-Upgraded webp to a MUCH newer version. Hoping it fixes some bugs in the process. Keeping old version just in case for now.

-Added ability to convert xml and tscn scenes to binary on export, makes loading of larger scenes faster

80 files changed, 15471 insertions, 12009 deletions

diff --git a/drivers/webp/SCsub b/drivers/webp/SCsub
index 5596edbe09..f65bd13dba 100644
--- a/drivers/webp/SCsub
+++ b/drivers/webp/SCsub
@@ -1,63 +1,115 @@
 Import('env')
 
-
 webp_sources = [
-	"webp/mux/muxedit.c",
-	"webp/mux/muxread.c",
-	"webp/mux/muxinternal.c",
-	"webp/mux/demux.c",
-	"webp/enc/tree.c",
-	"webp/enc/analysis.c",
-	"webp/enc/backward_references.c",
-	"webp/enc/alpha.c",
-	"webp/enc/picture.c",
-	"webp/enc/frame.c",
-	"webp/enc/webpenc.c",
-	"webp/enc/cost.c",
-	"webp/enc/filter.c",
-	"webp/enc/vp8l.c",
-	"webp/enc/quant.c",
-	"webp/enc/histogram.c",
-	"webp/enc/syntax.c",
-	"webp/enc/config.c",
-	"webp/enc/layer.c",
-	"webp/enc/iterator.c",
-	"webp/dsp/dec_sse2.c",
-	"webp/dsp/upsampling_sse2.c",
-	"webp/dsp/dec_neon.c",
-	"webp/dsp/enc.c",
-	"webp/dsp/enc_sse2.c",
-	"webp/dsp/upsampling.c",
-	"webp/dsp/lossless.c",
-	"webp/dsp/cpu.c",
-	"webp/dsp/dec.c",
-	"webp/dsp/yuv.c",
-	"webp/utils/bit_reader.c",
-	"webp/utils/filters.c",
-	"webp/utils/bit_writer.c",
-	"webp/utils/thread.c",
-	"webp/utils/quant_levels.c",
-	"webp/utils/color_cache.c",
-	"webp/utils/rescaler.c",
-	"webp/utils/utils.c",
-	"webp/utils/huffman.c",
-	"webp/utils/huffman_encode.c",
-	"webp/dec/tree.c",
-	"webp/dec/alpha.c",
-	"webp/dec/frame.c",
-	"webp/dec/vp8l.c",
-	"webp/dec/vp8.c",
-	"webp/dec/quant.c",
-	"webp/dec/webp.c",
-	"webp/dec/buffer.c",
-	"webp/dec/io.c",
-	"webp/dec/layer.c",
-	"webp/dec/idec.c",
-	"webp/image_loader_webp.cpp"
+"webp/enc/webpenc.c",\
+"webp/enc/near_lossless.c",\
+"webp/enc/frame.c",\
+"webp/enc/alpha.c",\
+"webp/enc/picture_csp.c",\
+"webp/enc/vp8l.c",\
+"webp/enc/picture_psnr.c",\
+"webp/enc/delta_palettization.c",\
+"webp/enc/syntax.c",\
+"webp/enc/backward_references.c",\
+"webp/enc/token.c",\
+"webp/enc/analysis.c",\
+"webp/enc/iterator.c",\
+"webp/enc/picture_tools.c",\
+"webp/enc/picture_rescale.c",\
+"webp/enc/config.c",\
+"webp/enc/tree.c",\
+"webp/enc/cost.c",\
+"webp/enc/picture.c",\
+"webp/enc/quant.c",\
+"webp/enc/filter.c",\
+"webp/enc/histogram.c",\
+"webp/image_loader_webp.cpp",\
+"webp/utils/rescaler.c",\
+"webp/utils/filters.c",\
+"webp/utils/quant_levels_dec.c",\
+"webp/utils/huffman.c",\
+"webp/utils/thread.c",\
+"webp/utils/quant_levels.c",\
+"webp/utils/bit_writer.c",\
+"webp/utils/bit_reader.c",\
+"webp/utils/random.c",\
+"webp/utils/utils.c",\
+"webp/utils/huffman_encode.c",\
+"webp/utils/color_cache.c",\
+"webp/mux/muxinternal.c",\
+"webp/mux/muxread.c",\
+"webp/mux/anim_encode.c",\
+"webp/mux/muxedit.c",\
+"webp/dec/webp.c",\
+"webp/dec/frame.c",\
+"webp/dec/alpha.c",\
+"webp/dec/vp8l.c",\
+"webp/dec/io.c",\
+"webp/dec/vp8.c",\
+"webp/dec/idec.c",\
+"webp/dec/tree.c",\
+"webp/dec/buffer.c",\
+"webp/dec/quant.c",\
+"webp/demux/demux.c",\
+"webp/demux/anim_decode.c",\
+"webp/dsp/yuv.c",\
+"webp/dsp/filters_sse2.c",\
+"webp/dsp/dec_sse41.c",\
+"webp/dsp/rescaler.c",\
+"webp/dsp/lossless_sse2.c",\
+"webp/dsp/alpha_processing_sse41.c",\
+"webp/dsp/alpha_processing_sse2.c",\
+"webp/dsp/filters.c",\
+"webp/dsp/upsampling_mips_dsp_r2.c",\
+"webp/dsp/dec_neon.c",\
+"webp/dsp/enc_neon.c",\
+"webp/dsp/lossless_enc_mips32.c",\
+"webp/dsp/lossless_enc_sse2.c",\
+"webp/dsp/upsampling.c",\
+"webp/dsp/lossless_enc_neon.c",\
+"webp/dsp/alpha_processing.c",\
+"webp/dsp/cost_sse2.c",\
+"webp/dsp/dec_mips32.c",\
+"webp/dsp/enc_avx2.c",\
+"webp/dsp/rescaler_mips32.c",\
+"webp/dsp/enc.c",\
+"webp/dsp/lossless_enc_sse41.c",\
+"webp/dsp/cost_mips32.c",\
+"webp/dsp/lossless_mips_dsp_r2.c",\
+"webp/dsp/filters_mips_dsp_r2.c",\
+"webp/dsp/upsampling_neon.c",\
+"webp/dsp/alpha_processing_mips_dsp_r2.c",\
+"webp/dsp/enc_mips_dsp_r2.c",\
+"webp/dsp/lossless.c",\
+"webp/dsp/yuv_mips_dsp_r2.c",\
+"webp/dsp/cost_mips_dsp_r2.c",\
+"webp/dsp/argb.c",\
+"webp/dsp/dec_sse2.c",\
+"webp/dsp/rescaler_sse2.c",\
+"webp/dsp/enc_sse41.c",\
+"webp/dsp/argb_mips_dsp_r2.c",\
+"webp/dsp/lossless_enc_mips_dsp_r2.c",\
+"webp/dsp/dec_clip_tables.c",\
+"webp/dsp/yuv_mips32.c",\
+"webp/dsp/cpu.c",\
+"webp/dsp/dec.c",\
+"webp/dsp/argb_sse2.c",\
+"webp/dsp/lossless_neon.c",\
+"webp/dsp/lossless_enc.c",\
+"webp/dsp/enc_mips32.c",\
+"webp/dsp/cost.c",\
+"webp/dsp/rescaler_mips_dsp_r2.c",\
+"webp/dsp/dec_mips_dsp_r2.c",\
+"webp/dsp/rescaler_neon.c",\
+"webp/dsp/yuv_sse2.c",\
+"webp/dsp/enc_sse2.c",\
+"webp/dsp/upsampling_sse2.c"
 ]
 
 env.drivers_sources+=webp_sources
 
 #env.add_source_files(env.drivers_sources, webp_sources)
 
+
 Export('env')
+
diff --git a/drivers/webp/dec/alpha.c b/drivers/webp/dec/alpha.c
index d1095fa555..52216fc4d6 100644
--- a/drivers/webp/dec/alpha.c
+++ b/drivers/webp/dec/alpha.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 decompression.
@@ -10,131 +12,156 @@
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include <stdlib.h>
+#include "./alphai.h"
 #include "./vp8i.h"
 #include "./vp8li.h"
-#include "../utils/filters.h"
-#include "../utils/quant_levels.h"
-#include "../format_constants.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-// 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;
+#include "../dsp/dsp.h"
+#include "../utils/quant_levels_dec.h"
+#include "../utils/utils.h"
+#include "../webp/format_constants.h"
+
+//------------------------------------------------------------------------------
+// ALPHDecoder object.
+
+ALPHDecoder* ALPHNew(void) {
+  ALPHDecoder* const dec = (ALPHDecoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
+  return dec;
+}
+
+void ALPHDelete(ALPHDecoder* const dec) {
+  if (dec != NULL) {
+    VP8LDelete(dec->vp8l_dec_);
+    dec->vp8l_dec_ = NULL;
+    WebPSafeFree(dec);
   }
 }
 
 //------------------------------------------------------------------------------
-// Decodes the compressed data 'data' of size 'data_size' into the 'output'.
-// The 'output' buffer should be pre-allocated and must be of the same
-// dimension 'height'x'stride', as that of the image.
-//
-// Returns 1 on successfully decoding the compressed alpha and
-//         0 if either:
-//           error in bit-stream header (invalid compression mode or filter), or
-//           error returned by appropriate compression method.
-
-static int DecodeAlpha(const uint8_t* data, size_t data_size,
-                       int width, int height, int stride, uint8_t* output) {
-  uint8_t* decoded_data = NULL;
-  const size_t decoded_size = height * width;
-  uint8_t* unfiltered_data = NULL;
-  WEBP_FILTER_TYPE filter;
-  int pre_processing;
-  int rsrv;
+// Decoding.
+
+// Initialize alpha decoding by parsing the alpha header and decoding the image
+// header for alpha data stored using lossless compression.
+// Returns false in case of error in alpha header (data too short, invalid
+// compression method or filter, error in lossless header data etc).
+static int ALPHInit(ALPHDecoder* const dec, const uint8_t* data,
+                    size_t data_size, int width, int height, uint8_t* output) {
   int ok = 0;
-  int method;
+  const uint8_t* const alpha_data = data + ALPHA_HEADER_LEN;
+  const size_t alpha_data_size = data_size - ALPHA_HEADER_LEN;
+  int rsrv;
 
-  assert(width > 0 && height > 0 && stride >= width);
+  assert(width > 0 && height > 0);
   assert(data != NULL && output != NULL);
 
+  dec->width_ = width;
+  dec->height_ = height;
+
   if (data_size <= ALPHA_HEADER_LEN) {
     return 0;
   }
 
-  method = (data[0] >> 0) & 0x03;
-  filter = (data[0] >> 2) & 0x03;
-  pre_processing = (data[0] >> 4) & 0x03;
+  dec->method_ = (data[0] >> 0) & 0x03;
+  dec->filter_ = (data[0] >> 2) & 0x03;
+  dec->pre_processing_ = (data[0] >> 4) & 0x03;
   rsrv = (data[0] >> 6) & 0x03;
-  if (method < ALPHA_NO_COMPRESSION ||
-      method > ALPHA_LOSSLESS_COMPRESSION ||
-      filter >= WEBP_FILTER_LAST ||
-      pre_processing > ALPHA_PREPROCESSED_LEVELS ||
+  if (dec->method_ < ALPHA_NO_COMPRESSION ||
+      dec->method_ > ALPHA_LOSSLESS_COMPRESSION ||
+      dec->filter_ >= WEBP_FILTER_LAST ||
+      dec->pre_processing_ > ALPHA_PREPROCESSED_LEVELS ||
       rsrv != 0) {
     return 0;
   }
 
-  if (method == ALPHA_NO_COMPRESSION) {
-    ok = (data_size >= decoded_size);
-    decoded_data = (uint8_t*)data + ALPHA_HEADER_LEN;
+  if (dec->method_ == ALPHA_NO_COMPRESSION) {
+    const size_t alpha_decoded_size = dec->width_ * dec->height_;
+    ok = (alpha_data_size >= alpha_decoded_size);
   } else {
-    decoded_data = (uint8_t*)malloc(decoded_size);
-    if (decoded_data == NULL) return 0;
-    ok = VP8LDecodeAlphaImageStream(width, height,
-                                    data + ALPHA_HEADER_LEN,
-                                    data_size - ALPHA_HEADER_LEN,
-                                    decoded_data);
+    assert(dec->method_ == ALPHA_LOSSLESS_COMPRESSION);
+    ok = VP8LDecodeAlphaHeader(dec, alpha_data, alpha_data_size, output);
   }
+  VP8FiltersInit();
+  return ok;
+}
 
-  if (ok) {
-    WebPFilterFunc unfilter_func = WebPUnfilters[filter];
-    if (unfilter_func != NULL) {
-      unfiltered_data = (uint8_t*)malloc(decoded_size);
-      if (unfiltered_data == NULL) {
-        ok = 0;
-        goto Error;
-      }
-      // TODO(vikas): Implement on-the-fly decoding & filter mechanism to decode
-      // and apply filter per image-row.
-      unfilter_func(decoded_data, width, height, 1, width, unfiltered_data);
-      // Construct raw_data (height x stride) from alpha data (height x width).
-      CopyPlane(unfiltered_data, width, output, stride, width, height);
-      free(unfiltered_data);
-    } else {
-      // Construct raw_data (height x stride) from alpha data (height x width).
-      CopyPlane(decoded_data, width, output, stride, width, height);
-    }
-    if (pre_processing == ALPHA_PREPROCESSED_LEVELS) {
-      ok = DequantizeLevels(decoded_data, width, height);
+// Decodes, unfilters and dequantizes *at least* 'num_rows' rows of alpha
+// starting from row number 'row'. It assumes that rows up to (row - 1) have
+// already been decoded.
+// Returns false in case of bitstream error.
+static int ALPHDecode(VP8Decoder* const dec, int row, int num_rows) {
+  ALPHDecoder* const alph_dec = dec->alph_dec_;
+  const int width = alph_dec->width_;
+  const int height = alph_dec->height_;
+  WebPUnfilterFunc unfilter_func = WebPUnfilters[alph_dec->filter_];
+  uint8_t* const output = dec->alpha_plane_;
+  if (alph_dec->method_ == ALPHA_NO_COMPRESSION) {
+    const size_t offset = row * width;
+    const size_t num_pixels = num_rows * width;
+    assert(dec->alpha_data_size_ >= ALPHA_HEADER_LEN + offset + num_pixels);
+    memcpy(dec->alpha_plane_ + offset,
+           dec->alpha_data_ + ALPHA_HEADER_LEN + offset, num_pixels);
+  } else {  // alph_dec->method_ == ALPHA_LOSSLESS_COMPRESSION
+    assert(alph_dec->vp8l_dec_ != NULL);
+    if (!VP8LDecodeAlphaImageStream(alph_dec, row + num_rows)) {
+      return 0;
     }
   }
 
- Error:
-  if (method != ALPHA_NO_COMPRESSION) {
-    free(decoded_data);
+  if (unfilter_func != NULL) {
+    unfilter_func(width, height, width, row, num_rows, output);
   }
-  return ok;
+
+  if (row + num_rows == dec->pic_hdr_.height_) {
+    dec->is_alpha_decoded_ = 1;
+  }
+  return 1;
 }
 
 //------------------------------------------------------------------------------
+// Main entry point.
 
 const uint8_t* VP8DecompressAlphaRows(VP8Decoder* const dec,
                                       int row, int num_rows) {
-  const int stride = dec->pic_hdr_.width_;
+  const int width = dec->pic_hdr_.width_;
+  const int height = dec->pic_hdr_.height_;
 
-  if (row < 0 || num_rows < 0 || row + num_rows > dec->pic_hdr_.height_) {
+  if (row < 0 || num_rows <= 0 || row + num_rows > height) {
     return NULL;    // sanity check.
   }
 
   if (row == 0) {
-    // Decode everything during the first call.
-    if (!DecodeAlpha(dec->alpha_data_, (size_t)dec->alpha_data_size_,
-                     dec->pic_hdr_.width_, dec->pic_hdr_.height_, stride,
-                     dec->alpha_plane_)) {
-      return NULL;  // Error.
+    // Initialize decoding.
+    assert(dec->alpha_plane_ != NULL);
+    dec->alph_dec_ = ALPHNew();
+    if (dec->alph_dec_ == NULL) return NULL;
+    if (!ALPHInit(dec->alph_dec_, dec->alpha_data_, dec->alpha_data_size_,
+                  width, height, dec->alpha_plane_)) {
+      ALPHDelete(dec->alph_dec_);
+      dec->alph_dec_ = NULL;
+      return NULL;
+    }
+    // if we allowed use of alpha dithering, check whether it's needed at all
+    if (dec->alph_dec_->pre_processing_ != ALPHA_PREPROCESSED_LEVELS) {
+      dec->alpha_dithering_ = 0;  // disable dithering
+    } else {
+      num_rows = height;          // decode everything in one pass
     }
   }
 
+  if (!dec->is_alpha_decoded_) {
+    int ok = 0;
+    assert(dec->alph_dec_ != NULL);
+    ok = ALPHDecode(dec, row, num_rows);
+    if (ok && dec->alpha_dithering_ > 0) {
+      ok = WebPDequantizeLevels(dec->alpha_plane_, width, height,
+                                dec->alpha_dithering_);
+    }
+    if (!ok || dec->is_alpha_decoded_) {
+      ALPHDelete(dec->alph_dec_);
+      dec->alph_dec_ = NULL;
+    }
+    if (!ok) return NULL;  // Error.
+  }
+
   // Return a pointer to the current decoded row.
-  return dec->alpha_plane_ + row * stride;
+  return dec->alpha_plane_ + row * width;
 }
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/buffer.c b/drivers/webp/dec/buffer.c
index c159f6f248..9ed2b3fe1a 100644
--- a/drivers/webp/dec/buffer.c
+++ b/drivers/webp/dec/buffer.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Everything about WebPDecBuffer
@@ -15,10 +17,6 @@
 #include "./webpi.h"
 #include "../utils/utils.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
 // WebPDecBuffer
 
@@ -35,6 +33,11 @@ static int IsValidColorspace(int webp_csp_mode) {
   return (webp_csp_mode >= MODE_RGB && webp_csp_mode < MODE_LAST);
 }
 
+// strictly speaking, the very last (or first, if flipped) row
+// doesn't require padding.
+#define MIN_BUFFER_SIZE(WIDTH, HEIGHT, STRIDE)       \
+    (uint64_t)(STRIDE) * ((HEIGHT) - 1) + (WIDTH)
+
 static VP8StatusCode CheckDecBuffer(const WebPDecBuffer* const buffer) {
   int ok = 1;
   const WEBP_CSP_MODE mode = buffer->colorspace;
@@ -44,33 +47,41 @@ static VP8StatusCode CheckDecBuffer(const WebPDecBuffer* const buffer) {
     ok = 0;
   } else if (!WebPIsRGBMode(mode)) {   // YUV checks
     const WebPYUVABuffer* const buf = &buffer->u.YUVA;
-    const uint64_t y_size = (uint64_t)buf->y_stride * height;
-    const uint64_t u_size = (uint64_t)buf->u_stride * ((height + 1) / 2);
-    const uint64_t v_size = (uint64_t)buf->v_stride * ((height + 1) / 2);
-    const uint64_t a_size = (uint64_t)buf->a_stride * height;
+    const int uv_width  = (width  + 1) / 2;
+    const int uv_height = (height + 1) / 2;
+    const int y_stride = abs(buf->y_stride);
+    const int u_stride = abs(buf->u_stride);
+    const int v_stride = abs(buf->v_stride);
+    const int a_stride = abs(buf->a_stride);
+    const uint64_t y_size = MIN_BUFFER_SIZE(width, height, y_stride);
+    const uint64_t u_size = MIN_BUFFER_SIZE(uv_width, uv_height, u_stride);
+    const uint64_t v_size = MIN_BUFFER_SIZE(uv_width, uv_height, v_stride);
+    const uint64_t a_size = MIN_BUFFER_SIZE(width, height, a_stride);
     ok &= (y_size <= buf->y_size);
     ok &= (u_size <= buf->u_size);
     ok &= (v_size <= buf->v_size);
-    ok &= (buf->y_stride >= width);
-    ok &= (buf->u_stride >= (width + 1) / 2);
-    ok &= (buf->v_stride >= (width + 1) / 2);
+    ok &= (y_stride >= width);
+    ok &= (u_stride >= uv_width);
+    ok &= (v_stride >= uv_width);
     ok &= (buf->y != NULL);
     ok &= (buf->u != NULL);
     ok &= (buf->v != NULL);
     if (mode == MODE_YUVA) {
-      ok &= (buf->a_stride >= width);
+      ok &= (a_stride >= width);
       ok &= (a_size <= buf->a_size);
       ok &= (buf->a != NULL);
     }
   } else {    // RGB checks
     const WebPRGBABuffer* const buf = &buffer->u.RGBA;
-    const uint64_t size = (uint64_t)buf->stride * height;
+    const int stride = abs(buf->stride);
+    const uint64_t size = MIN_BUFFER_SIZE(width, height, stride);
     ok &= (size <= buf->size);
-    ok &= (buf->stride >= width * kModeBpp[mode]);
+    ok &= (stride >= width * kModeBpp[mode]);
     ok &= (buf->rgba != NULL);
   }
   return ok ? VP8_STATUS_OK : VP8_STATUS_INVALID_PARAM;
 }
+#undef MIN_BUFFER_SIZE
 
 static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
   const int w = buffer->width;
@@ -133,9 +144,35 @@ static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
   return CheckDecBuffer(buffer);
 }
 
+VP8StatusCode WebPFlipBuffer(WebPDecBuffer* const buffer) {
+  if (buffer == NULL) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  if (WebPIsRGBMode(buffer->colorspace)) {
+    WebPRGBABuffer* const buf = &buffer->u.RGBA;
+    buf->rgba += (buffer->height - 1) * buf->stride;
+    buf->stride = -buf->stride;
+  } else {
+    WebPYUVABuffer* const buf = &buffer->u.YUVA;
+    const int H = buffer->height;
+    buf->y += (H - 1) * buf->y_stride;
+    buf->y_stride = -buf->y_stride;
+    buf->u += ((H - 1) >> 1) * buf->u_stride;
+    buf->u_stride = -buf->u_stride;
+    buf->v += ((H - 1) >> 1) * buf->v_stride;
+    buf->v_stride = -buf->v_stride;
+    if (buf->a != NULL) {
+      buf->a += (H - 1) * buf->a_stride;
+      buf->a_stride = -buf->a_stride;
+    }
+  }
+  return VP8_STATUS_OK;
+}
+
 VP8StatusCode WebPAllocateDecBuffer(int w, int h,
                                     const WebPDecoderOptions* const options,
                                     WebPDecBuffer* const out) {
+  VP8StatusCode status;
   if (out == NULL || w <= 0 || h <= 0) {
     return VP8_STATUS_INVALID_PARAM;
   }
@@ -152,18 +189,28 @@ VP8StatusCode WebPAllocateDecBuffer(int w, int h,
       h = ch;
     }
     if (options->use_scaling) {
-      if (options->scaled_width <= 0 || options->scaled_height <= 0) {
+      int scaled_width = options->scaled_width;
+      int scaled_height = options->scaled_height;
+      if (!WebPRescalerGetScaledDimensions(
+              w, h, &scaled_width, &scaled_height)) {
         return VP8_STATUS_INVALID_PARAM;
       }
-      w = options->scaled_width;
-      h = options->scaled_height;
+      w = scaled_width;
+      h = scaled_height;
     }
   }
   out->width = w;
   out->height = h;
 
-  // Then, allocate buffer for real
-  return AllocateBuffer(out);
+  // Then, allocate buffer for real.
+  status = AllocateBuffer(out);
+  if (status != VP8_STATUS_OK) return status;
+
+  // Use the stride trick if vertical flip is needed.
+  if (options != NULL && options->flip) {
+    status = WebPFlipBuffer(out);
+  }
+  return status;
 }
 
 //------------------------------------------------------------------------------
@@ -180,8 +227,9 @@ int WebPInitDecBufferInternal(WebPDecBuffer* buffer, int version) {
 
 void WebPFreeDecBuffer(WebPDecBuffer* buffer) {
   if (buffer != NULL) {
-    if (!buffer->is_external_memory)
-      free(buffer->private_memory);
+    if (!buffer->is_external_memory) {
+      WebPSafeFree(buffer->private_memory);
+    }
     buffer->private_memory = NULL;
   }
 }
@@ -210,6 +258,3 @@ void WebPGrabDecBuffer(WebPDecBuffer* const src, WebPDecBuffer* const dst) {
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/decode_vp8.h b/drivers/webp/dec/decode_vp8.h
index c26a9fc891..b9337bbec0 100644
--- a/drivers/webp/dec/decode_vp8.h
+++ b/drivers/webp/dec/decode_vp8.h
@@ -1,8 +1,10 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 //  Low-level API for VP8 decoder
@@ -12,9 +14,9 @@
 #ifndef WEBP_WEBP_DECODE_VP8_H_
 #define WEBP_WEBP_DECODE_VP8_H_
 
-#include "../decode.h"
+#include "../webp/decode.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -130,7 +132,8 @@ static WEBP_INLINE int VP8InitIo(VP8Io* const io) {
   return VP8InitIoInternal(io, WEBP_DECODER_ABI_VERSION);
 }
 
-// Start decoding a new picture. Returns true if ok.
+// Decode the VP8 frame header. Returns true if ok.
+// Note: 'io->data' must be pointing to the start of the VP8 frame header.
 int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io);
 
 // Decode a picture. Will call VP8GetHeaders() if it wasn't done already.
@@ -175,7 +178,7 @@ WEBP_EXTERN(int) VP8LGetInfo(
     const uint8_t* data, size_t data_size,  // data available so far
     int* const width, int* const height, int* const has_alpha);
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/dec/frame.c b/drivers/webp/dec/frame.c
index 9c91a48e17..b882133eab 100644
--- a/drivers/webp/dec/frame.c
+++ b/drivers/webp/dec/frame.c
@@ -1,8 +1,10 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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-reconstruction function. Memory allocation.
@@ -13,11 +15,180 @@
 #include "./vp8i.h"
 #include "../utils/utils.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+//------------------------------------------------------------------------------
+// Main reconstruction function.
+
+static const int kScan[16] = {
+  0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
+  0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
+  0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
+  0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS
+};
+
+static int CheckMode(int mb_x, int mb_y, int mode) {
+  if (mode == B_DC_PRED) {
+    if (mb_x == 0) {
+      return (mb_y == 0) ? B_DC_PRED_NOTOPLEFT : B_DC_PRED_NOLEFT;
+    } else {
+      return (mb_y == 0) ? B_DC_PRED_NOTOP : B_DC_PRED;
+    }
+  }
+  return mode;
+}
 
-#define ALIGN_MASK (32 - 1)
+static void Copy32b(uint8_t* const dst, const uint8_t* const src) {
+  memcpy(dst, src, 4);
+}
+
+static WEBP_INLINE void DoTransform(uint32_t bits, const int16_t* const src,
+                                    uint8_t* const dst) {
+  switch (bits >> 30) {
+    case 3:
+      VP8Transform(src, dst, 0);
+      break;
+    case 2:
+      VP8TransformAC3(src, dst);
+      break;
+    case 1:
+      VP8TransformDC(src, dst);
+      break;
+    default:
+      break;
+  }
+}
+
+static void DoUVTransform(uint32_t bits, const int16_t* const src,
+                          uint8_t* const dst) {
+  if (bits & 0xff) {    // any non-zero coeff at all?
+    if (bits & 0xaa) {  // any non-zero AC coefficient?
+      VP8TransformUV(src, dst);   // note we don't use the AC3 variant for U/V
+    } else {
+      VP8TransformDCUV(src, dst);
+    }
+  }
+}
+
+static void ReconstructRow(const VP8Decoder* const dec,
+                           const VP8ThreadContext* ctx) {
+  int j;
+  int mb_x;
+  const int mb_y = ctx->mb_y_;
+  const int cache_id = ctx->id_;
+  uint8_t* const y_dst = dec->yuv_b_ + Y_OFF;
+  uint8_t* const u_dst = dec->yuv_b_ + U_OFF;
+  uint8_t* const v_dst = dec->yuv_b_ + V_OFF;
+
+  // Initialize left-most block.
+  for (j = 0; j < 16; ++j) {
+    y_dst[j * BPS - 1] = 129;
+  }
+  for (j = 0; j < 8; ++j) {
+    u_dst[j * BPS - 1] = 129;
+    v_dst[j * BPS - 1] = 129;
+  }
+
+  // Init top-left sample on left column too.
+  if (mb_y > 0) {
+    y_dst[-1 - BPS] = u_dst[-1 - BPS] = v_dst[-1 - BPS] = 129;
+  } else {
+    // we only need to do this init once at block (0,0).
+    // Afterward, it remains valid for the whole topmost row.
+    memset(y_dst - BPS - 1, 127, 16 + 4 + 1);
+    memset(u_dst - BPS - 1, 127, 8 + 1);
+    memset(v_dst - BPS - 1, 127, 8 + 1);
+  }
+
+  // Reconstruct one row.
+  for (mb_x = 0; mb_x < dec->mb_w_; ++mb_x) {
+    const VP8MBData* const block = ctx->mb_data_ + mb_x;
+
+    // Rotate in the left samples from previously decoded block. We move four
+    // pixels at a time for alignment reason, and because of in-loop filter.
+    if (mb_x > 0) {
+      for (j = -1; j < 16; ++j) {
+        Copy32b(&y_dst[j * BPS - 4], &y_dst[j * BPS + 12]);
+      }
+      for (j = -1; j < 8; ++j) {
+        Copy32b(&u_dst[j * BPS - 4], &u_dst[j * BPS + 4]);
+        Copy32b(&v_dst[j * BPS - 4], &v_dst[j * BPS + 4]);
+      }
+    }
+    {
+      // bring top samples into the cache
+      VP8TopSamples* const top_yuv = dec->yuv_t_ + mb_x;
+      const int16_t* const coeffs = block->coeffs_;
+      uint32_t bits = block->non_zero_y_;
+      int n;
+
+      if (mb_y > 0) {
+        memcpy(y_dst - BPS, top_yuv[0].y, 16);
+        memcpy(u_dst - BPS, top_yuv[0].u, 8);
+        memcpy(v_dst - BPS, top_yuv[0].v, 8);
+      }
+
+      // predict and add residuals
+      if (block->is_i4x4_) {   // 4x4
+        uint32_t* const top_right = (uint32_t*)(y_dst - BPS + 16);
+
+        if (mb_y > 0) {
+          if (mb_x >= dec->mb_w_ - 1) {    // on rightmost border
+            memset(top_right, top_yuv[0].y[15], sizeof(*top_right));
+          } else {
+            memcpy(top_right, top_yuv[1].y, sizeof(*top_right));
+          }
+        }
+        // replicate the top-right pixels below
+        top_right[BPS] = top_right[2 * BPS] = top_right[3 * BPS] = top_right[0];
+
+        // predict and add residuals for all 4x4 blocks in turn.
+        for (n = 0; n < 16; ++n, bits <<= 2) {
+          uint8_t* const dst = y_dst + kScan[n];
+          VP8PredLuma4[block->imodes_[n]](dst);
+          DoTransform(bits, coeffs + n * 16, dst);
+        }
+      } else {    // 16x16
+        const int pred_func = CheckMode(mb_x, mb_y, block->imodes_[0]);
+        VP8PredLuma16[pred_func](y_dst);
+        if (bits != 0) {
+          for (n = 0; n < 16; ++n, bits <<= 2) {
+            DoTransform(bits, coeffs + n * 16, y_dst + kScan[n]);
+          }
+        }
+      }
+      {
+        // Chroma
+        const uint32_t bits_uv = block->non_zero_uv_;
+        const int pred_func = CheckMode(mb_x, mb_y, block->uvmode_);
+        VP8PredChroma8[pred_func](u_dst);
+        VP8PredChroma8[pred_func](v_dst);
+        DoUVTransform(bits_uv >> 0, coeffs + 16 * 16, u_dst);
+        DoUVTransform(bits_uv >> 8, coeffs + 20 * 16, v_dst);
+      }
+
+      // stash away top samples for next block
+      if (mb_y < dec->mb_h_ - 1) {
+        memcpy(top_yuv[0].y, y_dst + 15 * BPS, 16);
+        memcpy(top_yuv[0].u, u_dst +  7 * BPS,  8);
+        memcpy(top_yuv[0].v, v_dst +  7 * BPS,  8);
+      }
+    }
+    // Transfer reconstructed samples from yuv_b_ cache to final destination.
+    {
+      const int y_offset = cache_id * 16 * dec->cache_y_stride_;
+      const int uv_offset = cache_id * 8 * dec->cache_uv_stride_;
+      uint8_t* const y_out = dec->cache_y_ + mb_x * 16 + y_offset;
+      uint8_t* const u_out = dec->cache_u_ + mb_x * 8 + uv_offset;
+      uint8_t* const v_out = dec->cache_v_ + mb_x * 8 + uv_offset;
+      for (j = 0; j < 16; ++j) {
+        memcpy(y_out + j * dec->cache_y_stride_, y_dst + j * BPS, 16);
+      }
+      for (j = 0; j < 8; ++j) {
+        memcpy(u_out + j * dec->cache_uv_stride_, u_dst + j * BPS, 8);
+        memcpy(v_out + j * dec->cache_uv_stride_, v_dst + j * BPS, 8);
+      }
+    }
+  }
+}
 
 //------------------------------------------------------------------------------
 // Filtering
@@ -29,25 +200,18 @@ extern "C" {
 //                 U/V, so it's 8 samples total (because of the 2x upsampling).
 static const uint8_t kFilterExtraRows[3] = { 0, 2, 8 };
 
-static WEBP_INLINE int hev_thresh_from_level(int level, int keyframe) {
-  if (keyframe) {
-    return (level >= 40) ? 2 : (level >= 15) ? 1 : 0;
-  } else {
-    return (level >= 40) ? 3 : (level >= 20) ? 2 : (level >= 15) ? 1 : 0;
-  }
-}
-
 static void DoFilter(const VP8Decoder* const dec, int mb_x, int mb_y) {
   const VP8ThreadContext* const ctx = &dec->thread_ctx_;
+  const int cache_id = ctx->id_;
   const int y_bps = dec->cache_y_stride_;
-  VP8FInfo* const f_info = ctx->f_info_ + mb_x;
-  uint8_t* const y_dst = dec->cache_y_ + ctx->id_ * 16 * y_bps + mb_x * 16;
-  const int level = f_info->f_level_;
+  const VP8FInfo* const f_info = ctx->f_info_ + mb_x;
+  uint8_t* const y_dst = dec->cache_y_ + cache_id * 16 * y_bps + mb_x * 16;
   const int ilevel = f_info->f_ilevel_;
-  const int limit = 2 * level + ilevel;
-  if (level == 0) {
+  const int limit = f_info->f_limit_;
+  if (limit == 0) {
     return;
   }
+  assert(limit >= 3);
   if (dec->filter_type_ == 1) {   // simple
     if (mb_x > 0) {
       VP8SimpleHFilter16(y_dst, y_bps, limit + 4);
@@ -63,10 +227,9 @@ static void DoFilter(const VP8Decoder* const dec, int mb_x, int mb_y) {
     }
   } else {    // complex
     const int uv_bps = dec->cache_uv_stride_;
-    uint8_t* const u_dst = dec->cache_u_ + ctx->id_ * 8 * uv_bps + mb_x * 8;
-    uint8_t* const v_dst = dec->cache_v_ + ctx->id_ * 8 * uv_bps + mb_x * 8;
-    const int hev_thresh =
-        hev_thresh_from_level(level, dec->frm_hdr_.key_frame_);
+    uint8_t* const u_dst = dec->cache_u_ + cache_id * 8 * uv_bps + mb_x * 8;
+    uint8_t* const v_dst = dec->cache_v_ + cache_id * 8 * uv_bps + mb_x * 8;
+    const int hev_thresh = f_info->hev_thresh_;
     if (mb_x > 0) {
       VP8HFilter16(y_dst, y_bps, limit + 4, ilevel, hev_thresh);
       VP8HFilter8(u_dst, v_dst, uv_bps, limit + 4, ilevel, hev_thresh);
@@ -97,53 +260,138 @@ static void FilterRow(const VP8Decoder* const dec) {
 }
 
 //------------------------------------------------------------------------------
+// Precompute the filtering strength for each segment and each i4x4/i16x16 mode.
 
-void VP8StoreBlock(VP8Decoder* const dec) {
+static void PrecomputeFilterStrengths(VP8Decoder* const dec) {
   if (dec->filter_type_ > 0) {
-    VP8FInfo* const info = dec->f_info_ + dec->mb_x_;
-    const int skip = dec->mb_info_[dec->mb_x_].skip_;
-    int level = dec->filter_levels_[dec->segment_];
-    if (dec->filter_hdr_.use_lf_delta_) {
-      // TODO(skal): only CURRENT is handled for now.
-      level += dec->filter_hdr_.ref_lf_delta_[0];
-      if (dec->is_i4x4_) {
-        level += dec->filter_hdr_.mode_lf_delta_[0];
+    int s;
+    const VP8FilterHeader* const hdr = &dec->filter_hdr_;
+    for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+      int i4x4;
+      // First, compute the initial level
+      int base_level;
+      if (dec->segment_hdr_.use_segment_) {
+        base_level = dec->segment_hdr_.filter_strength_[s];
+        if (!dec->segment_hdr_.absolute_delta_) {
+          base_level += hdr->level_;
+        }
+      } else {
+        base_level = hdr->level_;
+      }
+      for (i4x4 = 0; i4x4 <= 1; ++i4x4) {
+        VP8FInfo* const info = &dec->fstrengths_[s][i4x4];
+        int level = base_level;
+        if (hdr->use_lf_delta_) {
+          level += hdr->ref_lf_delta_[0];
+          if (i4x4) {
+            level += hdr->mode_lf_delta_[0];
+          }
+        }
+        level = (level < 0) ? 0 : (level > 63) ? 63 : level;
+        if (level > 0) {
+          int ilevel = level;
+          if (hdr->sharpness_ > 0) {
+            if (hdr->sharpness_ > 4) {
+              ilevel >>= 2;
+            } else {
+              ilevel >>= 1;
+            }
+            if (ilevel > 9 - hdr->sharpness_) {
+              ilevel = 9 - hdr->sharpness_;
+            }
+          }
+          if (ilevel < 1) ilevel = 1;
+          info->f_ilevel_ = ilevel;
+          info->f_limit_ = 2 * level + ilevel;
+          info->hev_thresh_ = (level >= 40) ? 2 : (level >= 15) ? 1 : 0;
+        } else {
+          info->f_limit_ = 0;  // no filtering
+        }
+        info->f_inner_ = i4x4;
       }
     }
-    level = (level < 0) ? 0 : (level > 63) ? 63 : level;
-    info->f_level_ = level;
+  }
+}
 
-    if (dec->filter_hdr_.sharpness_ > 0) {
-      if (dec->filter_hdr_.sharpness_ > 4) {
-        level >>= 2;
-      } else {
-        level >>= 1;
+//------------------------------------------------------------------------------
+// Dithering
+
+#define DITHER_AMP_TAB_SIZE 12
+static const int kQuantToDitherAmp[DITHER_AMP_TAB_SIZE] = {
+  // roughly, it's dqm->uv_mat_[1]
+  8, 7, 6, 4, 4, 2, 2, 2, 1, 1, 1, 1
+};
+
+void VP8InitDithering(const WebPDecoderOptions* const options,
+                      VP8Decoder* const dec) {
+  assert(dec != NULL);
+  if (options != NULL) {
+    const int d = options->dithering_strength;
+    const int max_amp = (1 << VP8_RANDOM_DITHER_FIX) - 1;
+    const int f = (d < 0) ? 0 : (d > 100) ? max_amp : (d * max_amp / 100);
+    if (f > 0) {
+      int s;
+      int all_amp = 0;
+      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+        VP8QuantMatrix* const dqm = &dec->dqm_[s];
+        if (dqm->uv_quant_ < DITHER_AMP_TAB_SIZE) {
+          // TODO(skal): should we specially dither more for uv_quant_ < 0?
+          const int idx = (dqm->uv_quant_ < 0) ? 0 : dqm->uv_quant_;
+          dqm->dither_ = (f * kQuantToDitherAmp[idx]) >> 3;
+        }
+        all_amp |= dqm->dither_;
       }
-      if (level > 9 - dec->filter_hdr_.sharpness_) {
-        level = 9 - dec->filter_hdr_.sharpness_;
+      if (all_amp != 0) {
+        VP8InitRandom(&dec->dithering_rg_, 1.0f);
+        dec->dither_ = 1;
       }
     }
+    // potentially allow alpha dithering
+    dec->alpha_dithering_ = options->alpha_dithering_strength;
+    if (dec->alpha_dithering_ > 100) {
+      dec->alpha_dithering_ = 100;
+    } else if (dec->alpha_dithering_ < 0) {
+      dec->alpha_dithering_ = 0;
+    }
+  }
+}
 
-    info->f_ilevel_ = (level < 1) ? 1 : level;
-    info->f_inner_ = (!skip || dec->is_i4x4_);
+// minimal amp that will provide a non-zero dithering effect
+#define MIN_DITHER_AMP 4
+#define DITHER_DESCALE 4
+#define DITHER_DESCALE_ROUNDER (1 << (DITHER_DESCALE - 1))
+#define DITHER_AMP_BITS 8
+#define DITHER_AMP_CENTER (1 << DITHER_AMP_BITS)
+
+static void Dither8x8(VP8Random* const rg, uint8_t* dst, int bps, int amp) {
+  int i, j;
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 8; ++i) {
+      // TODO: could be made faster with SSE2
+      const int bits =
+          VP8RandomBits2(rg, DITHER_AMP_BITS + 1, amp) - DITHER_AMP_CENTER;
+      // Convert to range: [-2,2] for dither=50, [-4,4] for dither=100
+      const int delta = (bits + DITHER_DESCALE_ROUNDER) >> DITHER_DESCALE;
+      const int v = (int)dst[i] + delta;
+      dst[i] = (v < 0) ? 0 : (v > 255) ? 255u : (uint8_t)v;
+    }
+    dst += bps;
   }
-  {
-    // Transfer samples to row cache
-    int y;
-    const int y_offset = dec->cache_id_ * 16 * dec->cache_y_stride_;
-    const int uv_offset = dec->cache_id_ * 8 * dec->cache_uv_stride_;
-    uint8_t* const ydst = dec->cache_y_ + dec->mb_x_ * 16 + y_offset;
-    uint8_t* const udst = dec->cache_u_ + dec->mb_x_ * 8 + uv_offset;
-    uint8_t* const vdst = dec->cache_v_ + dec->mb_x_ * 8 + uv_offset;
-    for (y = 0; y < 16; ++y) {
-      memcpy(ydst + y * dec->cache_y_stride_,
-             dec->yuv_b_ + Y_OFF + y * BPS, 16);
-    }
-    for (y = 0; y < 8; ++y) {
-      memcpy(udst + y * dec->cache_uv_stride_,
-           dec->yuv_b_ + U_OFF + y * BPS, 8);
-      memcpy(vdst + y * dec->cache_uv_stride_,
-           dec->yuv_b_ + V_OFF + y * BPS, 8);
+}
+
+static void DitherRow(VP8Decoder* const dec) {
+  int mb_x;
+  assert(dec->dither_);
+  for (mb_x = dec->tl_mb_x_; mb_x < dec->br_mb_x_; ++mb_x) {
+    const VP8ThreadContext* const ctx = &dec->thread_ctx_;
+    const VP8MBData* const data = ctx->mb_data_ + mb_x;
+    const int cache_id = ctx->id_;
+    const int uv_bps = dec->cache_uv_stride_;
+    if (data->dither_ >= MIN_DITHER_AMP) {
+      uint8_t* const u_dst = dec->cache_u_ + cache_id * 8 * uv_bps + mb_x * 8;
+      uint8_t* const v_dst = dec->cache_v_ + cache_id * 8 * uv_bps + mb_x * 8;
+      Dither8x8(&dec->dithering_rg_, u_dst, uv_bps, data->dither_);
+      Dither8x8(&dec->dithering_rg_, v_dst, uv_bps, data->dither_);
     }
   }
 }
@@ -165,25 +413,35 @@ void VP8StoreBlock(VP8Decoder* const dec) {
 static int FinishRow(VP8Decoder* const dec, VP8Io* const io) {
   int ok = 1;
   const VP8ThreadContext* const ctx = &dec->thread_ctx_;
+  const int cache_id = ctx->id_;
   const int extra_y_rows = kFilterExtraRows[dec->filter_type_];
   const int ysize = extra_y_rows * dec->cache_y_stride_;
   const int uvsize = (extra_y_rows / 2) * dec->cache_uv_stride_;
-  const int y_offset = ctx->id_ * 16 * dec->cache_y_stride_;
-  const int uv_offset = ctx->id_ * 8 * dec->cache_uv_stride_;
+  const int y_offset = cache_id * 16 * dec->cache_y_stride_;
+  const int uv_offset = cache_id * 8 * dec->cache_uv_stride_;
   uint8_t* const ydst = dec->cache_y_ - ysize + y_offset;
   uint8_t* const udst = dec->cache_u_ - uvsize + uv_offset;
   uint8_t* const vdst = dec->cache_v_ - uvsize + uv_offset;
-  const int first_row = (ctx->mb_y_ == 0);
-  const int last_row = (ctx->mb_y_ >= dec->br_mb_y_ - 1);
-  int y_start = MACROBLOCK_VPOS(ctx->mb_y_);
-  int y_end = MACROBLOCK_VPOS(ctx->mb_y_ + 1);
+  const int mb_y = ctx->mb_y_;
+  const int is_first_row = (mb_y == 0);
+  const int is_last_row = (mb_y >= dec->br_mb_y_ - 1);
+
+  if (dec->mt_method_ == 2) {
+    ReconstructRow(dec, ctx);
+  }
 
   if (ctx->filter_row_) {
     FilterRow(dec);
   }
 
-  if (io->put) {
-    if (!first_row) {
+  if (dec->dither_) {
+    DitherRow(dec);
+  }
+
+  if (io->put != NULL) {
+    int y_start = MACROBLOCK_VPOS(mb_y);
+    int y_end = MACROBLOCK_VPOS(mb_y + 1);
+    if (!is_first_row) {
       y_start -= extra_y_rows;
       io->y = ydst;
       io->u = udst;
@@ -194,7 +452,7 @@ static int FinishRow(VP8Decoder* const dec, VP8Io* const io) {
       io->v = dec->cache_v_ + uv_offset;
     }
 
