, y_;                      // current macroblock
+  int y_stride_, uv_stride_;       // respective strides
+  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
+
+  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_ and yuv_in_/out_*
+  uint8_t yuv_left_mem_[17 + 16 + 16 + 8 + ALIGN_CST];     // memory for *_left_
+  uint8_t yuv_mem_[3 * YUV_SIZE + PRED_SIZE + ALIGN_CST];  // memory for yuv_*
+} 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* 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 final_delta_percent);
+// 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.
+#endif
+  int error_;         // true in case of malloc error
+} VP8TBuffer;
+
+void VP8TBufferInit(VP8TBuffer* const b);    // initialize an empty buffer
+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, int coeff_type, int first, int last,
+                         const int16_t* const coeffs,
+                         VP8TBuffer* const tokens);
+
+// Estimate the final coded size given a set of 'probas'.
+size_t VP8EstimateTokenSize(VP8TBuffer* const b, const uint8_t* const probas);
+
+// unused for now
+void VP8TokenToStats(const VP8TBuffer* const b, proba_t* const stats);
+
+#endif  // !DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+// VP8Encoder
+
+struct VP8Encoder {
+  const WebPConfig* config_;    // user configuration and parameters
+  WebPPicture* pic_;            // input / output picture
+
+  // headers
+  VP8FilterHeader   filter_hdr_;     // filtering information
+  VP8SegmentHeader  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_;
+
+  // enhancement layer
+  int use_layer_;
+  VP8BitWriter layer_bw_;
+  uint8_t* layer_data_;
+  size_t layer_data_size_;
+
+  // 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
+  VP8Proba 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 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)
+};
+
+//------------------------------------------------------------------------------
+// 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 VP8Proba* 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 VP8EncBands[16 + 1];
+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* const it, VP8ModeScore* 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
+
+  // in layer.c
+void VP8EncInitLayer(VP8Encoder* const enc);     // init everything
+void VP8EncCodeLayerBlock(VP8EncIterator* it);   // code one more macroblock
+int VP8EncFinishLayer(VP8Encoder* const enc);    // finalize coding
+void VP8EncDeleteLayer(VP8Encoder* enc);         // reclaim memory
+
+  // in filter.c
+
+// SSIM utils
+typedef struct {
+  double w, xm, ym, xxm, xym, yym;
+} DistoStats;
+void VP8SSIMAddStats(const DistoStats* const src, DistoStats* const dst);
+void VP8SSIMAccumulatePlane(const uint8_t* src1, int stride1,
+                            const uint8_t* src2, int stride2,
+                            int W, int H, DistoStats* const stats);
+double VP8SSIMGet(const DistoStats* const stats);
+double VP8SSIMGetSquaredError(const DistoStats* const stats);
+
+// 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);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_ENC_VP8ENCI_H_ */
diff --git a/src/3rdparty/libwebp/src/enc/vp8l.c b/src/3rdparty/libwebp/src/enc/vp8l.c
new file mode 100644
index 0000000..1572631
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/vp8l.c
@@ -0,0 +1,1168 @@
+// 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 <stdio.h>
+#include <stdlib.h>
+
+#include "./backward_references.h"
+#include "./vp8enci.h"
+#include "./vp8li.h"
+#include "../dsp/lossless.h"
+#include "../utils/bit_writer.h"
+#include "../utils/huffman_encode.h"
+#include "../utils/utils.h"
+#include "../webp/format_constants.h"
+
+#define PALETTE_KEY_RIGHT_SHIFT   22  // Key for 1K buffer.
+#define MAX_HUFF_IMAGE_SIZE       (16 * 1024 * 1024)
+#define MAX_COLORS_FOR_GRAPH      64
+
+// -----------------------------------------------------------------------------
+// Palette
+
+static int CompareColors(const void* p1, const void* p2) {
+  const uint32_t a = *(const uint32_t*)p1;
+  const uint32_t b = *(const uint32_t*)p2;
+  assert(a != b);
+  return (a < b) ? -1 : 1;
+}
+
+// If number of colors in the image is less than or equal to MAX_PALETTE_SIZE,
+// creates a palette and returns true, else returns false.
+static int AnalyzeAndCreatePalette(const WebPPicture* const pic,
+                                   uint32_t palette[MAX_PALETTE_SIZE],
+                                   int* const palette_size) {
+  int i, x, y, key;
+  int num_colors = 0;
+  uint8_t in_use[MAX_PALETTE_SIZE * 4] = { 0 };
+  uint32_t colors[MAX_PALETTE_SIZE * 4];
+  static const uint32_t kHashMul = 0x1e35a7bd;
+  const uint32_t* argb = pic->argb;
+  const int width = pic->width;
+  const int height = pic->height;
+  uint32_t last_pix = ~argb[0];   // so we're sure that last_pix != argb[0]
+
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      if (argb[x] == last_pix) {
+        continue;
+      }
+      last_pix = argb[x];
+      key = (kHashMul * last_pix) >> PALETTE_KEY_RIGHT_SHIFT;
+      while (1) {
+        if (!in_use[key]) {
+          colors[key] = last_pix;
+          in_use[key] = 1;
+          ++num_colors;
+          if (num_colors > MAX_PALETTE_SIZE) {
+            return 0;
+          }
+          break;
+        } else if (colors[key] == last_pix) {
+          // The color is already there.
+          break;
+        } else {
+          // Some other color sits there.
+          // Do linear conflict resolution.
+          ++key;
+          key &= (MAX_PALETTE_SIZE * 4 - 1);  // key mask for 1K buffer.