-    if (!last_row) {
+    if (!is_last_row) {
       y_end -= extra_y_rows;
     }
     if (y_end > io->crop_bottom) {
@@ -202,11 +460,8 @@ static int FinishRow(VP8Decoder* const dec, VP8Io* const io) {
     }
     io->a = NULL;
     if (dec->alpha_data_ != NULL && y_start < y_end) {
-      // TODO(skal): several things to correct here:
-      // * testing presence of alpha with dec->alpha_data_ is not a good idea
-      // * we're actually decompressing the full plane only once. It should be
-      //   more obvious from signature.
-      // * we could free alpha_data_ right after this call, but we don't own.
+      // TODO(skal): testing presence of alpha with dec->alpha_data_ is not a
+      // good idea.
       io->a = VP8DecompressAlphaRows(dec, y_start, y_end - y_start);
       if (io->a == NULL) {
         return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
@@ -238,8 +493,8 @@ static int FinishRow(VP8Decoder* const dec, VP8Io* const io) {
     }
   }
   // rotate top samples if needed
-  if (ctx->id_ + 1 == dec->num_caches_) {
-    if (!last_row) {
+  if (cache_id + 1 == dec->num_caches_) {
+    if (!is_last_row) {
       memcpy(dec->cache_y_ - ysize, ydst + 16 * dec->cache_y_stride_, ysize);
       memcpy(dec->cache_u_ - uvsize, udst + 8 * dec->cache_uv_stride_, uvsize);
       memcpy(dec->cache_v_ - uvsize, vdst + 8 * dec->cache_uv_stride_, uvsize);
@@ -256,27 +511,40 @@ static int FinishRow(VP8Decoder* const dec, VP8Io* const io) {
 int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io) {
   int ok = 1;
   VP8ThreadContext* const ctx = &dec->thread_ctx_;
-  if (!dec->use_threads_) {
+  const int filter_row =
+      (dec->filter_type_ > 0) &&
+      (dec->mb_y_ >= dec->tl_mb_y_) && (dec->mb_y_ <= dec->br_mb_y_);
+  if (dec->mt_method_ == 0) {
     // ctx->id_ and ctx->f_info_ are already set
     ctx->mb_y_ = dec->mb_y_;
-    ctx->filter_row_ = dec->filter_row_;
+    ctx->filter_row_ = filter_row;
+    ReconstructRow(dec, ctx);
     ok = FinishRow(dec, io);
   } else {
     WebPWorker* const worker = &dec->worker_;
     // Finish previous job *before* updating context
-    ok &= WebPWorkerSync(worker);
+    ok &= WebPGetWorkerInterface()->Sync(worker);
     assert(worker->status_ == OK);
     if (ok) {   // spawn a new deblocking/output job
       ctx->io_ = *io;
       ctx->id_ = dec->cache_id_;
       ctx->mb_y_ = dec->mb_y_;
-      ctx->filter_row_ = dec->filter_row_;
-      if (ctx->filter_row_) {    // just swap filter info
+      ctx->filter_row_ = filter_row;
+      if (dec->mt_method_ == 2) {  // swap macroblock data
+        VP8MBData* const tmp = ctx->mb_data_;
+        ctx->mb_data_ = dec->mb_data_;
+        dec->mb_data_ = tmp;
+      } else {
+        // perform reconstruction directly in main thread
+        ReconstructRow(dec, ctx);
+      }
+      if (filter_row) {            // swap filter info
         VP8FInfo* const tmp = ctx->f_info_;
         ctx->f_info_ = dec->f_info_;
         dec->f_info_ = tmp;
       }
-      WebPWorkerLaunch(worker);
+      // (reconstruct)+filter in parallel
+      WebPGetWorkerInterface()->Launch(worker);
       if (++dec->cache_id_ == dec->num_caches_) {
         dec->cache_id_ = 0;
       }
@@ -290,8 +558,8 @@ int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io) {
 
 VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io) {
   // Call setup() first. This may trigger additional decoding features on 'io'.
-  // Note: Afterward, we must call teardown() not matter what.
-  if (io->setup && !io->setup(io)) {
+  // Note: Afterward, we must call teardown() no matter what.
+  if (io->setup != NULL && !io->setup(io)) {
     VP8SetError(dec, VP8_STATUS_USER_ABORT, "Frame setup failed");
     return dec->status_;
   }
@@ -304,7 +572,7 @@ VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io) {
 
   // Define the area where we can skip in-loop filtering, in case of cropping.
   //
-  // 'Simple' filter reads two luma samples outside of the macroblock and
+  // 'Simple' filter reads two luma samples outside of the macroblock
   // and filters one. It doesn't filter the chroma samples. Hence, we can
   // avoid doing the in-loop filtering before crop_top/crop_left position.
   // For the 'Complex' filter, 3 samples are read and up to 3 are filtered.
@@ -339,16 +607,17 @@ VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io) {
       dec->br_mb_y_ = dec->mb_h_;
     }
   }
+  PrecomputeFilterStrengths(dec);
   return VP8_STATUS_OK;
 }
 
 int VP8ExitCritical(VP8Decoder* const dec, VP8Io* const io) {
   int ok = 1;
-  if (dec->use_threads_) {
-    ok = WebPWorkerSync(&dec->worker_);
+  if (dec->mt_method_ > 0) {
+    ok = WebPGetWorkerInterface()->Sync(&dec->worker_);
   }
 
-  if (io->teardown) {
+  if (io->teardown != NULL) {
     io->teardown(io);
   }
   return ok;
@@ -384,9 +653,9 @@ int VP8ExitCritical(VP8Decoder* const dec, VP8Io* const io) {
 // Initialize multi/single-thread worker
 static int InitThreadContext(VP8Decoder* const dec) {
   dec->cache_id_ = 0;
-  if (dec->use_threads_) {
+  if (dec->mt_method_ > 0) {
     WebPWorker* const worker = &dec->worker_;
-    if (!WebPWorkerReset(worker)) {
+    if (!WebPGetWorkerInterface()->Reset(worker)) {
       return VP8SetError(dec, VP8_STATUS_OUT_OF_MEMORY,
                          "thread initialization failed.");
     }
@@ -401,6 +670,28 @@ static int InitThreadContext(VP8Decoder* const dec) {
   return 1;
 }
 
+int VP8GetThreadMethod(const WebPDecoderOptions* const options,
+                       const WebPHeaderStructure* const headers,
+                       int width, int height) {
+  if (options == NULL || options->use_threads == 0) {
+    return 0;
+  }
+  (void)headers;
+  (void)width;
+  (void)height;
+  assert(headers == NULL || !headers->is_lossless);
+#if defined(WEBP_USE_THREAD)
+  if (width < MIN_WIDTH_FOR_THREADS) return 0;
+  // TODO(skal): tune the heuristic further
+#if 0
+  if (height < 2 * width) return 2;
+#endif
+  return 2;
+#else   // !WEBP_USE_THREAD
+  return 0;
+#endif
+}
+
 #undef MT_CACHE_LINES
 #undef ST_CACHE_LINES
 
@@ -412,14 +703,15 @@ static int AllocateMemory(VP8Decoder* const dec) {
   const int mb_w = dec->mb_w_;
   // Note: we use 'size_t' when there's no overflow risk, uint64_t otherwise.
   const size_t intra_pred_mode_size = 4 * mb_w * sizeof(uint8_t);
-  const size_t top_size = (16 + 8 + 8) * mb_w;
+  const size_t top_size = sizeof(VP8TopSamples) * mb_w;
   const size_t mb_info_size = (mb_w + 1) * sizeof(VP8MB);
   const size_t f_info_size =
       (dec->filter_type_ > 0) ?
-          mb_w * (dec->use_threads_ ? 2 : 1) * sizeof(VP8FInfo)
+          mb_w * (dec->mt_method_ > 0 ? 2 : 1) * sizeof(VP8FInfo)
         : 0;
   const size_t yuv_size = YUV_SIZE * sizeof(*dec->yuv_b_);
-  const size_t coeffs_size = 384 * sizeof(*dec->coeffs_);
+  const size_t mb_data_size =
+      (dec->mt_method_ == 2 ? 2 : 1) * mb_w * sizeof(*dec->mb_data_);
   const size_t cache_height = (16 * num_caches
                             + kFilterExtraRows[dec->filter_type_]) * 3 / 2;
   const size_t cache_size = top_size * cache_height;
@@ -428,13 +720,13 @@ static int AllocateMemory(VP8Decoder* const dec) {
       (uint64_t)dec->pic_hdr_.width_ * dec->pic_hdr_.height_ : 0ULL;
   const uint64_t needed = (uint64_t)intra_pred_mode_size
                         + top_size + mb_info_size + f_info_size
-                        + yuv_size + coeffs_size
-                        + cache_size + alpha_size + ALIGN_MASK;
+                        + yuv_size + mb_data_size
+                        + cache_size + alpha_size + WEBP_ALIGN_CST;
   uint8_t* mem;
 
   if (needed != (size_t)needed) return 0;  // check for overflow
   if (needed > dec->mem_size_) {
-    free(dec->mem_);
+    WebPSafeFree(dec->mem_);
     dec->mem_size_ = 0;
     dec->mem_ = WebPSafeMalloc(needed, sizeof(uint8_t));
     if (dec->mem_ == NULL) {
@@ -449,12 +741,8 @@ static int AllocateMemory(VP8Decoder* const dec) {
   dec->intra_t_ = (uint8_t*)mem;
   mem += intra_pred_mode_size;
 
-  dec->y_t_ = (uint8_t*)mem;
-  mem += 16 * mb_w;
-  dec->u_t_ = (uint8_t*)mem;
-  mem += 8 * mb_w;
-  dec->v_t_ = (uint8_t*)mem;
-  mem += 8 * mb_w;
+  dec->yuv_t_ = (VP8TopSamples*)mem;
+  mem += top_size;
 
   dec->mb_info_ = ((VP8MB*)mem) + 1;
   mem += mb_info_size;
@@ -463,20 +751,24 @@ static int AllocateMemory(VP8Decoder* const dec) {
   mem += f_info_size;
   dec->thread_ctx_.id_ = 0;
   dec->thread_ctx_.f_info_ = dec->f_info_;
-  if (dec->use_threads_) {
+  if (dec->mt_method_ > 0) {
     // secondary cache line. The deblocking process need to make use of the
     // filtering strength from previous macroblock row, while the new ones
     // are being decoded in parallel. We'll just swap the pointers.
     dec->thread_ctx_.f_info_ += mb_w;
   }
 
-  mem = (uint8_t*)((uintptr_t)(mem + ALIGN_MASK) & ~ALIGN_MASK);
-  assert((yuv_size & ALIGN_MASK) == 0);
+  mem = (uint8_t*)WEBP_ALIGN(mem);
+  assert((yuv_size & WEBP_ALIGN_CST) == 0);
   dec->yuv_b_ = (uint8_t*)mem;
   mem += yuv_size;
 
-  dec->coeffs_ = (int16_t*)mem;
-  mem += coeffs_size;
+  dec->mb_data_ = (VP8MBData*)mem;
+  dec->thread_ctx_.mb_data_ = (VP8MBData*)mem;
+  if (dec->mt_method_ == 2) {
+    dec->thread_ctx_.mb_data_ += mb_w;
+  }
+  mem += mb_data_size;
 
   dec->cache_y_stride_ = 16 * mb_w;
   dec->cache_uv_stride_ = 8 * mb_w;
@@ -496,9 +788,11 @@ static int AllocateMemory(VP8Decoder* const dec) {
   // alpha plane
   dec->alpha_plane_ = alpha_size ? (uint8_t*)mem : NULL;
   mem += alpha_size;
+  assert(mem <= (uint8_t*)dec->mem_ + dec->mem_size_);
 
-  // note: left-info is initialized once for all.
+  // note: left/top-info is initialized once for all.
   memset(dec->mb_info_ - 1, 0, mb_info_size);
+  VP8InitScanline(dec);   // initialize left too.
 
   // initialize top
   memset(dec->intra_t_, B_DC_PRED, intra_pred_mode_size);
@@ -517,7 +811,7 @@ static void InitIo(VP8Decoder* const dec, VP8Io* io) {
   io->a = NULL;
 }
 
-int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
+int VP8InitFrame(VP8Decoder* const dec, VP8Io* const io) {
   if (!InitThreadContext(dec)) return 0;  // call first. Sets dec->num_caches_.
   if (!AllocateMemory(dec)) return 0;
   InitIo(dec, io);
@@ -526,154 +820,3 @@ int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
 }
 
 //------------------------------------------------------------------------------
-// Main reconstruction function.
-
-static const int kScan[16] = {
-  0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
-  0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
-  0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
-  0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS
-};
-
-static WEBP_INLINE int CheckMode(VP8Decoder* const dec, int mode) {
-  if (mode == B_DC_PRED) {
-    if (dec->mb_x_ == 0) {
-      return (dec->mb_y_ == 0) ? B_DC_PRED_NOTOPLEFT : B_DC_PRED_NOLEFT;
-    } else {
-      return (dec->mb_y_ == 0) ? B_DC_PRED_NOTOP : B_DC_PRED;
-    }
-  }
-  return mode;
-}
-
-static WEBP_INLINE void Copy32b(uint8_t* dst, uint8_t* src) {
-  *(uint32_t*)dst = *(uint32_t*)src;
-}
-
-void VP8ReconstructBlock(VP8Decoder* const dec) {
-  uint8_t* const y_dst = dec->yuv_b_ + Y_OFF;
-  uint8_t* const u_dst = dec->yuv_b_ + U_OFF;
-  uint8_t* const v_dst = dec->yuv_b_ + V_OFF;
-
-  // Rotate in the left samples from previously decoded block. We move four
-  // pixels at a time for alignment reason, and because of in-loop filter.
-  if (dec->mb_x_ > 0) {
-    int j;
-    for (j = -1; j < 16; ++j) {
-      Copy32b(&y_dst[j * BPS - 4], &y_dst[j * BPS + 12]);
-    }
-    for (j = -1; j < 8; ++j) {
-      Copy32b(&u_dst[j * BPS - 4], &u_dst[j * BPS + 4]);
-      Copy32b(&v_dst[j * BPS - 4], &v_dst[j * BPS + 4]);
-    }
-  } else {
-    int j;
-    for (j = 0; j < 16; ++j) {
-      y_dst[j * BPS - 1] = 129;
-    }
-    for (j = 0; j < 8; ++j) {
-      u_dst[j * BPS - 1] = 129;
-      v_dst[j * BPS - 1] = 129;
-    }
-    // Init top-left sample on left column too
-    if (dec->mb_y_ > 0) {
-      y_dst[-1 - BPS] = u_dst[-1 - BPS] = v_dst[-1 - BPS] = 129;
-    }
-  }
-  {
-    // bring top samples into the cache
-    uint8_t* const top_y = dec->y_t_ + dec->mb_x_ * 16;
-    uint8_t* const top_u = dec->u_t_ + dec->mb_x_ * 8;
-    uint8_t* const top_v = dec->v_t_ + dec->mb_x_ * 8;
-    const int16_t* coeffs = dec->coeffs_;
-    int n;
-
-    if (dec->mb_y_ > 0) {
-      memcpy(y_dst - BPS, top_y, 16);
-      memcpy(u_dst - BPS, top_u, 8);
-      memcpy(v_dst - BPS, top_v, 8);
-    } else if (dec->mb_x_ == 0) {
-      // we only need to do this init once at block (0,0).
-      // Afterward, it remains valid for the whole topmost row.
-      memset(y_dst - BPS - 1, 127, 16 + 4 + 1);
-      memset(u_dst - BPS - 1, 127, 8 + 1);
-      memset(v_dst - BPS - 1, 127, 8 + 1);
-    }
-
-    // predict and add residuals
-
-    if (dec->is_i4x4_) {   // 4x4
-      uint32_t* const top_right = (uint32_t*)(y_dst - BPS + 16);
-
-      if (dec->mb_y_ > 0) {
-        if (dec->mb_x_ >= dec->mb_w_ - 1) {    // on rightmost border
-          top_right[0] = top_y[15] * 0x01010101u;
-        } else {
-          memcpy(top_right, top_y + 16, sizeof(*top_right));
-        }
-      }
-      // replicate the top-right pixels below
-      top_right[BPS] = top_right[2 * BPS] = top_right[3 * BPS] = top_right[0];
-
-      // predict and add residues for all 4x4 blocks in turn.
-      for (n = 0; n < 16; n++) {
-        uint8_t* const dst = y_dst + kScan[n];
-        VP8PredLuma4[dec->imodes_[n]](dst);
-        if (dec->non_zero_ac_ & (1 << n)) {
-          VP8Transform(coeffs + n * 16, dst, 0);
-        } else if (dec->non_zero_ & (1 << n)) {  // only DC is present
-          VP8TransformDC(coeffs + n * 16, dst);
-        }
-      }
-    } else {    // 16x16
-      const int pred_func = CheckMode(dec, dec->imodes_[0]);
-      VP8PredLuma16[pred_func](y_dst);
-      if (dec->non_zero_) {
-        for (n = 0; n < 16; n++) {
-          uint8_t* const dst = y_dst + kScan[n];
-          if (dec->non_zero_ac_ & (1 << n)) {
-            VP8Transform(coeffs + n * 16, dst, 0);
-          } else if (dec->non_zero_ & (1 << n)) {  // only DC is present
-            VP8TransformDC(coeffs + n * 16, dst);
-          }
-        }
-      }
-    }
-    {
-      // Chroma
-      const int pred_func = CheckMode(dec, dec->uvmode_);
-      VP8PredChroma8[pred_func](u_dst);
-      VP8PredChroma8[pred_func](v_dst);
-
-      if (dec->non_zero_ & 0x0f0000) {   // chroma-U
-        const int16_t* const u_coeffs = dec->coeffs_ + 16 * 16;
-        if (dec->non_zero_ac_ & 0x0f0000) {
-          VP8TransformUV(u_coeffs, u_dst);
-        } else {
-          VP8TransformDCUV(u_coeffs, u_dst);
-        }
-      }
-      if (dec->non_zero_ & 0xf00000) {   // chroma-V
-        const int16_t* const v_coeffs = dec->coeffs_ + 20 * 16;
-        if (dec->non_zero_ac_ & 0xf00000) {
-          VP8TransformUV(v_coeffs, v_dst);
-        } else {
-          VP8TransformDCUV(v_coeffs, v_dst);
-        }
-      }
-
-      // stash away top samples for next block
-      if (dec->mb_y_ < dec->mb_h_ - 1) {
-        memcpy(top_y, y_dst + 15 * BPS, 16);
-        memcpy(top_u, u_dst +  7 * BPS,  8);
-        memcpy(top_v, v_dst +  7 * BPS,  8);
-      }
-    }
-  }
-}
-
-//------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/idec.c b/drivers/webp/dec/idec.c
index 7df790ced8..abafb9f3d1 100644
--- a/drivers/webp/dec/idec.c
+++ b/drivers/webp/dec/idec.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Incremental decoding
@@ -13,14 +15,11 @@
 #include <string.h>
 #include <stdlib.h>
 
+#include "./alphai.h"
 #include "./webpi.h"
 #include "./vp8i.h"
 #include "../utils/utils.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 // In append mode, buffer allocations increase as multiples of this value.
 // Needs to be a power of 2.
 #define CHUNK_SIZE 4096
@@ -29,11 +28,13 @@ extern "C" {
 //------------------------------------------------------------------------------
 // Data structures for memory and states
 
-// Decoding states. State normally flows like HEADER->PARTS0->DATA->DONE.
+// Decoding states. State normally flows as:
+// WEBP_HEADER->VP8_HEADER->VP8_PARTS0->VP8_DATA->DONE for a lossy image, and
+// WEBP_HEADER->VP8L_HEADER->VP8L_DATA->DONE for a lossless image.
 // If there is any error the decoder goes into state ERROR.
 typedef enum {
-  STATE_PRE_VP8,  // All data before that of the first VP8 chunk.
-  STATE_VP8_FRAME_HEADER,  // For VP8 Frame header (within VP8 chunk).
+  STATE_WEBP_HEADER,  // All the data before that of the VP8/VP8L chunk.
+  STATE_VP8_HEADER,   // The VP8 Frame header (within the VP8 chunk).
   STATE_VP8_PARTS0,
   STATE_VP8_DATA,
   STATE_VP8L_HEADER,
@@ -71,32 +72,41 @@ struct WebPIDecoder {
   MemBuffer mem_;          // input memory buffer.
   WebPDecBuffer output_;   // output buffer (when no external one is supplied)
   size_t chunk_size_;      // Compressed VP8/VP8L size extracted from Header.
+
+  int last_mb_y_;          // last row reached for intra-mode decoding
 };
 
 // MB context to restore in case VP8DecodeMB() fails
 typedef struct {
   VP8MB left_;
   VP8MB info_;
-  uint8_t intra_t_[4];
-  uint8_t intra_l_[4];
-  VP8BitReader br_;
   VP8BitReader token_br_;
 } MBContext;
 
 //------------------------------------------------------------------------------
 // MemBuffer: incoming data handling
 
-static void RemapBitReader(VP8BitReader* const br, ptrdiff_t offset) {
-  if (br->buf_ != NULL) {
-    br->buf_ += offset;
-    br->buf_end_ += offset;
-  }
-}
-
 static WEBP_INLINE size_t MemDataSize(const MemBuffer* mem) {
   return (mem->end_ - mem->start_);
 }
 
+// Check if we need to preserve the compressed alpha data, as it may not have
+// been decoded yet.
+static int NeedCompressedAlpha(const WebPIDecoder* const idec) {
+  if (idec->state_ == STATE_WEBP_HEADER) {
+    // We haven't parsed the headers yet, so we don't know whether the image is
+    // lossy or lossless. This also means that we haven't parsed the ALPH chunk.
+    return 0;
+  }
+  if (idec->is_lossless_) {
+    return 0;  // ALPH chunk is not present for lossless images.
+  } else {
+    const VP8Decoder* const dec = (VP8Decoder*)idec->dec_;
+    assert(dec != NULL);  // Must be true as idec->state_ != STATE_WEBP_HEADER.
+    return (dec->alpha_data_ != NULL) && !dec->is_alpha_decoded_;
+  }
+}
+
 static void DoRemap(WebPIDecoder* const idec, ptrdiff_t offset) {
   MemBuffer* const mem = &idec->mem_;
   const uint8_t* const new_base = mem->buf_ + mem->start_;
@@ -112,16 +122,36 @@ static void DoRemap(WebPIDecoder* const idec, ptrdiff_t offset) {
       if (offset != 0) {
         int p;
         for (p = 0; p <= last_part; ++p) {
-          RemapBitReader(dec->parts_ + p, offset);
+          VP8RemapBitReader(dec->parts_ + p, offset);
         }
         // Remap partition #0 data pointer to new offset, but only in MAP
         // mode (in APPEND mode, partition #0 is copied into a fixed memory).
         if (mem->mode_ == MEM_MODE_MAP) {
-          RemapBitReader(&dec->br_, offset);
+          VP8RemapBitReader(&dec->br_, offset);
+        }
+      }
+      {
+        const uint8_t* const last_start = dec->parts_[last_part].buf_;
+        assert(last_part >= 0);
+        VP8BitReaderSetBuffer(&dec->parts_[last_part], last_start,
+                              mem->buf_ + mem->end_ - last_start);
+      }
+      if (NeedCompressedAlpha(idec)) {
+        ALPHDecoder* const alph_dec = dec->alph_dec_;
+        dec->alpha_data_ += offset;
+        if (alph_dec != NULL) {
+          if (alph_dec->method_ == ALPHA_LOSSLESS_COMPRESSION) {
+            VP8LDecoder* const alph_vp8l_dec = alph_dec->vp8l_dec_;
+            assert(alph_vp8l_dec != NULL);
+            assert(dec->alpha_data_size_ >= ALPHA_HEADER_LEN);
+            VP8LBitReaderSetBuffer(&alph_vp8l_dec->br_,
+                                   dec->alpha_data_ + ALPHA_HEADER_LEN,
+                                   dec->alpha_data_size_ - ALPHA_HEADER_LEN);
+          } else {  // alph_dec->method_ == ALPHA_NO_COMPRESSION
+            // Nothing special to do in this case.
+          }
         }
       }
-      assert(last_part >= 0);
-      dec->parts_[last_part].buf_end_ = mem->buf_ + mem->end_;
     } else {    // Resize lossless bitreader
       VP8LDecoder* const dec = (VP8LDecoder*)idec->dec_;
       VP8LBitReaderSetBuffer(&dec->br_, new_base, MemDataSize(mem));
@@ -133,8 +163,12 @@ static void DoRemap(WebPIDecoder* const idec, ptrdiff_t offset) {
 // size if required and also updates VP8BitReader's if new memory is allocated.
 static int AppendToMemBuffer(WebPIDecoder* const idec,
                              const uint8_t* const data, size_t data_size) {
+  VP8Decoder* const dec = (VP8Decoder*)idec->dec_;
   MemBuffer* const mem = &idec->mem_;
-  const uint8_t* const old_base = mem->buf_ + mem->start_;
+  const int need_compressed_alpha = NeedCompressedAlpha(idec);
+  const uint8_t* const old_start = mem->buf_ + mem->start_;
+  const uint8_t* const old_base =
+      need_compressed_alpha ? dec->alpha_data_ : old_start;
   assert(mem->mode_ == MEM_MODE_APPEND);
   if (data_size > MAX_CHUNK_PAYLOAD) {
     // security safeguard: trying to allocate more than what the format
@@ -143,17 +177,18 @@ static int AppendToMemBuffer(WebPIDecoder* const idec,
   }
 
   if (mem->end_ + data_size > mem->buf_size_) {  // Need some free memory
-    const size_t current_size = MemDataSize(mem);
+    const size_t new_mem_start = old_start - old_base;
+    const size_t current_size = MemDataSize(mem) + new_mem_start;
     const uint64_t new_size = (uint64_t)current_size + data_size;
     const uint64_t extra_size = (new_size + CHUNK_SIZE - 1) & ~(CHUNK_SIZE - 1);
     uint8_t* const new_buf =
         (uint8_t*)WebPSafeMalloc(extra_size, sizeof(*new_buf));
     if (new_buf == NULL) return 0;
     memcpy(new_buf, old_base, current_size);
-    free(mem->buf_);
+    WebPSafeFree(mem->buf_);
     mem->buf_ = new_buf;
     mem->buf_size_ = (size_t)extra_size;
-    mem->start_ = 0;
+    mem->start_ = new_mem_start;
     mem->end_ = current_size;
   }
 
@@ -161,14 +196,15 @@ static int AppendToMemBuffer(WebPIDecoder* const idec,
   mem->end_ += data_size;
   assert(mem->end_ <= mem->buf_size_);
 
-  DoRemap(idec, mem->buf_ + mem->start_ - old_base);
+  DoRemap(idec, mem->buf_ + mem->start_ - old_start);
   return 1;
 }
 
 static int RemapMemBuffer(WebPIDecoder* const idec,
                           const uint8_t* const data, size_t data_size) {
   MemBuffer* const mem = &idec->mem_;
-  const uint8_t* const old_base = mem->buf_ + mem->start_;
+  const uint8_t* const old_buf = mem->buf_;
+  const uint8_t* const old_start = old_buf + mem->start_;
   assert(mem->mode_ == MEM_MODE_MAP);
 
   if (data_size < mem->buf_size_) return 0;  // can't remap to a shorter buffer!
@@ -176,7 +212,7 @@ static int RemapMemBuffer(WebPIDecoder* const idec,
   mem->buf_ = (uint8_t*)data;
   mem->end_ = mem->buf_size_ = data_size;
 
-  DoRemap(idec, mem->buf_ + mem->start_ - old_base);
+  DoRemap(idec, mem->buf_ + mem->start_ - old_start);
   return 1;
 }
 
@@ -191,8 +227,8 @@ static void InitMemBuffer(MemBuffer* const mem) {
 static void ClearMemBuffer(MemBuffer* const mem) {
   assert(mem);
   if (mem->mode_ == MEM_MODE_APPEND) {
-    free(mem->buf_);
-    free((void*)mem->part0_buf_);
+    WebPSafeFree(mem->buf_);
+    WebPSafeFree((void*)mem->part0_buf_);
   }
 }
 
@@ -206,35 +242,34 @@ static int CheckMemBufferMode(MemBuffer* const mem, MemBufferMode expected) {
   return 1;
 }
 
+// To be called last.
+static VP8StatusCode FinishDecoding(WebPIDecoder* const idec) {
+  const WebPDecoderOptions* const options = idec->params_.options;
+  WebPDecBuffer* const output = idec->params_.output;
+
+  idec->state_ = STATE_DONE;
+  if (options != NULL && options->flip) {
+    return WebPFlipBuffer(output);
+  } else {
+    return VP8_STATUS_OK;
+  }
+}
+
 //------------------------------------------------------------------------------
 // Macroblock-decoding contexts
 
 static void SaveContext(const VP8Decoder* dec, const VP8BitReader* token_br,
                         MBContext* const context) {
-  const VP8BitReader* const br = &dec->br_;
-  const VP8MB* const left = dec->mb_info_ - 1;
-  const VP8MB* const info = dec->mb_info_ + dec->mb_x_;
-
-  context->left_ = *left;
-  context->info_ = *info;
-  context->br_ = *br;
+  context->left_ = dec->mb_info_[-1];
+  context->info_ = dec->mb_info_[dec->mb_x_];
   context->token_br_ = *token_br;
-  memcpy(context->intra_t_, dec->intra_t_ + 4 * dec->mb_x_, 4);
-  memcpy(context->intra_l_, dec->intra_l_, 4);
 }
 
 static void RestoreContext(const MBContext* context, VP8Decoder* const dec,
                            VP8BitReader* const token_br) {
-  VP8BitReader* const br = &dec->br_;
-  VP8MB* const left = dec->mb_info_ - 1;
-  VP8MB* const info = dec->mb_info_ + dec->mb_x_;
-
-  *left = context->left_;
-  *info = context->info_;
-  *br = context->br_;
+  dec->mb_info_[-1] = context->left_;
+  dec->mb_info_[dec->mb_x_] = context->info_;
   *token_br = context->token_br_;
-  memcpy(dec->intra_t_ + 4 * dec->mb_x_, context->intra_t_, 4);
-  memcpy(dec->intra_l_, context->intra_l_, 4);
 }
 
 //------------------------------------------------------------------------------
@@ -242,7 +277,7 @@ static void RestoreContext(const MBContext* context, VP8Decoder* const dec,
 static VP8StatusCode IDecError(WebPIDecoder* const idec, VP8StatusCode error) {
   if (idec->state_ == STATE_VP8_DATA) {
     VP8Io* const io = &idec->io_;
-    if (io->teardown) {
+    if (io->teardown != NULL) {
       io->teardown(io);
     }
   }
@@ -270,6 +305,7 @@ static VP8StatusCode DecodeWebPHeaders(WebPIDecoder* const idec) {
 
   headers.data = data;
   headers.data_size = curr_size;
+  headers.have_all_data = 0;
   status = WebPParseHeaders(&headers);
   if (status == VP8_STATUS_NOT_ENOUGH_DATA) {
     return VP8_STATUS_SUSPENDED;  // We haven't found a VP8 chunk yet.
@@ -285,15 +321,9 @@ static VP8StatusCode DecodeWebPHeaders(WebPIDecoder* const idec) {
       return VP8_STATUS_OUT_OF_MEMORY;
     }
     idec->dec_ = dec;
-#ifdef WEBP_USE_THREAD
-    dec->use_threads_ = (idec->params_.options != NULL) &&
-                        (idec->params_.options->use_threads > 0);
-#else
-    dec->use_threads_ = 0;
-#endif
     dec->alpha_data_ = headers.alpha_data;
     dec->alpha_data_size_ = headers.alpha_data_size;
-    ChangeState(idec, STATE_VP8_FRAME_HEADER, headers.offset);
+    ChangeState(idec, STATE_VP8_HEADER, headers.offset);
   } else {
     VP8LDecoder* const dec = VP8LNew();
     if (dec == NULL) {
@@ -308,13 +338,14 @@ static VP8StatusCode DecodeWebPHeaders(WebPIDecoder* const idec) {
 static VP8StatusCode DecodeVP8FrameHeader(WebPIDecoder* const idec) {
   const uint8_t* data = idec->mem_.buf_ + idec->mem_.start_;
   const size_t curr_size = MemDataSize(&idec->mem_);
+  int width, height;
   uint32_t bits;
 
   if (curr_size < VP8_FRAME_HEADER_SIZE) {
     // Not enough data bytes to extract VP8 Frame Header.
     return VP8_STATUS_SUSPENDED;
   }
-  if (!VP8GetInfo(data, curr_size, idec->chunk_size_, NULL, NULL)) {
+  if (!VP8GetInfo(data, curr_size, idec->chunk_size_, &width, &height)) {
     return IDecError(idec, VP8_STATUS_BITSTREAM_ERROR);
   }
 
@@ -328,30 +359,32 @@ static VP8StatusCode DecodeVP8FrameHeader(WebPIDecoder* const idec) {
 }
 
 // Partition #0
-static int CopyParts0Data(WebPIDecoder* const idec) {
+static VP8StatusCode CopyParts0Data(WebPIDecoder* const idec) {
   VP8Decoder* const dec = (VP8Decoder*)idec->dec_;
   VP8BitReader* const br = &dec->br_;
-  const size_t psize = br->buf_end_ - br->buf_;
+  const size_t part_size = br->buf_end_ - br->buf_;
   MemBuffer* const mem = &idec->mem_;
   assert(!idec->is_lossless_);
   assert(mem->part0_buf_ == NULL);
-  assert(psize > 0);
-  assert(psize <= mem->part0_size_);  // Format limit: no need for runtime check
+  // the following is a format limitation, no need for runtime check:
+  assert(part_size <= mem->part0_size_);
+  if (part_size == 0) {   // can't have zero-size partition #0
+    return VP8_STATUS_BITSTREAM_ERROR;
+  }
   if (mem->mode_ == MEM_MODE_APPEND) {
     // We copy and grab ownership of the partition #0 data.
-    uint8_t* const part0_buf = (uint8_t*)malloc(psize);
+    uint8_t* const part0_buf = (uint8_t*)WebPSafeMalloc(1ULL, part_size);
     if (part0_buf == NULL) {
-      return 0;
+      return VP8_STATUS_OUT_OF_MEMORY;
     }
-    memcpy(part0_buf, br->buf_, psize);
+    memcpy(part0_buf, br->buf_, part_size);
     mem->part0_buf_ = part0_buf;
-    br->buf_ = part0_buf;
-    br->buf_end_ = part0_buf + psize;
+    VP8BitReaderSetBuffer(br, part0_buf, part_size);
   } else {
     // Else: just keep pointers to the partition #0's data in dec_->br_.
   }
-  mem->start_ += psize;
-  return 1;
+  mem->start_ += part_size;
+  return VP8_STATUS_OK;
 }
 
 static VP8StatusCode DecodePartition0(WebPIDecoder* const idec) {
@@ -381,9 +414,14 @@ static VP8StatusCode DecodePartition0(WebPIDecoder* const idec) {
   if (dec->status_ != VP8_STATUS_OK) {
     return IDecError(idec, dec->status_);
   }
+  // This change must be done before calling VP8InitFrame()
+  dec->mt_method_ = VP8GetThreadMethod(params->options, NULL,
+                                       io->width, io->height);
+  VP8InitDithering(params->options, dec);
 
-  if (!CopyParts0Data(idec)) {
-    return IDecError(idec, VP8_STATUS_OUT_OF_MEMORY);
+  dec->status_ = CopyParts0Data(idec);
+  if (dec->status_ != VP8_STATUS_OK) {
+    return IDecError(idec, dec->status_);
   }
 
   // Finish setting up the decoding parameters. Will call io->setup().
@@ -407,50 +445,52 @@ static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
   VP8Io* const io = &idec->io_;
 
   assert(dec->ready_);
-
   for (; dec->mb_y_ < dec->mb_h_; ++dec->mb_y_) {
-    VP8BitReader* token_br = &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
-    if (dec->mb_x_ == 0) {
-      VP8InitScanline(dec);
+    if (idec->last_mb_y_ != dec->mb_y_) {
+      if (!VP8ParseIntraModeRow(&dec->br_, dec)) {
+        // note: normally, error shouldn't occur since we already have the whole
+        // partition0 available here in DecodeRemaining(). Reaching EOF while
+        // reading intra modes really means a BITSTREAM_ERROR.
+        return IDecError(idec, VP8_STATUS_BITSTREAM_ERROR);
+      }
+      idec->last_mb_y_ = dec->mb_y_;
     }
-    for (; dec->mb_x_ < dec->mb_w_;  dec->mb_x_++) {
+    for (; dec->mb_x_ < dec->mb_w_; ++dec->mb_x_) {
+      VP8BitReader* const token_br =
+          &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
       MBContext context;
       SaveContext(dec, token_br, &context);
-
       if (!VP8DecodeMB(dec, token_br)) {
-        RestoreContext(&context, dec, token_br);
         // We shouldn't fail when MAX_MB data was available
         if (dec->num_parts_ == 1 && MemDataSize(&idec->mem_) > MAX_MB_SIZE) {
           return IDecError(idec, VP8_STATUS_BITSTREAM_ERROR);
         }
+        RestoreContext(&context, dec, token_br);
         return VP8_STATUS_SUSPENDED;
       }
-      VP8ReconstructBlock(dec);
-      // Store data and save block's filtering params
-      VP8StoreBlock(dec);
-
       // Release buffer only if there is only one partition
       if (dec->num_parts_ == 1) {
         idec->mem_.start_ = token_br->buf_ - idec->mem_.buf_;
         assert(idec->mem_.start_ <= idec->mem_.end_);
       }
     }
+    VP8InitScanline(dec);   // Prepare for next scanline
+
+    // Reconstruct, filter and emit the row.
     if (!VP8ProcessRow(dec, io)) {
       return IDecError(idec, VP8_STATUS_USER_ABORT);
     }
-    dec->mb_x_ = 0;
   }
   // Synchronize the thread and check for errors.
   if (!VP8ExitCritical(dec, io)) {
     return IDecError(idec, VP8_STATUS_USER_ABORT);
   }
   dec->ready_ = 0;
-  idec->state_ = STATE_DONE;
-
-  return VP8_STATUS_OK;
+  return FinishDecoding(idec);
 }
 
-static int ErrorStatusLossless(WebPIDecoder* const idec, VP8StatusCode status) {
+static VP8StatusCode ErrorStatusLossless(WebPIDecoder* const idec,
+                                         VP8StatusCode status) {
   if (status == VP8_STATUS_SUSPENDED || status == VP8_STATUS_NOT_ENOUGH_DATA) {
     return VP8_STATUS_SUSPENDED;
   }
@@ -467,9 +507,15 @@ static VP8StatusCode DecodeVP8LHeader(WebPIDecoder* const idec) {
 
   // Wait until there's enough data for decoding header.
   if (curr_size < (idec->chunk_size_ >> 3)) {
-    return VP8_STATUS_SUSPENDED;
+    dec->status_ = VP8_STATUS_SUSPENDED;
+    return ErrorStatusLossless(idec, dec->status_);
   }
+
   if (!VP8LDecodeHeader(dec, io)) {
+    if (dec->status_ == VP8_STATUS_BITSTREAM_ERROR &&
+        curr_size < idec->chunk_size_) {
+      dec->status_ = VP8_STATUS_SUSPENDED;
+    }
     return ErrorStatusLossless(idec, dec->status_);
   }
   // Allocate/verify output buffer now.
@@ -488,33 +534,29 @@ static VP8StatusCode DecodeVP8LData(WebPIDecoder* const idec) {
   const size_t curr_size = MemDataSize(&idec->mem_);
   assert(idec->is_lossless_);
 
-  // At present Lossless decoder can't decode image incrementally. So wait till
-  // all the image data is aggregated before image can be decoded.
-  if (curr_size < idec->chunk_size_) {
-    return VP8_STATUS_SUSPENDED;
-  }
+  // Switch to incremental decoding if we don't have all the bytes available.
+  dec->incremental_ = (curr_size < idec->chunk_size_);
 
   if (!VP8LDecodeImage(dec)) {
     return ErrorStatusLossless(idec, dec->status_);
   }
-
-  idec->state_ = STATE_DONE;
-
-  return VP8_STATUS_OK;
+  assert(dec->status_ == VP8_STATUS_OK || dec->status_ == VP8_STATUS_SUSPENDED);
+  return (dec->status_ == VP8_STATUS_SUSPENDED) ? dec->status_
+                                                : FinishDecoding(idec);
 }
 
   // Main decoding loop
 static VP8StatusCode IDecode(WebPIDecoder* idec) {
   VP8StatusCode status = VP8_STATUS_SUSPENDED;
 
-  if (idec->state_ == STATE_PRE_VP8) {
+  if (idec->state_ == STATE_WEBP_HEADER) {
     status = DecodeWebPHeaders(idec);
   } else {
     if (idec->dec_ == NULL) {
       return VP8_STATUS_SUSPENDED;    // can't continue if we have no decoder.
     }
   }
-  if (idec->state_ == STATE_VP8_FRAME_HEADER) {
+  if (idec->state_ == STATE_VP8_HEADER) {
     status = DecodeVP8FrameHeader(idec);
   }
   if (idec->state_ == STATE_VP8_PARTS0) {
@@ -536,20 +578,23 @@ static VP8StatusCode IDecode(WebPIDecoder* idec) {
 // Public functions
 
 WebPIDecoder* WebPINewDecoder(WebPDecBuffer* output_buffer) {
-  WebPIDecoder* idec = (WebPIDecoder*)calloc(1, sizeof(*idec));
+  WebPIDecoder* idec = (WebPIDecoder*)WebPSafeCalloc(1ULL, sizeof(*idec));
   if (idec == NULL) {
     return NULL;
   }
 
-  idec->state_ = STATE_PRE_VP8;
+  idec->state_ = STATE_WEBP_HEADER;
   idec->chunk_size_ = 0;
 
+  idec->last_mb_y_ = -1;
+
   InitMemBuffer(&idec->mem_);
   WebPInitDecBuffer(&idec->output_);
   VP8InitIo(&idec->io_);
 
   WebPResetDecParams(&idec->params_);
-  idec->params_.output = output_buffer ? output_buffer : &idec->output_;
+  idec->params_.output = (output_buffer != NULL) ? output_buffer
+                                                 : &idec->output_;
   WebPInitCustomIo(&idec->params_, &idec->io_);  // Plug the I/O functions.
 
   return idec;
@@ -581,14 +626,18 @@ void WebPIDelete(WebPIDecoder* idec) {
   if (idec == NULL) return;
   if (idec->dec_ != NULL) {
     if (!idec->is_lossless_) {
-      VP8Delete(idec->dec_);
+      if (idec->state_ == STATE_VP8_DATA) {
+        // Synchronize the thread, clean-up and check for errors.
+        VP8ExitCritical((VP8Decoder*)idec->dec_, &idec->io_);
+      }
+      VP8Delete((VP8Decoder*)idec->dec_);
     } else {
-      VP8LDelete(idec->dec_);
+      VP8LDelete((VP8LDecoder*)idec->dec_);
     }
   }
   ClearMemBuffer(&idec->mem_);
   WebPFreeDecBuffer(&idec->output_);
-  free(idec);
+  WebPSafeFree(idec);
 }
 
 //------------------------------------------------------------------------------
@@ -596,12 +645,22 @@ void WebPIDelete(WebPIDecoder* idec) {
 
 WebPIDecoder* WebPINewRGB(WEBP_CSP_MODE mode, uint8_t* output_buffer,
                           size_t output_buffer_size, int output_stride) {
+  const int is_external_memory = (output_buffer != NULL);
   WebPIDecoder* idec;
+
   if (mode >= MODE_YUV) return NULL;
+  if (!is_external_memory) {    // Overwrite parameters to sane values.
+    output_buffer_size = 0;
+    output_stride = 0;
+  } else {  // A buffer was passed. Validate the other params.
+    if (output_stride == 0 || output_buffer_size == 0) {
+      return NULL;   // invalid parameter.
+    }
+  }
   idec = WebPINewDecoder(NULL);
   if (idec == NULL) return NULL;
   idec->output_.colorspace = mode;
-  idec->output_.is_external_memory = 1;
+  idec->output_.is_external_memory = is_external_memory;
   idec->output_.u.RGBA.rgba = output_buffer;
   idec->output_.u.RGBA.stride = output_stride;
   idec->output_.u.RGBA.size = output_buffer_size;
@@ -612,10 +671,30 @@ WebPIDecoder* WebPINewYUVA(uint8_t* luma, size_t luma_size, int luma_stride,
                            uint8_t* u, size_t u_size, int u_stride,
                            uint8_t* v, size_t v_size, int v_stride,
                            uint8_t* a, size_t a_size, int a_stride) {
-  WebPIDecoder* const idec = WebPINewDecoder(NULL);
+  const int is_external_memory = (luma != NULL);
+  WebPIDecoder* idec;
+  WEBP_CSP_MODE colorspace;
+
+  if (!is_external_memory) {    // Overwrite parameters to sane values.
+    luma_size = u_size = v_size = a_size = 0;
+    luma_stride = u_stride = v_stride = a_stride = 0;
+    u = v = a = NULL;
+    colorspace = MODE_YUVA;
+  } else {  // A luma buffer was passed. Validate the other parameters.
+    if (u == NULL || v == NULL) return NULL;
+    if (luma_size == 0 || u_size == 0 || v_size == 0) return NULL;
+    if (luma_stride == 0 || u_stride == 0 || v_stride == 0) return NULL;
+    if (a != NULL) {
+      if (a_size == 0 || a_stride == 0) return NULL;
+    }
+    colorspace = (a == NULL) ? MODE_YUV : MODE_YUVA;
+  }
+
+  idec = WebPINewDecoder(NULL);
   if (idec == NULL) return NULL;
-  idec->output_.colorspace = (a == NULL) ? MODE_YUV : MODE_YUVA;
-  idec->output_.is_external_memory = 1;
+
+  idec->output_.colorspace = colorspace;
+  idec->output_.is_external_memory = is_external_memory;
   idec->output_.u.YUVA.y = luma;
   idec->output_.u.YUVA.y_stride = luma_stride;
   idec->output_.u.YUVA.y_size = luma_size;
@@ -768,7 +847,7 @@ int WebPISetIOHooks(WebPIDecoder* const idec,
                     VP8IoSetupHook setup,
                     VP8IoTeardownHook teardown,
                     void* user_data) {
-  if (idec == NULL || idec->state_ > STATE_PRE_VP8) {
+  if (idec == NULL || idec->state_ > STATE_WEBP_HEADER) {
     return 0;
   }
 
@@ -779,7 +858,3 @@ int WebPISetIOHooks(WebPIDecoder* const idec,
 
   return 1;
 }
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/io.c b/drivers/webp/dec/io.c
index 594804c2e6..13e469ab73 100644
--- a/drivers/webp/dec/io.c
+++ b/drivers/webp/dec/io.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // functions for sample output.
@@ -15,10 +17,7 @@
 #include "./webpi.h"
 #include "../dsp/dsp.h"
 #include "../dsp/yuv.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 // Main YUV<->RGB conversion functions
@@ -46,57 +45,17 @@ static int EmitYUV(const VP8Io* const io, WebPDecParams* const p) {
 
 // Point-sampling U/V sampler.
 static int EmitSampledRGB(const VP8Io* const io, WebPDecParams* const p) {
-  WebPDecBuffer* output = p->output;
-  const WebPRGBABuffer* const buf = &output->u.RGBA;
-  uint8_t* dst = buf->rgba + io->mb_y * buf->stride;
-  const uint8_t* y_src = io->y;
-  const uint8_t* u_src = io->u;
-  const uint8_t* v_src = io->v;
-  const WebPSampleLinePairFunc sample = WebPSamplers[output->colorspace];
-  const int mb_w = io->mb_w;
-  const int last = io->mb_h - 1;
-  int j;
-  for (j = 0; j < last; j += 2) {
-    sample(y_src, y_src + io->y_stride, u_src, v_src,
-           dst, dst + buf->stride, mb_w);
-    y_src += 2 * io->y_stride;
-    u_src += io->uv_stride;
-    v_src += io->uv_stride;
-    dst += 2 * buf->stride;
-  }
-  if (j == last) {  // Just do the last line twice
-    sample(y_src, y_src, u_src, v_src, dst, dst, mb_w);
-  }
+  WebPDecBuffer* const output = p->output;
+  WebPRGBABuffer* const buf = &output->u.RGBA;
+  uint8_t* const dst = buf->rgba + io->mb_y * buf->stride;
+  WebPSamplerProcessPlane(io->y, io->y_stride,
+                          io->u, io->v, io->uv_stride,
+                          dst, buf->stride, io->mb_w, io->mb_h,
+                          WebPSamplers[output->colorspace]);
   return io->mb_h;
 }
 
 //------------------------------------------------------------------------------
-// YUV444 -> RGB conversion
-
-#if 0   // TODO(skal): this is for future rescaling.
-static int EmitRGB(const VP8Io* const io, WebPDecParams* const p) {
-  WebPDecBuffer* output = p->output;
-  const WebPRGBABuffer* const buf = &output->u.RGBA;
-  uint8_t* dst = buf->rgba + io->mb_y * buf->stride;
-  const uint8_t* y_src = io->y;
-  const uint8_t* u_src = io->u;
-  const uint8_t* v_src = io->v;
-  const WebPYUV444Converter convert = WebPYUV444Converters[output->colorspace];
-  const int mb_w = io->mb_w;
-  const int last = io->mb_h;
-  int j;
-  for (j = 0; j < last; ++j) {
-    convert(y_src, u_src, v_src, dst, mb_w);
-    y_src += io->y_stride;
-    u_src += io->uv_stride;
-    v_src += io->uv_stride;
-    dst += buf->stride;
-  }
-  return io->mb_h;
-}
-#endif
-
-//------------------------------------------------------------------------------
 // Fancy upsampling
 
 #ifdef FANCY_UPSAMPLING
@@ -117,7 +76,7 @@ static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
 
   if (y == 0) {
     // First line is special cased. We mirror the u/v samples at boundary.
-    upsample(NULL, cur_y, cur_u, cur_v, cur_u, cur_v, NULL, dst, mb_w);
+    upsample(cur_y, NULL, cur_u, cur_v, cur_u, cur_v, dst, NULL, mb_w);
   } else {
     // We can finish the left-over line from previous call.
     upsample(p->tmp_y, cur_y, top_u, top_v, cur_u, cur_v,
@@ -160,14 +119,16 @@ static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
 
 //------------------------------------------------------------------------------
 
-static int EmitAlphaYUV(const VP8Io* const io, WebPDecParams* const p) {
+static int EmitAlphaYUV(const VP8Io* const io, WebPDecParams* const p,
+                        int expected_num_lines_out) {
   const uint8_t* alpha = io->a;
   const WebPYUVABuffer* const buf = &p->output->u.YUVA;
   const int mb_w = io->mb_w;
   const int mb_h = io->mb_h;
   uint8_t* dst = buf->a + io->mb_y * buf->a_stride;
   int j;
-
+  (void)expected_num_lines_out;
+  assert(expected_num_lines_out == mb_h);
   if (alpha != NULL) {
     for (j = 0; j < mb_h; ++j) {
       memcpy(dst, alpha, mb_w * sizeof(*dst));
@@ -210,7 +171,8 @@ static int GetAlphaSourceRow(const VP8Io* const io,
   return start_y;
 }
 
-static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
+static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p,
+                        int expected_num_lines_out) {
   const uint8_t* alpha = io->a;
   if (alpha != NULL) {
     const int mb_w = io->mb_w;
@@ -221,21 +183,13 @@ static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
     int num_rows;
     const int start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
     uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
-    uint8_t* dst = base_rgba + (alpha_first ? 0 : 3);
-    uint32_t alpha_mask = 0xff;
-    int i, j;
-
-    for (j = 0; j < num_rows; ++j) {
-      for (i = 0; i < mb_w; ++i) {
-        const uint32_t alpha_value = alpha[i];
-        dst[4 * i] = alpha_value;
-        alpha_mask &= alpha_value;
-      }
-      alpha += io->width;
-      dst += buf->stride;
-    }
-    // alpha_mask is < 0xff if there's non-trivial alpha to premultiply with.
-    if (alpha_mask != 0xff && WebPIsPremultipliedMode(colorspace)) {
+    uint8_t* const dst = base_rgba + (alpha_first ? 0 : 3);
+    const int has_alpha = WebPDispatchAlpha(alpha, io->width, mb_w,
+                                            num_rows, dst, buf->stride);
+    (void)expected_num_lines_out;
+    assert(expected_num_lines_out == num_rows);
+    // has_alpha is true if there's non-trivial alpha to premultiply with.
+    if (has_alpha && WebPIsPremultipliedMode(colorspace)) {
       WebPApplyAlphaMultiply(base_rgba, alpha_first,
                              mb_w, num_rows, buf->stride);
     }
@@ -243,7 +197,8 @@ static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
   return 0;
 }
 
-static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
+static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p,
+                             int expected_num_lines_out) {
   const uint8_t* alpha = io->a;
   if (alpha != NULL) {
     const int mb_w = io->mb_w;
@@ -252,10 +207,13 @@ static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
     int num_rows;
     const int start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
     uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
+#ifdef WEBP_SWAP_16BIT_CSP
+    uint8_t* alpha_dst = base_rgba;
+#else
     uint8_t* alpha_dst = base_rgba + 1;
+#endif
     uint32_t alpha_mask = 0x0f;
     int i, j;
-
     for (j = 0; j < num_rows; ++j) {
       for (i = 0; i < mb_w; ++i) {
         // Fill in the alpha value (converted to 4 bits).
@@ -266,6 +224,8 @@ static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
       alpha += io->width;
       alpha_dst += buf->stride;
     }
+    (void)expected_num_lines_out;
+    assert(expected_num_lines_out == num_rows);
     if (alpha_mask != 0x0f && WebPIsPremultipliedMode(colorspace)) {
       WebPApplyAlphaMultiply4444(base_rgba, mb_w, num_rows, buf->stride);
     }
@@ -291,15 +251,35 @@ static int Rescale(const uint8_t* src, int src_stride,
 static int EmitRescaledYUV(const VP8Io* const io, WebPDecParams* const p) {
   const int mb_h = io->mb_h;
   const int uv_mb_h = (mb_h + 1) >> 1;
-  const int num_lines_out = Rescale(io->y, io->y_stride, mb_h, &p->scaler_y);
+  WebPRescaler* const scaler = &p->scaler_y;
+  int num_lines_out = 0;
+  if (WebPIsAlphaMode(p->output->colorspace) && io->a != NULL) {
+    // Before rescaling, we premultiply the luma directly into the io->y
+    // internal buffer. This is OK since these samples are not used for
+    // intra-prediction (the top samples are saved in cache_y_/u_/v_).
+    // But we need to cast the const away, though.
+    WebPMultRows((uint8_t*)io->y, io->y_stride,
+                 io->a, io->width, io->mb_w, mb_h, 0);
+  }
+  num_lines_out = Rescale(io->y, io->y_stride, mb_h, scaler);
   Rescale(io->u, io->uv_stride, uv_mb_h, &p->scaler_u);
   Rescale(io->v, io->uv_stride, uv_mb_h, &p->scaler_v);
   return num_lines_out;
 }
 
-static int EmitRescaledAlphaYUV(const VP8Io* const io, WebPDecParams* const p) {
+static int EmitRescaledAlphaYUV(const VP8Io* const io, WebPDecParams* const p,
+                                int expected_num_lines_out) {
   if (io->a != NULL) {
-    Rescale(io->a, io->width, io->mb_h, &p->scaler_a);
+    const WebPYUVABuffer* const buf = &p->output->u.YUVA;
+    uint8_t* dst_y = buf->y + p->last_y * buf->y_stride;
+    const uint8_t* src_a = buf->a + p->last_y * buf->a_stride;
+    const int num_lines_out = Rescale(io->a, io->width, io->mb_h, &p->scaler_a);
+    (void)expected_num_lines_out;
+    assert(expected_num_lines_out == num_lines_out);
+    if (num_lines_out > 0) {   // unmultiply the Y
+      WebPMultRows(dst_y, buf->y_stride, src_a, buf->a_stride,
+                   p->scaler_a.dst_width, num_lines_out, 1);
+    }
   }
   return 0;
 }
@@ -316,39 +296,34 @@ static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
   const size_t work_size = 2 * out_width;   // scratch memory for luma rescaler
   const size_t uv_work_size = 2 * uv_out_width;  // and for each u/v ones
   size_t tmp_size;
-  int32_t* work;
+  rescaler_t* work;
 
-  tmp_size = work_size + 2 * uv_work_size;
+  tmp_size = (work_size + 2 * uv_work_size) * sizeof(*work);
   if (has_alpha) {
-    tmp_size += work_size;
+    tmp_size += work_size * sizeof(*work);
   }
-  p->memory = calloc(1, tmp_size * sizeof(*work));
+  p->memory = WebPSafeMalloc(1ULL, tmp_size);
   if (p->memory == NULL) {
     return 0;   // memory error
   }
-  work = (int32_t*)p->memory;
+  work = (rescaler_t*)p->memory;
   WebPRescalerInit(&p->scaler_y, io->mb_w, io->mb_h,
                    buf->y, out_width, out_height, buf->y_stride, 1,
-                   io->mb_w, out_width, io->mb_h, out_height,
                    work);
   WebPRescalerInit(&p->scaler_u, uv_in_width, uv_in_height,
                    buf->u, uv_out_width, uv_out_height, buf->u_stride, 1,
-                   uv_in_width, uv_out_width,
-                   uv_in_height, uv_out_height,
                    work + work_size);
   WebPRescalerInit(&p->scaler_v, uv_in_width, uv_in_height,
                    buf->v, uv_out_width, uv_out_height, buf->v_stride, 1,
-                   uv_in_width, uv_out_width,
-                   uv_in_height, uv_out_height,
                    work + work_size + uv_work_size);
   p->emit = EmitRescaledYUV;
 
   if (has_alpha) {
     WebPRescalerInit(&p->scaler_a, io->mb_w, io->mb_h,
                      buf->a, out_width, out_height, buf->a_stride, 1,
-                     io->mb_w, out_width, io->mb_h, out_height,
                      work + work_size + 2 * uv_work_size);
     p->emit_alpha = EmitRescaledAlphaYUV;
+    WebPInitAlphaProcessing();
   }
   return 1;
 }
@@ -360,13 +335,13 @@ static int ExportRGB(WebPDecParams* const p, int y_pos) {
   const WebPYUV444Converter convert =
       WebPYUV444Converters[p->output->colorspace];
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* dst = buf->rgba + (p->last_y + y_pos) * buf->stride;
+  uint8_t* dst = buf->rgba + y_pos * buf->stride;
   int num_lines_out = 0;
   // For RGB rescaling, because of the YUV420, current scan position
   // U/V can be +1/-1 line from the Y one.  Hence the double test.
   while (WebPRescalerHasPendingOutput(&p->scaler_y) &&
          WebPRescalerHasPendingOutput(&p->scaler_u)) {
-    assert(p->last_y + y_pos + num_lines_out < p->output->height);
+    assert(y_pos + num_lines_out < p->output->height);
     assert(p->scaler_u.y_accum == p->scaler_v.y_accum);
     WebPRescalerExportRow(&p->scaler_y);
     WebPRescalerExportRow(&p->scaler_u);
@@ -388,65 +363,69 @@ static int EmitRescaledRGB(const VP8Io* const io, WebPDecParams* const p) {
     const int y_lines_in =
         WebPRescalerImport(&p->scaler_y, mb_h - j,
                            io->y + j * io->y_stride, io->y_stride);
-    const int u_lines_in =
-        WebPRescalerImport(&p->scaler_u, uv_mb_h - uv_j,
-                           io->u + uv_j * io->uv_stride, io->uv_stride);
-    const int v_lines_in =
-        WebPRescalerImport(&p->scaler_v, uv_mb_h - uv_j,
-                           io->v + uv_j * io->uv_stride, io->uv_stride);
-    (void)v_lines_in;   // remove a gcc warning
-    assert(u_lines_in == v_lines_in);
     j += y_lines_in;
-    uv_j += u_lines_in;
-    num_lines_out += ExportRGB(p, num_lines_out);
+    if (WebPRescaleNeededLines(&p->scaler_u, uv_mb_h - uv_j)) {
+      const int u_lines_in =
+          WebPRescalerImport(&p->scaler_u, uv_mb_h - uv_j,
+                             io->u + uv_j * io->uv_stride, io->uv_stride);
+      const int v_lines_in =
+          WebPRescalerImport(&p->scaler_v, uv_mb_h - uv_j,
+                             io->v + uv_j * io->uv_stride, io->uv_stride);
+      (void)v_lines_in;   // remove a gcc warning
+      assert(u_lines_in == v_lines_in);
+      uv_j += u_lines_in;
+    }
+    num_lines_out += ExportRGB(p, p->last_y + num_lines_out);
   }
   return num_lines_out;
 }
 
-static int ExportAlpha(WebPDecParams* const p, int y_pos) {
+static int ExportAlpha(WebPDecParams* const p, int y_pos, int max_lines_out) {
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* const base_rgba = buf->rgba + (p->last_y + y_pos) * buf->stride;
+  uint8_t* const base_rgba = buf->rgba + y_pos * buf->stride;
   const WEBP_CSP_MODE colorspace = p->output->colorspace;
   const int alpha_first =
       (colorspace == MODE_ARGB || colorspace == MODE_Argb);
   uint8_t* dst = base_rgba + (alpha_first ? 0 : 3);
   int num_lines_out = 0;
   const int is_premult_alpha = WebPIsPremultipliedMode(colorspace);
-  uint32_t alpha_mask = 0xff;
+  uint32_t non_opaque = 0;
   const int width = p->scaler_a.dst_width;
 
-  while (WebPRescalerHasPendingOutput(&p->scaler_a)) {
-    int i;
-    assert(p->last_y + y_pos + num_lines_out < p->output->height);
+  while (WebPRescalerHasPendingOutput(&p->scaler_a) &&
+         num_lines_out < max_lines_out) {
+    assert(y_pos + num_lines_out < p->output->height);
     WebPRescalerExportRow(&p->scaler_a);
-    for (i = 0; i < width; ++i) {
-      const uint32_t alpha_value = p->scaler_a.dst[i];
-      dst[4 * i] = alpha_value;
-      alpha_mask &= alpha_value;
-    }
+    non_opaque |= WebPDispatchAlpha(p->scaler_a.dst, 0, width, 1, dst, 0);
     dst += buf->stride;
     ++num_lines_out;
   }
-  if (is_premult_alpha && alpha_mask != 0xff) {
+  if (is_premult_alpha && non_opaque) {
     WebPApplyAlphaMultiply(base_rgba, alpha_first,
                            width, num_lines_out, buf->stride);
   }
   return num_lines_out;
 }
 
-static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
+static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos,
+                               int max_lines_out) {
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* const base_rgba = buf->rgba + (p->last_y + y_pos) * buf->stride;
+  uint8_t* const base_rgba = buf->rgba + y_pos * buf->stride;
+#ifdef WEBP_SWAP_16BIT_CSP
+  uint8_t* alpha_dst = base_rgba;
+#else
   uint8_t* alpha_dst = base_rgba + 1;
+#endif
   int num_lines_out = 0;
   const WEBP_CSP_MODE colorspace = p->output->colorspace;
   const int width = p->scaler_a.dst_width;
   const int is_premult_alpha = WebPIsPremultipliedMode(colorspace);
   uint32_t alpha_mask = 0x0f;
 
-  while (WebPRescalerHasPendingOutput(&p->scaler_a)) {
+  while (WebPRescalerHasPendingOutput(&p->scaler_a) &&
+         num_lines_out < max_lines_out) {
     int i;
-    assert(p->last_y + y_pos + num_lines_out < p->output->height);
+    assert(y_pos + num_lines_out < p->output->height);
     WebPRescalerExportRow(&p->scaler_a);
     for (i = 0; i < width; ++i) {
       // Fill in the alpha value (converted to 4 bits).
@@ -463,15 +442,17 @@ static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
   return num_lines_out;
 }
 
-static int EmitRescaledAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
+static int EmitRescaledAlphaRGB(const VP8Io* const io, WebPDecParams* const p,
+                                int expected_num_out_lines) {
   if (io->a != NULL) {
     WebPRescaler* const scaler = &p->scaler_a;
-    int j = 0;
-    int pos = 0;
-    while (j < io->mb_h) {
-      j += WebPRescalerImport(scaler, io->mb_h - j,
-                              io->a + j * io->width, io->width);
-      pos += p->emit_alpha_row(p, pos);
+    int lines_left = expected_num_out_lines;
+    const int y_end = p->last_y + lines_left;
+    while (lines_left > 0) {
+      const int row_offset = scaler->src_y - io->mb_y;
+      WebPRescalerImport(scaler, io->mb_h + io->mb_y - scaler->src_y,
+                         io->a + row_offset * io->width, io->width);
+      lines_left -= p->emit_alpha_row(p, y_end - lines_left, lines_left);
     }
   }
   return 0;
@@ -484,9 +465,9 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
   const int uv_in_width  = (io->mb_w + 1) >> 1;
   const int uv_in_height = (io->mb_h + 1) >> 1;
   const size_t work_size = 2 * out_width;   // scratch memory for one rescaler
-  int32_t* work;  // rescalers work area
+  rescaler_t* work;  // rescalers work area
   uint8_t* tmp;   // tmp storage for scaled YUV444 samples before RGB conversion
-  size_t tmp_size1, tmp_size2;
+  size_t tmp_size1, tmp_size2, total_size;
 
   tmp_size1 = 3 * work_size;
   tmp_size2 = 3 * out_width;
@@ -494,30 +475,28 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
     tmp_size1 += work_size;
     tmp_size2 += out_width;
   }
-  p->memory = calloc(1, tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp));
+  total_size = tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp);
+  p->memory = WebPSafeMalloc(1ULL, total_size);
   if (p->memory == NULL) {
     return 0;   // memory error
   }
-  work = (int32_t*)p->memory;
+  work = (rescaler_t*)p->memory;
   tmp = (uint8_t*)(work + tmp_size1);
   WebPRescalerInit(&p->scaler_y, io->mb_w, io->mb_h,
                    tmp + 0 * out_width, out_width, out_height, 0, 1,
-                   io->mb_w, out_width, io->mb_h, out_height,
                    work + 0 * work_size);
   WebPRescalerInit(&p->scaler_u, uv_in_width, uv_in_height,
                    tmp + 1 * out_width, out_width, out_height, 0, 1,
-                   io->mb_w, 2 * out_width, io->mb_h, 2 * out_height,
                    work + 1 * work_size);
   WebPRescalerInit(&p->scaler_v, uv_in_width, uv_in_height,
                    tmp + 2 * out_width, out_width, out_height, 0, 1,
-                   io->mb_w, 2 * out_width, io->mb_h, 2 * out_height,
                    work + 2 * work_size);
   p->emit = EmitRescaledRGB;
+  WebPInitYUV444Converters();
 
   if (has_alpha) {
     WebPRescalerInit(&p->scaler_a, io->mb_w, io->mb_h,
                      tmp + 3 * out_width, out_width, out_height, 0, 1,
-                     io->mb_w, out_width, io->mb_h, out_height,
                      work + 3 * work_size);
     p->emit_alpha = EmitRescaledAlphaRGB;
     if (p->output->colorspace == MODE_RGBA_4444 ||
@@ -526,6 +505,7 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
     } else {
       p->emit_alpha_row = ExportAlpha;
     }
+    WebPInitAlphaProcessing();
   }
   return 1;
 }
@@ -546,7 +526,9 @@ static int CustomSetup(VP8Io* io) {
   if (!WebPIoInitFromOptions(p->options, io, is_alpha ? MODE_YUV : MODE_YUVA)) {
     return 0;
   }
-
+  if (is_alpha && WebPIsPremultipliedMode(colorspace)) {
+    WebPInitUpsamplers();
+  }
   if (io->use_scaling) {
     const int ok = is_rgb ? InitRGBRescaler(io, p) : InitYUVRescaler(io, p);
     if (!ok) {
@@ -554,11 +536,12 @@ static int CustomSetup(VP8Io* io) {
     }
   } else {
     if (is_rgb) {
+      WebPInitSamplers();
       p->emit = EmitSampledRGB;   // default
-#ifdef FANCY_UPSAMPLING
       if (io->fancy_upsampling) {
+#ifdef FANCY_UPSAMPLING
         const int uv_width = (io->mb_w + 1) >> 1;
-        p->memory = malloc(io->mb_w + 2 * uv_width);
+        p->memory = WebPSafeMalloc(1ULL, (size_t)(io->mb_w + 2 * uv_width));
         if (p->memory == NULL) {
           return 0;   // memory error.
         }
@@ -567,18 +550,20 @@ static int CustomSetup(VP8Io* io) {
         p->tmp_v = p->tmp_u + uv_width;
         p->emit = EmitFancyRGB;
         WebPInitUpsamplers();
-      }
 #endif
+      }
     } else {
       p->emit = EmitYUV;
     }
     if (is_alpha) {  // need transparency output
-      if (WebPIsPremultipliedMode(colorspace)) WebPInitPremultiply();
       p->emit_alpha =
           (colorspace == MODE_RGBA_4444 || colorspace == MODE_rgbA_4444) ?
               EmitAlphaRGBA4444
           : is_rgb ? EmitAlphaRGB
           : EmitAlphaYUV;
+      if (is_rgb) {
+        WebPInitAlphaProcessing();
+      }
     }
   }
 
@@ -601,8 +586,8 @@ static int CustomPut(const VP8Io* io) {
     return 0;
   }
   num_lines_out = p->emit(io, p);
-  if (p->emit_alpha) {
-    p->emit_alpha(io, p);
+  if (p->emit_alpha != NULL) {
+    p->emit_alpha(io, p, num_lines_out);
   }
   p->last_y += num_lines_out;
   return 1;
@@ -612,7 +597,7 @@ static int CustomPut(const VP8Io* io) {
 
 static void CustomTeardown(const VP8Io* io) {
   WebPDecParams* const p = (WebPDecParams*)io->opaque;
-  free(p->memory);
+  WebPSafeFree(p->memory);
   p->memory = NULL;
 }
 
@@ -627,7 +612,3 @@ void WebPInitCustomIo(WebPDecParams* const params, VP8Io* const io) {
 }
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/layer.c b/drivers/webp/dec/layer.c
deleted file mode 100644
index a3a5bdcfe8..0000000000
--- a/drivers/webp/dec/layer.c
+++ /dev/null
@@ -1,35 +0,0 @@
-// Copyright 2011 Google Inc. All Rights Reserved.
-//
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
-// -----------------------------------------------------------------------------
-//
-// Enhancement layer (for YUV444/422)
-//
-// Author: Skal (pascal.massimino@gmail.com)
-
-#include <assert.h>
-#include <stdlib.h>
-
-#include "./vp8i.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-//------------------------------------------------------------------------------
-
-int VP8DecodeLayer(VP8Decoder* const dec) {
-  assert(dec);
-  assert(dec->layer_data_size_ > 0);
-  (void)dec;
-
-  // TODO: handle enhancement layer here.
-
-  return 1;
-}
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/quant.c b/drivers/webp/dec/quant.c
index d54097af0d..5b648f942c 100644
--- a/drivers/webp/dec/quant.c
+++ b/drivers/webp/dec/quant.c
@@ -1,8 +1,10 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Quantizer initialization
@@ -11,10 +13,6 @@
 
 #include "./vp8i.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 static WEBP_INLINE int clip(int v, int M) {
   return v < 0 ? 0 : v > M ? M : v;
 }
@@ -102,12 +100,11 @@ void VP8ParseQuant(VP8Decoder* const dec) {
 
       m->uv_mat_[0] = kDcTable[clip(q + dquv_dc, 117)];
       m->uv_mat_[1] = kAcTable[clip(q + dquv_ac, 127)];
+
+      m->uv_quant_ = q + dquv_ac;   // for dithering strength evaluation
     }
   }
 }
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/tree.c b/drivers/webp/dec/tree.c
index 82484e4c55..c2007ea733 100644
--- a/drivers/webp/dec/tree.c
+++ b/drivers/webp/dec/tree.c
@@ -1,22 +1,21 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 trees and probas
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
-#include "vp8i.h"
+#include "./vp8i.h"
+#include "../utils/bit_reader_inl.h"
 
 #define USE_GENERIC_TREE
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 #ifdef USE_GENERIC_TREE
 static const int8_t kYModesIntra4[18] = {
   -B_DC_PRED, 1,
@@ -31,61 +30,12 @@ static const int8_t kYModesIntra4[18] = {
 };
 #endif
 
-#ifndef ONLY_KEYFRAME_CODE
-
-// inter prediction modes
-enum {
-  LEFT4 = 0, ABOVE4 = 1, ZERO4 = 2, NEW4 = 3,
-  NEARESTMV, NEARMV, ZEROMV, NEWMV, SPLITMV };
-
-static const int8_t kYModesInter[8] = {
-  -DC_PRED, 1,
-    2, 3,
-      -V_PRED, -H_PRED,
-      -TM_PRED, -B_PRED
-};
-
-static const int8_t kMBSplit[6] = {
-  -3, 1,
-    -2, 2,
-      -0, -1
-};
-
-static const int8_t kMVRef[8] = {
-  -ZEROMV, 1,
-    -NEARESTMV, 2,
-      -NEARMV, 3,
-        -NEWMV, -SPLITMV
-};
-
-static const int8_t kMVRef4[6] = {
-  -LEFT4, 1,
-    -ABOVE4, 2,
-      -ZERO4, -NEW4
-};
-#endif
-
 //------------------------------------------------------------------------------
 // Default probabilities
 
-// Inter
-#ifndef ONLY_KEYFRAME_CODE
-static const uint8_t kYModeProbaInter0[4] = { 112, 86, 140, 37 };
-static const uint8_t kUVModeProbaInter0[3] = { 162, 101, 204 };
-static const uint8_t kMVProba0[2][NUM_MV_PROBAS] = {
-  { 162, 128, 225, 146, 172, 147, 214,  39,
-    156, 128, 129, 132,  75, 145, 178, 206,
-    239, 254, 254 },
-  { 164, 128, 204, 170, 119, 235, 140, 230,
-    228, 128, 130, 130,  74, 148, 180, 203,
-    236, 254, 254 }
-};
-#endif
-
 // Paragraph 13.5
 static const uint8_t
   CoeffsProba0[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS] = {
-  // genereated using vp8_default_coef_probs() in entropy.c:129
   { { { 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 }
@@ -326,28 +276,38 @@ static const uint8_t kBModesProba[NUM_BMODES][NUM_BMODES][NUM_BMODES - 1] = {
 
 void VP8ResetProba(VP8Proba* const proba) {
   memset(proba->segments_, 255u, sizeof(proba->segments_));
-  memcpy(proba->coeffs_, CoeffsProba0, sizeof(CoeffsProba0));
-#ifndef ONLY_KEYFRAME_CODE
-  memcpy(proba->mv_, kMVProba0, sizeof(kMVProba0));
-  memcpy(proba->ymode_, kYModeProbaInter0, sizeof(kYModeProbaInter0));
-  memcpy(proba->uvmode_, kUVModeProbaInter0, sizeof(kUVModeProbaInter0));
-#endif
+  // proba->bands_[][] is initialized later
 }
 
-void VP8ParseIntraMode(VP8BitReader* const br,  VP8Decoder* const dec) {
-  uint8_t* const top = dec->intra_t_ + 4 * dec->mb_x_;
+static void ParseIntraMode(VP8BitReader* const br,
+                           VP8Decoder* const dec, int mb_x) {
+  uint8_t* const top = dec->intra_t_ + 4 * mb_x;
   uint8_t* const left = dec->intra_l_;
-  // Hardcoded 16x16 intra-mode decision tree.
-  dec->is_i4x4_ = !VP8GetBit(br, 145);   // decide for B_PRED first
-  if (!dec->is_i4x4_) {
+  VP8MBData* const block = dec->mb_data_ + mb_x;
+
+  // Note: we don't save segment map (yet), as we don't expect
+  // to decode more than 1 keyframe.
+  if (dec->segment_hdr_.update_map_) {
+    // Hardcoded tree parsing
+    block->segment_ = !VP8GetBit(br, dec->proba_.segments_[0])
+                    ? VP8GetBit(br, dec->proba_.segments_[1])
+                    : 2 + VP8GetBit(br, dec->proba_.segments_[2]);
+  } else {
+    block->segment_ = 0;  // default for intra
+  }
+  if (dec->use_skip_proba_) block->skip_ = VP8GetBit(br, dec->skip_p_);
+
+  block->is_i4x4_ = !VP8GetBit(br, 145);   // decide for B_PRED first
+  if (!block->is_i4x4_) {
+    // Hardcoded 16x16 intra-mode decision tree.
     const int ymode =
         VP8GetBit(br, 156) ? (VP8GetBit(br, 128) ? TM_PRED : H_PRED)
                            : (VP8GetBit(br, 163) ? V_PRED : DC_PRED);
-    dec->imodes_[0] = ymode;
-    memset(top, ymode, 4 * sizeof(top[0]));
-    memset(left, ymode, 4 * sizeof(left[0]));
+    block->imodes_[0] = ymode;
+    memset(top, ymode, 4 * sizeof(*top));
+    memset(left, ymode, 4 * sizeof(*left));
   } else {
-    uint8_t* modes = dec->imodes_;
+    uint8_t* modes = block->imodes_;
     int y;
     for (y = 0; y < 4; ++y) {
       int ymode = left[y];
@@ -356,10 +316,10 @@ void VP8ParseIntraMode(VP8BitReader* const br,  VP8Decoder* const dec) {
         const uint8_t* const prob = kBModesProba[top[x]][ymode];
 #ifdef USE_GENERIC_TREE
         // Generic tree-parsing
-        int i = 0;
-        do {
+        int i = kYModesIntra4[VP8GetBit(br, prob[0])];
+        while (i > 0) {
           i = kYModesIntra4[2 * i + VP8GetBit(br, prob[i])];
-        } while (i > 0);
+        }
         ymode = -i;
 #else
         // Hardcoded tree parsing
@@ -374,15 +334,24 @@ void VP8ParseIntraMode(VP8BitReader* const br,  VP8Decoder* const dec) {
                             (!VP8GetBit(br, prob[8]) ? B_HD_PRED : B_HU_PRED)));
 #endif    // USE_GENERIC_TREE
         top[x] = ymode;
-        *modes++ = ymode;
       }
+      memcpy(modes, top, 4 * sizeof(*top));
+      modes += 4;
       left[y] = ymode;
     }
   }
   // Hardcoded UVMode decision tree
-  dec->uvmode_ = !VP8GetBit(br, 142) ? DC_PRED
-               : !VP8GetBit(br, 114) ? V_PRED
-               : VP8GetBit(br, 183) ? TM_PRED : H_PRED;
+  block->uvmode_ = !VP8GetBit(br, 142) ? DC_PRED
+                 : !VP8GetBit(br, 114) ? V_PRED
+                 : VP8GetBit(br, 183) ? TM_PRED : H_PRED;
+}
+
+int VP8ParseIntraModeRow(VP8BitReader* const br, VP8Decoder* const dec) {
+  int mb_x;
+  for (mb_x = 0; mb_x < dec->mb_w_; ++mb_x) {
+    ParseIntraMode(br, dec, mb_x);
+  }
+  return !dec->br_.eof_;
 }
 
 //------------------------------------------------------------------------------
@@ -524,18 +493,13 @@ static const uint8_t
   }
 };
 
-#ifndef ONLY_KEYFRAME_CODE
-static const uint8_t MVUpdateProba[2][NUM_MV_PROBAS] = {
-  { 237, 246, 253, 253, 254, 254, 254, 254,
-    254, 254, 254, 254, 254, 254, 250, 250,
-    252, 254, 254 },
-  { 231, 243, 245, 253, 254, 254, 254, 254,
-    254, 254, 254, 254, 254, 254, 251, 251,
-    254, 254, 254 }
+// Paragraph 9.9
+
+static const int kBands[16 + 1] = {
+  0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
+  0  // extra entry as sentinel
 };
-#endif
 
-// Paragraph 9.9
 void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec) {
   VP8Proba* const proba = &dec->proba_;
   int t, b, c, p;
@@ -543,47 +507,19 @@ void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec) {
     for (b = 0; b < NUM_BANDS; ++b) {
       for (c = 0; c < NUM_CTX; ++c) {
         for (p = 0; p < NUM_PROBAS; ++p) {
-          if (VP8GetBit(br, CoeffsUpdateProba[t][b][c][p])) {
-            proba->coeffs_[t][b][c][p] = VP8GetValue(br, 8);
-          }
+          const int v = VP8GetBit(br, CoeffsUpdateProba[t][b][c][p]) ?
+                        VP8GetValue(br, 8) : CoeffsProba0[t][b][c][p];
+          proba->bands_[t][b].probas_[c][p] = v;
         }
       }
     }
+    for (b = 0; b < 16 + 1; ++b) {
+      proba->bands_ptr_[t][b] = &proba->bands_[t][kBands[b]];
+    }
   }
   dec->use_skip_proba_ = VP8Get(br);
   if (dec->use_skip_proba_) {
     dec->skip_p_ = VP8GetValue(br, 8);
   }
-#ifndef ONLY_KEYFRAME_CODE
-  if (!dec->frm_hdr_.key_frame_) {
-    int i;
-    dec->intra_p_ = VP8GetValue(br, 8);
-    dec->last_p_ = VP8GetValue(br, 8);
-    dec->golden_p_ = VP8GetValue(br, 8);
-    if (VP8Get(br)) {   // update y-mode
-      for (i = 0; i < 4; ++i) {
-        proba->ymode_[i] = VP8GetValue(br, 8);
-      }
-    }
-    if (VP8Get(br)) {   // update uv-mode
-      for (i = 0; i < 3; ++i) {
-        proba->uvmode_[i] = VP8GetValue(br, 8);
-      }
-    }
-    // update MV
-    for (i = 0; i < 2; ++i) {
-      int k;
-      for (k = 0; k < NUM_MV_PROBAS; ++k) {
-        if (VP8GetBit(br, MVUpdateProba[i][k])) {
-          const int v = VP8GetValue(br, 7);
-          proba->mv_[i][k] = v ? v << 1 : 1;
-        }
-      }
-    }
-  }
-#endif
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/vp8.c b/drivers/webp/dec/vp8.c
index b0ccfa2a06..d89eb1c59e 100644
--- a/drivers/webp/dec/vp8.c
+++ b/drivers/webp/dec/vp8.c
@@ -1,8 +1,10 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 decoder
@@ -11,14 +13,12 @@
 
 #include <stdlib.h>
 
+#include "./alphai.h"
 #include "./vp8i.h"
 #include "./vp8li.h"
 #include "./webpi.h"
-#include "../utils/bit_reader.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "../utils/bit_reader_inl.h"
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 
@@ -45,10 +45,10 @@ int VP8InitIoInternal(VP8Io* const io, int version) {
 }
 
 VP8Decoder* VP8New(void) {
-  VP8Decoder* const dec = (VP8Decoder*)calloc(1, sizeof(*dec));
+  VP8Decoder* const dec = (VP8Decoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
   if (dec != NULL) {
     SetOk(dec);
-    WebPWorkerInit(&dec->worker_);
+    WebPGetWorkerInterface()->Init(&dec->worker_);
     dec->ready_ = 0;
     dec->num_parts_ = 1;
   }
@@ -69,16 +69,13 @@ const char* VP8StatusMessage(VP8Decoder* const dec) {
 void VP8Delete(VP8Decoder* const dec) {
   if (dec != NULL) {
     VP8Clear(dec);
-    free(dec);
+    WebPSafeFree(dec);
   }
 }
 
 int VP8SetError(VP8Decoder* const dec,
                 VP8StatusCode error, const char* const msg) {
-  // TODO This check would be unnecessary if alpha decompression was separated
-  // from VP8ProcessRow/FinishRow. This avoids setting 'dec->status_' to
-  // something other than VP8_STATUS_BITSTREAM_ERROR on alpha decompression
-  // failure.
+  // The oldest error reported takes precedence over the new one.
   if (dec->status_ == VP8_STATUS_OK) {
     dec->status_ = error;
     dec->error_msg_ = msg;
@@ -121,6 +118,9 @@ int VP8GetInfo(const uint8_t* data, size_t data_size, size_t chunk_size,
     if (((bits >> 5)) >= chunk_size) {  // partition_length
       return 0;         // inconsistent size information.
     }
+    if (w == 0 || h == 0) {
+      return 0;         // We don't support both width and height to be zero.
+    }
 
     if (width) {
       *width = w;
@@ -190,25 +190,27 @@ static VP8StatusCode ParsePartitions(VP8Decoder* const dec,
   const uint8_t* sz = buf;
   const uint8_t* buf_end = buf + size;
   const uint8_t* part_start;
-  int last_part;
-  int p;
+  size_t size_left = size;
+  size_t last_part;
+  size_t p;
 
   dec->num_parts_ = 1 << VP8GetValue(br, 2);
   last_part = dec->num_parts_ - 1;
-  part_start = buf + last_part * 3;
-  if (buf_end < part_start) {
+  if (size < 3 * last_part) {
     // we can't even read the sizes with sz[]! That's a failure.
     return VP8_STATUS_NOT_ENOUGH_DATA;
   }
+  part_start = buf + last_part * 3;
+  size_left -= last_part * 3;
   for (p = 0; p < last_part; ++p) {
-    const uint32_t psize = sz[0] | (sz[1] << 8) | (sz[2] << 16);
-    const uint8_t* part_end = part_start + psize;
-    if (part_end > buf_end) part_end = buf_end;
-    VP8InitBitReader(dec->parts_ + p, part_start, part_end);
-    part_start = part_end;
+    size_t psize = sz[0] | (sz[1] << 8) | (sz[2] << 16);
+    if (psize > size_left) psize = size_left;
+    VP8InitBitReader(dec->parts_ + p, part_start, psize);
+    part_start += psize;
+    size_left -= psize;
     sz += 3;
   }
-  VP8InitBitReader(dec->parts_ + last_part, part_start, buf_end);
+  VP8InitBitReader(dec->parts_ + last_part, part_start, size_left);
   return (part_start < buf_end) ? VP8_STATUS_OK :
            VP8_STATUS_SUSPENDED;   // Init is ok, but there's not enough data
 }
@@ -236,20 +238,6 @@ static int ParseFilterHeader(VP8BitReader* br, VP8Decoder* const dec) {
     }
   }
   dec->filter_type_ = (hdr->level_ == 0) ? 0 : hdr->simple_ ? 1 : 2;
-  if (dec->filter_type_ > 0) {    // precompute filter levels per segment
-    if (dec->segment_hdr_.use_segment_) {
-      int s;
-      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
-        int strength = dec->segment_hdr_.filter_strength_[s];
-        if (!dec->segment_hdr_.absolute_delta_) {
-          strength += hdr->level_;
-        }
-        dec->filter_levels_[s] = strength;
-      }
-    } else {
-      dec->filter_levels_[0] = hdr->level_;
-    }
-  }
   return !br->eof_;
 }
 
@@ -261,7 +249,6 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
   VP8PictureHeader* pic_hdr;
   VP8BitReader* br;
   VP8StatusCode status;
-  WebPHeaderStructure headers;
 
   if (dec == NULL) {
     return 0;
@@ -271,33 +258,8 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
     return VP8SetError(dec, VP8_STATUS_INVALID_PARAM,
                        "null VP8Io passed to VP8GetHeaders()");
   }
-
-  // Process Pre-VP8 chunks.
-  headers.data = io->data;
-  headers.data_size = io->data_size;
-  status = WebPParseHeaders(&headers);
-  if (status != VP8_STATUS_OK) {
-    return VP8SetError(dec, status, "Incorrect/incomplete header.");
-  }
-  if (headers.is_lossless) {
-    return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
-                       "Unexpected lossless format encountered.");
-  }
-
-  if (dec->alpha_data_ == NULL) {
-    assert(dec->alpha_data_size_ == 0);
-    // We have NOT set alpha data yet. Set it now.
-    // (This is to ensure that dec->alpha_data_ is NOT reset to NULL if
-    // WebPParseHeaders() is called more than once, as in incremental decoding
-    // case.)
-    dec->alpha_data_ = headers.alpha_data;
-    dec->alpha_data_size_ = headers.alpha_data_size;
-  }
-
-  // Process the VP8 frame header.
-  buf = headers.data + headers.offset;
-  buf_size = headers.data_size - headers.offset;
-  assert(headers.data_size >= headers.offset);  // WebPParseHeaders' guarantee
+  buf = io->data;
+  buf_size = io->data_size;
   if (buf_size < 4) {
     return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
                        "Truncated header.");
@@ -355,7 +317,6 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
 
     VP8ResetProba(&dec->proba_);
     ResetSegmentHeader(&dec->segment_hdr_);
-    dec->segment_ = 0;    // default for intra
   }
 
   // Check if we have all the partition #0 available, and initialize dec->br_
@@ -366,7 +327,7 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
   }
 
   br = &dec->br_;
-  VP8InitBitReader(br, buf, buf + frm_hdr->partition_length_);
+  VP8InitBitReader(br, buf, frm_hdr->partition_length_);
   buf += frm_hdr->partition_length_;
   buf_size -= frm_hdr->partition_length_;
 
@@ -393,63 +354,14 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
 
   // Frame buffer marking
   if (!frm_hdr->key_frame_) {
-    // Paragraph 9.7
-#ifndef ONLY_KEYFRAME_CODE
-    dec->buffer_flags_ = VP8Get(br) << 0;   // update golden
-    dec->buffer_flags_ |= VP8Get(br) << 1;  // update alt ref
-    if (!(dec->buffer_flags_ & 1)) {
-      dec->buffer_flags_ |= VP8GetValue(br, 2) << 2;
-    }
-    if (!(dec->buffer_flags_ & 2)) {
-      dec->buffer_flags_ |= VP8GetValue(br, 2) << 4;
-    }
-    dec->buffer_flags_ |= VP8Get(br) << 6;    // sign bias golden
-    dec->buffer_flags_ |= VP8Get(br) << 7;    // sign bias alt ref
-#else
     return VP8SetError(dec, VP8_STATUS_UNSUPPORTED_FEATURE,
                        "Not a key frame.");
-#endif
-  } else {
-    dec->buffer_flags_ = 0x003 | 0x100;
   }
 