+        }
+      }
+    }
+    argb += pic->argb_stride;
+  }
+
+  // TODO(skal): could we reuse in_use[] to speed up EncodePalette()?
+  num_colors = 0;
+  for (i = 0; i < (int)(sizeof(in_use) / sizeof(in_use[0])); ++i) {
+    if (in_use[i]) {
+      palette[num_colors] = colors[i];
+      ++num_colors;
+    }
+  }
+
+  qsort(palette, num_colors, sizeof(*palette), CompareColors);
+  *palette_size = num_colors;
+  return 1;
+}
+
+static int AnalyzeEntropy(const uint32_t* argb,
+                          int width, int height, int argb_stride,
+                          double* const nonpredicted_bits,
+                          double* const predicted_bits) {
+  int x, y;
+  const uint32_t* last_line = NULL;
+  uint32_t last_pix = argb[0];    // so we're sure that pix_diff == 0
+
+  VP8LHistogram* nonpredicted = NULL;
+  VP8LHistogram* predicted =
+      (VP8LHistogram*)malloc(2 * sizeof(*predicted));
+  if (predicted == NULL) return 0;
+  nonpredicted = predicted + 1;
+
+  VP8LHistogramInit(predicted, 0);
+  VP8LHistogramInit(nonpredicted, 0);
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      const uint32_t pix = argb[x];
+      const uint32_t pix_diff = VP8LSubPixels(pix, last_pix);
+      if (pix_diff == 0) continue;
+      if (last_line != NULL && pix == last_line[x]) {
+        continue;
+      }
+      last_pix = pix;
+      {
+        const PixOrCopy pix_token = PixOrCopyCreateLiteral(pix);
+        const PixOrCopy pix_diff_token = PixOrCopyCreateLiteral(pix_diff);
+        VP8LHistogramAddSinglePixOrCopy(nonpredicted, &pix_token);
+        VP8LHistogramAddSinglePixOrCopy(predicted, &pix_diff_token);
+      }
+    }
+    last_line = argb;
+    argb += argb_stride;
+  }
+  *nonpredicted_bits = VP8LHistogramEstimateBitsBulk(nonpredicted);
+  *predicted_bits = VP8LHistogramEstimateBitsBulk(predicted);
+  free(predicted);
+  return 1;
+}
+
+static int VP8LEncAnalyze(VP8LEncoder* const enc, WebPImageHint image_hint) {
+  const WebPPicture* const pic = enc->pic_;
+  assert(pic != NULL && pic->argb != NULL);
+
+  enc->use_palette_ =
+      AnalyzeAndCreatePalette(pic, enc->palette_, &enc->palette_size_);
+
+  if (image_hint == WEBP_HINT_GRAPH) {
+    if (enc->use_palette_ && enc->palette_size_ < MAX_COLORS_FOR_GRAPH) {
+      enc->use_palette_ = 0;
+    }
+  }
+
+  if (!enc->use_palette_) {
+    if (image_hint == WEBP_HINT_PHOTO) {
+      enc->use_predict_ = 1;
+      enc->use_cross_color_ = 1;
+    } else {
+      double non_pred_entropy, pred_entropy;
+      if (!AnalyzeEntropy(pic->argb, pic->width, pic->height, pic->argb_stride,
+                          &non_pred_entropy, &pred_entropy)) {
+        return 0;
+      }
+      if (pred_entropy < 0.95 * non_pred_entropy) {
+        enc->use_predict_ = 1;
+        enc->use_cross_color_ = 1;
+      }
+    }
+  }
+
+  return 1;
+}
+
+static int GetHuffBitLengthsAndCodes(
+    const VP8LHistogramSet* const histogram_image,
+    HuffmanTreeCode* const huffman_codes) {
+  int i, k;
+  int ok = 1;
+  uint64_t total_length_size = 0;
+  uint8_t* mem_buf = NULL;
+  const int histogram_image_size = histogram_image->size;
+
+  // 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];
+    for (k = 0; k < 5; ++k) {
+      const int num_symbols = (k == 0) ? VP8LHistogramNumCodes(histo)
+                            : (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) {
+      ok = 0;
+      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;
+    }
+  }
+
+  // Create Huffman trees.
+  for (i = 0; ok && (i < histogram_image_size); ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[5 * i];
+    VP8LHistogram* const histo = histogram_image->histograms[i];
+    ok = ok && VP8LCreateHuffmanTree(histo->literal_, 15, codes + 0);
+    ok = ok && VP8LCreateHuffmanTree(histo->red_, 15, codes + 1);
+    ok = ok && VP8LCreateHuffmanTree(histo->blue_, 15, codes + 2);
+    ok = ok && VP8LCreateHuffmanTree(histo->alpha_, 15, codes + 3);
+    ok = ok && VP8LCreateHuffmanTree(histo->distance_, 15, codes + 4);
+  }
+
+ End:
+  if (!ok) {
+    free(mem_buf);
+    // If one VP8LCreateHuffmanTree() above fails, we need to clean up behind.
+    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;
+    }
+  }
+  VP8LWriteBits(bw, 4, codes_to_store - 4);
+  for (i = 0; i < codes_to_store; ++i) {
+    VP8LWriteBits(bw, 3, code_length_bitdepth[kStorageOrder[i]]);
+  }
+}
+
+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;
+    VP8LWriteBits(bw, huffman_code->code_lengths[ix], huffman_code->codes[ix]);
+    switch (ix) {
+      case 16:
+        VP8LWriteBits(bw, 2, extra_bits);
+        break;
+      case 17:
+        VP8LWriteBits(bw, 3, extra_bits);
+        break;
+      case 18:
+        VP8LWriteBits(bw, 7, extra_bits);
+        break;
+    }
+  }
+}
+
+static int StoreFullHuffmanCode(VP8LBitWriter* const bw,
+                                const HuffmanTreeCode* const tree) {
+  int ok = 0;
+  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;
+  HuffmanTreeToken* const tokens =
+      (HuffmanTreeToken*)WebPSafeMalloc((uint64_t)max_tokens, sizeof(*tokens));
+  if (tokens == NULL) return 0;
+
+  huffman_code.num_symbols = CODE_LENGTH_CODES;
+  huffman_code.code_lengths = code_length_bitdepth;
+  huffman_code.codes = code_length_bitdepth_symbols;
+
+  VP8LWriteBits(bw, 1, 0);
+  num_tokens = VP8LCreateCompressedHuffmanTree(tree, tokens, max_tokens);
+  {
+    int histogram[CODE_LENGTH_CODES] = { 0 };
+    int i;
+    for (i = 0; i < num_tokens; ++i) {
+      ++histogram[tokens[i].code];
+    }
+
+    if (!VP8LCreateHuffmanTree(histogram, 7, &huffman_code)) {
+      goto End;
+    }
+  }
+
+  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;
+    VP8LWriteBits(bw, 1, write_trimmed_length);
+    if (write_trimmed_length) {
+      const int nbits = VP8LBitsLog2Ceiling(trimmed_length - 1);
+      const int nbitpairs = (nbits == 0) ? 1 : (nbits + 1) / 2;
+      VP8LWriteBits(bw, 3, nbitpairs - 1);
+      assert(trimmed_length >= 2);
+      VP8LWriteBits(bw, nbitpairs * 2, trimmed_length - 2);
+    }
+    StoreHuffmanTreeToBitMask(bw, tokens, length, &huffman_code);
+  }
+  ok = 1;
+ End:
+  free(tokens);
+  return ok;
+}
+
+static int StoreHuffmanCode(VP8LBitWriter* const bw,
+                            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
+    VP8LWriteBits(bw, 4, 0x01);
+    return 1;
+  } else if (count <= 2 && symbols[0] < kMaxSymbol && symbols[1] < kMaxSymbol) {
+    VP8LWriteBits(bw, 1, 1);  // Small tree marker to encode 1 or 2 symbols.
+    VP8LWriteBits(bw, 1, count - 1);
+    if (symbols[0] <= 1) {
+      VP8LWriteBits(bw, 1, 0);  // Code bit for small (1 bit) symbol value.
+      VP8LWriteBits(bw, 1, symbols[0]);
+    } else {
+      VP8LWriteBits(bw, 1, 1);
+      VP8LWriteBits(bw, 8, symbols[0]);
+    }
+    if (count == 2) {
+      VP8LWriteBits(bw, 8, symbols[1]);
+    }
+    return 1;
+  } else {
+    return StoreFullHuffmanCode(bw, huffman_code);
+  }
+}
+
+static 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];
+  VP8LWriteBits(bw, depth, symbol);
+}
+
+static void StoreImageToBitMask(
+    VP8LBitWriter* const bw, int width, int histo_bits,
+    const VP8LBackwardRefs* const refs,
+    const uint16_t* histogram_symbols,
+    const HuffmanTreeCode* const huffman_codes) {
+  // x and y trace the position in the image.
+  int x = 0;
+  int y = 0;
+  const int histo_xsize = histo_bits ? VP8LSubSampleSize(width, histo_bits) : 1;
+  int i;
+  for (i = 0; i < refs->size; ++i) {
+    const PixOrCopy* const v = &refs->refs[i];
+    const int histogram_ix = histogram_symbols[histo_bits ?
+                                               (y >> histo_bits) * histo_xsize +
+                                               (x >> histo_bits) : 0];
+    const HuffmanTreeCode* const codes = huffman_codes + 5 * histogram_ix;
+    if (PixOrCopyIsCacheIdx(v)) {
+      const int code = PixOrCopyCacheIdx(v);
+      const int literal_ix = 256 + NUM_LENGTH_CODES + code;
+      WriteHuffmanCode(bw, codes, literal_ix);
+    } else if (PixOrCopyIsLiteral(v)) {
+      static const int 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 {
+      int bits, n_bits;
+      int code, distance;
+
+      VP8LPrefixEncode(v->len, &code, &n_bits, &bits);
+      WriteHuffmanCode(bw, codes, 256 + code);
+      VP8LWriteBits(bw, n_bits, bits);
+
+      distance = PixOrCopyDistance(v);
+      VP8LPrefixEncode(distance, &code, &n_bits, &bits);
+      WriteHuffmanCode(bw, codes + 4, code);
+      VP8LWriteBits(bw, n_bits, bits);
+    }
+    x += PixOrCopyLength(v);
+    while (x >= width) {
+      x -= width;
+      ++y;
+    }
+  }
+}
+
+// Special case of EncodeImageInternal() for cache-bits=0, histo_bits=31
+static int EncodeImageNoHuffman(VP8LBitWriter* const bw,
+                                const uint32_t* const argb,
+                                int width, int height, int quality) {
+  int i;
+  int ok = 0;
+  VP8LBackwardRefs refs;
+  HuffmanTreeCode huffman_codes[5] = { { 0, NULL, NULL } };
+  const uint16_t histogram_symbols[1] = { 0 };    // only one tree, one symbol
+  VP8LHistogramSet* const histogram_image = VP8LAllocateHistogramSet(1, 0);
+  if (histogram_image == NULL) return 0;
+
+  // Calculate backward references from ARGB image.