-  // Paragraph 9.8
-#ifndef ONLY_KEYFRAME_CODE
-  dec->update_proba_ = VP8Get(br);
-  if (!dec->update_proba_) {    // save for later restore
-    dec->proba_saved_ = dec->proba_;
-  }
-  dec->buffer_flags_ &= 1 << 8;
-  dec->buffer_flags_ |=
-      (frm_hdr->key_frame_ || VP8Get(br)) << 8;    // refresh last frame
-#else
-  VP8Get(br);   // just ignore the value of update_proba_
-#endif
+  VP8Get(br);   // ignore the value of update_proba_
 
   VP8ParseProba(br, dec);
 
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-  // Extensions
-  if (dec->pic_hdr_.colorspace_) {
-    const size_t kTrailerSize = 8;
-    const uint8_t kTrailerMarker = 0x01;
-    const uint8_t* ext_buf = buf - kTrailerSize;
-    size_t size;
-
-    if (frm_hdr->partition_length_ < kTrailerSize ||
-        ext_buf[kTrailerSize - 1] != kTrailerMarker) {
-      return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
-                         "RIFF: Inconsistent extra information.");
-    }
-
-    // Layer
-    size = (ext_buf[0] << 0) | (ext_buf[1] << 8) | (ext_buf[2] << 16);
-    dec->layer_data_size_ = size;
-    dec->layer_data_ = NULL;  // will be set later
-    dec->layer_colorspace_ = ext_buf[3];
-  }
-#endif
-
   // sanitized state
   dec->ready_ = 1;
   return 1;
@@ -458,11 +370,6 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
 //------------------------------------------------------------------------------
 // Residual decoding (Paragraph 13.2 / 13.3)
 
-static const uint8_t kBands[16 + 1] = {
-  0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
-  0  // extra entry as sentinel
-};
-
 static const uint8_t kCat3[] = { 173, 148, 140, 0 };
 static const uint8_t kCat4[] = { 176, 155, 140, 135, 0 };
 static const uint8_t kCat5[] = { 180, 157, 141, 134, 130, 0 };
@@ -473,253 +380,226 @@ static const uint8_t kZigzag[16] = {
   0, 1, 4, 8,  5, 2, 3, 6,  9, 12, 13, 10,  7, 11, 14, 15
 };
 
-typedef const uint8_t (*ProbaArray)[NUM_CTX][NUM_PROBAS];  // for const-casting
+// See section 13-2: http://tools.ietf.org/html/rfc6386#section-13.2
+static int GetLargeValue(VP8BitReader* const br, const uint8_t* const p) {
+  int v;
+  if (!VP8GetBit(br, p[3])) {
+    if (!VP8GetBit(br, p[4])) {
+      v = 2;
+    } else {
+      v = 3 + VP8GetBit(br, p[5]);
+    }
+  } else {
+    if (!VP8GetBit(br, p[6])) {
+      if (!VP8GetBit(br, p[7])) {
+        v = 5 + VP8GetBit(br, 159);
+      } else {
+        v = 7 + 2 * VP8GetBit(br, 165);
+        v += VP8GetBit(br, 145);
+      }
+    } else {
+      const uint8_t* tab;
+      const int bit1 = VP8GetBit(br, p[8]);
+      const int bit0 = VP8GetBit(br, p[9 + bit1]);
+      const int cat = 2 * bit1 + bit0;
+      v = 0;
+      for (tab = kCat3456[cat]; *tab; ++tab) {
+        v += v + VP8GetBit(br, *tab);
+      }
+      v += 3 + (8 << cat);
+    }
+  }
+  return v;
+}
 
 // Returns the position of the last non-zero coeff plus one
-// (and 0 if there's no coeff at all)
-static int GetCoeffs(VP8BitReader* const br, ProbaArray prob,
+static int GetCoeffs(VP8BitReader* const br, const VP8BandProbas* const prob[],
                      int ctx, const quant_t dq, int n, int16_t* out) {
-  // n is either 0 or 1 here. kBands[n] is not necessary for extracting '*p'.
-  const uint8_t* p = prob[n][ctx];
-  if (!VP8GetBit(br, p[0])) {   // first EOB is more a 'CBP' bit.
-    return 0;
-  }
-  while (1) {
-    ++n;
-    if (!VP8GetBit(br, p[1])) {
-      p = prob[kBands[n]][0];
-    } else {  // non zero coeff
-      int v, j;
+  const uint8_t* p = prob[n]->probas_[ctx];
+  for (; n < 16; ++n) {
+    if (!VP8GetBit(br, p[0])) {
+      return n;  // previous coeff was last non-zero coeff
+    }
+    while (!VP8GetBit(br, p[1])) {       // sequence of zero coeffs
+      p = prob[++n]->probas_[0];
+      if (n == 16) return 16;
+    }
+    {        // non zero coeff
+      const VP8ProbaArray* const p_ctx = &prob[n + 1]->probas_[0];
+      int v;
       if (!VP8GetBit(br, p[2])) {
-        p = prob[kBands[n]][1];
         v = 1;
+        p = p_ctx[1];
       } else {
-        if (!VP8GetBit(br, p[3])) {
-          if (!VP8GetBit(br, p[4])) {
-            v = 2;
-          } else {
-            v = 3 + VP8GetBit(br, p[5]);
-          }
-        } else {
-          if (!VP8GetBit(br, p[6])) {
-            if (!VP8GetBit(br, p[7])) {
-              v = 5 + VP8GetBit(br, 159);
-            } else {
-              v = 7 + 2 * VP8GetBit(br, 165);
-              v += VP8GetBit(br, 145);
-            }
-          } else {
-            const uint8_t* tab;
-            const int bit1 = VP8GetBit(br, p[8]);
-            const int bit0 = VP8GetBit(br, p[9 + bit1]);
-            const int cat = 2 * bit1 + bit0;
-            v = 0;
-            for (tab = kCat3456[cat]; *tab; ++tab) {
-              v += v + VP8GetBit(br, *tab);
-            }
-            v += 3 + (8 << cat);
-          }
-        }
-        p = prob[kBands[n]][2];
+        v = GetLargeValue(br, p);
+        p = p_ctx[2];
       }
-      j = kZigzag[n - 1];
-      out[j] = VP8GetSigned(br, v) * dq[j > 0];
-      if (n == 16 || !VP8GetBit(br, p[0])) {   // EOB
-        return n;
-      }
-    }
-    if (n == 16) {
-      return 16;
+      out[kZigzag[n]] = VP8GetSigned(br, v) * dq[n > 0];
     }
   }
+  return 16;
 }
 
-// Alias-safe way of converting 4bytes to 32bits.
-typedef union {
-  uint8_t  i8[4];
-  uint32_t i32;
-} PackedNz;
-
-// Table to unpack four bits into four bytes
-static const PackedNz kUnpackTab[16] = {
-  {{0, 0, 0, 0}},  {{1, 0, 0, 0}},  {{0, 1, 0, 0}},  {{1, 1, 0, 0}},
-  {{0, 0, 1, 0}},  {{1, 0, 1, 0}},  {{0, 1, 1, 0}},  {{1, 1, 1, 0}},
-  {{0, 0, 0, 1}},  {{1, 0, 0, 1}},  {{0, 1, 0, 1}},  {{1, 1, 0, 1}},
-  {{0, 0, 1, 1}},  {{1, 0, 1, 1}},  {{0, 1, 1, 1}},  {{1, 1, 1, 1}} };
-
-// Macro to pack four LSB of four bytes into four bits.
-#if defined(__PPC__) || defined(_M_PPC) || defined(_ARCH_PPC) || \
-    defined(__BIG_ENDIAN__)
-#define PACK_CST 0x08040201U
-#else
-#define PACK_CST 0x01020408U
-#endif
-#define PACK(X, S) ((((X).i32 * PACK_CST) & 0xff000000) >> (S))
-
-static void ParseResiduals(VP8Decoder* const dec,
-                           VP8MB* const mb, VP8BitReader* const token_br) {
-  int out_t_nz, out_l_nz, first;
-  ProbaArray ac_prob;
-  const VP8QuantMatrix* q = &dec->dqm_[dec->segment_];
-  int16_t* dst = dec->coeffs_;
+static WEBP_INLINE uint32_t NzCodeBits(uint32_t nz_coeffs, int nz, int dc_nz) {
+  nz_coeffs <<= 2;
+  nz_coeffs |= (nz > 3) ? 3 : (nz > 1) ? 2 : dc_nz;
+  return nz_coeffs;
+}
+
+static int ParseResiduals(VP8Decoder* const dec,
+                          VP8MB* const mb, VP8BitReader* const token_br) {
+  const VP8BandProbas* (* const bands)[16 + 1] = dec->proba_.bands_ptr_;
+  const VP8BandProbas* const * ac_proba;
+  VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+  const VP8QuantMatrix* const q = &dec->dqm_[block->segment_];
+  int16_t* dst = block->coeffs_;
   VP8MB* const left_mb = dec->mb_info_ - 1;
-  PackedNz nz_ac, nz_dc;
-  PackedNz tnz, lnz;
-  uint32_t non_zero_ac = 0;
-  uint32_t non_zero_dc = 0;
+  uint8_t tnz, lnz;
+  uint32_t non_zero_y = 0;
+  uint32_t non_zero_uv = 0;
   int x, y, ch;
+  uint32_t out_t_nz, out_l_nz;
+  int first;
 
-  nz_dc.i32 = nz_ac.i32 = 0;
   memset(dst, 0, 384 * sizeof(*dst));
-  if (!dec->is_i4x4_) {    // parse DC
+  if (!block->is_i4x4_) {    // parse DC
     int16_t dc[16] = { 0 };
-    const int ctx = mb->dc_nz_ + left_mb->dc_nz_;
-    mb->dc_nz_ = left_mb->dc_nz_ =
-        (GetCoeffs(token_br, (ProbaArray)dec->proba_.coeffs_[1],
-                   ctx, q->y2_mat_, 0, dc) > 0);
+    const int ctx = mb->nz_dc_ + left_mb->nz_dc_;
+    const int nz = GetCoeffs(token_br, bands[1], ctx, q->y2_mat_, 0, dc);
+    mb->nz_dc_ = left_mb->nz_dc_ = (nz > 0);
+    if (nz > 1) {   // more than just the DC -> perform the full transform
+      VP8TransformWHT(dc, dst);
+    } else {        // only DC is non-zero -> inlined simplified transform
+      int i;
+      const int dc0 = (dc[0] + 3) >> 3;
+      for (i = 0; i < 16 * 16; i += 16) dst[i] = dc0;
+    }
     first = 1;
-    ac_prob = (ProbaArray)dec->proba_.coeffs_[0];
-    VP8TransformWHT(dc, dst);
+    ac_proba = bands[0];
   } else {
     first = 0;
-    ac_prob = (ProbaArray)dec->proba_.coeffs_[3];
+    ac_proba = bands[3];
   }
 
-  tnz = kUnpackTab[mb->nz_ & 0xf];
-  lnz = kUnpackTab[left_mb->nz_ & 0xf];
+  tnz = mb->nz_ & 0x0f;
+  lnz = left_mb->nz_ & 0x0f;
   for (y = 0; y < 4; ++y) {
-    int l = lnz.i8[y];
+    int l = lnz & 1;
+    uint32_t nz_coeffs = 0;
     for (x = 0; x < 4; ++x) {
-      const int ctx = l + tnz.i8[x];
-      const int nz = GetCoeffs(token_br, ac_prob, ctx,
-                               q->y1_mat_, first, dst);
-      tnz.i8[x] = l = (nz > 0);
-      nz_dc.i8[x] = (dst[0] != 0);
-      nz_ac.i8[x] = (nz > 1);
+      const int ctx = l + (tnz & 1);
+      const int nz = GetCoeffs(token_br, ac_proba, ctx, q->y1_mat_, first, dst);
+      l = (nz > first);
+      tnz = (tnz >> 1) | (l << 7);
+      nz_coeffs = NzCodeBits(nz_coeffs, nz, dst[0] != 0);
       dst += 16;
     }
-    lnz.i8[y] = l;
-    non_zero_dc |= PACK(nz_dc, 24 - y * 4);
-    non_zero_ac |= PACK(nz_ac, 24 - y * 4);
+    tnz >>= 4;
+    lnz = (lnz >> 1) | (l << 7);
+    non_zero_y = (non_zero_y << 8) | nz_coeffs;
   }
-  out_t_nz = PACK(tnz, 24);
-  out_l_nz = PACK(lnz, 24);
+  out_t_nz = tnz;
+  out_l_nz = lnz >> 4;
 
-  tnz = kUnpackTab[mb->nz_ >> 4];
-  lnz = kUnpackTab[left_mb->nz_ >> 4];
   for (ch = 0; ch < 4; ch += 2) {
+    uint32_t nz_coeffs = 0;
+    tnz = mb->nz_ >> (4 + ch);
+    lnz = left_mb->nz_ >> (4 + ch);
     for (y = 0; y < 2; ++y) {
-      int l = lnz.i8[ch + y];
+      int l = lnz & 1;
       for (x = 0; x < 2; ++x) {
-        const int ctx = l + tnz.i8[ch + x];
-        const int nz =
-            GetCoeffs(token_br, (ProbaArray)dec->proba_.coeffs_[2],
-                      ctx, q->uv_mat_, 0, dst);
-        tnz.i8[ch + x] = l = (nz > 0);
-        nz_dc.i8[y * 2 + x] = (dst[0] != 0);
-        nz_ac.i8[y * 2 + x] = (nz > 1);
+        const int ctx = l + (tnz & 1);
+        const int nz = GetCoeffs(token_br, bands[2], ctx, q->uv_mat_, 0, dst);
+        l = (nz > 0);
+        tnz = (tnz >> 1) | (l << 3);
+        nz_coeffs = NzCodeBits(nz_coeffs, nz, dst[0] != 0);
         dst += 16;
       }
-      lnz.i8[ch + y] = l;
+      tnz >>= 2;
+      lnz = (lnz >> 1) | (l << 5);
     }
-    non_zero_dc |= PACK(nz_dc, 8 - ch * 2);
-    non_zero_ac |= PACK(nz_ac, 8 - ch * 2);
+    // Note: we don't really need the per-4x4 details for U/V blocks.
+    non_zero_uv |= nz_coeffs << (4 * ch);
+    out_t_nz |= (tnz << 4) << ch;
+    out_l_nz |= (lnz & 0xf0) << ch;
   }
-  out_t_nz |= PACK(tnz, 20);
-  out_l_nz |= PACK(lnz, 20);
   mb->nz_ = out_t_nz;
   left_mb->nz_ = out_l_nz;
 
-  dec->non_zero_ac_ = non_zero_ac;
-  dec->non_zero_ = non_zero_ac | non_zero_dc;
-  mb->skip_ = !dec->non_zero_;
+  block->non_zero_y_ = non_zero_y;
+  block->non_zero_uv_ = non_zero_uv;
+
+  // We look at the mode-code of each block and check if some blocks have less
+  // than three non-zero coeffs (code < 2). This is to avoid dithering flat and
+  // empty blocks.
+  block->dither_ = (non_zero_uv & 0xaaaa) ? 0 : q->dither_;
+
+  return !(non_zero_y | non_zero_uv);  // will be used for further optimization
 }
-#undef PACK
 
 //------------------------------------------------------------------------------
 // Main loop
 
 int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br) {
-  VP8BitReader* const br = &dec->br_;
   VP8MB* const left = dec->mb_info_ - 1;
-  VP8MB* const info = dec->mb_info_ + dec->mb_x_;
-
-  // Note: we don't save segment map (yet), as we don't expect
-  // to decode more than 1 keyframe.
-  if (dec->segment_hdr_.update_map_) {
-    // Hardcoded tree parsing
-    dec->segment_ = !VP8GetBit(br, dec->proba_.segments_[0]) ?
-        VP8GetBit(br, dec->proba_.segments_[1]) :
-        2 + VP8GetBit(br, dec->proba_.segments_[2]);
-  }
-  info->skip_ = dec->use_skip_proba_ ? VP8GetBit(br, dec->skip_p_) : 0;
+  VP8MB* const mb = dec->mb_info_ + dec->mb_x_;
+  VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+  int skip = dec->use_skip_proba_ ? block->skip_ : 0;
 
-  VP8ParseIntraMode(br, dec);
-  if (br->eof_) {
-    return 0;
-  }
-
-  if (!info->skip_) {
-    ParseResiduals(dec, info, token_br);
+  if (!skip) {
+    skip = ParseResiduals(dec, mb, token_br);
   } else {
-    left->nz_ = info->nz_ = 0;
-    if (!dec->is_i4x4_) {
-      left->dc_nz_ = info->dc_nz_ = 0;
+    left->nz_ = mb->nz_ = 0;
+    if (!block->is_i4x4_) {
+      left->nz_dc_ = mb->nz_dc_ = 0;
     }
-    dec->non_zero_ = 0;
-    dec->non_zero_ac_ = 0;
+    block->non_zero_y_ = 0;
+    block->non_zero_uv_ = 0;
+    block->dither_ = 0;
+  }
+
+  if (dec->filter_type_ > 0) {  // store filter info
+    VP8FInfo* const finfo = dec->f_info_ + dec->mb_x_;
+    *finfo = dec->fstrengths_[block->segment_][block->is_i4x4_];
+    finfo->f_inner_ |= !skip;
   }
 
-  return (!token_br->eof_);
+  return !token_br->eof_;
 }
 
 void VP8InitScanline(VP8Decoder* const dec) {
   VP8MB* const left = dec->mb_info_ - 1;
   left->nz_ = 0;
-  left->dc_nz_ = 0;
+  left->nz_dc_ = 0;
   memset(dec->intra_l_, B_DC_PRED, sizeof(dec->intra_l_));
-  dec->filter_row_ =
-    (dec->filter_type_ > 0) &&
-    (dec->mb_y_ >= dec->tl_mb_y_) && (dec->mb_y_ <= dec->br_mb_y_);
+  dec->mb_x_ = 0;
 }
 
 static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
   for (dec->mb_y_ = 0; dec->mb_y_ < dec->br_mb_y_; ++dec->mb_y_) {
+    // Parse bitstream for this row.
     VP8BitReader* const token_br =
         &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
-    VP8InitScanline(dec);
-    for (dec->mb_x_ = 0; dec->mb_x_ < dec->mb_w_;  dec->mb_x_++) {
+    if (!VP8ParseIntraModeRow(&dec->br_, dec)) {
+      return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
+                         "Premature end-of-partition0 encountered.");
+    }
+    for (; dec->mb_x_ < dec->mb_w_; ++dec->mb_x_) {
       if (!VP8DecodeMB(dec, token_br)) {
         return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
                            "Premature end-of-file encountered.");
       }
-      VP8ReconstructBlock(dec);
-
-      // Store data and save block's filtering params
-      VP8StoreBlock(dec);
     }
+    VP8InitScanline(dec);   // Prepare for next scanline
+
+    // Reconstruct, filter and emit the row.
     if (!VP8ProcessRow(dec, io)) {
       return VP8SetError(dec, VP8_STATUS_USER_ABORT, "Output aborted.");
     }
   }
-  if (dec->use_threads_ && !WebPWorkerSync(&dec->worker_)) {
-    return 0;
-  }
-
-  // Finish
-#ifndef ONLY_KEYFRAME_CODE
-  if (!dec->update_proba_) {
-    dec->proba_ = dec->proba_saved_;
+  if (dec->mt_method_ > 0) {
+    if (!WebPGetWorkerInterface()->Sync(&dec->worker_)) return 0;
   }
-#endif
-
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-  if (dec->layer_data_size_ > 0) {
-    if (!VP8DecodeLayer(dec)) {
-      return 0;
-    }
-  }
-#endif
 
   return 1;
 }
@@ -768,12 +648,10 @@ void VP8Clear(VP8Decoder* const dec) {
   if (dec == NULL) {
     return;
   }
-  if (dec->use_threads_) {
-    WebPWorkerEnd(&dec->worker_);
-  }
-  if (dec->mem_) {
-    free(dec->mem_);
-  }
+  WebPGetWorkerInterface()->End(&dec->worker_);
+  ALPHDelete(dec->alph_dec_);
+  dec->alph_dec_ = NULL;
+  WebPSafeFree(dec->mem_);
   dec->mem_ = NULL;
   dec->mem_size_ = 0;
   memset(&dec->br_, 0, sizeof(dec->br_));
@@ -782,6 +660,3 @@ void VP8Clear(VP8Decoder* const dec) {
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/vp8i.h b/drivers/webp/dec/vp8i.h
index 4382edfd8e..b5f2b23009 100644
--- a/drivers/webp/dec/vp8i.h
+++ b/drivers/webp/dec/vp8i.h
@@ -1,8 +1,10 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // VP8 decoder: internal header.
@@ -13,12 +15,14 @@
 #define WEBP_DEC_VP8I_H_
 
 #include <string.h>     // for memcpy()
+#include "./common.h"
 #include "./vp8li.h"
 #include "../utils/bit_reader.h"
+#include "../utils/random.h"
 #include "../utils/thread.h"
 #include "../dsp/dsp.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -27,48 +31,10 @@ extern "C" {
 
 // version numbers
 #define DEC_MAJ_VERSION 0
-#define DEC_MIN_VERSION 2
-#define DEC_REV_VERSION 0
-
-#define ONLY_KEYFRAME_CODE      // to remove any code related to P-Frames
-
-// intra prediction modes
-enum { B_DC_PRED = 0,   // 4x4 modes
-       B_TM_PRED,
-       B_VE_PRED,
-       B_HE_PRED,
-       B_RD_PRED,
-       B_VR_PRED,
-       B_LD_PRED,
-       B_VL_PRED,
-       B_HD_PRED,
-       B_HU_PRED,
-       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,
-       B_PRED = NUM_BMODES,   // refined I4x4 mode
-
-       // special modes
-       B_DC_PRED_NOTOP = 4,
-       B_DC_PRED_NOLEFT = 5,
-       B_DC_PRED_NOTOPLEFT = 6,
-       NUM_B_DC_MODES = 7 };
-
-enum { MB_FEATURE_TREE_PROBS = 3,
-       NUM_MB_SEGMENTS = 4,
-       NUM_REF_LF_DELTAS = 4,
-       NUM_MODE_LF_DELTAS = 4,    // I4x4, ZERO, *, SPLIT
-       MAX_NUM_PARTITIONS = 8,
-       // Probabilities
-       NUM_TYPES = 4,
-       NUM_BANDS = 8,
-       NUM_CTX = 3,
-       NUM_PROBAS = 11,
-       NUM_MV_PROBAS = 19 };
-
-// YUV-cache parameters.
+#define DEC_MIN_VERSION 4
+#define DEC_REV_VERSION 4
+
+// YUV-cache parameters. Cache is 32-bytes wide (= one cacheline).
 // Constraints are: We need to store one 16x16 block of luma samples (y),
 // and two 8x8 chroma blocks (u/v). These are better be 16-bytes aligned,
 // in order to be SIMD-friendly. We also need to store the top, left and
@@ -90,14 +56,15 @@ enum { MB_FEATURE_TREE_PROBS = 3,
 //  'y' = y-samples   'u' = u-samples     'v' = u-samples
 //  '|' = left sample,   '-' = top sample,    '+' = top-left sample
 //  't' = extra top-right sample for 4x4 modes
-// With this layout, BPS (=Bytes Per Scan-line) is one cacheline size.
-#define BPS       32    // this is the common stride used by yuv[]
 #define YUV_SIZE (BPS * 17 + BPS * 9)
 #define Y_SIZE   (BPS * 17)
 #define Y_OFF    (BPS * 1 + 8)
 #define U_OFF    (Y_OFF + BPS * 16 + BPS)
 #define V_OFF    (U_OFF + 16)
 
+// minimal width under which lossy multi-threading is always disabled
+#define MIN_WIDTH_FOR_THREADS 512
+
 //------------------------------------------------------------------------------
 // Headers
 
@@ -126,15 +93,19 @@ typedef struct {
   int8_t filter_strength_[NUM_MB_SEGMENTS];  // filter strength for segments
 } VP8SegmentHeader;
 
+// probas associated to one of the contexts
+typedef uint8_t VP8ProbaArray[NUM_PROBAS];
+
+typedef struct {   // all the probas associated to one band
+  VP8ProbaArray probas_[NUM_CTX];
+} VP8BandProbas;
+
 // Struct collecting all frame-persistent probabilities.
 typedef struct {
   uint8_t segments_[MB_FEATURE_TREE_PROBS];
   // Type: 0:Intra16-AC  1:Intra16-DC   2:Chroma   3:Intra4
-  uint8_t coeffs_[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS];
-#ifndef ONLY_KEYFRAME_CODE
-  uint8_t ymode_[4], uvmode_[3];
-  uint8_t mv_[2][NUM_MV_PROBAS];
-#endif
+  VP8BandProbas bands_[NUM_TYPES][NUM_BANDS];
+  const VP8BandProbas* bands_ptr_[NUM_TYPES][16 + 1];
 } VP8Proba;
 
 // Filter parameters
@@ -151,32 +122,61 @@ typedef struct {
 // Informations about the macroblocks.
 
 typedef struct {  // filter specs
-  unsigned int f_level_:6;      // filter strength: 0..63
-  unsigned int f_ilevel_:6;     // inner limit: 1..63
-  unsigned int f_inner_:1;      // do inner filtering?
+  uint8_t f_limit_;      // filter limit in [3..189], or 0 if no filtering
+  uint8_t f_ilevel_;     // inner limit in [1..63]
+  uint8_t f_inner_;      // do inner filtering?
+  uint8_t hev_thresh_;   // high edge variance threshold in [0..2]
 } VP8FInfo;
 
-typedef struct {  // used for syntax-parsing
-  unsigned int nz_;          // non-zero AC/DC coeffs
-  unsigned int dc_nz_:1;     // non-zero DC coeffs
-  unsigned int skip_:1;      // block type
+typedef struct {  // Top/Left Contexts used for syntax-parsing
+  uint8_t nz_;        // non-zero AC/DC coeffs (4bit for luma + 4bit for chroma)
+  uint8_t nz_dc_;     // non-zero DC coeff (1bit)
 } VP8MB;
 
 // Dequantization matrices
 typedef int quant_t[2];      // [DC / AC].  Can be 'uint16_t[2]' too (~slower).
 typedef struct {
   quant_t y1_mat_, y2_mat_, uv_mat_;
+
+  int uv_quant_;   // U/V quantizer value
+  int dither_;     // dithering amplitude (0 = off, max=255)
 } VP8QuantMatrix;
 
+// Data needed to reconstruct a macroblock
+typedef struct {
+  int16_t coeffs_[384];   // 384 coeffs = (16+4+4) * 4*4
+  uint8_t is_i4x4_;       // true if intra4x4
+  uint8_t imodes_[16];    // one 16x16 mode (#0) or sixteen 4x4 modes
+  uint8_t uvmode_;        // chroma prediction mode
+  // bit-wise info about the content of each sub-4x4 blocks (in decoding order).
+  // Each of the 4x4 blocks for y/u/v is associated with a 2b code according to:
+  //   code=0 -> no coefficient
+  //   code=1 -> only DC
+  //   code=2 -> first three coefficients are non-zero
+  //   code=3 -> more than three coefficients are non-zero
+  // This allows to call specialized transform functions.
+  uint32_t non_zero_y_;
+  uint32_t non_zero_uv_;
+  uint8_t dither_;      // local dithering strength (deduced from non_zero_*)
+  uint8_t skip_;
+  uint8_t segment_;
+} VP8MBData;
+
 // Persistent information needed by the parallel processing
 typedef struct {
-  int id_;            // cache row to process (in [0..2])
-  int mb_y_;          // macroblock position of the row
-  int filter_row_;    // true if row-filtering is needed
-  VP8FInfo* f_info_;  // filter strengths
-  VP8Io io_;          // copy of the VP8Io to pass to put()
+  int id_;              // cache row to process (in [0..2])
+  int mb_y_;            // macroblock position of the row
+  int filter_row_;      // true if row-filtering is needed
+  VP8FInfo* f_info_;    // filter strengths (swapped with dec->f_info_)
+  VP8MBData* mb_data_;  // reconstruction data (swapped with dec->mb_data_)
+  VP8Io io_;            // copy of the VP8Io to pass to put()
 } VP8ThreadContext;
 
+// Saved top samples, per macroblock. Fits into a cache-line.
+typedef struct {
+  uint8_t y[16], u[8], v[8];
+} VP8TopSamples;
+
 //------------------------------------------------------------------------------
 // VP8Decoder: the main opaque structure handed over to user
 
@@ -196,7 +196,8 @@ struct VP8Decoder {
 
   // Worker
   WebPWorker worker_;
-  int use_threads_;    // use multi-thread
+  int mt_method_;      // multi-thread method: 0=off, 1=[parse+recon][filter]
+                       // 2=[parse][recon+filter]
   int cache_id_;       // current cache row
   int num_caches_;     // number of cached rows of 16 pixels (1, 2 or 3)
   VP8ThreadContext thread_ctx_;  // Thread context
@@ -213,12 +214,9 @@ struct VP8Decoder {
   // per-partition boolean decoders.
   VP8BitReader parts_[MAX_NUM_PARTITIONS];
 
-  // buffer refresh flags
-  //   bit 0: refresh Gold, bit 1: refresh Alt
-  //   bit 2-3: copy to Gold, bit 4-5: copy to Alt
-  //   bit 6: Gold sign bias, bit 7: Alt sign bias
-  //   bit 8: refresh last frame
-  uint32_t buffer_flags_;
+  // Dithering strength, deduced from decoding options
+  int dither_;                // whether to use dithering or not
+  VP8Random dithering_rg_;    // random generator for dithering
 
   // dequantization (one set of DC/AC dequant factor per segment)
   VP8QuantMatrix dqm_[NUM_MB_SEGMENTS];
@@ -227,24 +225,18 @@ struct VP8Decoder {
   VP8Proba proba_;
   int use_skip_proba_;
   uint8_t skip_p_;
-#ifndef ONLY_KEYFRAME_CODE
-  uint8_t intra_p_, last_p_, golden_p_;
-  VP8Proba proba_saved_;
-  int update_proba_;
-#endif
 
   // Boundary data cache and persistent buffers.
-  uint8_t* intra_t_;     // top intra modes values: 4 * mb_w_
-  uint8_t  intra_l_[4];  // left intra modes values
-  uint8_t* y_t_;         // top luma samples: 16 * mb_w_
-  uint8_t* u_t_, *v_t_;  // top u/v samples: 8 * mb_w_ each
+  uint8_t* intra_t_;      // top intra modes values: 4 * mb_w_
+  uint8_t  intra_l_[4];   // left intra modes values
 
-  VP8MB* mb_info_;       // contextual macroblock info (mb_w_ + 1)
-  VP8FInfo* f_info_;     // filter strength info
-  uint8_t* yuv_b_;       // main block for Y/U/V (size = YUV_SIZE)
-  int16_t* coeffs_;      // 384 coeffs = (16+8+8) * 4*4
+  VP8TopSamples* yuv_t_;  // top y/u/v samples
 
-  uint8_t* cache_y_;     // macroblock row for storing unfiltered samples
+  VP8MB* mb_info_;        // contextual macroblock info (mb_w_ + 1)
+  VP8FInfo* f_info_;      // filter strength info
+  uint8_t* yuv_b_;        // main block for Y/U/V (size = YUV_SIZE)
+
+  uint8_t* cache_y_;      // macroblock row for storing unfiltered samples
   uint8_t* cache_u_;
   uint8_t* cache_v_;
   int cache_y_stride_;
@@ -256,31 +248,19 @@ struct VP8Decoder {
 
   // Per macroblock non-persistent infos.
   int mb_x_, mb_y_;       // current position, in macroblock units
-  uint8_t is_i4x4_;       // true if intra4x4
-  uint8_t imodes_[16];    // one 16x16 mode (#0) or sixteen 4x4 modes
-  uint8_t uvmode_;        // chroma prediction mode
-  uint8_t segment_;       // block's segment
-
-  // bit-wise info about the content of each sub-4x4 blocks: there are 16 bits
-  // for luma (bits #0->#15), then 4 bits for chroma-u (#16->#19) and 4 bits for
-  // chroma-v (#20->#23), each corresponding to one 4x4 block in decoding order.
-  // If the bit is set, the 4x4 block contains some non-zero coefficients.
-  uint32_t non_zero_;
-  uint32_t non_zero_ac_;
+  VP8MBData* mb_data_;    // parsed reconstruction data
 
   // Filtering side-info
-  int filter_type_;                         // 0=off, 1=simple, 2=complex
-  int filter_row_;                          // per-row flag
-  uint8_t filter_levels_[NUM_MB_SEGMENTS];  // precalculated per-segment
+  int filter_type_;                          // 0=off, 1=simple, 2=complex
+  VP8FInfo fstrengths_[NUM_MB_SEGMENTS][2];  // precalculated per-segment/type
 
-  // extensions
-  const uint8_t* alpha_data_;   // compressed alpha data (if present)
+  // Alpha
+  struct ALPHDecoder* alph_dec_;  // alpha-plane decoder object
+  const uint8_t* alpha_data_;     // compressed alpha data (if present)
   size_t alpha_data_size_;
-  uint8_t* alpha_plane_;        // output. Persistent, contains the whole data.
-
-  int layer_colorspace_;
-  const uint8_t* layer_data_;   // compressed layer data (if present)
-  size_t layer_data_size_;
+  int is_alpha_decoded_;  // true if alpha_data_ is decoded in alpha_plane_
+  uint8_t* alpha_plane_;  // output. Persistent, contains the whole data.
+  int alpha_dithering_;   // derived from decoding options (0=off, 100=full).
 };
 
 //------------------------------------------------------------------------------
@@ -293,15 +273,14 @@ int VP8SetError(VP8Decoder* const dec,
 // in tree.c
 void VP8ResetProba(VP8Proba* const proba);
 void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec);
-void VP8ParseIntraMode(VP8BitReader* const br,  VP8Decoder* const dec);
+// parses one row of intra mode data in partition 0, returns !eof
+int VP8ParseIntraModeRow(VP8BitReader* const br, VP8Decoder* const dec);
 
 // in quant.c
 void VP8ParseQuant(VP8Decoder* const dec);
 
 // in frame.c
-int VP8InitFrame(VP8Decoder* const dec, VP8Io* io);
-// Predict a block and add residual
-void VP8ReconstructBlock(VP8Decoder* const dec);
+int VP8InitFrame(VP8Decoder* const dec, VP8Io* const io);
 // Call io->setup() and finish setting up scan parameters.
 // After this call returns, one must always call VP8ExitCritical() with the
 // same parameters. Both functions should be used in pair. Returns VP8_STATUS_OK
@@ -310,10 +289,16 @@ VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io);
 // Must always be called in pair with VP8EnterCritical().
 // Returns false in case of error.
 int VP8ExitCritical(VP8Decoder* const dec, VP8Io* const io);
-// Process the last decoded row (filtering + output)
+// Return the multi-threading method to use (0=off), depending
+// on options and bitstream size. Only for lossy decoding.
+int VP8GetThreadMethod(const WebPDecoderOptions* const options,
+                       const WebPHeaderStructure* const headers,
+                       int width, int height);
+// Initialize dithering post-process if needed.
+void VP8InitDithering(const WebPDecoderOptions* const options,
+                      VP8Decoder* const dec);
+// Process the last decoded row (filtering + output).
 int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io);
-// Store a block, along with filtering params
-void VP8StoreBlock(VP8Decoder* const dec);
 // To be called at the start of a new scanline, to initialize predictors.
 void VP8InitScanline(VP8Decoder* const dec);
 // Decode one macroblock. Returns false if there is not enough data.
@@ -323,12 +308,9 @@ int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br);
 const uint8_t* VP8DecompressAlphaRows(VP8Decoder* const dec,
                                       int row, int num_rows);
 
-// in layer.c
-int VP8DecodeLayer(VP8Decoder* const dec);
-
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/dec/vp8l.c b/drivers/webp/dec/vp8l.c
index 897e4395c7..19665a007d 100644
--- a/drivers/webp/dec/vp8l.c
+++ b/drivers/webp/dec/vp8l.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 decoder
@@ -10,18 +12,17 @@
 // Authors: Vikas Arora (vikaas.arora@gmail.com)
 //          Jyrki Alakuijala (jyrki@google.com)
 
-#include <stdio.h>
 #include <stdlib.h>
+
+#include "./alphai.h"
 #include "./vp8li.h"
+#include "../dsp/dsp.h"
 #include "../dsp/lossless.h"
 #include "../dsp/yuv.h"
+#include "../utils/endian_inl.h"
 #include "../utils/huffman.h"
 #include "../utils/utils.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 #define NUM_ARGB_CACHE_ROWS          16
 
 static const int kCodeLengthLiterals = 16;
@@ -50,6 +51,9 @@ static const uint16_t kAlphabetSize[HUFFMAN_CODES_PER_META_CODE] = {
   NUM_DISTANCE_CODES
 };
 
+static const uint8_t kLiteralMap[HUFFMAN_CODES_PER_META_CODE] = {
+  0, 1, 1, 1, 0
+};
 
 #define NUM_CODE_LENGTH_CODES       19
 static const uint8_t kCodeLengthCodeOrder[NUM_CODE_LENGTH_CODES] = {
@@ -57,19 +61,43 @@ static const uint8_t kCodeLengthCodeOrder[NUM_CODE_LENGTH_CODES] = {
 };
 
 #define CODE_TO_PLANE_CODES        120
-static const uint8_t code_to_plane_lut[CODE_TO_PLANE_CODES] = {
-   0x18, 0x07, 0x17, 0x19, 0x28, 0x06, 0x27, 0x29, 0x16, 0x1a,
-   0x26, 0x2a, 0x38, 0x05, 0x37, 0x39, 0x15, 0x1b, 0x36, 0x3a,
-   0x25, 0x2b, 0x48, 0x04, 0x47, 0x49, 0x14, 0x1c, 0x35, 0x3b,
-   0x46, 0x4a, 0x24, 0x2c, 0x58, 0x45, 0x4b, 0x34, 0x3c, 0x03,
-   0x57, 0x59, 0x13, 0x1d, 0x56, 0x5a, 0x23, 0x2d, 0x44, 0x4c,
-   0x55, 0x5b, 0x33, 0x3d, 0x68, 0x02, 0x67, 0x69, 0x12, 0x1e,
-   0x66, 0x6a, 0x22, 0x2e, 0x54, 0x5c, 0x43, 0x4d, 0x65, 0x6b,
-   0x32, 0x3e, 0x78, 0x01, 0x77, 0x79, 0x53, 0x5d, 0x11, 0x1f,
-   0x64, 0x6c, 0x42, 0x4e, 0x76, 0x7a, 0x21, 0x2f, 0x75, 0x7b,
-   0x31, 0x3f, 0x63, 0x6d, 0x52, 0x5e, 0x00, 0x74, 0x7c, 0x41,
-   0x4f, 0x10, 0x20, 0x62, 0x6e, 0x30, 0x73, 0x7d, 0x51, 0x5f,
-   0x40, 0x72, 0x7e, 0x61, 0x6f, 0x50, 0x71, 0x7f, 0x60, 0x70
+static const uint8_t kCodeToPlane[CODE_TO_PLANE_CODES] = {
+  0x18, 0x07, 0x17, 0x19, 0x28, 0x06, 0x27, 0x29, 0x16, 0x1a,
+  0x26, 0x2a, 0x38, 0x05, 0x37, 0x39, 0x15, 0x1b, 0x36, 0x3a,
+  0x25, 0x2b, 0x48, 0x04, 0x47, 0x49, 0x14, 0x1c, 0x35, 0x3b,
+  0x46, 0x4a, 0x24, 0x2c, 0x58, 0x45, 0x4b, 0x34, 0x3c, 0x03,
+  0x57, 0x59, 0x13, 0x1d, 0x56, 0x5a, 0x23, 0x2d, 0x44, 0x4c,
+  0x55, 0x5b, 0x33, 0x3d, 0x68, 0x02, 0x67, 0x69, 0x12, 0x1e,
+  0x66, 0x6a, 0x22, 0x2e, 0x54, 0x5c, 0x43, 0x4d, 0x65, 0x6b,
+  0x32, 0x3e, 0x78, 0x01, 0x77, 0x79, 0x53, 0x5d, 0x11, 0x1f,
+  0x64, 0x6c, 0x42, 0x4e, 0x76, 0x7a, 0x21, 0x2f, 0x75, 0x7b,
+  0x31, 0x3f, 0x63, 0x6d, 0x52, 0x5e, 0x00, 0x74, 0x7c, 0x41,
+  0x4f, 0x10, 0x20, 0x62, 0x6e, 0x30, 0x73, 0x7d, 0x51, 0x5f,
+  0x40, 0x72, 0x7e, 0x61, 0x6f, 0x50, 0x71, 0x7f, 0x60, 0x70
+};
+
+// Memory needed for lookup tables of one Huffman tree group. Red, blue, alpha
+// and distance alphabets are constant (256 for red, blue and alpha, 40 for
+// distance) and lookup table sizes for them in worst case are 630 and 410
+// respectively. Size of green alphabet depends on color cache size and is equal
+// to 256 (green component values) + 24 (length prefix values)
+// + color_cache_size (between 0 and 2048).
+// All values computed for 8-bit first level lookup with Mark Adler's tool:
+// http://www.hdfgroup.org/ftp/lib-external/zlib/zlib-1.2.5/examples/enough.c
+#define FIXED_TABLE_SIZE (630 * 3 + 410)
+static const int kTableSize[12] = {
+  FIXED_TABLE_SIZE + 654,
+  FIXED_TABLE_SIZE + 656,
+  FIXED_TABLE_SIZE + 658,
+  FIXED_TABLE_SIZE + 662,
+  FIXED_TABLE_SIZE + 670,
+  FIXED_TABLE_SIZE + 686,
+  FIXED_TABLE_SIZE + 718,
+  FIXED_TABLE_SIZE + 782,
+  FIXED_TABLE_SIZE + 912,
+  FIXED_TABLE_SIZE + 1168,
+  FIXED_TABLE_SIZE + 1680,
+  FIXED_TABLE_SIZE + 2704
 };
 
 static int DecodeImageStream(int xsize, int ysize,
@@ -80,27 +108,28 @@ static int DecodeImageStream(int xsize, int ysize,
 //------------------------------------------------------------------------------
 
 int VP8LCheckSignature(const uint8_t* const data, size_t size) {
-  return (size >= 1) && (data[0] == VP8L_MAGIC_BYTE);
+  return (size >= VP8L_FRAME_HEADER_SIZE &&
+          data[0] == VP8L_MAGIC_BYTE &&
+          (data[4] >> 5) == 0);  // version
 }
 
 static int ReadImageInfo(VP8LBitReader* const br,
                          int* const width, int* const height,
                          int* const has_alpha) {
-  const uint8_t signature = VP8LReadBits(br, 8);
-  if (!VP8LCheckSignature(&signature, 1)) {
-    return 0;
-  }
+  if (VP8LReadBits(br, 8) != VP8L_MAGIC_BYTE) return 0;
   *width = VP8LReadBits(br, VP8L_IMAGE_SIZE_BITS) + 1;
   *height = VP8LReadBits(br, VP8L_IMAGE_SIZE_BITS) + 1;
   *has_alpha = VP8LReadBits(br, 1);
-  VP8LReadBits(br, VP8L_VERSION_BITS);  // Read/ignore the version number.
-  return 1;
+  if (VP8LReadBits(br, VP8L_VERSION_BITS) != 0) return 0;
+  return !br->eos_;
 }
 
 int VP8LGetInfo(const uint8_t* data, size_t data_size,
                 int* const width, int* const height, int* const has_alpha) {
   if (data == NULL || data_size < VP8L_FRAME_HEADER_SIZE) {
     return 0;         // not enough data
+  } else if (!VP8LCheckSignature(data, data_size)) {
+    return 0;         // bad signature
   } else {
     int w, h, a;
     VP8LBitReader br;
@@ -138,39 +167,80 @@ static WEBP_INLINE int PlaneCodeToDistance(int xsize, int plane_code) {
   if (plane_code > CODE_TO_PLANE_CODES) {
     return plane_code - CODE_TO_PLANE_CODES;
   } else {
-    const int dist_code = code_to_plane_lut[plane_code - 1];
+    const int dist_code = kCodeToPlane[plane_code - 1];
     const int yoffset = dist_code >> 4;
     const int xoffset = 8 - (dist_code & 0xf);
     const int dist = yoffset * xsize + xoffset;
-    return (dist >= 1) ? dist : 1;
+    return (dist >= 1) ? dist : 1;  // dist<1 can happen if xsize is very small
   }
 }
 