+  if (!VP8LGetBackwardReferences(width, height, argb, quality, 0, 1, &refs)) {
+    goto Error;
+  }
+  // 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)) {
+    goto Error;
+  }
+
+  // No color cache, no Huffman image.
+  VP8LWriteBits(bw, 1, 0);
+
+  // Store Huffman codes.
+  for (i = 0; i < 5; ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[i];
+    if (!StoreHuffmanCode(bw, codes)) {
+      goto Error;
+    }
+    ClearHuffmanTreeIfOnlyOneSymbol(codes);
+  }
+
+  // Store actual literals.
+  StoreImageToBitMask(bw, width, 0, &refs, histogram_symbols, huffman_codes);
+  ok = 1;
+
+ Error:
+  free(histogram_image);
+  VP8LClearBackwardRefs(&refs);
+  free(huffman_codes[0].codes);
+  return ok;
+}
+
+static int EncodeImageInternal(VP8LBitWriter* const bw,
+                               const uint32_t* const argb,
+                               int width, int height, int quality,
+                               int cache_bits, int histogram_bits) {
+  int ok = 0;
+  const int use_2d_locality = 1;
+  const int use_color_cache = (cache_bits > 0);
+  const uint32_t histogram_image_xysize =
+      VP8LSubSampleSize(width, histogram_bits) *
+      VP8LSubSampleSize(height, histogram_bits);
+  VP8LHistogramSet* histogram_image =
+      VP8LAllocateHistogramSet(histogram_image_xysize, 0);
+  int histogram_image_size = 0;
+  size_t bit_array_size = 0;
+  HuffmanTreeCode* huffman_codes = NULL;
+  VP8LBackwardRefs refs;
+  uint16_t* const histogram_symbols =
+      (uint16_t*)WebPSafeMalloc((uint64_t)histogram_image_xysize,
+                                sizeof(*histogram_symbols));
+  assert(histogram_bits >= MIN_HUFFMAN_BITS);
+  assert(histogram_bits <= MAX_HUFFMAN_BITS);
+
+  if (histogram_image == NULL || histogram_symbols == NULL) {
+    free(histogram_image);
+    free(histogram_symbols);
+    return 0;
+  }
+
+  // Calculate backward references from ARGB image.
+  if (!VP8LGetBackwardReferences(width, height, argb, quality, cache_bits,
+                                 use_2d_locality, &refs)) {
+    goto Error;
+  }
+  // Build histogram image and symbols from backward references.
+  if (!VP8LGetHistoImageSymbols(width, height, &refs,
+                                quality, histogram_bits, cache_bits,
+                                histogram_image,
+                                histogram_symbols)) {
+    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));
+  if (huffman_codes == NULL ||
+      !GetHuffBitLengthsAndCodes(histogram_image, huffman_codes)) {
+    goto Error;
+  }
+  // Free combined histograms.
+  free(histogram_image);
+  histogram_image = NULL;
+
+  // Color Cache parameters.
+  VP8LWriteBits(bw, 1, use_color_cache);
+  if (use_color_cache) {
+    VP8LWriteBits(bw, 4, cache_bits);
+  }
+
+  // Huffman image + meta huffman.
+  {
+    const int write_histogram_image = (histogram_image_size > 1);
+    VP8LWriteBits(bw, 1, write_histogram_image);
+    if (write_histogram_image) {
+      uint32_t* const histogram_argb =
+          (uint32_t*)WebPSafeMalloc((uint64_t)histogram_image_xysize,
+                                    sizeof(*histogram_argb));
+      int max_index = 0;
+      uint32_t i;
+      if (histogram_argb == NULL) goto Error;
+      for (i = 0; i < histogram_image_xysize; ++i) {
+        const int symbol_index = histogram_symbols[i] & 0xffff;
+        histogram_argb[i] = 0xff000000 | (symbol_index << 8);
+        if (symbol_index >= max_index) {
+          max_index = symbol_index + 1;
+        }
+      }
+      histogram_image_size = max_index;
+
+      VP8LWriteBits(bw, 3, histogram_bits - 2);
+      ok = EncodeImageNoHuffman(bw, histogram_argb,
+                                VP8LSubSampleSize(width, histogram_bits),
+                                VP8LSubSampleSize(height, histogram_bits),
+                                quality);
+      free(histogram_argb);
+      if (!ok) goto Error;
+    }
+  }
+
+  // Store Huffman codes.
+  {
+    int i;
+    for (i = 0; i < 5 * histogram_image_size; ++i) {
+      HuffmanTreeCode* const codes = &huffman_codes[i];
+      if (!StoreHuffmanCode(bw, codes)) goto Error;
+      ClearHuffmanTreeIfOnlyOneSymbol(codes);
+    }
+  }
+
+  // Store actual literals.
+  StoreImageToBitMask(bw, width, histogram_bits, &refs,
+                      histogram_symbols, huffman_codes);
+  ok = 1;
+
+ Error:
+  free(histogram_image);
+
+  VP8LClearBackwardRefs(&refs);
+  if (huffman_codes != NULL) {
+    free(huffman_codes->codes);
+    free(huffman_codes);
+  }
+  free(histogram_symbols);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+// Transforms
+
+// Check if it would be a good idea to subtract green from red and blue. We
+// only impact entropy in red/blue components, don't bother to look at others.
+static int EvalAndApplySubtractGreen(VP8LEncoder* const enc,
+                                     int width, int height,
+                                     VP8LBitWriter* const bw) {
+  if (!enc->use_palette_) {
+    int i;
+    const uint32_t* const argb = enc->argb_;
+    double bit_cost_before, bit_cost_after;
+    VP8LHistogram* const histo = (VP8LHistogram*)malloc(sizeof(*histo));
+    if (histo == NULL) return 0;
+
+    VP8LHistogramInit(histo, 1);
+    for (i = 0; i < width * height; ++i) {
+      const uint32_t c = argb[i];
+      ++histo->red_[(c >> 16) & 0xff];
+      ++histo->blue_[(c >> 0) & 0xff];
+    }
+    bit_cost_before = VP8LHistogramEstimateBits(histo);
+
+    VP8LHistogramInit(histo, 1);
+    for (i = 0; i < width * height; ++i) {
+      const uint32_t c = argb[i];
+      const int green = (c >> 8) & 0xff;
+      ++histo->red_[((c >> 16) - green) & 0xff];
+      ++histo->blue_[((c >> 0) - green) & 0xff];
+    }
+    bit_cost_after = VP8LHistogramEstimateBits(histo);
+    free(histo);
+
+    // Check if subtracting green yields low entropy.
+    enc->use_subtract_green_ = (bit_cost_after < bit_cost_before);
+    if (enc->use_subtract_green_) {
+      VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
+      VP8LWriteBits(bw, 2, SUBTRACT_GREEN);
+      VP8LSubtractGreenFromBlueAndRed(enc->argb_, width * height);
+    }
+  }
+  return 1;
+}
+
+static int ApplyPredictFilter(const VP8LEncoder* const enc,
+                              int width, int height, int quality,
+                              VP8LBitWriter* const bw) {
+  const int pred_bits = enc->transform_bits_;
+  const int transform_width = VP8LSubSampleSize(width, pred_bits);
+  const int transform_height = VP8LSubSampleSize(height, pred_bits);
+
+  VP8LResidualImage(width, height, pred_bits, enc->argb_, enc->argb_scratch_,
+                    enc->transform_data_);
+  VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
+  VP8LWriteBits(bw, 2, PREDICTOR_TRANSFORM);
+  assert(pred_bits >= 2);
+  VP8LWriteBits(bw, 3, pred_bits - 2);
+  if (!EncodeImageNoHuffman(bw, enc->transform_data_,
+                            transform_width, transform_height, quality)) {
+    return 0;
+  }
+  return 1;
+}
+
+static int ApplyCrossColorFilter(const VP8LEncoder* const enc,
+                                 int width, int height, int quality,
+                                 VP8LBitWriter* const bw) {
+  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);
+  const int step = (quality < 25) ? 32 : (quality > 50) ? 8 : 16;
+
+  VP8LColorSpaceTransform(width, height, ccolor_transform_bits, step,
+                          enc->argb_, enc->transform_data_);
+  VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
+  VP8LWriteBits(bw, 2, CROSS_COLOR_TRANSFORM);
+  assert(ccolor_transform_bits >= 2);
+  VP8LWriteBits(bw, 3, ccolor_transform_bits - 2);
+  if (!EncodeImageNoHuffman(bw, enc->transform_data_,
+                            transform_width, transform_height, quality)) {
+    return 0;
+  }
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+
+static WebPEncodingError 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);
+  if (!pic->writer(riff, sizeof(riff), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+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);
+
+  VP8LWriteBits(bw, VP8L_IMAGE_SIZE_BITS, width);
+  VP8LWriteBits(bw, VP8L_IMAGE_SIZE_BITS, height);
+  return !bw->error_;
+}
+
+static int WriteRealAlphaAndVersion(VP8LBitWriter* const bw, int has_alpha) {
+  VP8LWriteBits(bw, 1, has_alpha);
+  VP8LWriteBits(bw, VP8L_VERSION_BITS, VP8L_VERSION);
+  return !bw->error_;
+}
+
+static WebPEncodingError WriteImage(const WebPPicture* const pic,
+                                    VP8LBitWriter* const bw,
+                                    size_t* const coded_size) {
+  WebPEncodingError err = VP8_ENC_OK;
+  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;
+
+  err = WriteRiffHeader(pic, riff_size, vp8l_size);
+  if (err != VP8_ENC_OK) goto Error;
+
+  if (!pic->writer(webpll_data, webpll_size, pic)) {
+    err = VP8_ENC_ERROR_BAD_WRITE;
+    goto Error;
+  }
+
+  if (pad) {
+    const uint8_t pad_byte[1] = { 0 };
+    if (!pic->writer(pad_byte, 1, pic)) {
+      err = VP8_ENC_ERROR_BAD_WRITE;
+      goto Error;
+    }
+  }
+  *coded_size = CHUNK_HEADER_SIZE + riff_size;
+  return VP8_ENC_OK;
+
+ Error:
+  return err;
+}
+
+// -----------------------------------------------------------------------------
+
+// Allocates the memory for argb (W x H) buffer, 2 rows of context for
+// prediction and transform data.