 //------------------------------------------------------------------------------
 // Decodes the next Huffman code from bit-stream.
 // FillBitWindow(br) needs to be called at minimum every second call
-// to ReadSymbolUnsafe.
-static int ReadSymbolUnsafe(const HuffmanTree* tree, VP8LBitReader* const br) {
-  const HuffmanTreeNode* node = tree->root_;
-  assert(node != NULL);
-  while (!HuffmanTreeNodeIsLeaf(node)) {
-    node = HuffmanTreeNextNode(node, VP8LReadOneBitUnsafe(br));
-  }
-  return node->symbol_;
+// to ReadSymbol, in order to pre-fetch enough bits.
+static WEBP_INLINE int ReadSymbol(const HuffmanCode* table,
+                                  VP8LBitReader* const br) {
+  int nbits;
+  uint32_t val = VP8LPrefetchBits(br);
+  table += val & HUFFMAN_TABLE_MASK;
+  nbits = table->bits - HUFFMAN_TABLE_BITS;
+  if (nbits > 0) {
+    VP8LSetBitPos(br, br->bit_pos_ + HUFFMAN_TABLE_BITS);
+    val = VP8LPrefetchBits(br);
+    table += table->value;
+    table += val & ((1 << nbits) - 1);
+  }
+  VP8LSetBitPos(br, br->bit_pos_ + table->bits);
+  return table->value;
 }
 
-static WEBP_INLINE int ReadSymbol(const HuffmanTree* tree,
-                                  VP8LBitReader* const br) {
-  const int read_safe = (br->pos_ + 8 > br->len_);
-  if (!read_safe) {
-    return ReadSymbolUnsafe(tree, br);
+// Reads packed symbol depending on GREEN channel
+#define BITS_SPECIAL_MARKER 0x100  // something large enough (and a bit-mask)
+#define PACKED_NON_LITERAL_CODE 0  // must be < NUM_LITERAL_CODES
+static WEBP_INLINE int ReadPackedSymbols(const HTreeGroup* group,
+                                         VP8LBitReader* const br,
+                                         uint32_t* const dst) {
+  const uint32_t val = VP8LPrefetchBits(br) & (HUFFMAN_PACKED_TABLE_SIZE - 1);
+  const HuffmanCode32 code = group->packed_table[val];
+  assert(group->use_packed_table);
+  if (code.bits < BITS_SPECIAL_MARKER) {
+    VP8LSetBitPos(br, br->bit_pos_ + code.bits);
+    *dst = code.value;
+    return PACKED_NON_LITERAL_CODE;
   } else {
-    const HuffmanTreeNode* node = tree->root_;
-    assert(node != NULL);
-    while (!HuffmanTreeNodeIsLeaf(node)) {
-      node = HuffmanTreeNextNode(node, VP8LReadOneBit(br));
+    VP8LSetBitPos(br, br->bit_pos_ + code.bits - BITS_SPECIAL_MARKER);
+    assert(code.value >= NUM_LITERAL_CODES);
+    return code.value;
+  }
+}
+
+static int AccumulateHCode(HuffmanCode hcode, int shift,
+                           HuffmanCode32* const huff) {
+  huff->bits += hcode.bits;
+  huff->value |= (uint32_t)hcode.value << shift;
+  assert(huff->bits <= HUFFMAN_TABLE_BITS);
+  return hcode.bits;
+}
+
+static void BuildPackedTable(HTreeGroup* const htree_group) {
+  uint32_t code;
+  for (code = 0; code < HUFFMAN_PACKED_TABLE_SIZE; ++code) {
+    uint32_t bits = code;
+    HuffmanCode32* const huff = &htree_group->packed_table[bits];
+    HuffmanCode hcode = htree_group->htrees[GREEN][bits];
+    if (hcode.value >= NUM_LITERAL_CODES) {
+      huff->bits = hcode.bits + BITS_SPECIAL_MARKER;
+      huff->value = hcode.value;
+    } else {
+      huff->bits = 0;
+      huff->value = 0;
+      bits >>= AccumulateHCode(hcode, 8, huff);
+      bits >>= AccumulateHCode(htree_group->htrees[RED][bits], 16, huff);
+      bits >>= AccumulateHCode(htree_group->htrees[BLUE][bits], 0, huff);
+      bits >>= AccumulateHCode(htree_group->htrees[ALPHA][bits], 24, huff);
+      (void)bits;
     }
-    return node->symbol_;
   }
 }
 
@@ -182,19 +252,18 @@ static int ReadHuffmanCodeLengths(
   int symbol;
   int max_symbol;
   int prev_code_len = DEFAULT_CODE_LENGTH;
-  HuffmanTree tree;
+  HuffmanCode table[1 << LENGTHS_TABLE_BITS];
 
-  if (!HuffmanTreeBuildImplicit(&tree, code_length_code_lengths,
-                                NUM_CODE_LENGTH_CODES)) {
-    dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
-    return 0;
+  if (!VP8LBuildHuffmanTable(table, LENGTHS_TABLE_BITS,
+                             code_length_code_lengths,
+                             NUM_CODE_LENGTH_CODES)) {
+    goto End;
   }
 
   if (VP8LReadBits(br, 1)) {    // use length
     const int length_nbits = 2 + 2 * VP8LReadBits(br, 3);
     max_symbol = 2 + VP8LReadBits(br, length_nbits);
     if (max_symbol > num_symbols) {
-      dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
       goto End;
     }
   } else {
@@ -203,10 +272,13 @@ static int ReadHuffmanCodeLengths(
 
   symbol = 0;
   while (symbol < num_symbols) {
+    const HuffmanCode* p;
     int code_len;
     if (max_symbol-- == 0) break;
     VP8LFillBitWindow(br);
-    code_len = ReadSymbol(&tree, br);
+    p = &table[VP8LPrefetchBits(br) & LENGTHS_TABLE_MASK];
+    VP8LSetBitPos(br, br->bit_pos_ + p->bits);
+    code_len = p->value;
     if (code_len < kCodeLengthLiterals) {
       code_lengths[symbol++] = code_len;
       if (code_len != 0) prev_code_len = code_len;
@@ -217,7 +289,6 @@ static int ReadHuffmanCodeLengths(
       const int repeat_offset = kCodeLengthRepeatOffsets[slot];
       int repeat = VP8LReadBits(br, extra_bits) + repeat_offset;
       if (symbol + repeat > num_symbols) {
-        dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
         goto End;
       } else {
         const int length = use_prev ? prev_code_len : 0;
@@ -228,36 +299,34 @@ static int ReadHuffmanCodeLengths(
   ok = 1;
 
  End:
-  HuffmanTreeRelease(&tree);
+  if (!ok) dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
   return ok;
 }
 
+// 'code_lengths' is pre-allocated temporary buffer, used for creating Huffman
+// tree.
 static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
-                           HuffmanTree* const tree) {
+                           int* const code_lengths, HuffmanCode* const table) {
   int ok = 0;
+  int size = 0;
   VP8LBitReader* const br = &dec->br_;
   const int simple_code = VP8LReadBits(br, 1);
 
+  memset(code_lengths, 0, alphabet_size * sizeof(*code_lengths));
+
   if (simple_code) {  // Read symbols, codes & code lengths directly.
-    int symbols[2];
-    int codes[2];
-    int code_lengths[2];
     const int num_symbols = VP8LReadBits(br, 1) + 1;
     const int first_symbol_len_code = VP8LReadBits(br, 1);
     // The first code is either 1 bit or 8 bit code.
-    symbols[0] = VP8LReadBits(br, (first_symbol_len_code == 0) ? 1 : 8);
-    codes[0] = 0;
-    code_lengths[0] = num_symbols - 1;
+    int symbol = VP8LReadBits(br, (first_symbol_len_code == 0) ? 1 : 8);
+    code_lengths[symbol] = 1;
     // The second code (if present), is always 8 bit long.
     if (num_symbols == 2) {
-      symbols[1] = VP8LReadBits(br, 8);
-      codes[1] = 1;
-      code_lengths[1] = num_symbols - 1;
+      symbol = VP8LReadBits(br, 8);
+      code_lengths[symbol] = 1;
     }
-    ok = HuffmanTreeBuildExplicit(tree, code_lengths, codes, symbols,
-                                  alphabet_size, num_symbols);
+    ok = 1;
   } else {  // Decode Huffman-coded code lengths.
-    int* code_lengths = NULL;
     int i;
     int code_length_code_lengths[NUM_CODE_LENGTH_CODES] = { 0 };
     const int num_codes = VP8LReadBits(br, 4) + 4;
@@ -266,42 +335,23 @@ static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
       return 0;
     }
 
-    code_lengths =
-        (int*)WebPSafeCalloc((uint64_t)alphabet_size, sizeof(*code_lengths));
-    if (code_lengths == NULL) {
-      dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
-      return 0;
-    }
-
     for (i = 0; i < num_codes; ++i) {
       code_length_code_lengths[kCodeLengthCodeOrder[i]] = VP8LReadBits(br, 3);
     }
     ok = ReadHuffmanCodeLengths(dec, code_length_code_lengths, alphabet_size,
                                 code_lengths);
-    if (ok) {
-      ok = HuffmanTreeBuildImplicit(tree, code_lengths, alphabet_size);
-    }
-    free(code_lengths);
   }
-  ok = ok && !br->error_;
-  if (!ok) {
+
+  ok = ok && !br->eos_;
+  if (ok) {
+    size = VP8LBuildHuffmanTable(table, HUFFMAN_TABLE_BITS,
+                                 code_lengths, alphabet_size);
+  }
+  if (!ok || size == 0) {
     dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
     return 0;
   }
-  return 1;
-}
-
-static void DeleteHtreeGroups(HTreeGroup* htree_groups, int num_htree_groups) {
-  if (htree_groups != NULL) {
-    int i, j;
-    for (i = 0; i < num_htree_groups; ++i) {
-      HuffmanTree* const htrees = htree_groups[i].htrees_;
-      for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
-        HuffmanTreeRelease(&htrees[j]);
-      }
-    }
-    free(htree_groups);
-  }
+  return size;
 }
 
 static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
@@ -311,7 +361,12 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
   VP8LMetadata* const hdr = &dec->hdr_;
   uint32_t* huffman_image = NULL;
   HTreeGroup* htree_groups = NULL;
+  HuffmanCode* huffman_tables = NULL;
+  HuffmanCode* next = NULL;
   int num_htree_groups = 1;
+  int max_alphabet_size = 0;
+  int* code_lengths = NULL;
+  const int table_size = kTableSize[color_cache_bits];
 
   if (allow_recursion && VP8LReadBits(br, 1)) {
     // use meta Huffman codes.
@@ -321,51 +376,108 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
     const int huffman_pixs = huffman_xsize * huffman_ysize;
     if (!DecodeImageStream(huffman_xsize, huffman_ysize, 0, dec,
                            &huffman_image)) {
-      dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
       goto Error;
     }
     hdr->huffman_subsample_bits_ = huffman_precision;
     for (i = 0; i < huffman_pixs; ++i) {
       // The huffman data is stored in red and green bytes.
-      const int index = (huffman_image[i] >> 8) & 0xffff;
-      huffman_image[i] = index;
-      if (index >= num_htree_groups) {
-        num_htree_groups = index + 1;
+      const int group = (huffman_image[i] >> 8) & 0xffff;
+      huffman_image[i] = group;
+      if (group >= num_htree_groups) {
+        num_htree_groups = group + 1;
       }
     }
   }
 
-  if (br->error_) goto Error;
+  if (br->eos_) goto Error;
+
+  // Find maximum alphabet size for the htree group.
+  for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
+    int alphabet_size = kAlphabetSize[j];
+    if (j == 0 && color_cache_bits > 0) {
+      alphabet_size += 1 << color_cache_bits;
+    }
+    if (max_alphabet_size < alphabet_size) {
+      max_alphabet_size = alphabet_size;
+    }
+  }
+
+  huffman_tables = (HuffmanCode*)WebPSafeMalloc(num_htree_groups * table_size,
+                                                sizeof(*huffman_tables));
+  htree_groups = VP8LHtreeGroupsNew(num_htree_groups);
+  code_lengths = (int*)WebPSafeCalloc((uint64_t)max_alphabet_size,
+                                      sizeof(*code_lengths));
 
-  assert(num_htree_groups <= 0x10000);
-  htree_groups =
-      (HTreeGroup*)WebPSafeCalloc((uint64_t)num_htree_groups,
-                                  sizeof(*htree_groups));
-  if (htree_groups == NULL) {
+  if (htree_groups == NULL || code_lengths == NULL || huffman_tables == NULL) {
     dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
     goto Error;
   }
 
+  next = huffman_tables;
   for (i = 0; i < num_htree_groups; ++i) {
-    HuffmanTree* const htrees = htree_groups[i].htrees_;
+    HTreeGroup* const htree_group = &htree_groups[i];
+    HuffmanCode** const htrees = htree_group->htrees;
+    int size;
+    int total_size = 0;
+    int is_trivial_literal = 1;
+    int max_bits = 0;
     for (j = 0; j < HUFFMAN_CODES_PER_META_CODE; ++j) {
       int alphabet_size = kAlphabetSize[j];
+      htrees[j] = next;
       if (j == 0 && color_cache_bits > 0) {
         alphabet_size += 1 << color_cache_bits;
       }
-      if (!ReadHuffmanCode(alphabet_size, dec, htrees + j)) goto Error;
+      size = ReadHuffmanCode(alphabet_size, dec, code_lengths, next);
+      if (is_trivial_literal && kLiteralMap[j] == 1) {
+        is_trivial_literal = (next->bits == 0);
+      }
+      total_size += next->bits;
+      next += size;
+      if (size == 0) {
+        goto Error;
+      }
+      if (j <= ALPHA) {
+        int local_max_bits = code_lengths[0];
+        int k;
+        for (k = 1; k < alphabet_size; ++k) {
+          if (code_lengths[k] > local_max_bits) {
+            local_max_bits = code_lengths[k];
+          }
+        }
+        max_bits += local_max_bits;
+      }
     }
+    htree_group->is_trivial_literal = is_trivial_literal;
+    htree_group->is_trivial_code = 0;
+    if (is_trivial_literal) {
+      const int red = htrees[RED][0].value;
+      const int blue = htrees[BLUE][0].value;
+      const int alpha = htrees[ALPHA][0].value;
+      htree_group->literal_arb =
+          ((uint32_t)alpha << 24) | (red << 16) | blue;
+      if (total_size == 0 && htrees[GREEN][0].value < NUM_LITERAL_CODES) {
+        htree_group->is_trivial_code = 1;
+        htree_group->literal_arb |= htrees[GREEN][0].value << 8;
+      }
+    }
+    htree_group->use_packed_table = !htree_group->is_trivial_code &&
+                                    (max_bits < HUFFMAN_PACKED_BITS);
+    if (htree_group->use_packed_table) BuildPackedTable(htree_group);
   }
+  WebPSafeFree(code_lengths);
 
   // All OK. Finalize pointers and return.
   hdr->huffman_image_ = huffman_image;
   hdr->num_htree_groups_ = num_htree_groups;
   hdr->htree_groups_ = htree_groups;
+  hdr->huffman_tables_ = huffman_tables;
   return 1;
 
  Error:
-  free(huffman_image);
-  DeleteHtreeGroups(htree_groups, num_htree_groups);
+  WebPSafeFree(code_lengths);
+  WebPSafeFree(huffman_image);
+  WebPSafeFree(huffman_tables);
+  VP8LHtreeGroupsFree(htree_groups);
   return 0;
 }
 
@@ -379,13 +491,13 @@ static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
   const int in_height = io->mb_h;
   const int out_height = io->scaled_height;
   const uint64_t work_size = 2 * num_channels * (uint64_t)out_width;
-  int32_t* work;        // Rescaler work area.
-  const uint64_t scaled_data_size = num_channels * (uint64_t)out_width;
+  rescaler_t* work;        // Rescaler work area.
+  const uint64_t scaled_data_size = (uint64_t)out_width;
   uint32_t* scaled_data;  // Temporary storage for scaled BGRA data.
   const uint64_t memory_size = sizeof(*dec->rescaler) +
                                work_size * sizeof(*work) +
                                scaled_data_size * sizeof(*scaled_data);
-  uint8_t* memory = (uint8_t*)WebPSafeCalloc(memory_size, sizeof(*memory));
+  uint8_t* memory = (uint8_t*)WebPSafeMalloc(memory_size, sizeof(*memory));
   if (memory == NULL) {
     dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
     return 0;
@@ -395,13 +507,12 @@ static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
 
   dec->rescaler = (WebPRescaler*)memory;
   memory += sizeof(*dec->rescaler);
-  work = (int32_t*)memory;
+  work = (rescaler_t*)memory;
   memory += work_size * sizeof(*work);
   scaled_data = (uint32_t*)memory;
 
   WebPRescalerInit(dec->rescaler, in_width, in_height, (uint8_t*)scaled_data,
-                   out_width, out_height, 0, num_channels,
-                   in_width, out_width, in_height, out_height, work);
+                   out_width, out_height, 0, num_channels, work);
   return 1;
 }
 
@@ -411,12 +522,13 @@ static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
 // We have special "export" function since we need to convert from BGRA
 static int Export(WebPRescaler* const rescaler, WEBP_CSP_MODE colorspace,
                   int rgba_stride, uint8_t* const rgba) {
-  const uint32_t* const src = (const uint32_t*)rescaler->dst;
+  uint32_t* const src = (uint32_t*)rescaler->dst;
   const int dst_width = rescaler->dst_width;
   int num_lines_out = 0;
   while (WebPRescalerHasPendingOutput(rescaler)) {
     uint8_t* const dst = rgba + num_lines_out * rgba_stride;
     WebPRescalerExportRow(rescaler);
+    WebPMultARGBRow(src, dst_width, 1);
     VP8LConvertFromBGRA(src, dst_width, colorspace, dst);
     ++num_lines_out;
   }
@@ -424,18 +536,22 @@ static int Export(WebPRescaler* const rescaler, WEBP_CSP_MODE colorspace,
 }
 
 // Emit scaled rows.
-static int EmitRescaledRows(const VP8LDecoder* const dec,
-                            const uint32_t* const data, int in_stride, int mb_h,
-                            uint8_t* const out, int out_stride) {
+static int EmitRescaledRowsRGBA(const VP8LDecoder* const dec,
+                                uint8_t* in, int in_stride, int mb_h,
+                                uint8_t* const out, int out_stride) {
   const WEBP_CSP_MODE colorspace = dec->output_->colorspace;
-  const uint8_t* const in = (const uint8_t*)data;
   int num_lines_in = 0;
   int num_lines_out = 0;
   while (num_lines_in < mb_h) {
-    const uint8_t* const row_in = in + num_lines_in * in_stride;
+    uint8_t* const row_in = in + num_lines_in * in_stride;
     uint8_t* const row_out = out + num_lines_out * out_stride;
-    num_lines_in += WebPRescalerImport(dec->rescaler, mb_h - num_lines_in,
-                                       row_in, in_stride);
+    const int lines_left = mb_h - num_lines_in;
+    const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
+    assert(needed_lines > 0 && needed_lines <= lines_left);
+    WebPMultARGBRows(row_in, in_stride,
+                     dec->rescaler->src_width, needed_lines, 0);
+    WebPRescalerImport(dec->rescaler, lines_left, row_in, in_stride);
+    num_lines_in += needed_lines;
     num_lines_out += Export(dec->rescaler, colorspace, out_stride, row_out);
   }
   return num_lines_out;
@@ -443,11 +559,10 @@ static int EmitRescaledRows(const VP8LDecoder* const dec,
 
 // Emit rows without any scaling.
 static int EmitRows(WEBP_CSP_MODE colorspace,
-                    const uint32_t* const data, int in_stride,
+                    const uint8_t* row_in, int in_stride,
                     int mb_w, int mb_h,
                     uint8_t* const out, int out_stride) {
   int lines = mb_h;
-  const uint8_t* row_in = (const uint8_t*)data;
   uint8_t* row_out = out;
   while (lines-- > 0) {
     VP8LConvertFromBGRA((const uint32_t*)row_in, mb_w, colorspace, row_out);
@@ -463,72 +578,37 @@ static int EmitRows(WEBP_CSP_MODE colorspace,
 static void ConvertToYUVA(const uint32_t* const src, int width, int y_pos,
                           const WebPDecBuffer* const output) {
   const WebPYUVABuffer* const buf = &output->u.YUVA;
+
   // first, the luma plane
-  {
-    int i;
-    uint8_t* const y = buf->y + y_pos * buf->y_stride;
-    for (i = 0; i < width; ++i) {
-      const uint32_t p = src[i];
-      y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff);
-    }
-  }
+  WebPConvertARGBToY(src, buf->y + y_pos * buf->y_stride, width);
 
   // then U/V planes
   {
     uint8_t* const u = buf->u + (y_pos >> 1) * buf->u_stride;
     uint8_t* const v = buf->v + (y_pos >> 1) * buf->v_stride;
-    const int uv_width = width >> 1;
-    int i;
-    for (i = 0; i < uv_width; ++i) {
-      const uint32_t v0 = src[2 * i + 0];
-      const uint32_t v1 = src[2 * i + 1];
-      // VP8RGBToU/V expects four accumulated pixels. Hence we need to
-      // scale r/g/b value by a factor 2. We just shift v0/v1 one bit less.
-      const int r = ((v0 >> 15) & 0x1fe) + ((v1 >> 15) & 0x1fe);
-      const int g = ((v0 >>  7) & 0x1fe) + ((v1 >>  7) & 0x1fe);
-      const int b = ((v0 <<  1) & 0x1fe) + ((v1 <<  1) & 0x1fe);
-      if (!(y_pos & 1)) {  // even lines: store values
-        u[i] = VP8RGBToU(r, g, b);
-        v[i] = VP8RGBToV(r, g, b);
-      } else {             // odd lines: average with previous values
-        const int tmp_u = VP8RGBToU(r, g, b);
-        const int tmp_v = VP8RGBToV(r, g, b);
-        // Approximated average-of-four. But it's an acceptable diff.
-        u[i] = (u[i] + tmp_u + 1) >> 1;
-        v[i] = (v[i] + tmp_v + 1) >> 1;
-      }
-    }
-    if (width & 1) {       // last pixel
-      const uint32_t v0 = src[2 * i + 0];
-      const int r = (v0 >> 14) & 0x3fc;
-      const int g = (v0 >>  6) & 0x3fc;
-      const int b = (v0 <<  2) & 0x3fc;
-      if (!(y_pos & 1)) {  // even lines
-        u[i] = VP8RGBToU(r, g, b);
-        v[i] = VP8RGBToV(r, g, b);
-      } else {             // odd lines (note: we could just skip this)
-        const int tmp_u = VP8RGBToU(r, g, b);
-        const int tmp_v = VP8RGBToV(r, g, b);
-        u[i] = (u[i] + tmp_u + 1) >> 1;
-        v[i] = (v[i] + tmp_v + 1) >> 1;
-      }
-    }
+    // even lines: store values
+    // odd lines: average with previous values
+    WebPConvertARGBToUV(src, u, v, width, !(y_pos & 1));
   }
   // Lastly, store alpha if needed.
   if (buf->a != NULL) {
-    int i;
     uint8_t* const a = buf->a + y_pos * buf->a_stride;
-    for (i = 0; i < width; ++i) a[i] = (src[i] >> 24);
+#if defined(WORDS_BIGENDIAN)
+    WebPExtractAlpha((uint8_t*)src + 0, 0, width, 1, a, 0);
+#else
+    WebPExtractAlpha((uint8_t*)src + 3, 0, width, 1, a, 0);
+#endif
   }
 }
 
 static int ExportYUVA(const VP8LDecoder* const dec, int y_pos) {
   WebPRescaler* const rescaler = dec->rescaler;
-  const uint32_t* const src = (const uint32_t*)rescaler->dst;
+  uint32_t* const src = (uint32_t*)rescaler->dst;
   const int dst_width = rescaler->dst_width;
   int num_lines_out = 0;
   while (WebPRescalerHasPendingOutput(rescaler)) {
     WebPRescalerExportRow(rescaler);
+    WebPMultARGBRow(src, dst_width, 1);
     ConvertToYUVA(src, dst_width, y_pos, dec->output_);
     ++y_pos;
     ++num_lines_out;
@@ -537,28 +617,28 @@ static int ExportYUVA(const VP8LDecoder* const dec, int y_pos) {
 }
 
 static int EmitRescaledRowsYUVA(const VP8LDecoder* const dec,
-                                const uint32_t* const data,
-                                int in_stride, int mb_h) {
-  const uint8_t* const in = (const uint8_t*)data;
+                                uint8_t* in, int in_stride, int mb_h) {
   int num_lines_in = 0;
   int y_pos = dec->last_out_row_;
   while (num_lines_in < mb_h) {
-    const uint8_t* const row_in = in + num_lines_in * in_stride;
-    num_lines_in += WebPRescalerImport(dec->rescaler, mb_h - num_lines_in,
-                                       row_in, in_stride);
+    const int lines_left = mb_h - num_lines_in;
+    const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
+    WebPMultARGBRows(in, in_stride, dec->rescaler->src_width, needed_lines, 0);
+    WebPRescalerImport(dec->rescaler, lines_left, in, in_stride);
+    num_lines_in += needed_lines;
+    in += needed_lines * in_stride;
     y_pos += ExportYUVA(dec, y_pos);
   }
   return y_pos;
 }
 
 static int EmitRowsYUVA(const VP8LDecoder* const dec,
-                        const uint32_t* const data, int in_stride,
+                        const uint8_t* in, int in_stride,
                         int mb_w, int num_rows) {
   int y_pos = dec->last_out_row_;
-  const uint8_t* row_in = (const uint8_t*)data;
   while (num_rows-- > 0) {
-    ConvertToYUVA((const uint32_t*)row_in, mb_w, y_pos, dec->output_);
-    row_in += in_stride;
+    ConvertToYUVA((const uint32_t*)in, mb_w, y_pos, dec->output_);
+    in += in_stride;
     ++y_pos;
   }
   return y_pos;
@@ -569,11 +649,11 @@ static int EmitRowsYUVA(const VP8LDecoder* const dec,
 
 // Sets io->mb_y, io->mb_h & io->mb_w according to start row, end row and
 // crop options. Also updates the input data pointer, so that it points to the
-// start of the cropped window.
-// Note that 'pixel_stride' is in units of 'uint32_t' (and not 'bytes).
+// start of the cropped window. Note that pixels are in ARGB format even if
+// 'in_data' is uint8_t*.
 // Returns true if the crop window is not empty.
 static int SetCropWindow(VP8Io* const io, int y_start, int y_end,
-                         const uint32_t** const in_data, int pixel_stride) {
+                         uint8_t** const in_data, int pixel_stride) {
   assert(y_start < y_end);
   assert(io->crop_left < io->crop_right);
   if (y_end > io->crop_bottom) {
@@ -582,11 +662,11 @@ static int SetCropWindow(VP8Io* const io, int y_start, int y_end,
   if (y_start < io->crop_top) {
     const int delta = io->crop_top - y_start;
     y_start = io->crop_top;
-    *in_data += pixel_stride * delta;
+    *in_data += delta * pixel_stride;
   }
   if (y_start >= y_end) return 0;  // Crop window is empty.
 
-  *in_data += io->crop_left;
+  *in_data += io->crop_left * sizeof(uint32_t);
 
   io->mb_y = y_start - io->crop_top;
   io->mb_w = io->crop_right - io->crop_left;
@@ -634,10 +714,24 @@ static void ApplyInverseTransforms(VP8LDecoder* const dec, int num_rows,
   }
 }
 
+// Special method for paletted alpha data.
+static void ApplyInverseTransformsAlpha(VP8LDecoder* const dec, int num_rows,
+                                        const uint8_t* const rows) {
+  const int start_row = dec->last_row_;
+  const int end_row = start_row + num_rows;
+  const uint8_t* rows_in = rows;
+  uint8_t* rows_out = (uint8_t*)dec->io_->opaque + dec->io_->width * start_row;
+  VP8LTransform* const transform = &dec->transforms_[0];
+  assert(dec->next_transform_ == 1);
+  assert(transform->type_ == COLOR_INDEXING_TRANSFORM);
+  VP8LColorIndexInverseTransformAlpha(transform, start_row, end_row, rows_in,
+                                      rows_out);
+}
+
 // Processes (transforms, scales & color-converts) the rows decoded after the
 // last call.
 static void ProcessRows(VP8LDecoder* const dec, int row) {
-  const uint32_t* const rows = dec->argb_ + dec->width_ * dec->last_row_;
+  const uint32_t* const rows = dec->pixels_ + dec->width_ * dec->last_row_;
   const int num_rows = row - dec->last_row_;
 
   if (num_rows <= 0) return;  // Nothing to be done.
@@ -646,18 +740,18 @@ static void ProcessRows(VP8LDecoder* const dec, int row) {
   // Emit output.
   {
     VP8Io* const io = dec->io_;
-    const uint32_t* rows_data = dec->argb_cache_;
-    if (!SetCropWindow(io, dec->last_row_, row, &rows_data, io->width)) {
+    uint8_t* rows_data = (uint8_t*)dec->argb_cache_;
+    const int in_stride = io->width * sizeof(uint32_t);  // in unit of RGBA
+    if (!SetCropWindow(io, dec->last_row_, row, &rows_data, in_stride)) {
       // Nothing to output (this time).
     } else {
       const WebPDecBuffer* const output = dec->output_;
-      const int in_stride = io->width * sizeof(*rows_data);
-      if (output->colorspace < MODE_YUV) {  // convert to RGBA
+      if (WebPIsRGBMode(output->colorspace)) {  // convert to RGBA
         const WebPRGBABuffer* const buf = &output->u.RGBA;
         uint8_t* const rgba = buf->rgba + dec->last_out_row_ * buf->stride;
         const int num_rows_out = io->use_scaling ?
-            EmitRescaledRows(dec, rows_data, in_stride, io->mb_h,
-                             rgba, buf->stride) :
+            EmitRescaledRowsRGBA(dec, rows_data, in_stride, io->mb_h,
+                                 rgba, buf->stride) :
             EmitRows(output->colorspace, rows_data, in_stride,
                      io->mb_w, io->mb_h, rgba, buf->stride);
         // Update 'last_out_row_'.
@@ -676,50 +770,317 @@ static void ProcessRows(VP8LDecoder* const dec, int row) {
   assert(dec->last_row_ <= dec->height_);
 }
 
-static int DecodeImageData(VP8LDecoder* const dec,
-                           uint32_t* const data, int width, int height,
-                           ProcessRowsFunc process_func) {
+// Row-processing for the special case when alpha data contains only one
+// transform (color indexing), and trivial non-green literals.
+static int Is8bOptimizable(const VP8LMetadata* const hdr) {
+  int i;
+  if (hdr->color_cache_size_ > 0) return 0;
+  // When the Huffman tree contains only one symbol, we can skip the
+  // call to ReadSymbol() for red/blue/alpha channels.
+  for (i = 0; i < hdr->num_htree_groups_; ++i) {
+    HuffmanCode** const htrees = hdr->htree_groups_[i].htrees;
+    if (htrees[RED][0].bits > 0) return 0;
+    if (htrees[BLUE][0].bits > 0) return 0;
+    if (htrees[ALPHA][0].bits > 0) return 0;
+  }
+  return 1;
+}
+
+static void ExtractPalettedAlphaRows(VP8LDecoder* const dec, int row) {
+  const int num_rows = row - dec->last_row_;
+  const uint8_t* const in =
+      (uint8_t*)dec->pixels_ + dec->width_ * dec->last_row_;
+  if (num_rows > 0) {
+    ApplyInverseTransformsAlpha(dec, num_rows, in);
+  }
+  dec->last_row_ = dec->last_out_row_ = row;
+}
+
+//------------------------------------------------------------------------------
+// Helper functions for fast pattern copy (8b and 32b)
+
+// cyclic rotation of pattern word
+static WEBP_INLINE uint32_t Rotate8b(uint32_t V) {
+#if defined(WORDS_BIGENDIAN)
+  return ((V & 0xff000000u) >> 24) | (V << 8);
+#else
+  return ((V & 0xffu) << 24) | (V >> 8);
+#endif
+}
+
+// copy 1, 2 or 4-bytes pattern
+static WEBP_INLINE void CopySmallPattern8b(const uint8_t* src, uint8_t* dst,
+                                           int length, uint32_t pattern) {
+  int i;
+  // align 'dst' to 4-bytes boundary. Adjust the pattern along the way.
+  while ((uintptr_t)dst & 3) {
+    *dst++ = *src++;
+    pattern = Rotate8b(pattern);
+    --length;
+  }
+  // Copy the pattern 4 bytes at a time.
+  for (i = 0; i < (length >> 2); ++i) {
+    ((uint32_t*)dst)[i] = pattern;
+  }
+  // Finish with left-overs. 'pattern' is still correctly positioned,
+  // so no Rotate8b() call is needed.
+  for (i <<= 2; i < length; ++i) {
+    dst[i] = src[i];
+  }
+}
+
+static WEBP_INLINE void CopyBlock8b(uint8_t* const dst, int dist, int length) {
+  const uint8_t* src = dst - dist;
+  if (length >= 8) {
+    uint32_t pattern = 0;
+    switch (dist) {
+      case 1:
+        pattern = src[0];
+#if defined(__arm__) || defined(_M_ARM)   // arm doesn't like multiply that much
+        pattern |= pattern << 8;
+        pattern |= pattern << 16;
+#elif defined(WEBP_USE_MIPS_DSP_R2)
+        __asm__ volatile ("replv.qb %0, %0" : "+r"(pattern));
+#else
+        pattern = 0x01010101u * pattern;
+#endif
+        break;
+      case 2:
+        memcpy(&pattern, src, sizeof(uint16_t));
+#if defined(__arm__) || defined(_M_ARM)
+        pattern |= pattern << 16;
+#elif defined(WEBP_USE_MIPS_DSP_R2)
+        __asm__ volatile ("replv.ph %0, %0" : "+r"(pattern));
+#else
+        pattern = 0x00010001u * pattern;
+#endif
+        break;
+      case 4:
+        memcpy(&pattern, src, sizeof(uint32_t));
+        break;
+      default:
+        goto Copy;
+        break;
+    }
+    CopySmallPattern8b(src, dst, length, pattern);
+    return;
+  }
+ Copy:
+  if (dist >= length) {  // no overlap -> use memcpy()
+    memcpy(dst, src, length * sizeof(*dst));
+  } else {
+    int i;
+    for (i = 0; i < length; ++i) dst[i] = src[i];
+  }
+}
+
+// copy pattern of 1 or 2 uint32_t's
+static WEBP_INLINE void CopySmallPattern32b(const uint32_t* src,
+                                            uint32_t* dst,
+                                            int length, uint64_t pattern) {
+  int i;
+  if ((uintptr_t)dst & 4) {           // Align 'dst' to 8-bytes boundary.
+    *dst++ = *src++;
+    pattern = (pattern >> 32) | (pattern << 32);
+    --length;
+  }
+  assert(0 == ((uintptr_t)dst & 7));
+  for (i = 0; i < (length >> 1); ++i) {
+    ((uint64_t*)dst)[i] = pattern;    // Copy the pattern 8 bytes at a time.
+  }
+  if (length & 1) {                   // Finish with left-over.
+    dst[i << 1] = src[i << 1];
+  }
+}
+
+static WEBP_INLINE void CopyBlock32b(uint32_t* const dst,
+                                     int dist, int length) {
+  const uint32_t* const src = dst - dist;
+  if (dist <= 2 && length >= 4 && ((uintptr_t)dst & 3) == 0) {
+    uint64_t pattern;
+    if (dist == 1) {
+      pattern = (uint64_t)src[0];
+      pattern |= pattern << 32;
+    } else {
+      memcpy(&pattern, src, sizeof(pattern));
+    }
+    CopySmallPattern32b(src, dst, length, pattern);
+  } else if (dist >= length) {  // no overlap
+    memcpy(dst, src, length * sizeof(*dst));
+  } else {
+    int i;
+    for (i = 0; i < length; ++i) dst[i] = src[i];
+  }
+}
+
+//------------------------------------------------------------------------------
+
+static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
+                           int width, int height, int last_row) {
   int ok = 1;
-  int col = 0, row = 0;
+  int row = dec->last_pixel_ / width;
+  int col = dec->last_pixel_ % width;
   VP8LBitReader* const br = &dec->br_;
   VP8LMetadata* const hdr = &dec->hdr_;
-  HTreeGroup* htree_group = hdr->htree_groups_;
-  uint32_t* src = data;
-  uint32_t* last_cached = data;
-  uint32_t* const src_end = data + width * height;
+  const HTreeGroup* htree_group = GetHtreeGroupForPos(hdr, col, row);
+  int pos = dec->last_pixel_;         // current position
+  const int end = width * height;     // End of data
+  const int last = width * last_row;  // Last pixel to decode
   const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES;
-  const int color_cache_limit = len_code_limit + hdr->color_cache_size_;
-  VP8LColorCache* const color_cache =
-      (hdr->color_cache_size_ > 0) ? &hdr->color_cache_ : NULL;
   const int mask = hdr->huffman_mask_;
-
   assert(htree_group != NULL);
+  assert(pos < end);
+  assert(last_row <= height);
+  assert(Is8bOptimizable(hdr));
 
-  while (!br->eos_ && src < src_end) {
+  while (!br->eos_ && pos < last) {
     int code;
-    // Only update when changing tile. Note we could use the following test:
-    //   if "((((prev_col ^ col) | prev_row ^ row)) > mask)" -> tile changed
-    // but that's actually slower and requires storing the previous col/row
+    // Only update when changing tile.
     if ((col & mask) == 0) {
       htree_group = GetHtreeGroupForPos(hdr, col, row);
     }
     VP8LFillBitWindow(br);
-    code = ReadSymbol(&htree_group->htrees_[GREEN], br);
-    if (code < NUM_LITERAL_CODES) {   // Literal.
-      int red, green, blue, alpha;
-      red = ReadSymbol(&htree_group->htrees_[RED], br);
-      green = code;
+    code = ReadSymbol(htree_group->htrees[GREEN], br);
+    if (code < NUM_LITERAL_CODES) {  // Literal
+      data[pos] = code;
+      ++pos;
+      ++col;
+      if (col >= width) {
+        col = 0;
+        ++row;
+        if (row % NUM_ARGB_CACHE_ROWS == 0) {
+          ExtractPalettedAlphaRows(dec, row);
+        }
+      }
+    } else if (code < len_code_limit) {  // Backward reference
+      int dist_code, dist;
+      const int length_sym = code - NUM_LITERAL_CODES;
+      const int length = GetCopyLength(length_sym, br);
+      const int dist_symbol = ReadSymbol(htree_group->htrees[DIST], br);
       VP8LFillBitWindow(br);
-      blue = ReadSymbol(&htree_group->htrees_[BLUE], br);
-      alpha = ReadSymbol(&htree_group->htrees_[ALPHA], br);
-      *src = (alpha << 24) + (red << 16) + (green << 8) + blue;
- AdvanceByOne:
+      dist_code = GetCopyDistance(dist_symbol, br);
+      dist = PlaneCodeToDistance(width, dist_code);
+      if (pos >= dist && end - pos >= length) {
+        CopyBlock8b(data + pos, dist, length);
+      } else {
+        ok = 0;
+        goto End;
+      }
+      pos += length;
+      col += length;
+      while (col >= width) {
+        col -= width;
+        ++row;
+        if (row % NUM_ARGB_CACHE_ROWS == 0) {
+          ExtractPalettedAlphaRows(dec, row);
+        }
+      }
+      if (pos < last && (col & mask)) {
+        htree_group = GetHtreeGroupForPos(hdr, col, row);
+      }
+    } else {  // Not reached
+      ok = 0;
+      goto End;
+    }
+    assert(br->eos_ == VP8LIsEndOfStream(br));
+  }
+  // Process the remaining rows corresponding to last row-block.
+  ExtractPalettedAlphaRows(dec, row);
+
+ End:
+  if (!ok || (br->eos_ && pos < end)) {
+    ok = 0;
+    dec->status_ = br->eos_ ? VP8_STATUS_SUSPENDED
+                            : VP8_STATUS_BITSTREAM_ERROR;
+  } else {
+    dec->last_pixel_ = pos;
+  }
+  return ok;
+}
+
+static void SaveState(VP8LDecoder* const dec, int last_pixel) {
+  assert(dec->incremental_);
+  dec->saved_br_ = dec->br_;
+  dec->saved_last_pixel_ = last_pixel;
+  if (dec->hdr_.color_cache_size_ > 0) {
+    VP8LColorCacheCopy(&dec->hdr_.color_cache_, &dec->hdr_.saved_color_cache_);
+  }
+}
+
+static void RestoreState(VP8LDecoder* const dec) {
+  assert(dec->br_.eos_);
+  dec->status_ = VP8_STATUS_SUSPENDED;
+  dec->br_ = dec->saved_br_;
+  dec->last_pixel_ = dec->saved_last_pixel_;
+  if (dec->hdr_.color_cache_size_ > 0) {
+    VP8LColorCacheCopy(&dec->hdr_.saved_color_cache_, &dec->hdr_.color_cache_);
+  }
+}
+
+#define SYNC_EVERY_N_ROWS 8  // minimum number of rows between check-points
+static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
+                           int width, int height, int last_row,
+                           ProcessRowsFunc process_func) {
+  int row = dec->last_pixel_ / width;
+  int col = dec->last_pixel_ % width;
+  VP8LBitReader* const br = &dec->br_;
+  VP8LMetadata* const hdr = &dec->hdr_;
+  HTreeGroup* htree_group = GetHtreeGroupForPos(hdr, col, row);
+  uint32_t* src = data + dec->last_pixel_;
+  uint32_t* last_cached = src;
+  uint32_t* const src_end = data + width * height;     // End of data
+  uint32_t* const src_last = data + width * last_row;  // Last pixel to decode
+  const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES;
+  const int color_cache_limit = len_code_limit + hdr->color_cache_size_;
+  int next_sync_row = dec->incremental_ ? row : 1 << 24;
+  VP8LColorCache* const color_cache =
+      (hdr->color_cache_size_ > 0) ? &hdr->color_cache_ : NULL;
+  const int mask = hdr->huffman_mask_;
+  assert(htree_group != NULL);
+  assert(src < src_end);
+  assert(src_last <= src_end);
+
+  while (src < src_last) {
+    int code;
+    if (row >= next_sync_row) {
+      SaveState(dec, (int)(src - data));
+      next_sync_row = row + SYNC_EVERY_N_ROWS;
+    }
+    // Only update when changing tile. Note we could use this test:
+    // if "((((prev_col ^ col) | prev_row ^ row)) > mask)" -> tile changed
+    // but that's actually slower and needs storing the previous col/row.
+    if ((col & mask) == 0) htree_group = GetHtreeGroupForPos(hdr, col, row);
+    if (htree_group->is_trivial_code) {
+      *src = htree_group->literal_arb;
+      goto AdvanceByOne;
+    }
+    VP8LFillBitWindow(br);
+    if (htree_group->use_packed_table) {
+      code = ReadPackedSymbols(htree_group, br, src);
+      if (code == PACKED_NON_LITERAL_CODE) goto AdvanceByOne;
+    } else {
+      code = ReadSymbol(htree_group->htrees[GREEN], br);
+    }
+    if (br->eos_) break;  // early out
+    if (code < NUM_LITERAL_CODES) {  // Literal
+      if (htree_group->is_trivial_literal) {
+        *src = htree_group->literal_arb | (code << 8);
+      } else {
+        int red, blue, alpha;
+        red = ReadSymbol(htree_group->htrees[RED], br);
+        VP8LFillBitWindow(br);
+        blue = ReadSymbol(htree_group->htrees[BLUE], br);
+        alpha = ReadSymbol(htree_group->htrees[ALPHA], br);
+        if (br->eos_) break;
+        *src = ((uint32_t)alpha << 24) | (red << 16) | (code << 8) | blue;
+      }
+    AdvanceByOne:
       ++src;
       ++col;
       if (col >= width) {
         col = 0;
         ++row;
-        if ((process_func != NULL) && (row % NUM_ARGB_CACHE_ROWS == 0)) {
+        if ((row % NUM_ARGB_CACHE_ROWS == 0) && (process_func != NULL)) {
           process_func(dec, row);
         }
         if (color_cache != NULL) {
@@ -728,40 +1089,39 @@ static int DecodeImageData(VP8LDecoder* const dec,
           }
         }
       }
-    } else if (code < len_code_limit) {           // Backward reference
+    } else if (code < len_code_limit) {  // Backward reference
       int dist_code, dist;
       const int length_sym = code - NUM_LITERAL_CODES;
       const int length = GetCopyLength(length_sym, br);
-      const int dist_symbol = ReadSymbol(&htree_group->htrees_[DIST], br);
+      const int dist_symbol = ReadSymbol(htree_group->htrees[DIST], br);
       VP8LFillBitWindow(br);
       dist_code = GetCopyDistance(dist_symbol, br);
       dist = PlaneCodeToDistance(width, dist_code);
-      if (src - data < dist || src_end - src < length) {
-        ok = 0;
-        goto End;
-      }
-      {
-        int i;
-        for (i = 0; i < length; ++i) src[i] = src[i - dist];
-        src += length;
+      if (br->eos_) break;
+      if (src - data < (ptrdiff_t)dist || src_end - src < (ptrdiff_t)length) {
+        goto Error;
+      } else {
+        CopyBlock32b(src, dist, length);
       }
+      src += length;
       col += length;
       while (col >= width) {
         col -= width;
         ++row;
-        if ((process_func != NULL) && (row % NUM_ARGB_CACHE_ROWS == 0)) {
+        if ((row % NUM_ARGB_CACHE_ROWS == 0) && (process_func != NULL)) {
           process_func(dec, row);
         }
       }
-      if (src < src_end) {
-        htree_group = GetHtreeGroupForPos(hdr, col, row);
-        if (color_cache != NULL) {
-          while (last_cached < src) {
-            VP8LColorCacheInsert(color_cache, *last_cached++);
-          }
+      // Because of the check done above (before 'src' was incremented by
+      // 'length'), the following holds true.
+      assert(src <= src_end);
+      if (col & mask) htree_group = GetHtreeGroupForPos(hdr, col, row);
+      if (color_cache != NULL) {
+        while (last_cached < src) {
+          VP8LColorCacheInsert(color_cache, *last_cached++);
         }
       }
-    } else if (code < color_cache_limit) {    // Color cache.
+    } else if (code < color_cache_limit) {  // Color cache
       const int key = code - len_code_limit;
       assert(color_cache != NULL);
       while (last_cached < src) {
@@ -769,33 +1129,38 @@ static int DecodeImageData(VP8LDecoder* const dec,
       }
       *src = VP8LColorCacheLookup(color_cache, key);
       goto AdvanceByOne;
-    } else {    // Not reached.
-      ok = 0;
-      goto End;
+    } else {  // Not reached
+      goto Error;
     }
-    ok = !br->error_;
-    if (!ok) goto End;
+    assert(br->eos_ == VP8LIsEndOfStream(br));
   }
-  // Process the remaining rows corresponding to last row-block.
-  if (process_func != NULL) process_func(dec, row);
 
- End:
-  if (br->error_ || !ok || (br->eos_ && src < src_end)) {
-    ok = 0;
-    dec->status_ = (!br->eos_) ?
-        VP8_STATUS_BITSTREAM_ERROR : VP8_STATUS_SUSPENDED;
-  } else if (src == src_end) {
-    dec->state_ = READ_DATA;
+  if (dec->incremental_ && br->eos_ && src < src_end) {
+    RestoreState(dec);
+  } else if (!br->eos_) {
+    // Process the remaining rows corresponding to last row-block.
+    if (process_func != NULL) {
+      process_func(dec, row);
+    }
+    dec->status_ = VP8_STATUS_OK;
+    dec->last_pixel_ = (int)(src - data);  // end-of-scan marker
+  } else {
+    // if not incremental, and we are past the end of buffer (eos_=1), then this
+    // is a real bitstream error.
+    goto Error;
   }
+  return 1;
 
-  return ok;
+ Error:
+  dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+  return 0;
 }
 
 // -----------------------------------------------------------------------------
 // VP8LTransform
 
 static void ClearTransform(VP8LTransform* const transform) {
-  free(transform->data_);
+  WebPSafeFree(transform->data_);
   transform->data_ = NULL;
 }
 
@@ -819,7 +1184,7 @@ static int ExpandColorMap(int num_colors, VP8LTransform* const transform) {
     }
     for (; i < 4 * final_num_colors; ++i)
       new_data[i] = 0;  // black tail.
-    free(transform->data_);
+    WebPSafeFree(transform->data_);
     transform->data_ = new_color_map;
   }
   return 1;
@@ -882,16 +1247,18 @@ static int ReadTransform(int* const xsize, int const* ysize,
 // VP8LMetadata
 
 static void InitMetadata(VP8LMetadata* const hdr) {
-  assert(hdr);
+  assert(hdr != NULL);
   memset(hdr, 0, sizeof(*hdr));
 }
 
 static void ClearMetadata(VP8LMetadata* const hdr) {
-  assert(hdr);
+  assert(hdr != NULL);
 
-  free(hdr->huffman_image_);
-  DeleteHtreeGroups(hdr->htree_groups_, hdr->num_htree_groups_);
+  WebPSafeFree(hdr->huffman_image_);
+  WebPSafeFree(hdr->huffman_tables_);
+  VP8LHtreeGroupsFree(hdr->htree_groups_);
   VP8LColorCacheClear(&hdr->color_cache_);
+  VP8LColorCacheClear(&hdr->saved_color_cache_);
   InitMetadata(hdr);
 }
 
@@ -899,11 +1266,13 @@ static void ClearMetadata(VP8LMetadata* const hdr) {
 // VP8LDecoder
 
 VP8LDecoder* VP8LNew(void) {
-  VP8LDecoder* const dec = (VP8LDecoder*)calloc(1, sizeof(*dec));
+  VP8LDecoder* const dec = (VP8LDecoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
   if (dec == NULL) return NULL;
   dec->status_ = VP8_STATUS_OK;
-  dec->action_ = READ_DIM;
   dec->state_ = READ_DIM;
+
+  VP8LDspInit();  // Init critical function pointers.
+
   return dec;
 }
 
@@ -912,15 +1281,15 @@ void VP8LClear(VP8LDecoder* const dec) {
   if (dec == NULL) return;
   ClearMetadata(&dec->hdr_);
 
-  free(dec->argb_);
-  dec->argb_ = NULL;
+  WebPSafeFree(dec->pixels_);
+  dec->pixels_ = NULL;
   for (i = 0; i < dec->next_transform_; ++i) {
     ClearTransform(&dec->transforms_[i]);
   }
   dec->next_transform_ = 0;
   dec->transforms_seen_ = 0;
 
-  free(dec->rescaler_memory);
+  WebPSafeFree(dec->rescaler_memory);
   dec->rescaler_memory = NULL;
 
   dec->output_ = NULL;   // leave no trace behind
@@ -929,7 +1298,7 @@ void VP8LClear(VP8LDecoder* const dec) {
 void VP8LDelete(VP8LDecoder* const dec) {
   if (dec != NULL) {
     VP8LClear(dec);
-    free(dec);
+    WebPSafeFree(dec);
   }
 }
 
@@ -1009,19 +1378,14 @@ static int DecodeImageStream(int xsize, int ysize,
   }
 
   // Use the Huffman trees to decode the LZ77 encoded data.
-  ok = DecodeImageData(dec, data, transform_xsize, transform_ysize, NULL);
-  ok = ok && !br->error_;
+  ok = DecodeImageData(dec, data, transform_xsize, transform_ysize,
+                       transform_ysize, NULL);
+  ok = ok && !br->eos_;
 
  End:
-
   if (!ok) {
-    free(data);
+    WebPSafeFree(data);
     ClearMetadata(hdr);
-    // If not enough data (br.eos_) resulted in BIT_STREAM_ERROR, update the
-    // status appropriately.
-    if (dec->status_ == VP8_STATUS_BITSTREAM_ERROR && dec->br_.eos_) {
-      dec->status_ = VP8_STATUS_SUSPENDED;
-    }
   } else {
     if (decoded_data != NULL) {
       *decoded_data = data;
@@ -1031,41 +1395,52 @@ static int DecodeImageStream(int xsize, int ysize,
       assert(data == NULL);
       assert(is_level0);
     }
+    dec->last_pixel_ = 0;  // Reset for future DECODE_DATA_FUNC() calls.
     if (!is_level0) ClearMetadata(hdr);  // Clean up temporary data behind.
   }
   return ok;
 }
 
 //------------------------------------------------------------------------------
-// Allocate dec->argb_ and dec->argb_cache_ using dec->width_ and dec->height_
-
-static int AllocateARGBBuffers(VP8LDecoder* const dec, int final_width) {
+// Allocate internal buffers dec->pixels_ and dec->argb_cache_.
+static int AllocateInternalBuffers32b(VP8LDecoder* const dec, int final_width) {
   const uint64_t num_pixels = (uint64_t)dec->width_ * dec->height_;
   // Scratch buffer corresponding to top-prediction row for transforming the
-  // first row in the row-blocks.
-  const uint64_t cache_top_pixels = final_width;
-  // Scratch buffer for temporary BGRA storage.
+  // first row in the row-blocks. Not needed for paletted alpha.
+  const uint64_t cache_top_pixels = (uint16_t)final_width;
+  // Scratch buffer for temporary BGRA storage. Not needed for paletted alpha.
   const uint64_t cache_pixels = (uint64_t)final_width * NUM_ARGB_CACHE_ROWS;
   const uint64_t total_num_pixels =
       num_pixels + cache_top_pixels + cache_pixels;
 
   assert(dec->width_ <= final_width);
-  dec->argb_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(*dec->argb_));
-  if (dec->argb_ == NULL) {
+  dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint32_t));
+  if (dec->pixels_ == NULL) {
     dec->argb_cache_ = NULL;    // for sanity check
     dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
     return 0;
   }
-  dec->argb_cache_ = dec->argb_ + num_pixels + cache_top_pixels;
+  dec->argb_cache_ = dec->pixels_ + num_pixels + cache_top_pixels;
+  return 1;
+}
+
+static int AllocateInternalBuffers8b(VP8LDecoder* const dec) {
+  const uint64_t total_num_pixels = (uint64_t)dec->width_ * dec->height_;
+  dec->argb_cache_ = NULL;    // for sanity check
+  dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint8_t));
+  if (dec->pixels_ == NULL) {
+    dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+    return 0;
+  }
   return 1;
 }
 
 //------------------------------------------------------------------------------
-// Special row-processing that only stores the alpha data.
 
+// Special row-processing that only stores the alpha data.
 static void ExtractAlphaRows(VP8LDecoder* const dec, int row) {
   const int num_rows = row - dec->last_row_;
-  const uint32_t* const in = dec->argb_ + dec->width_ * dec->last_row_;
+  const uint32_t* const in = dec->pixels_ + dec->width_ * dec->last_row_;
 
   if (num_rows <= 0) return;  // Nothing to be done.
   ApplyInverseTransforms(dec, num_rows, in);
@@ -1079,44 +1454,77 @@ static void ExtractAlphaRows(VP8LDecoder* const dec, int row) {
     int i;
     for (i = 0; i < cache_pixs; ++i) dst[i] = (src[i] >> 8) & 0xff;
   }
-
   dec->last_row_ = dec->last_out_row_ = row;
 }
 
-int VP8LDecodeAlphaImageStream(int width, int height, const uint8_t* const data,
-                               size_t data_size, uint8_t* const output) {
-  VP8Io io;
+int VP8LDecodeAlphaHeader(ALPHDecoder* const alph_dec,
+                          const uint8_t* const data, size_t data_size,
+                          uint8_t* const output) {
   int ok = 0;
-  VP8LDecoder* const dec = VP8LNew();
-  if (dec == NULL) return 0;
-
-  dec->width_ = width;
-  dec->height_ = height;
-  dec->io_ = &io;
+  VP8LDecoder* dec;
+  VP8Io* io;
+  assert(alph_dec != NULL);
+  alph_dec->vp8l_dec_ = VP8LNew();
+  if (alph_dec->vp8l_dec_ == NULL) return 0;
+  dec = alph_dec->vp8l_dec_;
+
+  dec->width_ = alph_dec->width_;
+  dec->height_ = alph_dec->height_;
+  dec->io_ = &alph_dec->io_;
+  io = dec->io_;
 
-  VP8InitIo(&io);
-  WebPInitCustomIo(NULL, &io);    // Just a sanity Init. io won't be used.
-  io.opaque = output;
-  io.width = width;
-  io.height = height;
+  VP8InitIo(io);
+  WebPInitCustomIo(NULL, io);  // Just a sanity Init. io won't be used.
+  io->opaque = output;
+  io->width = alph_dec->width_;
+  io->height = alph_dec->height_;
 
   dec->status_ = VP8_STATUS_OK;
   VP8LInitBitReader(&dec->br_, data, data_size);
 
-  dec->action_ = READ_HDR;
-  if (!DecodeImageStream(width, height, 1, dec, NULL)) goto Err;
+  if (!DecodeImageStream(alph_dec->width_, alph_dec->height_, 1, dec, NULL)) {
+    goto Err;
+  }
 
-  // Allocate output (note that dec->width_ may have changed here).
-  if (!AllocateARGBBuffers(dec, width)) goto Err;
+  // Special case: if alpha data uses only the color indexing transform and
+  // doesn't use color cache (a frequent case), we will use DecodeAlphaData()
+  // method that only needs allocation of 1 byte per pixel (alpha channel).
+  if (dec->next_transform_ == 1 &&
+      dec->transforms_[0].type_ == COLOR_INDEXING_TRANSFORM &&
+      Is8bOptimizable(&dec->hdr_)) {
+    alph_dec->use_8b_decode = 1;
+    ok = AllocateInternalBuffers8b(dec);
+  } else {
+    // Allocate internal buffers (note that dec->width_ may have changed here).
+    alph_dec->use_8b_decode = 0;
+    ok = AllocateInternalBuffers32b(dec, alph_dec->width_);
+  }
 
-  // Decode (with special row processing).
-  dec->action_ = READ_DATA;
-  ok = DecodeImageData(dec, dec->argb_, dec->width_, dec->height_,
-                       ExtractAlphaRows);
+  if (!ok) goto Err;
+
+  return 1;
 
  Err:
-  VP8LDelete(dec);
-  return ok;
+  VP8LDelete(alph_dec->vp8l_dec_);
+  alph_dec->vp8l_dec_ = NULL;
+  return 0;
+}
+
+int VP8LDecodeAlphaImageStream(ALPHDecoder* const alph_dec, int last_row) {
+  VP8LDecoder* const dec = alph_dec->vp8l_dec_;
+  assert(dec != NULL);
+  assert(last_row <= dec->height_);
+
+  if (dec->last_pixel_ == dec->width_ * dec->height_) {
+    return 1;  // done
+  }
+
+  // Decode (with special row processing).
+  return alph_dec->use_8b_decode ?
+      DecodeAlphaData(dec, (uint8_t*)dec->pixels_, dec->width_, dec->height_,
+                      last_row) :
+      DecodeImageData(dec, dec->pixels_, dec->width_, dec->height_,
+                      last_row, ExtractAlphaRows);
 }
 
 //------------------------------------------------------------------------------
@@ -1141,14 +1549,13 @@ int VP8LDecodeHeader(VP8LDecoder* const dec, VP8Io* const io) {
   io->width = width;
   io->height = height;
 
-  dec->action_ = READ_HDR;
   if (!DecodeImageStream(width, height, 1, dec, NULL)) goto Error;
   return 1;
 
  Error:
-   VP8LClear(dec);
-   assert(dec->status_ != VP8_STATUS_OK);
-   return 0;
+  VP8LClear(dec);
+  assert(dec->status_ != VP8_STATUS_OK);
+  return 0;
 }
 
 int VP8LDecodeImage(VP8LDecoder* const dec) {
@@ -1158,33 +1565,57 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
   // Sanity checks.
   if (dec == NULL) return 0;
 
+  assert(dec->hdr_.huffman_tables_ != NULL);
+  assert(dec->hdr_.htree_groups_ != NULL);
+  assert(dec->hdr_.num_htree_groups_ > 0);
+
   io = dec->io_;
   assert(io != NULL);
   params = (WebPDecParams*)io->opaque;
   assert(params != NULL);
-  dec->output_ = params->output;
-  assert(dec->output_ != NULL);
 
   // Initialization.
-  if (!WebPIoInitFromOptions(params->options, io, MODE_BGRA)) {
-    dec->status_ = VP8_STATUS_INVALID_PARAM;
-    goto Err;
-  }
+  if (dec->state_ != READ_DATA) {
+    dec->output_ = params->output;
+    assert(dec->output_ != NULL);
+
+    if (!WebPIoInitFromOptions(params->options, io, MODE_BGRA)) {
+      dec->status_ = VP8_STATUS_INVALID_PARAM;
+      goto Err;
+    }
 
-  if (!AllocateARGBBuffers(dec, io->width)) goto Err;
+    if (!AllocateInternalBuffers32b(dec, io->width)) goto Err;
 
-  if (io->use_scaling && !AllocateAndInitRescaler(dec, io)) goto Err;
+    if (io->use_scaling && !AllocateAndInitRescaler(dec, io)) goto Err;
+
+    if (io->use_scaling || WebPIsPremultipliedMode(dec->output_->colorspace)) {
+      // need the alpha-multiply functions for premultiplied output or rescaling
+      WebPInitAlphaProcessing();
+    }
+    if (!WebPIsRGBMode(dec->output_->colorspace)) {
+      WebPInitConvertARGBToYUV();
+      if (dec->output_->u.YUVA.a != NULL) WebPInitAlphaProcessing();
+    }
+    if (dec->incremental_) {
+      if (dec->hdr_.color_cache_size_ > 0 &&
+          dec->hdr_.saved_color_cache_.colors_ == NULL) {
+        if (!VP8LColorCacheInit(&dec->hdr_.saved_color_cache_,
+                                dec->hdr_.color_cache_.hash_bits_)) {
+          dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+          goto Err;
+        }
+      }
+    }
+    dec->state_ = READ_DATA;
+  }
 
   // Decode.
-  dec->action_ = READ_DATA;
-  if (!DecodeImageData(dec, dec->argb_, dec->width_, dec->height_,
-                       ProcessRows)) {
+  if (!DecodeImageData(dec, dec->pixels_, dec->width_, dec->height_,
+                       dec->height_, ProcessRows)) {
     goto Err;
   }
 
-  // Cleanup.
   params->last_y = dec->last_out_row_;
-  VP8LClear(dec);
   return 1;
 
  Err:
@@ -1194,7 +1625,3 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
 }
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/vp8li.h b/drivers/webp/dec/vp8li.h
index 5f6cd6a01c..8886e47f62 100644
--- a/drivers/webp/dec/vp8li.h
+++ b/drivers/webp/dec/vp8li.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 decoder: internal header.
@@ -18,9 +20,8 @@
 #include "../utils/bit_reader.h"
 #include "../utils/color_cache.h"
 #include "../utils/huffman.h"
-#include "../format_constants.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -40,12 +41,9 @@ struct VP8LTransform {
 };
 
 typedef struct {
-  HuffmanTree htrees_[HUFFMAN_CODES_PER_META_CODE];
-} HTreeGroup;
-
-typedef struct {
   int             color_cache_size_;
   VP8LColorCache  color_cache_;
+  VP8LColorCache  saved_color_cache_;  // for incremental
 
   int             huffman_mask_;
   int             huffman_subsample_bits_;
@@ -53,24 +51,32 @@ typedef struct {
   uint32_t       *huffman_image_;
   int             num_htree_groups_;
   HTreeGroup     *htree_groups_;
+  HuffmanCode    *huffman_tables_;
 } VP8LMetadata;
 
-typedef struct {
+typedef struct VP8LDecoder VP8LDecoder;
+struct VP8LDecoder {
   VP8StatusCode    status_;
-  VP8LDecodeState  action_;
   VP8LDecodeState  state_;
   VP8Io           *io_;
 
   const WebPDecBuffer *output_;    // shortcut to io->opaque->output
 
-  uint32_t        *argb_;          // Internal data: always in BGRA color mode.
+  uint32_t        *pixels_;        // Internal data: either uint8_t* for alpha
+                                   // or uint32_t* for BGRA.
   uint32_t        *argb_cache_;    // Scratch buffer for temporary BGRA storage.
 
   VP8LBitReader    br_;
+  int              incremental_;   // if true, incremental decoding is expected
+  VP8LBitReader    saved_br_;      // note: could be local variables too
+  int              saved_last_pixel_;
 
   int              width_;
   int              height_;
   int              last_row_;      // last input row decoded so far.
+  int              last_pixel_;    // last pixel decoded so far. However, it may
+                                   // not be transformed, scaled and
+                                   // color-converted yet.
   int              last_out_row_;  // last row output so far.
 
   VP8LMetadata     hdr_;
@@ -82,18 +88,27 @@ typedef struct {
 
   uint8_t         *rescaler_memory;  // Working memory for rescaling work.
   WebPRescaler    *rescaler;         // Common rescaler for all channels.
-} VP8LDecoder;
+};
 
 //------------------------------------------------------------------------------
 // internal functions. Not public.
 
+struct ALPHDecoder;  // Defined in dec/alphai.h.
+
 // in vp8l.c
 
-// Decodes a raw image stream (without header) and store the alpha data
-// into *output, which must be of size width x height. Returns false in case
-// of error.
-int VP8LDecodeAlphaImageStream(int width, int height, const uint8_t* const data,
-                               size_t data_size, uint8_t* const output);
+// Decodes image header for alpha data stored using lossless compression.
+// Returns false in case of error.
+int VP8LDecodeAlphaHeader(struct ALPHDecoder* const alph_dec,
+                          const uint8_t* const data, size_t data_size,
+                          uint8_t* const output);
+
+// Decodes *at least* 'last_row' rows of alpha. If some of the initial rows are
+// already decoded in previous call(s), it will resume decoding from where it
+// was paused.
+// Returns false in case of bitstream error.
+int VP8LDecodeAlphaImageStream(struct ALPHDecoder* const alph_dec,
+                               int last_row);
 
 // Allocates and initialize a new lossless decoder instance.
 VP8LDecoder* VP8LNew(void);
@@ -114,7 +129,7 @@ void VP8LDelete(VP8LDecoder* const dec);
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/dec/webp.c b/drivers/webp/dec/webp.c
index f44bc2b8ae..952178fa89 100644
--- a/drivers/webp/dec/webp.c
+++ b/drivers/webp/dec/webp.c
@@ -1,8 +1,10 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 decoding functions for WEBP images.
@@ -14,11 +16,8 @@
 #include "./vp8i.h"
 #include "./vp8li.h"
 #include "./webpi.h"
-#include "../format_constants.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "../utils/utils.h"
+#include "../webp/mux_types.h"  // ALPHA_FLAG
 
 //------------------------------------------------------------------------------
 // RIFF layout is:
@@ -40,27 +39,20 @@ extern "C" {
 //   20..23  VP8X flags bit-map corresponding to the chunk-types present.
 //   24..26  Width of the Canvas Image.
 //   27..29  Height of the Canvas Image.
-// There can be extra chunks after the "VP8X" chunk (ICCP, TILE, FRM, VP8,
-// META  ...)
+// There can be extra chunks after the "VP8X" chunk (ICCP, FRGM, ANMF, VP8,
+// VP8L, XMP, EXIF  ...)
 // All sizes are in little-endian order.
 // Note: chunk data size must be padded to multiple of 2 when written.
 
-static WEBP_INLINE uint32_t get_le24(const uint8_t* const data) {
-  return data[0] | (data[1] << 8) | (data[2] << 16);
-}
-
-static WEBP_INLINE uint32_t get_le32(const uint8_t* const data) {
-  return (uint32_t)get_le24(data) | (data[3] << 24);
-}
-
 // Validates the RIFF container (if detected) and skips over it.
-// If a RIFF container is detected,
-// Returns VP8_STATUS_BITSTREAM_ERROR for invalid header, and
-//         VP8_STATUS_OK otherwise.
+// If a RIFF container is detected, returns:
+//     VP8_STATUS_BITSTREAM_ERROR for invalid header,
+//     VP8_STATUS_NOT_ENOUGH_DATA for truncated data if have_all_data is true,
+// and VP8_STATUS_OK otherwise.
 // In case there are not enough bytes (partial RIFF container), return 0 for
 // *riff_size. Else return the RIFF size extracted from the header.
 static VP8StatusCode ParseRIFF(const uint8_t** const data,
-                               size_t* const data_size,
+                               size_t* const data_size, int have_all_data,
                                size_t* const riff_size) {
   assert(data != NULL);
   assert(data_size != NULL);
@@ -71,11 +63,17 @@ static VP8StatusCode ParseRIFF(const uint8_t** const data,
     if (memcmp(*data + 8, "WEBP", TAG_SIZE)) {
       return VP8_STATUS_BITSTREAM_ERROR;  // Wrong image file signature.
     } else {
-      const uint32_t size = get_le32(*data + TAG_SIZE);
+      const uint32_t size = GetLE32(*data + TAG_SIZE);
       // Check that we have at least one chunk (i.e "WEBP" + "VP8?nnnn").
       if (size < TAG_SIZE + CHUNK_HEADER_SIZE) {
         return VP8_STATUS_BITSTREAM_ERROR;
       }
+      if (size > MAX_CHUNK_PAYLOAD) {
+        return VP8_STATUS_BITSTREAM_ERROR;
+      }
+      if (have_all_data && (size > *data_size - CHUNK_HEADER_SIZE)) {
+        return VP8_STATUS_NOT_ENOUGH_DATA;  // Truncated bitstream.
+      }
       // We have a RIFF container. Skip it.
       *riff_size = size;
       *data += RIFF_HEADER_SIZE;
@@ -111,7 +109,7 @@ static VP8StatusCode ParseVP8X(const uint8_t** const data,
   if (!memcmp(*data, "VP8X", TAG_SIZE)) {
     int width, height;
     uint32_t flags;
-    const uint32_t chunk_size = get_le32(*data + TAG_SIZE);
+    const uint32_t chunk_size = GetLE32(*data + TAG_SIZE);
     if (chunk_size != VP8X_CHUNK_SIZE) {
       return VP8_STATUS_BITSTREAM_ERROR;  // Wrong chunk size.
     }
@@ -120,9 +118,9 @@ static VP8StatusCode ParseVP8X(const uint8_t** const data,
     if (*data_size < vp8x_size) {
       return VP8_STATUS_NOT_ENOUGH_DATA;  // Insufficient data.
     }
-    flags = get_le32(*data + 8);
-    width = 1 + get_le24(*data + 12);
-    height = 1 + get_le24(*data + 15);
+    flags = GetLE32(*data + 8);
+    width = 1 + GetLE24(*data + 12);
+    height = 1 + GetLE24(*data + 15);
     if (width * (uint64_t)height >= MAX_IMAGE_AREA) {
       return VP8_STATUS_BITSTREAM_ERROR;  // image is too large
     }
@@ -176,7 +174,10 @@ static VP8StatusCode ParseOptionalChunks(const uint8_t** const data,
       return VP8_STATUS_NOT_ENOUGH_DATA;
     }
 
-    chunk_size = get_le32(buf + TAG_SIZE);
+    chunk_size = GetLE32(buf + TAG_SIZE);
+    if (chunk_size > MAX_CHUNK_PAYLOAD) {
+      return VP8_STATUS_BITSTREAM_ERROR;          // Not a valid chunk size.
+    }
     // For odd-sized chunk-payload, there's one byte padding at the end.
     disk_chunk_size = (CHUNK_HEADER_SIZE + chunk_size + 1) & ~1;
     total_size += disk_chunk_size;
@@ -186,6 +187,15 @@ static VP8StatusCode ParseOptionalChunks(const uint8_t** const data,
       return VP8_STATUS_BITSTREAM_ERROR;          // Not a valid chunk size.
     }
 
+    // Start of a (possibly incomplete) VP8/VP8L chunk implies that we have
+    // parsed all the optional chunks.
+    // Note: This check must occur before the check 'buf_size < disk_chunk_size'
+    // below to allow incomplete VP8/VP8L chunks.
+    if (!memcmp(buf, "VP8 ", TAG_SIZE) ||
+        !memcmp(buf, "VP8L", TAG_SIZE)) {
+      return VP8_STATUS_OK;
+    }
+
     if (buf_size < disk_chunk_size) {             // Insufficient data.
       return VP8_STATUS_NOT_ENOUGH_DATA;
     }
@@ -193,9 +203,6 @@ static VP8StatusCode ParseOptionalChunks(const uint8_t** const data,
     if (!memcmp(buf, "ALPH", TAG_SIZE)) {         // A valid ALPH header.
       *alpha_data = buf + CHUNK_HEADER_SIZE;
       *alpha_size = chunk_size;
-    } else if (!memcmp(buf, "VP8 ", TAG_SIZE) ||
-               !memcmp(buf, "VP8L", TAG_SIZE)) {  // A valid VP8/VP8L header.
-      return VP8_STATUS_OK;  // Found.
     }
 
     // We have a full and valid chunk; skip it.
@@ -213,9 +220,8 @@ static VP8StatusCode ParseOptionalChunks(const uint8_t** const data,
 // extracted from the VP8/VP8L chunk header.
 // The flag '*is_lossless' is set to 1 in case of VP8L chunk / raw VP8L data.
 static VP8StatusCode ParseVP8Header(const uint8_t** const data_ptr,
-                                    size_t* const data_size,
-                                    size_t riff_size,
-                                    size_t* const chunk_size,
+                                    size_t* const data_size, int have_all_data,
+                                    size_t riff_size, size_t* const chunk_size,
                                     int* const is_lossless) {
   const uint8_t* const data = *data_ptr;
   const int is_vp8 = !memcmp(data, "VP8 ", TAG_SIZE);
@@ -234,10 +240,13 @@ static VP8StatusCode ParseVP8Header(const uint8_t** const data_ptr,
 
   if (is_vp8 || is_vp8l) {
     // Bitstream contains VP8/VP8L header.
-    const uint32_t size = get_le32(data + TAG_SIZE);
+    const uint32_t size = GetLE32(data + TAG_SIZE);
     if ((riff_size >= minimal_size) && (size > riff_size - minimal_size)) {
       return VP8_STATUS_BITSTREAM_ERROR;  // Inconsistent size information.
     }
+    if (have_all_data && (size > *data_size - CHUNK_HEADER_SIZE)) {
+      return VP8_STATUS_NOT_ENOUGH_DATA;  // Truncated bitstream.
+    }
     // Skip over CHUNK_HEADER_SIZE bytes from VP8/VP8L Header.
     *chunk_size = size;
     *data_ptr += CHUNK_HEADER_SIZE;
@@ -270,9 +279,19 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
                                           int* const width,
                                           int* const height,
                                           int* const has_alpha,
+                                          int* const has_animation,
+                                          int* const format,
                                           WebPHeaderStructure* const headers) {
+  int canvas_width = 0;
+  int canvas_height = 0;
+  int image_width = 0;
+  int image_height = 0;
   int found_riff = 0;
   int found_vp8x = 0;
+  int animation_present = 0;
+  int fragments_present = 0;
+  const int have_all_data = (headers != NULL) ? headers->have_all_data : 0;
+
   VP8StatusCode status;
   WebPHeaderStructure hdrs;
 
@@ -284,7 +303,7 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
   hdrs.data_size = data_size;
 
   // Skip over RIFF header.
-  status = ParseRIFF(&data, &data_size, &hdrs.riff_size);
+  status = ParseRIFF(&data, &data_size, have_all_data, &hdrs.riff_size);
   if (status != VP8_STATUS_OK) {
     return status;   // Wrong RIFF header / insufficient data.
   }
@@ -293,22 +312,35 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
   // Skip over VP8X.
   {
     uint32_t flags = 0;
-    status = ParseVP8X(&data, &data_size, &found_vp8x, width, height, &flags);
+    status = ParseVP8X(&data, &data_size, &found_vp8x,
+                       &canvas_width, &canvas_height, &flags);
     if (status != VP8_STATUS_OK) {
       return status;  // Wrong VP8X / insufficient data.
     }
+    animation_present = !!(flags & ANIMATION_FLAG);
+    fragments_present = !!(flags & FRAGMENTS_FLAG);
     if (!found_riff && found_vp8x) {
       // Note: This restriction may be removed in the future, if it becomes
       // necessary to send VP8X chunk to the decoder.
       return VP8_STATUS_BITSTREAM_ERROR;
     }
-    if (has_alpha != NULL) *has_alpha = !!(flags & ALPHA_FLAG_BIT);
-    if (found_vp8x && headers == NULL) {
-      return VP8_STATUS_OK;  // Return features from VP8X header.
+    if (has_alpha != NULL) *has_alpha = !!(flags & ALPHA_FLAG);
+    if (has_animation != NULL) *has_animation = animation_present;
+    if (format != NULL) *format = 0;   // default = undefined
+
+    image_width = canvas_width;
+    image_height = canvas_height;
+    if (found_vp8x && (animation_present || fragments_present) &&
+        headers == NULL) {
+      status = VP8_STATUS_OK;
+      goto ReturnWidthHeight;  // Just return features from VP8X header.
     }
   }
 
-  if (data_size < TAG_SIZE) return VP8_STATUS_NOT_ENOUGH_DATA;
+  if (data_size < TAG_SIZE) {
+    status = VP8_STATUS_NOT_ENOUGH_DATA;
+    goto ReturnWidthHeight;
+  }
 
   // Skip over optional chunks if data started with "RIFF + VP8X" or "ALPH".
   if ((found_riff && found_vp8x) ||
@@ -316,43 +348,49 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
     status = ParseOptionalChunks(&data, &data_size, hdrs.riff_size,
                                  &hdrs.alpha_data, &hdrs.alpha_data_size);
     if (status != VP8_STATUS_OK) {
-      return status;  // Found an invalid chunk size / insufficient data.
+      goto ReturnWidthHeight;  // Invalid chunk size / insufficient data.
     }
   }
 
   // Skip over VP8/VP8L header.
-  status = ParseVP8Header(&data, &data_size, hdrs.riff_size,
+  status = ParseVP8Header(&data, &data_size, have_all_data, hdrs.riff_size,
                           &hdrs.compressed_size, &hdrs.is_lossless);
   if (status != VP8_STATUS_OK) {
-    return status;  // Wrong VP8/VP8L chunk-header / insufficient data.
+    goto ReturnWidthHeight;  // Wrong VP8/VP8L chunk-header / insufficient data.
   }
   if (hdrs.compressed_size > MAX_CHUNK_PAYLOAD) {
     return VP8_STATUS_BITSTREAM_ERROR;
   }
 
+  if (format != NULL && !(animation_present || fragments_present)) {
+    *format = hdrs.is_lossless ? 2 : 1;
+  }
+
   if (!hdrs.is_lossless) {
     if (data_size < VP8_FRAME_HEADER_SIZE) {
-      return VP8_STATUS_NOT_ENOUGH_DATA;
+      status = VP8_STATUS_NOT_ENOUGH_DATA;
+      goto ReturnWidthHeight;
     }
     // Validates raw VP8 data.
-    if (!VP8GetInfo(data, data_size,
-                    (uint32_t)hdrs.compressed_size, width, height)) {
+    if (!VP8GetInfo(data, data_size, (uint32_t)hdrs.compressed_size,
+                    &image_width, &image_height)) {
       return VP8_STATUS_BITSTREAM_ERROR;
     }
   } else {
     if (data_size < VP8L_FRAME_HEADER_SIZE) {
-      return VP8_STATUS_NOT_ENOUGH_DATA;
+      status = VP8_STATUS_NOT_ENOUGH_DATA;
+      goto ReturnWidthHeight;
     }
     // Validates raw VP8L data.
-    if (!VP8LGetInfo(data, data_size, width, height, has_alpha)) {
+    if (!VP8LGetInfo(data, data_size, &image_width, &image_height, has_alpha)) {
       return VP8_STATUS_BITSTREAM_ERROR;
     }
   }
-
-  if (has_alpha != NULL) {
-    // If the data did not contain a VP8X/VP8L chunk the only definitive way
-    // to set this is by looking for alpha data (from an ALPH chunk).
-    *has_alpha |= (hdrs.alpha_data != NULL);
+  // Validates image size coherency.
+  if (found_vp8x) {
+    if (canvas_width != image_width || canvas_height != image_height) {
+      return VP8_STATUS_BITSTREAM_ERROR;
+    }
   }
   if (headers != NULL) {
     *headers = hdrs;
@@ -360,21 +398,44 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
     assert((uint64_t)(data - headers->data) < MAX_CHUNK_PAYLOAD);
     assert(headers->offset == headers->data_size - data_size);
   }
-  return VP8_STATUS_OK;  // Return features from VP8 header.
+ ReturnWidthHeight:
+  if (status == VP8_STATUS_OK ||
+      (status == VP8_STATUS_NOT_ENOUGH_DATA && found_vp8x && headers == NULL)) {
+    if (has_alpha != NULL) {
+      // If the data did not contain a VP8X/VP8L chunk the only definitive way
+      // to set this is by looking for alpha data (from an ALPH chunk).
+      *has_alpha |= (hdrs.alpha_data != NULL);
+    }
+    if (width != NULL) *width = image_width;
+    if (height != NULL) *height = image_height;
+    return VP8_STATUS_OK;
+  } else {
+    return status;
+  }
 }
 
 VP8StatusCode WebPParseHeaders(WebPHeaderStructure* const headers) {
+  VP8StatusCode status;
+  int has_animation = 0;
   assert(headers != NULL);
   // fill out headers, ignore width/height/has_alpha.
-  return ParseHeadersInternal(headers->data, headers->data_size,
-                              NULL, NULL, NULL, headers);
+  status = ParseHeadersInternal(headers->data, headers->data_size,
+                                NULL, NULL, NULL, &has_animation,
+                                NULL, headers);
+  if (status == VP8_STATUS_OK || status == VP8_STATUS_NOT_ENOUGH_DATA) {
+    // TODO(jzern): full support of animation frames will require API additions.
+    if (has_animation) {
+      status = VP8_STATUS_UNSUPPORTED_FEATURE;
+    }
+  }
+  return status;
 }
 
 //------------------------------------------------------------------------------
 // WebPDecParams
 
 void WebPResetDecParams(WebPDecParams* const params) {
-  if (params) {
+  if (params != NULL) {
     memset(params, 0, sizeof(*params));
   }
 }
@@ -391,6 +452,7 @@ static VP8StatusCode DecodeInto(const uint8_t* const data, size_t data_size,
 
   headers.data = data;
   headers.data_size = data_size;
+  headers.have_all_data = 1;
   status = WebPParseHeaders(&headers);   // Process Pre-VP8 chunks.
   if (status != VP8_STATUS_OK) {
     return status;
@@ -407,11 +469,6 @@ static VP8StatusCode DecodeInto(const uint8_t* const data, size_t data_size,
     if (dec == NULL) {
       return VP8_STATUS_OUT_OF_MEMORY;
     }
-#ifdef WEBP_USE_THREAD
-    dec->use_threads_ = params->options && (params->options->use_threads > 0);
-#else
-    dec->use_threads_ = 0;
-#endif
     dec->alpha_data_ = headers.alpha_data;
     dec->alpha_data_size_ = headers.alpha_data_size;
 
@@ -423,6 +480,10 @@ static VP8StatusCode DecodeInto(const uint8_t* const data, size_t data_size,
       status = WebPAllocateDecBuffer(io.width, io.height, params->options,
                                      params->output);
       if (status == VP8_STATUS_OK) {  // Decode
+        // This change must be done before calling VP8Decode()
+        dec->mt_method_ = VP8GetThreadMethod(params->options, &headers,
+                                             io.width, io.height);
+        VP8InitDithering(params->options, dec);
         if (!VP8Decode(dec, &io)) {
           status = dec->status_;
         }
@@ -452,6 +513,10 @@ static VP8StatusCode DecodeInto(const uint8_t* const data, size_t data_size,
   if (status != VP8_STATUS_OK) {
     WebPFreeDecBuffer(params->output);
   }
+
+  if (params->options != NULL && params->options->flip) {
+    status = WebPFlipBuffer(params->output);
+  }
   return status;
 }
 
@@ -609,7 +674,6 @@ uint8_t* WebPDecodeYUV(const uint8_t* data, size_t data_size,
 static void DefaultFeatures(WebPBitstreamFeatures* const features) {
   assert(features != NULL);
   memset(features, 0, sizeof(*features));
-  features->bitstream_version = 0;
 }
 
 static VP8StatusCode GetFeatures(const uint8_t* const data, size_t data_size,
@@ -619,10 +683,11 @@ static VP8StatusCode GetFeatures(const uint8_t* const data, size_t data_size,
   }
   DefaultFeatures(features);
 
-  // Only parse enough of the data to retrieve width/height/has_alpha.
+  // Only parse enough of the data to retrieve the features.
   return ParseHeadersInternal(data, data_size,
                               &features->width, &features->height,
-                              &features->has_alpha, NULL);
+                              &features->has_alpha, &features->has_animation,
+                              &features->format, NULL);
 }
 
 //------------------------------------------------------------------------------
@@ -666,19 +731,13 @@ int WebPInitDecoderConfigInternal(WebPDecoderConfig* config,
 VP8StatusCode WebPGetFeaturesInternal(const uint8_t* data, size_t data_size,
                                       WebPBitstreamFeatures* features,
                                       int version) {
-  VP8StatusCode status;
   if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_DECODER_ABI_VERSION)) {
     return VP8_STATUS_INVALID_PARAM;   // version mismatch
   }
   if (features == NULL) {
     return VP8_STATUS_INVALID_PARAM;
   }
-
-  status = GetFeatures(data, data_size, features);
-  if (status == VP8_STATUS_NOT_ENOUGH_DATA) {
-    return VP8_STATUS_BITSTREAM_ERROR;  // Not-enough-data treated as error.
-  }
-  return status;
+  return GetFeatures(data, data_size, features);
 }
 
 VP8StatusCode WebPDecode(const uint8_t* data, size_t data_size,
@@ -722,9 +781,9 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
     h = options->crop_height;
     x = options->crop_left;
     y = options->crop_top;
-    if (!WebPIsRGBMode(src_colorspace)) {   // only snap for YUV420 or YUV422
+    if (!WebPIsRGBMode(src_colorspace)) {   // only snap for YUV420
       x &= ~1;
-      y &= ~1;    // TODO(later): only for YUV420, not YUV422.
+      y &= ~1;
     }
     if (x < 0 || y < 0 || w <= 0 || h <= 0 || x + w > W || y + h > H) {
       return 0;  // out of frame boundary error
@@ -740,11 +799,13 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
   // Scaling
   io->use_scaling = (options != NULL) && (options->use_scaling > 0);
   if (io->use_scaling) {
-    if (options->scaled_width <= 0 || options->scaled_height <= 0) {
+    int scaled_width = options->scaled_width;
+    int scaled_height = options->scaled_height;
+    if (!WebPRescalerGetScaledDimensions(w, h, &scaled_width, &scaled_height)) {
       return 0;
     }
-    io->scaled_width = options->scaled_width;
-    io->scaled_height = options->scaled_height;
+    io->scaled_width = scaled_width;
+    io->scaled_height = scaled_height;
   }
 
   // Filter
@@ -766,6 +827,3 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dec/webpi.h b/drivers/webp/dec/webpi.h
index 44e5744411..c75a2e4a5b 100644
--- a/drivers/webp/dec/webpi.h
+++ b/drivers/webp/dec/webpi.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Internal header: WebP decoding parameters and custom IO on buffer
@@ -12,7 +14,7 @@
 #ifndef WEBP_DEC_WEBPI_H_
 #define WEBP_DEC_WEBPI_H_
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -24,7 +26,10 @@ extern "C" {
 
 typedef struct WebPDecParams WebPDecParams;
 typedef int (*OutputFunc)(const VP8Io* const io, WebPDecParams* const p);
-typedef int (*OutputRowFunc)(WebPDecParams* const p, int y_pos);
+typedef int (*OutputAlphaFunc)(const VP8Io* const io, WebPDecParams* const p,
+                               int expected_num_out_lines);
+typedef int (*OutputRowFunc)(WebPDecParams* const p, int y_pos,
+                             int max_out_lines);
 
 struct WebPDecParams {
   WebPDecBuffer* output;             // output buffer.
@@ -38,7 +43,7 @@ struct WebPDecParams {
   void* memory;                  // overall scratch memory for the output work.
 