+static WebPEncodingError AllocateTransformBuffer(VP8LEncoder* const enc,
+                                                 int width, int height) {
+  WebPEncodingError err = VP8_ENC_OK;
+  const int tile_size = 1 << enc->transform_bits_;
+  const uint64_t image_size = width * height;
+  const uint64_t argb_scratch_size = tile_size * width + width;
+  const uint64_t transform_data_size =
+      (uint64_t)VP8LSubSampleSize(width, enc->transform_bits_) *
+      (uint64_t)VP8LSubSampleSize(height, enc->transform_bits_);
+  const uint64_t total_size =
+      image_size + argb_scratch_size + transform_data_size;
+  uint32_t* mem = (uint32_t*)WebPSafeMalloc(total_size, sizeof(*mem));
+  if (mem == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+  enc->argb_ = mem;
+  mem += image_size;
+  enc->argb_scratch_ = mem;
+  mem += argb_scratch_size;
+  enc->transform_data_ = mem;
+  enc->current_width_ = width;
+
+ Error:
+  return err;
+}
+
+static void ApplyPalette(uint32_t* src, uint32_t* dst,
+                         uint32_t src_stride, uint32_t dst_stride,
+                         const uint32_t* palette, int palette_size,
+                         int width, int height, int xbits, uint8_t* row) {
+  int i, x, y;
+  int use_LUT = 1;
+  for (i = 0; i < palette_size; ++i) {
+    if ((palette[i] & 0xffff00ffu) != 0) {
+      use_LUT = 0;
+      break;
+    }
+  }
+
+  if (use_LUT) {
+    uint8_t inv_palette[MAX_PALETTE_SIZE] = { 0 };
+    for (i = 0; i < palette_size; ++i) {
+      const int color = (palette[i] >> 8) & 0xff;
+      inv_palette[color] = i;
+    }
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const int color = (src[x] >> 8) & 0xff;
+        row[x] = inv_palette[color];
+      }
+      VP8LBundleColorMap(row, width, xbits, dst);
+      src += src_stride;
+      dst += dst_stride;
+    }
+  } else {
+    // Use 1 pixel cache for ARGB pixels.
+    uint32_t last_pix = palette[0];
+    int last_idx = 0;
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const uint32_t pix = src[x];
+        if (pix != last_pix) {
+          for (i = 0; i < palette_size; ++i) {
+            if (pix == palette[i]) {
+              last_idx = i;
+              last_pix = pix;
+              break;
+            }
+          }
+        }
+        row[x] = last_idx;
+      }
+      VP8LBundleColorMap(row, width, xbits, dst);
+      src += src_stride;
+      dst += dst_stride;
+    }
+  }
+}
+
+// Note: Expects "enc->palette_" to be set properly.
+// Also, "enc->palette_" will be modified after this call and should not be used
+// later.
+static WebPEncodingError EncodePalette(VP8LBitWriter* const bw,
+                                       VP8LEncoder* const enc, int quality) {
+  WebPEncodingError err = VP8_ENC_OK;
+  int i;
+  const WebPPicture* const pic = enc->pic_;
+  uint32_t* src = pic->argb;
+  uint32_t* dst;
+  const int width = pic->width;
+  const int height = pic->height;
+  uint32_t* const palette = enc->palette_;
+  const int palette_size = enc->palette_size_;
+  uint8_t* row = NULL;
+  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;
+  }
+
+  err = AllocateTransformBuffer(enc, VP8LSubSampleSize(width, xbits), height);
+  if (err != VP8_ENC_OK) goto Error;
+  dst = enc->argb_;
+
+  row = (uint8_t*)WebPSafeMalloc((uint64_t)width, sizeof(*row));
+  if (row == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
+
+  ApplyPalette(src, dst, pic->argb_stride, enc->current_width_,
+               palette, palette_size, width, height, xbits, row);
+
+  // Save palette to bitstream.
+  VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
+  VP8LWriteBits(bw, 2, COLOR_INDEXING_TRANSFORM);
+  assert(palette_size >= 1);
+  VP8LWriteBits(bw, 8, palette_size - 1);
+  for (i = palette_size - 1; i >= 1; --i) {
+    palette[i] = VP8LSubPixels(palette[i], palette[i - 1]);
+  }
+  if (!EncodeImageNoHuffman(bw, palette, palette_size, 1, quality)) {
+    err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
+    goto Error;
+  }
+
+ Error:
+  free(row);
+  return err;
+}
+
+// -----------------------------------------------------------------------------
+
+static int GetHistoBits(int method, int use_palette, int width, int height) {
+  const uint64_t hist_size = sizeof(VP8LHistogram);
+  // Make tile size a function of encoding method (Range: 0 to 6).
+  int histo_bits = (use_palette ? 9 : 7) - method;
+  while (1) {
+    const uint64_t huff_image_size = VP8LSubSampleSize(width, histo_bits) *
+                                     VP8LSubSampleSize(height, histo_bits) *
+                                     hist_size;
+    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 void FinishEncParams(VP8LEncoder* const enc) {
+  const WebPConfig* const config = enc->config_;
+  const WebPPicture* const pic = enc->pic_;
+  const int method = config->method;
+  const float quality = config->quality;
+  const int use_palette = enc->use_palette_;
+  enc->transform_bits_ = (method < 4) ? 5 : (method > 4) ? 3 : 4;
+  enc->histo_bits_ = GetHistoBits(method, use_palette, pic->width, pic->height);
+  enc->cache_bits_ = (quality <= 25.f) ? 0 : 7;
+}
+
+// -----------------------------------------------------------------------------
+// VP8LEncoder
+
+static VP8LEncoder* VP8LEncoderNew(const WebPConfig* const config,
+                                   const WebPPicture* const picture) {
+  VP8LEncoder* const enc = (VP8LEncoder*)calloc(1, sizeof(*enc));
+  if (enc == NULL) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return NULL;
+  }
+  enc->config_ = config;
+  enc->pic_ = picture;
+
+  VP8LDspInit();
+
+  return enc;
+}
+
+static void VP8LEncoderDelete(VP8LEncoder* enc) {
+  free(enc->argb_);
+  free(enc);
+}
+
+// -----------------------------------------------------------------------------
+// Main call
+
+WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
+                                   const WebPPicture* const picture,
+                                   VP8LBitWriter* const bw) {
+  WebPEncodingError err = VP8_ENC_OK;
+  const int quality = (int)config->quality;
+  const int width = picture->width;
+  const int height = picture->height;
+  VP8LEncoder* const enc = VP8LEncoderNew(config, picture);
+  const size_t byte_position = VP8LBitWriterNumBytes(bw);
+
+  if (enc == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  // ---------------------------------------------------------------------------
+  // Analyze image (entropy, num_palettes etc)
+
+  if (!VP8LEncAnalyze(enc, config->image_hint)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  FinishEncParams(enc);
+
+  if (enc->use_palette_) {
+    err = EncodePalette(bw, enc, quality);
+    if (err != VP8_ENC_OK) goto Error;
+    // Color cache is disabled for palette.