   OutputFunc emit;               // output RGB or YUV samples
-  OutputFunc emit_alpha;         // output alpha channel
+  OutputAlphaFunc emit_alpha;    // output alpha channel
   OutputRowFunc emit_alpha_row;  // output one line of rescaled alpha values
 };
 
@@ -52,6 +57,7 @@ void WebPResetDecParams(WebPDecParams* const params);
 typedef struct {
   const uint8_t* data;         // input buffer
   size_t data_size;            // input buffer size
+  int have_all_data;           // true if all data is known to be available
   size_t offset;               // offset to main data chunk (VP8 or VP8L)
   const uint8_t* alpha_data;   // points to alpha chunk (if present)
   size_t alpha_data_size;      // alpha chunk size
@@ -61,10 +67,10 @@ typedef struct {
 } WebPHeaderStructure;
 
 // Skips over all valid chunks prior to the first VP8/VP8L frame header.
-// Returns VP8_STATUS_OK on success,
-//         VP8_STATUS_BITSTREAM_ERROR if an invalid header/chunk is found, and
-//         VP8_STATUS_NOT_ENOUGH_DATA if case of insufficient data.
-// In 'headers', compressed_size, offset, alpha_data, alpha_size and lossless
+// Returns: VP8_STATUS_OK, VP8_STATUS_BITSTREAM_ERROR (invalid header/chunk),
+// VP8_STATUS_NOT_ENOUGH_DATA (partial input) or VP8_STATUS_UNSUPPORTED_FEATURE
+// in the case of non-decodable features (animation for instance).
+// In 'headers', compressed_size, offset, alpha_data, alpha_size, and lossless
 // fields are updated appropriately upon success.
 VP8StatusCode WebPParseHeaders(WebPHeaderStructure* const headers);
 
@@ -91,10 +97,15 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
 // dimension / etc.). If *options is not NULL, also verify that the options'
 // parameters are valid and apply them to the width/height dimensions of the
 // output buffer. This takes cropping / scaling / rotation into account.
+// Also incorporates the options->flip flag to flip the buffer parameters if
+// needed.
 VP8StatusCode WebPAllocateDecBuffer(int width, int height,
                                     const WebPDecoderOptions* const options,
                                     WebPDecBuffer* const buffer);
 
+// Flip buffer vertically by negating the various strides.
+VP8StatusCode WebPFlipBuffer(WebPDecBuffer* const buffer);
+
 // Copy 'src' into 'dst' buffer, making sure 'dst' is not marked as owner of the
 // memory (still held by 'src').
 void WebPCopyDecBuffer(const WebPDecBuffer* const src,
@@ -103,11 +114,9 @@ void WebPCopyDecBuffer(const WebPDecBuffer* const src,
 // Copy and transfer ownership from src to dst (beware of parameter order!)
 void WebPGrabDecBuffer(WebPDecBuffer* const src, WebPDecBuffer* const dst);
 
-
-
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/decode.h b/drivers/webp/decode.h
index 43b6c58f4f..fa4b13411d 100644
--- a/drivers/webp/decode.h
+++ b/drivers/webp/decode.h
@@ -1,8 +1,10 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 decoding functions for WebP images.
@@ -14,11 +16,23 @@
 
 #include "./types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-#define WEBP_DECODER_ABI_VERSION 0x0200    // MAJOR(8b) + MINOR(8b)
+#define WEBP_DECODER_ABI_VERSION 0x0207    // MAJOR(8b) + MINOR(8b)
+
+// Note: forward declaring enumerations is not allowed in (strict) C and C++,
+// the types are left here for reference.
+// typedef enum VP8StatusCode VP8StatusCode;
+// typedef enum WEBP_CSP_MODE WEBP_CSP_MODE;
+typedef struct WebPRGBABuffer WebPRGBABuffer;
+typedef struct WebPYUVABuffer WebPYUVABuffer;
+typedef struct WebPDecBuffer WebPDecBuffer;
+typedef struct WebPIDecoder WebPIDecoder;
+typedef struct WebPBitstreamFeatures WebPBitstreamFeatures;
+typedef struct WebPDecoderOptions WebPDecoderOptions;
+typedef struct WebPDecoderConfig WebPDecoderConfig;
 
 // Return the decoder's version number, packed in hexadecimal using 8bits for
 // each of major/minor/revision. E.g: v2.5.7 is 0x020507.
@@ -34,7 +48,7 @@ WEBP_EXTERN(int) WebPGetInfo(const uint8_t* data, size_t data_size,
 // Decodes WebP images pointed to by 'data' and returns RGBA samples, along
 // with the dimensions in *width and *height. The ordering of samples in
 // memory is R, G, B, A, R, G, B, A... in scan order (endian-independent).
-// The returned pointer should be deleted calling free().
+// The returned pointer should be deleted calling WebPFree().
 // Returns NULL in case of error.
 WEBP_EXTERN(uint8_t*) WebPDecodeRGBA(const uint8_t* data, size_t data_size,
                                      int* width, int* height);
@@ -59,9 +73,9 @@ WEBP_EXTERN(uint8_t*) WebPDecodeBGR(const uint8_t* data, size_t data_size,
 
 // Decode WebP images pointed to by 'data' to Y'UV format(*). The pointer
 // returned is the Y samples buffer. Upon return, *u and *v will point to
-// the U and V chroma data. These U and V buffers need NOT be free()'d,
-// unlike the returned Y luma one. The dimension of the U and V planes
-// are both (*width + 1) / 2 and (*height + 1)/ 2.
+// the U and V chroma data. These U and V buffers need NOT be passed to
+// WebPFree(), unlike the returned Y luma one. The dimension of the U and V
+// planes are both (*width + 1) / 2 and (*height + 1)/ 2.
 // Upon return, the Y buffer has a stride returned as '*stride', while U and V
 // have a common stride returned as '*uv_stride'.
 // Return NULL in case of error.
@@ -71,6 +85,9 @@ WEBP_EXTERN(uint8_t*) WebPDecodeYUV(const uint8_t* data, size_t data_size,
                                     uint8_t** u, uint8_t** v,
                                     int* stride, int* uv_stride);
 
+// Releases memory returned by the WebPDecode*() functions above.
+WEBP_EXTERN(void) WebPFree(void* ptr);
+
 // These five functions are variants of the above ones, that decode the image
 // directly into a pre-allocated buffer 'output_buffer'. The maximum storage
 // available in this buffer is indicated by 'output_buffer_size'. If this
@@ -118,20 +135,28 @@ WEBP_EXTERN(uint8_t*) WebPDecodeYUVInto(
 // Note: the naming describes the byte-ordering of packed samples in memory.
 // For instance, MODE_BGRA relates to samples ordered as B,G,R,A,B,G,R,A,...
 // Non-capital names (e.g.:MODE_Argb) relates to pre-multiplied RGB channels.
-// RGB-565 and RGBA-4444 are also endian-agnostic and byte-oriented.
-typedef enum { MODE_RGB = 0, MODE_RGBA = 1,
-               MODE_BGR = 2, MODE_BGRA = 3,
-               MODE_ARGB = 4, MODE_RGBA_4444 = 5,
-               MODE_RGB_565 = 6,
-               // RGB-premultiplied transparent modes (alpha value is preserved)
-               MODE_rgbA = 7,
-               MODE_bgrA = 8,
-               MODE_Argb = 9,
-               MODE_rgbA_4444 = 10,
-               // YUV modes must come after RGB ones.
-               MODE_YUV = 11, MODE_YUVA = 12,  // yuv 4:2:0
-               MODE_LAST = 13
-             } WEBP_CSP_MODE;
+// RGBA-4444 and RGB-565 colorspaces are represented by following byte-order:
+// RGBA-4444: [r3 r2 r1 r0 g3 g2 g1 g0], [b3 b2 b1 b0 a3 a2 a1 a0], ...
+// RGB-565: [r4 r3 r2 r1 r0 g5 g4 g3], [g2 g1 g0 b4 b3 b2 b1 b0], ...
+// In the case WEBP_SWAP_16BITS_CSP is defined, the bytes are swapped for
+// these two modes:
+// RGBA-4444: [b3 b2 b1 b0 a3 a2 a1 a0], [r3 r2 r1 r0 g3 g2 g1 g0], ...
+// RGB-565: [g2 g1 g0 b4 b3 b2 b1 b0], [r4 r3 r2 r1 r0 g5 g4 g3], ...
+
+typedef enum WEBP_CSP_MODE {
+  MODE_RGB = 0, MODE_RGBA = 1,
+  MODE_BGR = 2, MODE_BGRA = 3,
+  MODE_ARGB = 4, MODE_RGBA_4444 = 5,
+  MODE_RGB_565 = 6,
+  // RGB-premultiplied transparent modes (alpha value is preserved)
+  MODE_rgbA = 7,
+  MODE_bgrA = 8,
+  MODE_Argb = 9,
+  MODE_rgbA_4444 = 10,
+  // YUV modes must come after RGB ones.
+  MODE_YUV = 11, MODE_YUVA = 12,  // yuv 4:2:0
+  MODE_LAST = 13
+} WEBP_CSP_MODE;
 
 // Some useful macros:
 static WEBP_INLINE int WebPIsPremultipliedMode(WEBP_CSP_MODE mode) {
@@ -152,13 +177,13 @@ static WEBP_INLINE int WebPIsRGBMode(WEBP_CSP_MODE mode) {
 //------------------------------------------------------------------------------
 // WebPDecBuffer: Generic structure for describing the output sample buffer.
 
-typedef struct {    // view as RGBA
+struct WebPRGBABuffer {    // view as RGBA
   uint8_t* rgba;    // pointer to RGBA samples
   int stride;       // stride in bytes from one scanline to the next.
   size_t size;      // total size of the *rgba buffer.
-} WebPRGBABuffer;
+};
 
-typedef struct {              // view as YUVA
+struct WebPYUVABuffer {              // view as YUVA
   uint8_t* y, *u, *v, *a;     // pointer to luma, chroma U/V, alpha samples
   int y_stride;               // luma stride
   int u_stride, v_stride;     // chroma strides
@@ -166,10 +191,10 @@ typedef struct {              // view as YUVA
   size_t y_size;              // luma plane size
   size_t u_size, v_size;      // chroma planes size
   size_t a_size;              // alpha-plane size
-} WebPYUVABuffer;
+};
 
 // Output buffer
-typedef struct {
+struct WebPDecBuffer {
   WEBP_CSP_MODE colorspace;  // Colorspace.
   int width, height;         // Dimensions.
   int is_external_memory;    // If true, 'internal_memory' pointer is not used.
@@ -182,7 +207,7 @@ typedef struct {
   uint8_t* private_memory;   // Internally allocated memory (only when
                              // is_external_memory is false). Should not be used
                              // externally, but accessed via the buffer union.
-} WebPDecBuffer;
+};
 
 // Internal, version-checked, entry point
 WEBP_EXTERN(int) WebPInitDecBufferInternal(WebPDecBuffer*, int);
@@ -200,7 +225,7 @@ WEBP_EXTERN(void) WebPFreeDecBuffer(WebPDecBuffer* buffer);
 //------------------------------------------------------------------------------
 // Enumeration of the status codes
 
-typedef enum {
+typedef enum VP8StatusCode {
   VP8_STATUS_OK = 0,
   VP8_STATUS_OUT_OF_MEMORY,
   VP8_STATUS_INVALID_PARAM,
@@ -237,13 +262,17 @@ typedef enum {
 //   }
 //   WebPIDelete(idec);
 
-typedef struct WebPIDecoder WebPIDecoder;
-
 // Creates a new incremental decoder with the supplied buffer parameter.
 // This output_buffer can be passed NULL, in which case a default output buffer
 // is used (with MODE_RGB). Otherwise, an internal reference to 'output_buffer'
 // is kept, which means that the lifespan of 'output_buffer' must be larger than
 // that of the returned WebPIDecoder object.
+// The supplied 'output_buffer' content MUST NOT be changed between calls to
+// WebPIAppend() or WebPIUpdate() unless 'output_buffer.is_external_memory' is
+// set to 1. In such a case, it is allowed to modify the pointers, size and
+// stride of output_buffer.u.RGBA or output_buffer.u.YUVA, provided they remain
+// within valid bounds.
+// All other fields of WebPDecBuffer MUST remain constant between calls.
 // Returns NULL if the allocation failed.
 WEBP_EXTERN(WebPIDecoder*) WebPINewDecoder(WebPDecBuffer* output_buffer);
 
@@ -251,19 +280,27 @@ WEBP_EXTERN(WebPIDecoder*) WebPINewDecoder(WebPDecBuffer* output_buffer);
 // will output the RGB/A samples specified by 'csp' into a preallocated
 // buffer 'output_buffer'. The size of this buffer is at least
 // 'output_buffer_size' and the stride (distance in bytes between two scanlines)
-// is specified by 'output_stride'. Returns NULL if the allocation failed.
+// is specified by 'output_stride'.
+// Additionally, output_buffer can be passed NULL in which case the output
+// buffer will be allocated automatically when the decoding starts. The
+// colorspace 'csp' is taken into account for allocating this buffer. All other
+// parameters are ignored.
+// Returns NULL if the allocation failed, or if some parameters are invalid.
 WEBP_EXTERN(WebPIDecoder*) WebPINewRGB(
     WEBP_CSP_MODE csp,
     uint8_t* output_buffer, size_t output_buffer_size, int output_stride);
 
 // This function allocates and initializes an incremental-decoder object, which
-// will output the raw luma/chroma samples into a preallocated planes. The luma
-// plane is specified by its pointer 'luma', its size 'luma_size' and its stride
-// 'luma_stride'. Similarly, the chroma-u plane is specified by the 'u',
-// 'u_size' and 'u_stride' parameters, and the chroma-v plane by 'v'
-// and 'v_size'. And same for the alpha-plane. The 'a' pointer can be pass
-// NULL in case one is not interested in the transparency plane.
-// Returns NULL if the allocation failed.
+// will output the raw luma/chroma samples into a preallocated planes if
+// supplied. The luma plane is specified by its pointer 'luma', its size
+// 'luma_size' and its stride 'luma_stride'. Similarly, the chroma-u plane
+// is specified by the 'u', 'u_size' and 'u_stride' parameters, and the chroma-v
+// plane by 'v' and 'v_size'. And same for the alpha-plane. The 'a' pointer
+// can be pass NULL in case one is not interested in the transparency plane.
+// Conversely, 'luma' can be passed NULL if no preallocated planes are supplied.
+// In this case, the output buffer will be automatically allocated (using
+// MODE_YUVA) when decoding starts. All parameters are then ignored.
+// Returns NULL if the allocation failed or if a parameter is invalid.
 WEBP_EXTERN(WebPIDecoder*) WebPINewYUVA(
     uint8_t* luma, size_t luma_size, int luma_stride,
     uint8_t* u, size_t u_size, int u_stride,
@@ -344,7 +381,7 @@ WEBP_EXTERN(const WebPDecBuffer*) WebPIDecodedArea(
      CHECK(WebPGetFeatures(data, data_size, &config.input) == VP8_STATUS_OK);
 
      // C) Adjust 'config', if needed
-     config.no_fancy = 1;
+     config.no_fancy_upsampling = 1;
      config.output.colorspace = MODE_BGRA;
      // etc.
 
@@ -365,19 +402,15 @@ WEBP_EXTERN(const WebPDecBuffer*) WebPIDecodedArea(
 */
 
 // Features gathered from the bitstream
-typedef struct {
-  int width;        // Width in pixels, as read from the bitstream.
-  int height;       // Height in pixels, as read from the bitstream.
-  int has_alpha;    // True if the bitstream contains an alpha channel.
-
-  // Unused for now:
-  int bitstream_version;        // should be 0 for now. TODO(later)
-  int no_incremental_decoding;  // if true, using incremental decoding is not
-                                // recommended.
-  int rotate;                   // TODO(later)
-  int uv_sampling;              // should be 0 for now. TODO(later)
-  uint32_t pad[3];              // padding for later use
-} WebPBitstreamFeatures;
+struct WebPBitstreamFeatures {
+  int width;          // Width in pixels, as read from the bitstream.
+  int height;         // Height in pixels, as read from the bitstream.
+  int has_alpha;      // True if the bitstream contains an alpha channel.
+  int has_animation;  // True if the bitstream is an animation.
+  int format;         // 0 = undefined (/mixed), 1 = lossy, 2 = lossless
+
+  uint32_t pad[5];    // padding for later use
+};
 
 // Internal, version-checked, entry point
 WEBP_EXTERN(VP8StatusCode) WebPGetFeaturesInternal(
@@ -385,8 +418,9 @@ WEBP_EXTERN(VP8StatusCode) WebPGetFeaturesInternal(
 
 // Retrieve features from the bitstream. The *features structure is filled
 // with information gathered from the bitstream.
-// Returns false in case of error or version mismatch.
-// In case of error, features->bitstream_status will reflect the error code.
+// Returns VP8_STATUS_OK when the features are successfully retrieved. Returns
+// VP8_STATUS_NOT_ENOUGH_DATA when more data is needed to retrieve the
+// features from headers. Returns error in other cases.
 static WEBP_INLINE VP8StatusCode WebPGetFeatures(
     const uint8_t* data, size_t data_size,
     WebPBitstreamFeatures* features) {
@@ -395,7 +429,7 @@ static WEBP_INLINE VP8StatusCode WebPGetFeatures(
 }
 
 // Decoding options
-typedef struct {
+struct WebPDecoderOptions {
   int bypass_filtering;               // if true, skip the in-loop filtering
   int no_fancy_upsampling;            // if true, use faster pointwise upsampler
   int use_cropping;                   // if true, cropping is applied _first_
@@ -405,19 +439,19 @@ typedef struct {
   int use_scaling;                    // if true, scaling is applied _afterward_
   int scaled_width, scaled_height;    // final resolution
   int use_threads;                    // if true, use multi-threaded decoding
+  int dithering_strength;             // dithering strength (0=Off, 100=full)
+  int flip;                           // flip output vertically
+  int alpha_dithering_strength;       // alpha dithering strength in [0..100]
 
-  // Unused for now:
-  int force_rotation;                 // forced rotation (to be applied _last_)
-  int no_enhancement;                 // if true, discard enhancement layer
-  uint32_t pad[6];                    // padding for later use
-} WebPDecoderOptions;
+  uint32_t pad[5];                    // padding for later use
+};
 
 // Main object storing the configuration for advanced decoding.
-typedef struct {
+struct WebPDecoderConfig {
   WebPBitstreamFeatures input;  // Immutable bitstream features (optional)
   WebPDecBuffer output;         // Output buffer (can point to external mem)
   WebPDecoderOptions options;   // Decoding options
-} WebPDecoderConfig;
+};
 
 // Internal, version-checked, entry point
 WEBP_EXTERN(int) WebPInitDecoderConfigInternal(WebPDecoderConfig*, int);
@@ -447,7 +481,7 @@ WEBP_EXTERN(WebPIDecoder*) WebPIDecode(const uint8_t* data, size_t data_size,
 WEBP_EXTERN(VP8StatusCode) WebPDecode(const uint8_t* data, size_t data_size,
                                       WebPDecoderConfig* config);
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/dsp/cpu.c b/drivers/webp/dsp/cpu.c
index 0228734457..35c2af7f58 100644
--- a/drivers/webp/dsp/cpu.c
+++ b/drivers/webp/dsp/cpu.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // CPU detection
@@ -11,14 +13,10 @@
 
 #include "./dsp.h"
 
-#if defined(__ANDROID__)
+#if defined(WEBP_ANDROID_NEON)
 #include <cpu-features.h>
 #endif
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
 // SSE2 detection.
 //
@@ -31,22 +29,66 @@ static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) {
     "cpuid\n"
     "xchg %%edi, %%ebx\n"
     : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type));
+    : "a"(info_type), "c"(0));
 }
 #elif defined(__i386__) || defined(__x86_64__)
 static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) {
   __asm__ volatile (
     "cpuid\n"
     : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
-    : "a"(info_type));
+    : "a"(info_type), "c"(0));
 }
+#elif (defined(_M_X64) || defined(_M_IX86)) && \
+      defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 150030729  // >= VS2008 SP1
+#include <intrin.h>
+#define GetCPUInfo(info, type) __cpuidex(info, type, 0)  // set ecx=0
 #elif defined(WEBP_MSC_SSE2)
 #define GetCPUInfo __cpuid
 #endif
 
+// NaCl has no support for xgetbv or the raw opcode.
+#if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__))
+static WEBP_INLINE uint64_t xgetbv(void) {
+  const uint32_t ecx = 0;
+  uint32_t eax, edx;
+  // Use the raw opcode for xgetbv for compatibility with older toolchains.
+  __asm__ volatile (
+    ".byte 0x0f, 0x01, 0xd0\n"
+    : "=a"(eax), "=d"(edx) : "c" (ecx));
+  return ((uint64_t)edx << 32) | eax;
+}
+#elif (defined(_M_X64) || defined(_M_IX86)) && \
+      defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219  // >= VS2010 SP1
+#include <immintrin.h>
+#define xgetbv() _xgetbv(0)
+#elif defined(_MSC_VER) && defined(_M_IX86)
+static WEBP_INLINE uint64_t xgetbv(void) {
+  uint32_t eax_, edx_;
+  __asm {
+    xor ecx, ecx  // ecx = 0
+    // Use the raw opcode for xgetbv for compatibility with older toolchains.
+    __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0
+    mov eax_, eax
+    mov edx_, edx
+  }
+  return ((uint64_t)edx_ << 32) | eax_;
+}
+#else
+#define xgetbv() 0U  // no AVX for older x64 or unrecognized toolchains.
+#endif
+
 #if defined(__i386__) || defined(__x86_64__) || defined(WEBP_MSC_SSE2)
 static int x86CPUInfo(CPUFeature feature) {
+  int max_cpuid_value;
   int cpu_info[4];
+
+  // get the highest feature value cpuid supports
+  GetCPUInfo(cpu_info, 0);
+  max_cpuid_value = cpu_info[0];
+  if (max_cpuid_value < 1) {
+    return 0;
+  }
+
   GetCPUInfo(cpu_info, 1);
   if (feature == kSSE2) {
     return 0 != (cpu_info[3] & 0x04000000);
@@ -54,10 +96,26 @@ static int x86CPUInfo(CPUFeature feature) {
   if (feature == kSSE3) {
     return 0 != (cpu_info[2] & 0x00000001);
   }
+  if (feature == kSSE4_1) {
+    return 0 != (cpu_info[2] & 0x00080000);
+  }
+  if (feature == kAVX) {
+    // bits 27 (OSXSAVE) & 28 (256-bit AVX)
+    if ((cpu_info[2] & 0x18000000) == 0x18000000) {
+      // XMM state and YMM state enabled by the OS.
+      return (xgetbv() & 0x6) == 0x6;
+    }
+  }
+  if (feature == kAVX2) {
+    if (x86CPUInfo(kAVX) && max_cpuid_value >= 7) {
+      GetCPUInfo(cpu_info, 7);
+      return ((cpu_info[1] & 0x00000020) == 0x00000020);
+    }
+  }
   return 0;
 }
 VP8CPUInfo VP8GetCPUInfo = x86CPUInfo;
-#elif defined(WEBP_ANDROID_NEON)
+#elif defined(WEBP_ANDROID_NEON)  // NB: needs to be before generic NEON test.
 static int AndroidCPUInfo(CPUFeature feature) {
   const AndroidCpuFamily cpu_family = android_getCpuFamily();
   const uint64_t cpu_features = android_getCpuFeatures();
@@ -68,7 +126,7 @@ static int AndroidCPUInfo(CPUFeature feature) {
   return 0;
 }
 VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo;
-#elif defined(__ARM_NEON__)
+#elif defined(WEBP_USE_NEON)
 // define a dummy function to enable turning off NEON at runtime by setting
 // VP8DecGetCPUInfo = NULL
 static int armCPUInfo(CPUFeature feature) {
@@ -76,10 +134,17 @@ static int armCPUInfo(CPUFeature feature) {
   return 1;
 }
 VP8CPUInfo VP8GetCPUInfo = armCPUInfo;
+#elif defined(WEBP_USE_MIPS32) || defined(WEBP_USE_MIPS_DSP_R2)
+static int mipsCPUInfo(CPUFeature feature) {
+  if ((feature == kMIPS32) || (feature == kMIPSdspR2)) {
+    return 1;
+  } else {
+    return 0;
+  }
+
+}
+VP8CPUInfo VP8GetCPUInfo = mipsCPUInfo;
 #else
 VP8CPUInfo VP8GetCPUInfo = NULL;
 #endif
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dsp/dec.c b/drivers/webp/dsp/dec.c
index 9ae7b6fa76..77a00381c5 100644
--- a/drivers/webp/dsp/dec.c
+++ b/drivers/webp/dsp/dec.c
@@ -1,53 +1,20 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
-// Speed-critical decoding functions.
+// Speed-critical decoding functions, default plain-C implementations.
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include "./dsp.h"
 #include "../dec/vp8i.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
-// run-time tables (~4k)
-
-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]
-
-// We declare this variable 'volatile' to prevent instruction reordering
-// and make sure it's set to true _last_ (so as to be thread-safe)
-static volatile int tables_ok = 0;
-
-static void DspInitTables(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;
-  }
-}
 
 static WEBP_INLINE uint8_t clip_8b(int v) {
   return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
@@ -59,9 +26,16 @@ static WEBP_INLINE uint8_t clip_8b(int v) {
 #define STORE(x, y, v) \
   dst[x + y * BPS] = clip_8b(dst[x + y * BPS] + ((v) >> 3))
 
-static const int kC1 = 20091 + (1 << 16);
-static const int kC2 = 35468;
-#define MUL(a, b) (((a) * (b)) >> 16)
+#define STORE2(y, dc, d, c) do {    \
+  const int DC = (dc);              \
+  STORE(0, y, DC + (d));            \
+  STORE(1, y, DC + (c));            \
+  STORE(2, y, DC - (c));            \
+  STORE(3, y, DC - (d));            \
+} while (0)
+
+#define MUL1(a) ((((a) * 20091) >> 16) + (a))
+#define MUL2(a) (((a) * 35468) >> 16)
 
 static void TransformOne(const int16_t* in, uint8_t* dst) {
   int C[4 * 4], *tmp;
@@ -70,8 +44,8 @@ static void TransformOne(const int16_t* in, uint8_t* dst) {
   for (i = 0; i < 4; ++i) {    // vertical pass
     const int a = in[0] + in[8];    // [-4096, 4094]
     const int b = in[0] - in[8];    // [-4095, 4095]
-    const int c = MUL(in[4], kC2) - MUL(in[12], kC1);   // [-3783, 3783]
-    const int d = MUL(in[4], kC1) + MUL(in[12], kC2);   // [-3785, 3781]
+    const int c = MUL2(in[4]) - MUL1(in[12]);   // [-3783, 3783]
+    const int d = MUL1(in[4]) + MUL2(in[12]);   // [-3785, 3781]
     tmp[0] = a + d;   // [-7881, 7875]
     tmp[1] = b + c;   // [-7878, 7878]
     tmp[2] = b - c;   // [-7878, 7878]
@@ -80,7 +54,7 @@ static void TransformOne(const int16_t* in, uint8_t* dst) {
     in++;
   }
   // Each pass is expanding the dynamic range by ~3.85 (upper bound).
-  // The exact value is (2. + (kC1 + kC2) / 65536).
+  // The exact value is (2. + (20091 + 35468) / 65536).
   // After the second pass, maximum interval is [-3794, 3794], assuming
   // an input in [-2048, 2047] interval. We then need to add a dst value
   // in the [0, 255] range.
@@ -91,8 +65,8 @@ static void TransformOne(const int16_t* in, uint8_t* dst) {
     const int dc = tmp[0] + 4;
     const int a =  dc +  tmp[8];
     const int b =  dc -  tmp[8];
-    const int c = MUL(tmp[4], kC2) - MUL(tmp[12], kC1);
-    const int d = MUL(tmp[4], kC1) + MUL(tmp[12], kC2);
+    const int c = MUL2(tmp[4]) - MUL1(tmp[12]);
+    const int d = MUL1(tmp[4]) + MUL2(tmp[12]);
     STORE(0, 0, a + d);
     STORE(1, 0, b + c);
     STORE(2, 0, b - c);
@@ -101,7 +75,22 @@ static void TransformOne(const int16_t* in, uint8_t* dst) {
     dst += BPS;
   }
 }
-#undef MUL
+
+// Simplified transform when only in[0], in[1] and in[4] are non-zero
+static void TransformAC3(const int16_t* in, uint8_t* dst) {
+  const int a = in[0] + 4;
+  const int c4 = MUL2(in[4]);
+  const int d4 = MUL1(in[4]);
+  const int c1 = MUL2(in[1]);
+  const int d1 = MUL1(in[1]);
+  STORE2(0, a + d4, d1, c1);
+  STORE2(1, a + c4, d1, c1);
+  STORE2(2, a - c4, d1, c1);
+  STORE2(3, a - d4, d1, c1);
+}
+#undef MUL1
+#undef MUL2
+#undef STORE2
 
 static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
   TransformOne(in, dst);
@@ -115,7 +104,7 @@ static void TransformUV(const int16_t* in, uint8_t* dst) {
   VP8Transform(in + 2 * 16, dst + 4 * BPS, 1);
 }
 
-static void TransformDC(const int16_t *in, uint8_t* dst) {
+static void TransformDC(const int16_t* in, uint8_t* dst) {
   const int DC = in[0] + 4;
   int i, j;
   for (j = 0; j < 4; ++j) {
@@ -126,10 +115,10 @@ static void TransformDC(const int16_t *in, uint8_t* dst) {
 }
 
 static void TransformDCUV(const int16_t* in, uint8_t* dst) {
-  if (in[0 * 16]) TransformDC(in + 0 * 16, dst);
-  if (in[1 * 16]) TransformDC(in + 1 * 16, dst + 4);
-  if (in[2 * 16]) TransformDC(in + 2 * 16, dst + 4 * BPS);
-  if (in[3 * 16]) TransformDC(in + 3 * 16, dst + 4 * BPS + 4);
+  if (in[0 * 16]) VP8TransformDC(in + 0 * 16, dst);
+  if (in[1 * 16]) VP8TransformDC(in + 1 * 16, dst + 4);
+  if (in[2 * 16]) VP8TransformDC(in + 2 * 16, dst + 4 * BPS);
+  if (in[3 * 16]) VP8TransformDC(in + 3 * 16, dst + 4 * BPS + 4);
 }
 
 #undef STORE
@@ -164,16 +153,16 @@ static void TransformWHT(const int16_t* in, int16_t* out) {
   }
 }
 
-void (*VP8TransformWHT)(const int16_t* in, int16_t* out) = TransformWHT;
+void (*VP8TransformWHT)(const int16_t* in, int16_t* out);
 
 //------------------------------------------------------------------------------
 // Intra predictions
 
 #define DST(x, y) dst[(x) + (y) * BPS]
 
-static WEBP_INLINE void TrueMotion(uint8_t *dst, int size) {
+static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) {
   const uint8_t* top = dst - BPS;
-  const uint8_t* const clip0 = clip1 + 255 - top[-1];
+  const uint8_t* const clip0 = VP8kclip1 - top[-1];
   int y;
   for (y = 0; y < size; ++y) {
     const uint8_t* const clip = clip0 + dst[-1];
@@ -184,21 +173,21 @@ static WEBP_INLINE void TrueMotion(uint8_t *dst, int size) {
     dst += BPS;
   }
 }
-static void TM4(uint8_t *dst)   { TrueMotion(dst, 4); }
-static void TM8uv(uint8_t *dst) { TrueMotion(dst, 8); }
-static void TM16(uint8_t *dst)  { TrueMotion(dst, 16); }
+static void TM4(uint8_t* dst)   { TrueMotion(dst, 4); }
+static void TM8uv(uint8_t* dst) { TrueMotion(dst, 8); }
+static void TM16(uint8_t* dst)  { TrueMotion(dst, 16); }
 
 //------------------------------------------------------------------------------
 // 16x16
 
-static void VE16(uint8_t *dst) {     // vertical
+static void VE16(uint8_t* dst) {     // vertical
   int j;
   for (j = 0; j < 16; ++j) {
     memcpy(dst + j * BPS, dst - BPS, 16);
   }
 }
 
-static void HE16(uint8_t *dst) {     // horizontal
+static void HE16(uint8_t* dst) {     // horizontal
   int j;
   for (j = 16; j > 0; --j) {
     memset(dst, dst[-1], 16);
@@ -213,7 +202,7 @@ static WEBP_INLINE void Put16(int v, uint8_t* dst) {
   }
 }
 
-static void DC16(uint8_t *dst) {    // DC
+static void DC16(uint8_t* dst) {    // DC
   int DC = 16;
   int j;
   for (j = 0; j < 16; ++j) {
@@ -222,7 +211,7 @@ static void DC16(uint8_t *dst) {    // DC
   Put16(DC >> 5, dst);
 }
 
-static void DC16NoTop(uint8_t *dst) {   // DC with top samples not available
+static void DC16NoTop(uint8_t* dst) {   // DC with top samples not available
   int DC = 8;
   int j;
   for (j = 0; j < 16; ++j) {
@@ -231,7 +220,7 @@ static void DC16NoTop(uint8_t *dst) {   // DC with top samples not available
   Put16(DC >> 4, dst);
 }
 
-static void DC16NoLeft(uint8_t *dst) {  // DC with left samples not available
+static void DC16NoLeft(uint8_t* dst) {  // DC with left samples not available
   int DC = 8;
   int i;
   for (i = 0; i < 16; ++i) {
@@ -240,17 +229,19 @@ static void DC16NoLeft(uint8_t *dst) {  // DC with left samples not available
   Put16(DC >> 4, dst);
 }
 
-static void DC16NoTopLeft(uint8_t *dst) {  // DC with no top and left samples
+static void DC16NoTopLeft(uint8_t* dst) {  // DC with no top and left samples
   Put16(0x80, dst);
 }
 
+VP8PredFunc VP8PredLuma16[NUM_B_DC_MODES];
+
 //------------------------------------------------------------------------------
 // 4x4
 
 #define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
 
-static void VE4(uint8_t *dst) {    // vertical
+static void VE4(uint8_t* dst) {    // vertical
   const uint8_t* top = dst - BPS;
   const uint8_t vals[4] = {
     AVG3(top[-1], top[0], top[1]),
@@ -264,7 +255,7 @@ static void VE4(uint8_t *dst) {    // vertical
   }
 }
 
-static void HE4(uint8_t *dst) {    // horizontal
+static void HE4(uint8_t* dst) {    // horizontal
   const int A = dst[-1 - BPS];
   const int B = dst[-1];
   const int C = dst[-1 + BPS];
@@ -276,7 +267,7 @@ static void HE4(uint8_t *dst) {    // horizontal
   *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(D, E, E);
 }
 
-static void DC4(uint8_t *dst) {   // DC
+static void DC4(uint8_t* dst) {   // DC
   uint32_t dc = 4;
   int i;
   for (i = 0; i < 4; ++i) dc += dst[i - BPS] + dst[-1 + i * BPS];
@@ -284,7 +275,7 @@ static void DC4(uint8_t *dst) {   // DC
   for (i = 0; i < 4; ++i) memset(dst + i * BPS, dc, 4);
 }
 
-static void RD4(uint8_t *dst) {   // Down-right
+static void RD4(uint8_t* dst) {   // Down-right
   const int I = dst[-1 + 0 * BPS];
   const int J = dst[-1 + 1 * BPS];
   const int K = dst[-1 + 2 * BPS];
@@ -295,15 +286,15 @@ static void RD4(uint8_t *dst) {   // Down-right
   const int C = dst[2 - BPS];
   const int D = dst[3 - BPS];
   DST(0, 3)                                     = AVG3(J, K, L);
-  DST(0, 2) = DST(1, 3)                         = AVG3(I, J, K);
-  DST(0, 1) = DST(1, 2) = DST(2, 3)             = AVG3(X, I, J);
-  DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I);
-  DST(1, 0) = DST(2, 1) = DST(3, 2)             = AVG3(B, A, X);
-  DST(2, 0) = DST(3, 1)                         = AVG3(C, B, A);
-  DST(3, 0)                                     = AVG3(D, C, B);
+  DST(1, 3) = DST(0, 2)                         = AVG3(I, J, K);
+  DST(2, 3) = DST(1, 2) = DST(0, 1)             = AVG3(X, I, J);
+  DST(3, 3) = DST(2, 2) = DST(1, 1) = DST(0, 0) = AVG3(A, X, I);
+              DST(3, 2) = DST(2, 1) = DST(1, 0) = AVG3(B, A, X);
+                          DST(3, 1) = DST(2, 0) = AVG3(C, B, A);
+                                      DST(3, 0) = AVG3(D, C, B);
 }
 
-static void LD4(uint8_t *dst) {   // Down-Left
+static void LD4(uint8_t* dst) {   // Down-Left
   const int A = dst[0 - BPS];
   const int B = dst[1 - BPS];
   const int C = dst[2 - BPS];
@@ -316,12 +307,12 @@ static void LD4(uint8_t *dst) {   // Down-Left
   DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
   DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
   DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
-  DST(3, 1) = DST(2, 2) = DST(1, 3)             = AVG3(E, F, G);
-  DST(3, 2) = DST(2, 3)                         = AVG3(F, G, H);
-  DST(3, 3)                                     = AVG3(G, H, H);
+              DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G);
+                          DST(3, 2) = DST(2, 3) = AVG3(F, G, H);
+                                      DST(3, 3) = AVG3(G, H, H);
 }
 
-static void VR4(uint8_t *dst) {   // Vertical-Right
+static void VR4(uint8_t* dst) {   // Vertical-Right
   const int I = dst[-1 + 0 * BPS];
   const int J = dst[-1 + 1 * BPS];
   const int K = dst[-1 + 2 * BPS];
@@ -343,7 +334,7 @@ static void VR4(uint8_t *dst) {   // Vertical-Right
   DST(3, 1) =             AVG3(B, C, D);
 }
 
-static void VL4(uint8_t *dst) {   // Vertical-Left
+static void VL4(uint8_t* dst) {   // Vertical-Left
   const int A = dst[0 - BPS];
   const int B = dst[1 - BPS];
   const int C = dst[2 - BPS];
@@ -365,7 +356,7 @@ static void VL4(uint8_t *dst) {   // Vertical-Left
               DST(3, 3) = AVG3(F, G, H);
 }
 
-static void HU4(uint8_t *dst) {   // Horizontal-Up
+static void HU4(uint8_t* dst) {   // Horizontal-Up
   const int I = dst[-1 + 0 * BPS];
   const int J = dst[-1 + 1 * BPS];
   const int K = dst[-1 + 2 * BPS];
@@ -380,7 +371,7 @@ static void HU4(uint8_t *dst) {   // Horizontal-Up
     DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
 }
 
-static void HD4(uint8_t *dst) {  // Horizontal-Down
+static void HD4(uint8_t* dst) {  // Horizontal-Down
   const int I = dst[-1 + 0 * BPS];
   const int J = dst[-1 + 1 * BPS];
   const int K = dst[-1 + 2 * BPS];
@@ -407,17 +398,19 @@ static void HD4(uint8_t *dst) {  // Horizontal-Down
 #undef AVG3
 #undef AVG2
 
+VP8PredFunc VP8PredLuma4[NUM_BMODES];
+
 //------------------------------------------------------------------------------
 // Chroma
 
-static void VE8uv(uint8_t *dst) {    // vertical
+static void VE8uv(uint8_t* dst) {    // vertical
   int j;
   for (j = 0; j < 8; ++j) {
     memcpy(dst + j * BPS, dst - BPS, 8);
   }
 }
 
-static void HE8uv(uint8_t *dst) {    // horizontal
+static void HE8uv(uint8_t* dst) {    // horizontal
   int j;
   for (j = 0; j < 8; ++j) {
     memset(dst, dst[-1], 8);
@@ -426,60 +419,45 @@ static void HE8uv(uint8_t *dst) {    // horizontal
 }
 
 // helper for chroma-DC predictions
-static WEBP_INLINE void Put8x8uv(uint64_t v, uint8_t* dst) {
+static WEBP_INLINE void Put8x8uv(uint8_t value, uint8_t* dst) {
   int j;
   for (j = 0; j < 8; ++j) {
-    *(uint64_t*)(dst + j * BPS) = v;
+    memset(dst + j * BPS, value, 8);
   }
 }
 
-static void DC8uv(uint8_t *dst) {     // DC
+static void DC8uv(uint8_t* dst) {     // DC
   int dc0 = 8;
   int i;
   for (i = 0; i < 8; ++i) {
     dc0 += dst[i - BPS] + dst[-1 + i * BPS];
   }
-  Put8x8uv((uint64_t)((dc0 >> 4) * 0x0101010101010101ULL), dst);
+  Put8x8uv(dc0 >> 4, dst);
 }
 
-static void DC8uvNoLeft(uint8_t *dst) {   // DC with no left samples
+static void DC8uvNoLeft(uint8_t* dst) {   // DC with no left samples
   int dc0 = 4;
   int i;
   for (i = 0; i < 8; ++i) {
     dc0 += dst[i - BPS];
   }
-  Put8x8uv((uint64_t)((dc0 >> 3) * 0x0101010101010101ULL), dst);
+  Put8x8uv(dc0 >> 3, dst);
 }
 
-static void DC8uvNoTop(uint8_t *dst) {  // DC with no top samples
+static void DC8uvNoTop(uint8_t* dst) {  // DC with no top samples
   int dc0 = 4;
   int i;
   for (i = 0; i < 8; ++i) {
     dc0 += dst[-1 + i * BPS];
   }
-  Put8x8uv((uint64_t)((dc0 >> 3) * 0x0101010101010101ULL), dst);
+  Put8x8uv(dc0 >> 3, dst);
 }
 
-static void DC8uvNoTopLeft(uint8_t *dst) {    // DC with nothing
-  Put8x8uv(0x8080808080808080ULL, dst);
+static void DC8uvNoTopLeft(uint8_t* dst) {    // DC with nothing
+  Put8x8uv(0x80, dst);
 }
 
-//------------------------------------------------------------------------------
-// default C implementations
-
-const VP8PredFunc VP8PredLuma4[NUM_BMODES] = {
-  DC4, TM4, VE4, HE4, RD4, VR4, LD4, VL4, HD4, HU4
-};
-
-const VP8PredFunc VP8PredLuma16[NUM_B_DC_MODES] = {
-  DC16, TM16, VE16, HE16,
-  DC16NoTop, DC16NoLeft, DC16NoTopLeft
-};
-
-const VP8PredFunc VP8PredChroma8[NUM_B_DC_MODES] = {
-  DC8uv, TM8uv, VE8uv, HE8uv,
-  DC8uvNoTop, DC8uvNoLeft, DC8uvNoTopLeft
-};
+VP8PredFunc VP8PredChroma8[NUM_B_DC_MODES];
 
 //------------------------------------------------------------------------------
 // Edge filtering functions
@@ -487,61 +465,62 @@ const VP8PredFunc VP8PredChroma8[NUM_B_DC_MODES] = {
 // 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];
+  const int a = 3 * (q0 - p0) + VP8ksclip1[p1 - q1];  // in [-893,892]
+  const int a1 = VP8ksclip2[(a + 4) >> 3];            // in [-16,15]
+  const int a2 = VP8ksclip2[(a + 3) >> 3];
+  p[-step] = VP8kclip1[p0 + a2];
+  p[    0] = VP8kclip1[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 a1 = VP8ksclip2[(a + 4) >> 3];
+  const int a2 = VP8ksclip2[(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];
+  p[-2*step] = VP8kclip1[p1 + a3];
+  p[-  step] = VP8kclip1[p0 + a2];
+  p[      0] = VP8kclip1[q0 - a1];
+  p[   step] = VP8kclip1[q1 - a3];
 }
 
 // 6 pixels in, 6 pixels out
 static WEBP_INLINE void do_filter6(uint8_t* p, int step) {
   const int p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step];
   const int q0 = p[0], q1 = p[step], q2 = p[2*step];
-  const int a = sclip1[1020 + 3 * (q0 - p0) + sclip1[1020 + p1 - q1]];
+  const int a = VP8ksclip1[3 * (q0 - p0) + VP8ksclip1[p1 - q1]];
+  // a is in [-128,127], a1 in [-27,27], a2 in [-18,18] and a3 in [-9,9]
   const int a1 = (27 * a + 63) >> 7;  // eq. to ((3 * a + 7) * 9) >> 7
   const int a2 = (18 * a + 63) >> 7;  // eq. to ((2 * a + 7) * 9) >> 7
   const int a3 = (9  * a + 63) >> 7;  // eq. to ((1 * a + 7) * 9) >> 7
-  p[-3*step] = clip1[255 + p2 + a3];
-  p[-2*step] = clip1[255 + p1 + a2];
-  p[-  step] = clip1[255 + p0 + a1];
-  p[      0] = clip1[255 + q0 - a1];
-  p[   step] = clip1[255 + q1 - a2];
-  p[ 2*step] = clip1[255 + q2 - a3];
+  p[-3*step] = VP8kclip1[p2 + a3];
+  p[-2*step] = VP8kclip1[p1 + a2];
+  p[-  step] = VP8kclip1[p0 + a1];
+  p[      0] = VP8kclip1[q0 - a1];
+  p[   step] = VP8kclip1[q1 - a2];
+  p[ 2*step] = VP8kclip1[q2 - a3];
 }
 
 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);
+  return (VP8kabs0[p1 - p0] > thresh) || (VP8kabs0[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_filter(const uint8_t* p, int step, int t) {
+  const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return ((4 * VP8kabs0[p0 - q0] + VP8kabs0[p1 - q1]) <= t);
 }
 
 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;
+  const int p3 = p[-4 * step], p2 = p[-3 * step], p1 = p[-2 * step];
+  const int p0 = p[-step], q0 = p[0];
+  const int q1 = p[step], q2 = p[2 * step], q3 = p[3 * step];
+  if ((4 * VP8kabs0[p0 - q0] + VP8kabs0[p1 - q1]) > t) return 0;
+  return VP8kabs0[p3 - p2] <= it && VP8kabs0[p2 - p1] <= it &&
+         VP8kabs0[p1 - p0] <= it && VP8kabs0[q3 - q2] <= it &&
+         VP8kabs0[q2 - q1] <= it && VP8kabs0[q1 - q0] <= it;
 }
 
 //------------------------------------------------------------------------------
@@ -549,8 +528,9 @@ static WEBP_INLINE int needs_filter2(const uint8_t* p,
 
 static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
   int i;
+  const int thresh2 = 2 * thresh + 1;
   for (i = 0; i < 16; ++i) {
-    if (needs_filter(p + i, stride, thresh)) {
+    if (needs_filter(p + i, stride, thresh2)) {
       do_filter2(p + i, stride);
     }
   }
@@ -558,8 +538,9 @@ static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
 
 static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
   int i;
+  const int thresh2 = 2 * thresh + 1;
   for (i = 0; i < 16; ++i) {
-    if (needs_filter(p + i * stride, 1, thresh)) {
+    if (needs_filter(p + i * stride, 1, thresh2)) {
       do_filter2(p + i * stride, 1);
     }
   }
@@ -587,8 +568,9 @@ static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
 static WEBP_INLINE void FilterLoop26(uint8_t* p,
                                      int hstride, int vstride, int size,
                                      int thresh, int ithresh, int hev_thresh) {
+  const int thresh2 = 2 * thresh + 1;
   while (size-- > 0) {
-    if (needs_filter2(p, hstride, thresh, ithresh)) {
+    if (needs_filter2(p, hstride, thresh2, ithresh)) {
       if (hev(p, hstride, hev_thresh)) {
         do_filter2(p, hstride);
       } else {
@@ -602,8 +584,9 @@ static WEBP_INLINE void FilterLoop26(uint8_t* p,
 static WEBP_INLINE void FilterLoop24(uint8_t* p,
                                      int hstride, int vstride, int size,
                                      int thresh, int ithresh, int hev_thresh) {
+  const int thresh2 = 2 * thresh + 1;
   while (size-- > 0) {
-    if (needs_filter2(p, hstride, thresh, ithresh)) {
+    if (needs_filter2(p, hstride, thresh2, ithresh)) {
       if (hev(p, hstride, hev_thresh)) {
         do_filter2(p, hstride);
       } else {
@@ -672,6 +655,7 @@ static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
 //------------------------------------------------------------------------------
 