+    enc->cache_bits_ = 0;
+  }
+
+  // In case image is not packed.
+  if (enc->argb_ == NULL) {
+    int y;
+    err = AllocateTransformBuffer(enc, width, height);
+    if (err != VP8_ENC_OK) goto Error;
+    for (y = 0; y < height; ++y) {
+      memcpy(enc->argb_ + y * width,
+             picture->argb + y * picture->argb_stride,
+             width * sizeof(*enc->argb_));
+    }
+    enc->current_width_ = width;
+  }
+
+  // ---------------------------------------------------------------------------
+  // Apply transforms and write transform data.
+
+  if (!EvalAndApplySubtractGreen(enc, enc->current_width_, height, bw)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  if (enc->use_predict_) {
+    if (!ApplyPredictFilter(enc, enc->current_width_, height, quality, bw)) {
+      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
+      goto Error;
+    }
+  }
+
+  if (enc->use_cross_color_) {
+    if (!ApplyCrossColorFilter(enc, enc->current_width_, height, quality, bw)) {
+      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
+      goto Error;
+    }
+  }
+
+  VP8LWriteBits(bw, 1, !TRANSFORM_PRESENT);  // No more transforms.
+
+  // ---------------------------------------------------------------------------
+  // Estimate the color cache size.
+
+  if (enc->cache_bits_ > 0) {
+    if (!VP8LCalculateEstimateForCacheSize(enc->argb_, enc->current_width_,
+                                           height, &enc->cache_bits_)) {
+      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
+      goto Error;
+    }
+  }
+
+  // ---------------------------------------------------------------------------
+  // Encode and write the transformed image.
+
+  if (!EncodeImageInternal(bw, enc->argb_, enc->current_width_, height,
+                           quality, enc->cache_bits_, enc->histo_bits_)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  if (picture->stats != NULL) {
+    WebPAuxStats* const stats = picture->stats;
+    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)(VP8LBitWriterNumBytes(bw) - byte_position);
+  }
+
+ Error:
+  VP8LEncoderDelete(enc);
+  return err;
+}
+
+int VP8LEncodeImage(const WebPConfig* const config,
+                    const WebPPicture* const picture) {
+  int width, height;
+  int has_alpha;
+  size_t coded_size;
+  int percent = 0;
+  WebPEncodingError err = VP8_ENC_OK;
+  VP8LBitWriter bw;
+
+  if (picture == NULL) return 0;
+
+  if (config == NULL || picture->argb == NULL) {
+    err = VP8_ENC_ERROR_NULL_PARAMETER;
+    WebPEncodingSetError(picture, err);
+    return 0;
+  }
+
+  width = picture->width;
+  height = picture->height;
+  if (!VP8LBitWriterInit(&bw, (width * height) >> 1)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  if (!WebPReportProgress(picture, 1, &percent)) {
+ UserAbort:
+    err = 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)) {
+    err = 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)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  if (!WebPReportProgress(picture, 5, &percent)) goto UserAbort;
+
+  // Encode main image stream.
+  err = VP8LEncodeStream(config, picture, &bw);
+  if (err != VP8_ENC_OK) goto Error;
+
+  // TODO(skal): have a fine-grained progress report in VP8LEncodeStream().
+  if (!WebPReportProgress(picture, 90, &percent)) goto UserAbort;
+
+  // Finish the RIFF chunk.
+  err = WriteImage(picture, &bw, &coded_size);
+  if (err != VP8_ENC_OK) goto Error;
+
+  if (!WebPReportProgress(picture, 100, &percent)) goto UserAbort;
+
+  // Save size.
+  if (picture->stats != NULL) {
+    picture->stats->coded_size += (int)coded_size;
+    picture->stats->lossless_size = (int)coded_size;
+  }
+
+  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_) err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+  VP8LBitWriterDestroy(&bw);
+  if (err != VP8_ENC_OK) {
+    WebPEncodingSetError(picture, err);
+    return 0;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/src/3rdparty/libwebp/src/enc/vp8li.h b/src/3rdparty/libwebp/src/enc/vp8li.h
new file mode 100644
index 0000000..96d6fae
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/vp8li.h
@@ -0,0 +1,70 @@
+// 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_H_
+#define WEBP_ENC_VP8LI_H_
+
+#include "./histogram.h"
+#include "../utils/bit_writer.h"
+#include "../webp/encode.h"
+#include "../webp/format_constants.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  const WebPConfig* config_;    // user configuration and parameters
+  const WebPPicture* pic_;      // input picture.
+
+  uint32_t* argb_;              // Transformed argb image data.
+  uint32_t* argb_scratch_;      // Scratch memory for argb rows
+                                // (used for prediction).
+  uint32_t* transform_data_;    // Scratch memory for transform data.
+  int       current_width_;     // Corresponds to packed image width.
+
+  // Encoding parameters derived from quality parameter.
+  int histo_bits_;
+  int 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];
+} 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.
+WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
+                                   const WebPPicture* const picture,
+                                   VP8LBitWriter* const bw);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_ENC_VP8LI_H_ */
diff --git a/src/3rdparty/libwebp/src/enc/webpenc.c b/src/3rdparty/libwebp/src/enc/webpenc.c
new file mode 100644
index 0000000..207cce6
--- /dev/null
+++ b/src/3rdparty/libwebp/src/enc/webpenc.c
@@ -0,0 +1,404 @@
+// 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 "./vp8enci.h"
+#include "./vp8li.h"
+#include "../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;
+}
+
+//------------------------------------------------------------------------------
+// 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;
+}
+
+//------------------------------------------------------------------------------
+// VP8Encoder
+//------------------------------------------------------------------------------
+
+static void ResetSegmentHeader(VP8Encoder* const enc) {
+  VP8SegmentHeader* 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) {
+  VP8FilterHeader* 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 |
+//-------------------+---+---+---+---+---+---+---+
+// basic rd-opt      |   |   |   | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// disto-score i4/16 |   |   | 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.
+
+  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 (768x510 picture)
+// Memory used:
+//              encoder: 33919
+//          block cache: 2880
+//                 info: 3072
+//                preds: 24897
+//          top samples: 1623
+//             non-zero: 196
+//             lf-stats: 2048
+//                total: 68635
+// Transient object sizes:
+//       VP8EncIterator: 352
+//         VP8ModeScore: 912
+//       VP8SegmentInfo: 532
+//             VP8Proba: 31032
+//              LFStats: 2048
+// Picture size (yuv): 589824
+
+static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
+                                  WebPPicture* const picture) {
+  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(uint8_t);
+  const int top_stride = mb_w * 16;
+  const size_t nz_size = (mb_w + 1) * sizeof(uint32_t) + ALIGN_CST;
+  const size_t info_size = mb_w * mb_h * sizeof(VP8MBInfo);
+  const size_t samples_size = 2 * top_stride * sizeof(uint8_t)  // top-luma/u/v
+                            + ALIGN_CST;                        // align all
+  const size_t lf_stats_size =
+      config->autofilter ? sizeof(LFStats) + ALIGN_CST : 0;
+  VP8Encoder* enc;
+  uint8_t* mem;
+  const uint64_t size = (uint64_t)sizeof(VP8Encoder)   // main struct
+                      + ALIGN_CST                      // cache alignment
+                      + info_size                      // modes info
+                      + preds_size                     // prediction modes
+                      + samples_size                   // top/left samples
+                      + 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"
+         "            non-zero: %ld\n"
+         "            lf-stats: %ld\n"
+         "               total: %ld\n",
+         sizeof(VP8Encoder) + ALIGN_CST, info_size,
+         preds_size, samples_size, nz_size, lf_stats_size, size);
+  printf("Transient object sizes:\n"
+         "      VP8EncIterator: %ld\n"
+         "        VP8ModeScore: %ld\n"
+         "      VP8SegmentInfo: %ld\n"
+         "            VP8Proba: %ld\n"
+         "             LFStats: %ld\n",
+         sizeof(VP8EncIterator), sizeof(VP8ModeScore),
+         sizeof(VP8SegmentInfo), sizeof(VP8Proba),
+         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*)DO_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_ = ((uint8_t*)mem) + 1 + enc->preds_w_;
+  mem += preds_w * preds_h * sizeof(uint8_t);
+  enc->nz_ = 1 + (uint32_t*)DO_ALIGN(mem);
+  mem += nz_size;
+  enc->lf_stats_ = lf_stats_size ? (LFStats*)DO_ALIGN(mem) : NULL;
+  mem += lf_stats_size;
+
+  // top samples (all 16-aligned)
+  mem = (uint8_t*)DO_ALIGN(mem);
+  enc->y_top_ = (uint8_t*)mem;
+  enc->uv_top_ = enc->y_top_ + top_stride;
+  mem += 2 * top_stride;
+  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);
+
+  VP8EncInitAlpha(enc);
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  VP8EncInitLayer(enc);
+#endif
+
+  VP8TBufferInit(&enc->tokens_);
+  return enc;
+}
+
+static int DeleteVP8Encoder(VP8Encoder* enc) {
+  int ok = 1;
+  if (enc != NULL) {
+    ok = VP8EncDeleteAlpha(enc);
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    VP8EncDeleteLayer(enc);
+#endif
+    VP8TBufferClear(&enc->tokens_);
+    free(enc);
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+
+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);
+}
+
+static void StoreStats(VP8Encoder* const enc) {
+  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];
+    }
+  }
+  WebPReportProgress(enc->pic_, 100, &enc->percent_);  // done!
+}
+
+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 (pic->width <= 0 || pic->height <= 0)
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_BAD_DIMENSION);
+  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->y == NULL || pic->u == NULL || pic->v == NULL) {
+      // Make sure we have YUVA samples.
+      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;
+      }
+    }
+
+    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);
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    ok = ok && VP8EncFinishLayer(enc);
+#endif
+
+    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;
+    }
+
+    ok = VP8LEncodeImage(config, pic);  // Sets pic->error in case of problem.
+  }
+
+  return ok;
+}
+