 VP8DecIdct2 VP8Transform;
+VP8DecIdct VP8TransformAC3;
 VP8DecIdct VP8TransformUV;
 VP8DecIdct VP8TransformDC;
 VP8DecIdct VP8TransformDCUV;
@@ -690,15 +674,25 @@ VP8SimpleFilterFunc VP8SimpleVFilter16i;
 VP8SimpleFilterFunc VP8SimpleHFilter16i;
 
 extern void VP8DspInitSSE2(void);
+extern void VP8DspInitSSE41(void);
 extern void VP8DspInitNEON(void);
+extern void VP8DspInitMIPS32(void);
+extern void VP8DspInitMIPSdspR2(void);
+
+static volatile VP8CPUInfo dec_last_cpuinfo_used =
+    (VP8CPUInfo)&dec_last_cpuinfo_used;
 
-void VP8DspInit(void) {
-  DspInitTables();
+WEBP_TSAN_IGNORE_FUNCTION void VP8DspInit(void) {
+  if (dec_last_cpuinfo_used == VP8GetCPUInfo) return;
 
+  VP8InitClipTables();
+
+  VP8TransformWHT = TransformWHT;
   VP8Transform = TransformTwo;
   VP8TransformUV = TransformUV;
   VP8TransformDC = TransformDC;
   VP8TransformDCUV = TransformDCUV;
+  VP8TransformAC3 = TransformAC3;
 
   VP8VFilter16 = VFilter16;
   VP8HFilter16 = HFilter16;
@@ -713,20 +707,60 @@ void VP8DspInit(void) {
   VP8SimpleVFilter16i = SimpleVFilter16i;
   VP8SimpleHFilter16i = SimpleHFilter16i;
 
+  VP8PredLuma4[0] = DC4;
+  VP8PredLuma4[1] = TM4;
+  VP8PredLuma4[2] = VE4;
+  VP8PredLuma4[3] = HE4;
+  VP8PredLuma4[4] = RD4;
+  VP8PredLuma4[5] = VR4;
+  VP8PredLuma4[6] = LD4;
+  VP8PredLuma4[7] = VL4;
+  VP8PredLuma4[8] = HD4;
+  VP8PredLuma4[9] = HU4;
+
+  VP8PredLuma16[0] = DC16;
+  VP8PredLuma16[1] = TM16;
+  VP8PredLuma16[2] = VE16;
+  VP8PredLuma16[3] = HE16;
+  VP8PredLuma16[4] = DC16NoTop;
+  VP8PredLuma16[5] = DC16NoLeft;
+  VP8PredLuma16[6] = DC16NoTopLeft;
+
+  VP8PredChroma8[0] = DC8uv;
+  VP8PredChroma8[1] = TM8uv;
+  VP8PredChroma8[2] = VE8uv;
+  VP8PredChroma8[3] = HE8uv;
+  VP8PredChroma8[4] = DC8uvNoTop;
+  VP8PredChroma8[5] = DC8uvNoLeft;
+  VP8PredChroma8[6] = DC8uvNoTopLeft;
+
   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
-  if (VP8GetCPUInfo) {
+  if (VP8GetCPUInfo != NULL) {
 #if defined(WEBP_USE_SSE2)
     if (VP8GetCPUInfo(kSSE2)) {
       VP8DspInitSSE2();
+#if defined(WEBP_USE_SSE41)
+      if (VP8GetCPUInfo(kSSE4_1)) {
+        VP8DspInitSSE41();
+      }
+#endif
     }
-#elif defined(WEBP_USE_NEON)
+#endif
+#if defined(WEBP_USE_NEON)
     if (VP8GetCPUInfo(kNEON)) {
       VP8DspInitNEON();
     }
 #endif
+#if defined(WEBP_USE_MIPS32)
+    if (VP8GetCPUInfo(kMIPS32)) {
+      VP8DspInitMIPS32();
+    }
+#endif
+#if defined(WEBP_USE_MIPS_DSP_R2)
+    if (VP8GetCPUInfo(kMIPSdspR2)) {
+      VP8DspInitMIPSdspR2();
+    }
+#endif
   }
+  dec_last_cpuinfo_used = VP8GetCPUInfo;
 }
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dsp/dec_neon.c b/drivers/webp/dsp/dec_neon.c
index ec824b790b..a63f43fe17 100644
--- a/drivers/webp/dsp/dec_neon.c
+++ b/drivers/webp/dsp/dec_neon.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // ARM NEON version of dsp functions and loop filtering.
@@ -14,13 +16,535 @@
 
 #if defined(WEBP_USE_NEON)
 
+#include "./neon.h"
 #include "../dec/vp8i.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+//------------------------------------------------------------------------------
+// NxM Loading functions
+
+// Load/Store vertical edge
+#define LOAD8x4(c1, c2, c3, c4, b1, b2, stride)                                \
+  "vld4.8 {" #c1 "[0]," #c2 "[0]," #c3 "[0]," #c4 "[0]}," #b1 "," #stride "\n" \
+  "vld4.8 {" #c1 "[1]," #c2 "[1]," #c3 "[1]," #c4 "[1]}," #b2 "," #stride "\n" \
+  "vld4.8 {" #c1 "[2]," #c2 "[2]," #c3 "[2]," #c4 "[2]}," #b1 "," #stride "\n" \
+  "vld4.8 {" #c1 "[3]," #c2 "[3]," #c3 "[3]," #c4 "[3]}," #b2 "," #stride "\n" \
+  "vld4.8 {" #c1 "[4]," #c2 "[4]," #c3 "[4]," #c4 "[4]}," #b1 "," #stride "\n" \
+  "vld4.8 {" #c1 "[5]," #c2 "[5]," #c3 "[5]," #c4 "[5]}," #b2 "," #stride "\n" \
+  "vld4.8 {" #c1 "[6]," #c2 "[6]," #c3 "[6]," #c4 "[6]}," #b1 "," #stride "\n" \
+  "vld4.8 {" #c1 "[7]," #c2 "[7]," #c3 "[7]," #c4 "[7]}," #b2 "," #stride "\n"
+
+#define STORE8x2(c1, c2, p, stride)                                            \
+  "vst2.8   {" #c1 "[0], " #c2 "[0]}," #p "," #stride " \n"                    \
+  "vst2.8   {" #c1 "[1], " #c2 "[1]}," #p "," #stride " \n"                    \
+  "vst2.8   {" #c1 "[2], " #c2 "[2]}," #p "," #stride " \n"                    \
+  "vst2.8   {" #c1 "[3], " #c2 "[3]}," #p "," #stride " \n"                    \
+  "vst2.8   {" #c1 "[4], " #c2 "[4]}," #p "," #stride " \n"                    \
+  "vst2.8   {" #c1 "[5], " #c2 "[5]}," #p "," #stride " \n"                    \
+  "vst2.8   {" #c1 "[6], " #c2 "[6]}," #p "," #stride " \n"                    \
+  "vst2.8   {" #c1 "[7], " #c2 "[7]}," #p "," #stride " \n"
+
+#if !defined(WORK_AROUND_GCC)
+
+// This intrinsics version makes gcc-4.6.3 crash during Load4x??() compilation
+// (register alloc, probably). The variants somewhat mitigate the problem, but
+// not quite. HFilter16i() remains problematic.
+static WEBP_INLINE uint8x8x4_t Load4x8(const uint8_t* const src, int stride) {
+  const uint8x8_t zero = vdup_n_u8(0);
+  uint8x8x4_t out;
+  INIT_VECTOR4(out, zero, zero, zero, zero);
+  out = vld4_lane_u8(src + 0 * stride, out, 0);
+  out = vld4_lane_u8(src + 1 * stride, out, 1);
+  out = vld4_lane_u8(src + 2 * stride, out, 2);
+  out = vld4_lane_u8(src + 3 * stride, out, 3);
+  out = vld4_lane_u8(src + 4 * stride, out, 4);
+  out = vld4_lane_u8(src + 5 * stride, out, 5);
+  out = vld4_lane_u8(src + 6 * stride, out, 6);
+  out = vld4_lane_u8(src + 7 * stride, out, 7);
+  return out;
+}
+
+static WEBP_INLINE void Load4x16(const uint8_t* const src, int stride,
+                                 uint8x16_t* const p1, uint8x16_t* const p0,
+                                 uint8x16_t* const q0, uint8x16_t* const q1) {
+  // row0 = p1[0..7]|p0[0..7]|q0[0..7]|q1[0..7]
+  // row8 = p1[8..15]|p0[8..15]|q0[8..15]|q1[8..15]
+  const uint8x8x4_t row0 = Load4x8(src - 2 + 0 * stride, stride);
+  const uint8x8x4_t row8 = Load4x8(src - 2 + 8 * stride, stride);
+  *p1 = vcombine_u8(row0.val[0], row8.val[0]);
+  *p0 = vcombine_u8(row0.val[1], row8.val[1]);
+  *q0 = vcombine_u8(row0.val[2], row8.val[2]);
+  *q1 = vcombine_u8(row0.val[3], row8.val[3]);
+}
+
+#else  // WORK_AROUND_GCC
+
+#define LOADQ_LANE_32b(VALUE, LANE) do {                             \
+  (VALUE) = vld1q_lane_u32((const uint32_t*)src, (VALUE), (LANE));   \
+  src += stride;                                                     \
+} while (0)
+
+static WEBP_INLINE void Load4x16(const uint8_t* src, int stride,
+                                 uint8x16_t* const p1, uint8x16_t* const p0,
+                                 uint8x16_t* const q0, uint8x16_t* const q1) {
+  const uint32x4_t zero = vdupq_n_u32(0);
+  uint32x4x4_t in;
+  INIT_VECTOR4(in, zero, zero, zero, zero);
+  src -= 2;
+  LOADQ_LANE_32b(in.val[0], 0);
+  LOADQ_LANE_32b(in.val[1], 0);
+  LOADQ_LANE_32b(in.val[2], 0);
+  LOADQ_LANE_32b(in.val[3], 0);
+  LOADQ_LANE_32b(in.val[0], 1);
+  LOADQ_LANE_32b(in.val[1], 1);
+  LOADQ_LANE_32b(in.val[2], 1);
+  LOADQ_LANE_32b(in.val[3], 1);
+  LOADQ_LANE_32b(in.val[0], 2);
+  LOADQ_LANE_32b(in.val[1], 2);
+  LOADQ_LANE_32b(in.val[2], 2);
+  LOADQ_LANE_32b(in.val[3], 2);
+  LOADQ_LANE_32b(in.val[0], 3);
+  LOADQ_LANE_32b(in.val[1], 3);
+  LOADQ_LANE_32b(in.val[2], 3);
+  LOADQ_LANE_32b(in.val[3], 3);
+  // Transpose four 4x4 parts:
+  {
+    const uint8x16x2_t row01 = vtrnq_u8(vreinterpretq_u8_u32(in.val[0]),
+                                        vreinterpretq_u8_u32(in.val[1]));
+    const uint8x16x2_t row23 = vtrnq_u8(vreinterpretq_u8_u32(in.val[2]),
+                                        vreinterpretq_u8_u32(in.val[3]));
+    const uint16x8x2_t row02 = vtrnq_u16(vreinterpretq_u16_u8(row01.val[0]),
+                                         vreinterpretq_u16_u8(row23.val[0]));
+    const uint16x8x2_t row13 = vtrnq_u16(vreinterpretq_u16_u8(row01.val[1]),
+                                         vreinterpretq_u16_u8(row23.val[1]));
+    *p1 = vreinterpretq_u8_u16(row02.val[0]);
+    *p0 = vreinterpretq_u8_u16(row13.val[0]);
+    *q0 = vreinterpretq_u8_u16(row02.val[1]);
+    *q1 = vreinterpretq_u8_u16(row13.val[1]);
+  }
+}
+#undef LOADQ_LANE_32b
+
+#endif  // !WORK_AROUND_GCC
+
+static WEBP_INLINE void Load8x16(const uint8_t* const src, int stride,
+                                 uint8x16_t* const p3, uint8x16_t* const p2,
+                                 uint8x16_t* const p1, uint8x16_t* const p0,
+                                 uint8x16_t* const q0, uint8x16_t* const q1,
+                                 uint8x16_t* const q2, uint8x16_t* const q3) {
+  Load4x16(src - 2, stride, p3, p2, p1, p0);
+  Load4x16(src + 2, stride, q0, q1, q2, q3);
+}
+
+static WEBP_INLINE void Load16x4(const uint8_t* const src, int stride,
+                                 uint8x16_t* const p1, uint8x16_t* const p0,
+                                 uint8x16_t* const q0, uint8x16_t* const q1) {
+  *p1 = vld1q_u8(src - 2 * stride);
+  *p0 = vld1q_u8(src - 1 * stride);
+  *q0 = vld1q_u8(src + 0 * stride);
+  *q1 = vld1q_u8(src + 1 * stride);
+}
+
+static WEBP_INLINE void Load16x8(const uint8_t* const src, int stride,
+                                 uint8x16_t* const p3, uint8x16_t* const p2,
+                                 uint8x16_t* const p1, uint8x16_t* const p0,
+                                 uint8x16_t* const q0, uint8x16_t* const q1,
+                                 uint8x16_t* const q2, uint8x16_t* const q3) {
+  Load16x4(src - 2  * stride, stride, p3, p2, p1, p0);
+  Load16x4(src + 2  * stride, stride, q0, q1, q2, q3);
+}
+
+static WEBP_INLINE void Load8x8x2(const uint8_t* const u,
+                                  const uint8_t* const v,
+                                  int stride,
+                                  uint8x16_t* const p3, uint8x16_t* const p2,
+                                  uint8x16_t* const p1, uint8x16_t* const p0,
+                                  uint8x16_t* const q0, uint8x16_t* const q1,
+                                  uint8x16_t* const q2, uint8x16_t* const q3) {
+  // We pack the 8x8 u-samples in the lower half of the uint8x16_t destination
+  // and the v-samples on the higher half.
+  *p3 = vcombine_u8(vld1_u8(u - 4 * stride), vld1_u8(v - 4 * stride));
+  *p2 = vcombine_u8(vld1_u8(u - 3 * stride), vld1_u8(v - 3 * stride));
+  *p1 = vcombine_u8(vld1_u8(u - 2 * stride), vld1_u8(v - 2 * stride));
+  *p0 = vcombine_u8(vld1_u8(u - 1 * stride), vld1_u8(v - 1 * stride));
+  *q0 = vcombine_u8(vld1_u8(u + 0 * stride), vld1_u8(v + 0 * stride));
+  *q1 = vcombine_u8(vld1_u8(u + 1 * stride), vld1_u8(v + 1 * stride));
+  *q2 = vcombine_u8(vld1_u8(u + 2 * stride), vld1_u8(v + 2 * stride));
+  *q3 = vcombine_u8(vld1_u8(u + 3 * stride), vld1_u8(v + 3 * stride));
+}
+
+#if !defined(WORK_AROUND_GCC)
+
+#define LOAD_UV_8(ROW) \
+  vcombine_u8(vld1_u8(u - 4 + (ROW) * stride), vld1_u8(v - 4 + (ROW) * stride))
+
+static WEBP_INLINE void Load8x8x2T(const uint8_t* const u,
+                                   const uint8_t* const v,
+                                   int stride,
+                                   uint8x16_t* const p3, uint8x16_t* const p2,
+                                   uint8x16_t* const p1, uint8x16_t* const p0,
+                                   uint8x16_t* const q0, uint8x16_t* const q1,
+                                   uint8x16_t* const q2, uint8x16_t* const q3) {
+  // We pack the 8x8 u-samples in the lower half of the uint8x16_t destination
+  // and the v-samples on the higher half.
+  const uint8x16_t row0 = LOAD_UV_8(0);
+  const uint8x16_t row1 = LOAD_UV_8(1);
+  const uint8x16_t row2 = LOAD_UV_8(2);
+  const uint8x16_t row3 = LOAD_UV_8(3);
+  const uint8x16_t row4 = LOAD_UV_8(4);
+  const uint8x16_t row5 = LOAD_UV_8(5);
+  const uint8x16_t row6 = LOAD_UV_8(6);
+  const uint8x16_t row7 = LOAD_UV_8(7);
+  // Perform two side-by-side 8x8 transposes
+  // u00 u01 u02 u03 u04 u05 u06 u07 | v00 v01 v02 v03 v04 v05 v06 v07
+  // u10 u11 u12 u13 u14 u15 u16 u17 | v10 v11 v12 ...
+  // u20 u21 u22 u23 u24 u25 u26 u27 | v20 v21 ...
+  // u30 u31 u32 u33 u34 u35 u36 u37 | ...
+  // u40 u41 u42 u43 u44 u45 u46 u47 | ...
+  // u50 u51 u52 u53 u54 u55 u56 u57 | ...
+  // u60 u61 u62 u63 u64 u65 u66 u67 | v60 ...
+  // u70 u71 u72 u73 u74 u75 u76 u77 | v70 v71 v72 ...
+  const uint8x16x2_t row01 = vtrnq_u8(row0, row1);  // u00 u10 u02 u12 ...
+                                                    // u01 u11 u03 u13 ...
+  const uint8x16x2_t row23 = vtrnq_u8(row2, row3);  // u20 u30 u22 u32 ...
+                                                    // u21 u31 u23 u33 ...
+  const uint8x16x2_t row45 = vtrnq_u8(row4, row5);  // ...
+  const uint8x16x2_t row67 = vtrnq_u8(row6, row7);  // ...
+  const uint16x8x2_t row02 = vtrnq_u16(vreinterpretq_u16_u8(row01.val[0]),
+                                       vreinterpretq_u16_u8(row23.val[0]));
+  const uint16x8x2_t row13 = vtrnq_u16(vreinterpretq_u16_u8(row01.val[1]),
+                                       vreinterpretq_u16_u8(row23.val[1]));
+  const uint16x8x2_t row46 = vtrnq_u16(vreinterpretq_u16_u8(row45.val[0]),
+                                       vreinterpretq_u16_u8(row67.val[0]));
+  const uint16x8x2_t row57 = vtrnq_u16(vreinterpretq_u16_u8(row45.val[1]),
+                                       vreinterpretq_u16_u8(row67.val[1]));
+  const uint32x4x2_t row04 = vtrnq_u32(vreinterpretq_u32_u16(row02.val[0]),
+                                       vreinterpretq_u32_u16(row46.val[0]));
+  const uint32x4x2_t row26 = vtrnq_u32(vreinterpretq_u32_u16(row02.val[1]),
+                                       vreinterpretq_u32_u16(row46.val[1]));
+  const uint32x4x2_t row15 = vtrnq_u32(vreinterpretq_u32_u16(row13.val[0]),
+                                       vreinterpretq_u32_u16(row57.val[0]));
+  const uint32x4x2_t row37 = vtrnq_u32(vreinterpretq_u32_u16(row13.val[1]),
+                                       vreinterpretq_u32_u16(row57.val[1]));
+  *p3 = vreinterpretq_u8_u32(row04.val[0]);
+  *p2 = vreinterpretq_u8_u32(row15.val[0]);
+  *p1 = vreinterpretq_u8_u32(row26.val[0]);
+  *p0 = vreinterpretq_u8_u32(row37.val[0]);
+  *q0 = vreinterpretq_u8_u32(row04.val[1]);
+  *q1 = vreinterpretq_u8_u32(row15.val[1]);
+  *q2 = vreinterpretq_u8_u32(row26.val[1]);
+  *q3 = vreinterpretq_u8_u32(row37.val[1]);
+}
+#undef LOAD_UV_8
+
+#endif  // !WORK_AROUND_GCC
+
+static WEBP_INLINE void Store2x8(const uint8x8x2_t v,
+                                 uint8_t* const dst, int stride) {
+  vst2_lane_u8(dst + 0 * stride, v, 0);
+  vst2_lane_u8(dst + 1 * stride, v, 1);
+  vst2_lane_u8(dst + 2 * stride, v, 2);
+  vst2_lane_u8(dst + 3 * stride, v, 3);
+  vst2_lane_u8(dst + 4 * stride, v, 4);
+  vst2_lane_u8(dst + 5 * stride, v, 5);
+  vst2_lane_u8(dst + 6 * stride, v, 6);
+  vst2_lane_u8(dst + 7 * stride, v, 7);
+}
 
-#define QRegs "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",                  \
+static WEBP_INLINE void Store2x16(const uint8x16_t p0, const uint8x16_t q0,
+                                  uint8_t* const dst, int stride) {
+  uint8x8x2_t lo, hi;
+  lo.val[0] = vget_low_u8(p0);
+  lo.val[1] = vget_low_u8(q0);
+  hi.val[0] = vget_high_u8(p0);
+  hi.val[1] = vget_high_u8(q0);
+  Store2x8(lo, dst - 1 + 0 * stride, stride);
+  Store2x8(hi, dst - 1 + 8 * stride, stride);
+}
+
+#if !defined(WORK_AROUND_GCC)
+static WEBP_INLINE void Store4x8(const uint8x8x4_t v,
+                                 uint8_t* const dst, int stride) {
+  vst4_lane_u8(dst + 0 * stride, v, 0);
+  vst4_lane_u8(dst + 1 * stride, v, 1);
+  vst4_lane_u8(dst + 2 * stride, v, 2);
+  vst4_lane_u8(dst + 3 * stride, v, 3);
+  vst4_lane_u8(dst + 4 * stride, v, 4);
+  vst4_lane_u8(dst + 5 * stride, v, 5);
+  vst4_lane_u8(dst + 6 * stride, v, 6);
+  vst4_lane_u8(dst + 7 * stride, v, 7);
+}
+
+static WEBP_INLINE void Store4x16(const uint8x16_t p1, const uint8x16_t p0,
+                                  const uint8x16_t q0, const uint8x16_t q1,
+                                  uint8_t* const dst, int stride) {
+  uint8x8x4_t lo, hi;
+  INIT_VECTOR4(lo,
+               vget_low_u8(p1), vget_low_u8(p0),
+               vget_low_u8(q0), vget_low_u8(q1));
+  INIT_VECTOR4(hi,
+               vget_high_u8(p1), vget_high_u8(p0),
+               vget_high_u8(q0), vget_high_u8(q1));
+  Store4x8(lo, dst - 2 + 0 * stride, stride);
+  Store4x8(hi, dst - 2 + 8 * stride, stride);
+}
+#endif  // !WORK_AROUND_GCC
+
+static WEBP_INLINE void Store16x2(const uint8x16_t p0, const uint8x16_t q0,
+                                  uint8_t* const dst, int stride) {
+  vst1q_u8(dst - stride, p0);
+  vst1q_u8(dst, q0);
+}
+
+static WEBP_INLINE void Store16x4(const uint8x16_t p1, const uint8x16_t p0,
+                                  const uint8x16_t q0, const uint8x16_t q1,
+                                  uint8_t* const dst, int stride) {
+  Store16x2(p1, p0, dst - stride, stride);
+  Store16x2(q0, q1, dst + stride, stride);
+}
+
+static WEBP_INLINE void Store8x2x2(const uint8x16_t p0, const uint8x16_t q0,
+                                   uint8_t* const u, uint8_t* const v,
+                                   int stride) {
+  // p0 and q0 contain the u+v samples packed in low/high halves.
+  vst1_u8(u - stride, vget_low_u8(p0));
+  vst1_u8(u,          vget_low_u8(q0));
+  vst1_u8(v - stride, vget_high_u8(p0));
+  vst1_u8(v,          vget_high_u8(q0));
+}
+
+static WEBP_INLINE void Store8x4x2(const uint8x16_t p1, const uint8x16_t p0,
+                                   const uint8x16_t q0, const uint8x16_t q1,
+                                   uint8_t* const u, uint8_t* const v,
+                                   int stride) {
+  // The p1...q1 registers contain the u+v samples packed in low/high halves.
+  Store8x2x2(p1, p0, u - stride, v - stride, stride);
+  Store8x2x2(q0, q1, u + stride, v + stride, stride);
+}
+
+#if !defined(WORK_AROUND_GCC)
+
+#define STORE6_LANE(DST, VAL0, VAL1, LANE) do {   \
+  vst3_lane_u8((DST) - 3, (VAL0), (LANE));        \
+  vst3_lane_u8((DST) + 0, (VAL1), (LANE));        \
+  (DST) += stride;                                \
+} while (0)
+
+static WEBP_INLINE void Store6x8x2(const uint8x16_t p2, const uint8x16_t p1,
+                                   const uint8x16_t p0, const uint8x16_t q0,
+                                   const uint8x16_t q1, const uint8x16_t q2,
+                                   uint8_t* u, uint8_t* v,
+                                   int stride) {
+  uint8x8x3_t u0, u1, v0, v1;
+  INIT_VECTOR3(u0, vget_low_u8(p2), vget_low_u8(p1), vget_low_u8(p0));
+  INIT_VECTOR3(u1, vget_low_u8(q0), vget_low_u8(q1), vget_low_u8(q2));
+  INIT_VECTOR3(v0, vget_high_u8(p2), vget_high_u8(p1), vget_high_u8(p0));
+  INIT_VECTOR3(v1, vget_high_u8(q0), vget_high_u8(q1), vget_high_u8(q2));
+  STORE6_LANE(u, u0, u1, 0);
+  STORE6_LANE(u, u0, u1, 1);
+  STORE6_LANE(u, u0, u1, 2);
+  STORE6_LANE(u, u0, u1, 3);
+  STORE6_LANE(u, u0, u1, 4);
+  STORE6_LANE(u, u0, u1, 5);
+  STORE6_LANE(u, u0, u1, 6);
+  STORE6_LANE(u, u0, u1, 7);
+  STORE6_LANE(v, v0, v1, 0);
+  STORE6_LANE(v, v0, v1, 1);
+  STORE6_LANE(v, v0, v1, 2);
+  STORE6_LANE(v, v0, v1, 3);
+  STORE6_LANE(v, v0, v1, 4);
+  STORE6_LANE(v, v0, v1, 5);
+  STORE6_LANE(v, v0, v1, 6);
+  STORE6_LANE(v, v0, v1, 7);
+}
+#undef STORE6_LANE
+
+static WEBP_INLINE void Store4x8x2(const uint8x16_t p1, const uint8x16_t p0,
+                                   const uint8x16_t q0, const uint8x16_t q1,
+                                   uint8_t* const u, uint8_t* const v,
+                                   int stride) {
+  uint8x8x4_t u0, v0;
+  INIT_VECTOR4(u0,
+               vget_low_u8(p1), vget_low_u8(p0),
+               vget_low_u8(q0), vget_low_u8(q1));
+  INIT_VECTOR4(v0,
+               vget_high_u8(p1), vget_high_u8(p0),
+               vget_high_u8(q0), vget_high_u8(q1));
+  vst4_lane_u8(u - 2 + 0 * stride, u0, 0);
+  vst4_lane_u8(u - 2 + 1 * stride, u0, 1);
+  vst4_lane_u8(u - 2 + 2 * stride, u0, 2);
+  vst4_lane_u8(u - 2 + 3 * stride, u0, 3);
+  vst4_lane_u8(u - 2 + 4 * stride, u0, 4);
+  vst4_lane_u8(u - 2 + 5 * stride, u0, 5);
+  vst4_lane_u8(u - 2 + 6 * stride, u0, 6);
+  vst4_lane_u8(u - 2 + 7 * stride, u0, 7);
+  vst4_lane_u8(v - 2 + 0 * stride, v0, 0);
+  vst4_lane_u8(v - 2 + 1 * stride, v0, 1);
+  vst4_lane_u8(v - 2 + 2 * stride, v0, 2);
+  vst4_lane_u8(v - 2 + 3 * stride, v0, 3);
+  vst4_lane_u8(v - 2 + 4 * stride, v0, 4);
+  vst4_lane_u8(v - 2 + 5 * stride, v0, 5);
+  vst4_lane_u8(v - 2 + 6 * stride, v0, 6);
+  vst4_lane_u8(v - 2 + 7 * stride, v0, 7);
+}
+
+#endif  // !WORK_AROUND_GCC
+
+// Zero extend 'v' to an int16x8_t.
+static WEBP_INLINE int16x8_t ConvertU8ToS16(uint8x8_t v) {
+  return vreinterpretq_s16_u16(vmovl_u8(v));
+}
+
+// Performs unsigned 8b saturation on 'dst01' and 'dst23' storing the result
+// to the corresponding rows of 'dst'.
+static WEBP_INLINE void SaturateAndStore4x4(uint8_t* const dst,
+                                            const int16x8_t dst01,
+                                            const int16x8_t dst23) {
+  // Unsigned saturate to 8b.
+  const uint8x8_t dst01_u8 = vqmovun_s16(dst01);
+  const uint8x8_t dst23_u8 = vqmovun_s16(dst23);
+
+  // Store the results.
+  vst1_lane_u32((uint32_t*)(dst + 0 * BPS), vreinterpret_u32_u8(dst01_u8), 0);
+  vst1_lane_u32((uint32_t*)(dst + 1 * BPS), vreinterpret_u32_u8(dst01_u8), 1);
+  vst1_lane_u32((uint32_t*)(dst + 2 * BPS), vreinterpret_u32_u8(dst23_u8), 0);
+  vst1_lane_u32((uint32_t*)(dst + 3 * BPS), vreinterpret_u32_u8(dst23_u8), 1);
+}
+
+static WEBP_INLINE void Add4x4(const int16x8_t row01, const int16x8_t row23,
+                               uint8_t* const dst) {
+  uint32x2_t dst01 = vdup_n_u32(0);
+  uint32x2_t dst23 = vdup_n_u32(0);
+
+  // Load the source pixels.
+  dst01 = vld1_lane_u32((uint32_t*)(dst + 0 * BPS), dst01, 0);
+  dst23 = vld1_lane_u32((uint32_t*)(dst + 2 * BPS), dst23, 0);
+  dst01 = vld1_lane_u32((uint32_t*)(dst + 1 * BPS), dst01, 1);
+  dst23 = vld1_lane_u32((uint32_t*)(dst + 3 * BPS), dst23, 1);
+
+  {
+    // Convert to 16b.
+    const int16x8_t dst01_s16 = ConvertU8ToS16(vreinterpret_u8_u32(dst01));
+    const int16x8_t dst23_s16 = ConvertU8ToS16(vreinterpret_u8_u32(dst23));
+
+    // Descale with rounding.
+    const int16x8_t out01 = vrsraq_n_s16(dst01_s16, row01, 3);
+    const int16x8_t out23 = vrsraq_n_s16(dst23_s16, row23, 3);
+    // Add the inverse transform.
+    SaturateAndStore4x4(dst, out01, out23);
+  }
+}
+
+//-----------------------------------------------------------------------------
+// Simple In-loop filtering (Paragraph 15.2)
+
+static uint8x16_t NeedsFilter(const uint8x16_t p1, const uint8x16_t p0,
+                              const uint8x16_t q0, const uint8x16_t q1,
+                              int thresh) {
+  const uint8x16_t thresh_v = vdupq_n_u8((uint8_t)thresh);
+  const uint8x16_t a_p0_q0 = vabdq_u8(p0, q0);               // abs(p0-q0)
+  const uint8x16_t a_p1_q1 = vabdq_u8(p1, q1);               // abs(p1-q1)
+  const uint8x16_t a_p0_q0_2 = vqaddq_u8(a_p0_q0, a_p0_q0);  // 2 * abs(p0-q0)
+  const uint8x16_t a_p1_q1_2 = vshrq_n_u8(a_p1_q1, 1);       // abs(p1-q1) / 2
+  const uint8x16_t sum = vqaddq_u8(a_p0_q0_2, a_p1_q1_2);
+  const uint8x16_t mask = vcgeq_u8(thresh_v, sum);
+  return mask;
+}
+
+static int8x16_t FlipSign(const uint8x16_t v) {
+  const uint8x16_t sign_bit = vdupq_n_u8(0x80);
+  return vreinterpretq_s8_u8(veorq_u8(v, sign_bit));
+}
+
+static uint8x16_t FlipSignBack(const int8x16_t v) {
+  const int8x16_t sign_bit = vdupq_n_s8(0x80);
+  return vreinterpretq_u8_s8(veorq_s8(v, sign_bit));
+}
+
+static int8x16_t GetBaseDelta(const int8x16_t p1, const int8x16_t p0,
+                              const int8x16_t q0, const int8x16_t q1) {
+  const int8x16_t q0_p0 = vqsubq_s8(q0, p0);      // (q0-p0)
+  const int8x16_t p1_q1 = vqsubq_s8(p1, q1);      // (p1-q1)
+  const int8x16_t s1 = vqaddq_s8(p1_q1, q0_p0);   // (p1-q1) + 1 * (q0 - p0)
+  const int8x16_t s2 = vqaddq_s8(q0_p0, s1);      // (p1-q1) + 2 * (q0 - p0)
+  const int8x16_t s3 = vqaddq_s8(q0_p0, s2);      // (p1-q1) + 3 * (q0 - p0)
+  return s3;
+}
+
+static int8x16_t GetBaseDelta0(const int8x16_t p0, const int8x16_t q0) {
+  const int8x16_t q0_p0 = vqsubq_s8(q0, p0);      // (q0-p0)
+  const int8x16_t s1 = vqaddq_s8(q0_p0, q0_p0);   // 2 * (q0 - p0)
+  const int8x16_t s2 = vqaddq_s8(q0_p0, s1);      // 3 * (q0 - p0)
+  return s2;
+}
+
+//------------------------------------------------------------------------------
+
+static void ApplyFilter2NoFlip(const int8x16_t p0s, const int8x16_t q0s,
+                               const int8x16_t delta,
+                               int8x16_t* const op0, int8x16_t* const oq0) {
+  const int8x16_t kCst3 = vdupq_n_s8(0x03);
+  const int8x16_t kCst4 = vdupq_n_s8(0x04);
+  const int8x16_t delta_p3 = vqaddq_s8(delta, kCst3);
+  const int8x16_t delta_p4 = vqaddq_s8(delta, kCst4);
+  const int8x16_t delta3 = vshrq_n_s8(delta_p3, 3);
+  const int8x16_t delta4 = vshrq_n_s8(delta_p4, 3);
+  *op0 = vqaddq_s8(p0s, delta3);
+  *oq0 = vqsubq_s8(q0s, delta4);
+}
+
+#if defined(WEBP_USE_INTRINSICS)
+
+static void ApplyFilter2(const int8x16_t p0s, const int8x16_t q0s,
+                         const int8x16_t delta,
+                         uint8x16_t* const op0, uint8x16_t* const oq0) {
+  const int8x16_t kCst3 = vdupq_n_s8(0x03);
+  const int8x16_t kCst4 = vdupq_n_s8(0x04);
+  const int8x16_t delta_p3 = vqaddq_s8(delta, kCst3);
+  const int8x16_t delta_p4 = vqaddq_s8(delta, kCst4);
+  const int8x16_t delta3 = vshrq_n_s8(delta_p3, 3);
+  const int8x16_t delta4 = vshrq_n_s8(delta_p4, 3);
+  const int8x16_t sp0 = vqaddq_s8(p0s, delta3);
+  const int8x16_t sq0 = vqsubq_s8(q0s, delta4);
+  *op0 = FlipSignBack(sp0);
+  *oq0 = FlipSignBack(sq0);
+}
+
+static void DoFilter2(const uint8x16_t p1, const uint8x16_t p0,
+                      const uint8x16_t q0, const uint8x16_t q1,
+                      const uint8x16_t mask,
+                      uint8x16_t* const op0, uint8x16_t* const oq0) {
+  const int8x16_t p1s = FlipSign(p1);
+  const int8x16_t p0s = FlipSign(p0);
+  const int8x16_t q0s = FlipSign(q0);
+  const int8x16_t q1s = FlipSign(q1);
+  const int8x16_t delta0 = GetBaseDelta(p1s, p0s, q0s, q1s);
+  const int8x16_t delta1 = vandq_s8(delta0, vreinterpretq_s8_u8(mask));
+  ApplyFilter2(p0s, q0s, delta1, op0, oq0);
+}
+
+static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
+  uint8x16_t p1, p0, q0, q1, op0, oq0;
+  Load16x4(p, stride, &p1, &p0, &q0, &q1);
+  {
+    const uint8x16_t mask = NeedsFilter(p1, p0, q0, q1, thresh);
+    DoFilter2(p1, p0, q0, q1, mask, &op0, &oq0);
+  }
+  Store16x2(op0, oq0, p, stride);
+}
+
+static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
+  uint8x16_t p1, p0, q0, q1, oq0, op0;
+  Load4x16(p, stride, &p1, &p0, &q0, &q1);
+  {
+    const uint8x16_t mask = NeedsFilter(p1, p0, q0, q1, thresh);
+    DoFilter2(p1, p0, q0, q1, mask, &op0, &oq0);
+  }
+  Store2x16(op0, oq0, p, stride);
+}
+
+#else
+
+#define QRegs "q0", "q1", "q2", "q3",                                          \
               "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
 
 #define FLIP_SIGN_BIT2(a, b, s)                                                \
@@ -68,40 +592,16 @@ extern "C" {
   DO_SIMPLE_FILTER(p0, q0, q9)                 /* apply filter */              \
   FLIP_SIGN_BIT2(p0, q0, q10)
 
-// Load/Store vertical edge
-#define LOAD8x4(c1, c2, c3, c4, b1, b2, stride)                                \
-  "vld4.8   {" #c1"[0], " #c2"[0], " #c3"[0], " #c4"[0]}," #b1 "," #stride"\n" \
-  "vld4.8   {" #c1"[1], " #c2"[1], " #c3"[1], " #c4"[1]}," #b2 "," #stride"\n" \
-  "vld4.8   {" #c1"[2], " #c2"[2], " #c3"[2], " #c4"[2]}," #b1 "," #stride"\n" \
-  "vld4.8   {" #c1"[3], " #c2"[3], " #c3"[3], " #c4"[3]}," #b2 "," #stride"\n" \
-  "vld4.8   {" #c1"[4], " #c2"[4], " #c3"[4], " #c4"[4]}," #b1 "," #stride"\n" \
-  "vld4.8   {" #c1"[5], " #c2"[5], " #c3"[5], " #c4"[5]}," #b2 "," #stride"\n" \
-  "vld4.8   {" #c1"[6], " #c2"[6], " #c3"[6], " #c4"[6]}," #b1 "," #stride"\n" \
-  "vld4.8   {" #c1"[7], " #c2"[7], " #c3"[7], " #c4"[7]}," #b2 "," #stride"\n"
-
-#define STORE8x2(c1, c2, p,stride)                                             \
-  "vst2.8   {" #c1"[0], " #c2"[0]}," #p "," #stride " \n"                      \
-  "vst2.8   {" #c1"[1], " #c2"[1]}," #p "," #stride " \n"                      \
-  "vst2.8   {" #c1"[2], " #c2"[2]}," #p "," #stride " \n"                      \
-  "vst2.8   {" #c1"[3], " #c2"[3]}," #p "," #stride " \n"                      \
-  "vst2.8   {" #c1"[4], " #c2"[4]}," #p "," #stride " \n"                      \
-  "vst2.8   {" #c1"[5], " #c2"[5]}," #p "," #stride " \n"                      \
-  "vst2.8   {" #c1"[6], " #c2"[6]}," #p "," #stride " \n"                      \
-  "vst2.8   {" #c1"[7], " #c2"[7]}," #p "," #stride " \n"
-
-//-----------------------------------------------------------------------------
-// Simple In-loop filtering (Paragraph 15.2)
-
-static void SimpleVFilter16NEON(uint8_t* p, int stride, int thresh) {
+static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
   __asm__ volatile (
     "sub        %[p], %[p], %[stride], lsl #1  \n"  // p -= 2 * stride
 
     "vld1.u8    {q1}, [%[p]], %[stride]        \n"  // p1
     "vld1.u8    {q2}, [%[p]], %[stride]        \n"  // p0
     "vld1.u8    {q3}, [%[p]], %[stride]        \n"  // q0
-    "vld1.u8    {q4}, [%[p]]                   \n"  // q1
+    "vld1.u8    {q12}, [%[p]]                  \n"  // q1
 
-    DO_FILTER2(q1, q2, q3, q4, %[thresh])
+    DO_FILTER2(q1, q2, q3, q12, %[thresh])
 
     "sub        %[p], %[p], %[stride], lsl #1  \n"  // p -= 2 * stride
 
@@ -113,25 +613,25 @@ static void SimpleVFilter16NEON(uint8_t* p, int stride, int thresh) {
   );
 }
 
-static void SimpleHFilter16NEON(uint8_t* p, int stride, int thresh) {
+static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
   __asm__ volatile (
     "sub        r4, %[p], #2                   \n"  // base1 = p - 2
     "lsl        r6, %[stride], #1              \n"  // r6 = 2 * stride
     "add        r5, r4, %[stride]              \n"  // base2 = base1 + stride
 
     LOAD8x4(d2, d3, d4, d5, [r4], [r5], r6)
-    LOAD8x4(d6, d7, d8, d9, [r4], [r5], r6)
-    "vswp       d3, d6                         \n"  // p1:q1 p0:q3
-    "vswp       d5, d8                         \n"  // q0:q2 q1:q4
-    "vswp       q2, q3                         \n"  // p1:q1 p0:q2 q0:q3 q1:q4
+    LOAD8x4(d24, d25, d26, d27, [r4], [r5], r6)
+    "vswp       d3, d24                        \n"  // p1:q1 p0:q3
+    "vswp       d5, d26                        \n"  // q0:q2 q1:q4
+    "vswp       q2, q12                        \n"  // p1:q1 p0:q2 q0:q3 q1:q4
 
-    DO_FILTER2(q1, q2, q3, q4, %[thresh])
+    DO_FILTER2(q1, q2, q12, q13, %[thresh])
 
     "sub        %[p], %[p], #1                 \n"  // p - 1
 
-    "vswp        d5, d6                        \n"
+    "vswp        d5, d24                       \n"
     STORE8x2(d4, d5, [%[p]], %[stride])
-    STORE8x2(d6, d7, [%[p]], %[stride])
+    STORE8x2(d24, d25, [%[p]], %[stride])
 
     : [p] "+r"(p)
     : [stride] "r"(stride), [thresh] "r"(thresh)
@@ -139,44 +639,408 @@ static void SimpleHFilter16NEON(uint8_t* p, int stride, int thresh) {
   );
 }
 
-static void SimpleVFilter16iNEON(uint8_t* p, int stride, int thresh) {
-  int k;
-  for (k = 3; k > 0; --k) {
+#endif    // WEBP_USE_INTRINSICS
+
+static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
+  uint32_t k;
+  for (k = 3; k != 0; --k) {
     p += 4 * stride;
-    SimpleVFilter16NEON(p, stride, thresh);
+    SimpleVFilter16(p, stride, thresh);
   }
 }
 
-static void SimpleHFilter16iNEON(uint8_t* p, int stride, int thresh) {
-  int k;
-  for (k = 3; k > 0; --k) {
+static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
+  uint32_t k;
+  for (k = 3; k != 0; --k) {
     p += 4;
-    SimpleHFilter16NEON(p, stride, thresh);
+    SimpleHFilter16(p, stride, thresh);
   }
 }
 
-static void TransformOneNEON(const int16_t *in, uint8_t *dst) {
-  const int kBPS = BPS;
-  const int16_t constants[] = {20091, 17734, 0, 0};
-  /* kC1, kC2. Padded because vld1.16 loads 8 bytes
-   * Technically these are unsigned but vqdmulh is only available in signed.
-   * vqdmulh returns high half (effectively >> 16) but also doubles the value,
-   * changing the >> 16 to >> 15 and requiring an additional >> 1.
-   * We use this to our advantage with kC2. The canonical value is 35468.
-   * However, the high bit is set so treating it as signed will give incorrect
-   * results. We avoid this by down shifting by 1 here to clear the highest bit.
-   * Combined with the doubling effect of vqdmulh we get >> 16.
-   * This can not be applied to kC1 because the lowest bit is set. Down shifting
-   * the constant would reduce precision.
-   */
-
-  /* libwebp uses a trick to avoid some extra addition that libvpx does.
-   * Instead of:
-   * temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1) >> 16);
-   * libwebp adds 1 << 16 to cospi8sqrt2minus1 (kC1). However, this causes the
-   * same issue with kC1 and vqdmulh that we work around by down shifting kC2
-   */
+//------------------------------------------------------------------------------
+// Complex In-loop filtering (Paragraph 15.3)
+
+static uint8x16_t NeedsHev(const uint8x16_t p1, const uint8x16_t p0,
+                           const uint8x16_t q0, const uint8x16_t q1,
+                           int hev_thresh) {
+  const uint8x16_t hev_thresh_v = vdupq_n_u8((uint8_t)hev_thresh);
+  const uint8x16_t a_p1_p0 = vabdq_u8(p1, p0);  // abs(p1 - p0)
+  const uint8x16_t a_q1_q0 = vabdq_u8(q1, q0);  // abs(q1 - q0)
+  const uint8x16_t mask1 = vcgtq_u8(a_p1_p0, hev_thresh_v);
+  const uint8x16_t mask2 = vcgtq_u8(a_q1_q0, hev_thresh_v);
+  const uint8x16_t mask = vorrq_u8(mask1, mask2);
+  return mask;
+}
+
+static uint8x16_t NeedsFilter2(const uint8x16_t p3, const uint8x16_t p2,
+                               const uint8x16_t p1, const uint8x16_t p0,
+                               const uint8x16_t q0, const uint8x16_t q1,
+                               const uint8x16_t q2, const uint8x16_t q3,
+                               int ithresh, int thresh) {
+  const uint8x16_t ithresh_v = vdupq_n_u8((uint8_t)ithresh);
+  const uint8x16_t a_p3_p2 = vabdq_u8(p3, p2);  // abs(p3 - p2)
+  const uint8x16_t a_p2_p1 = vabdq_u8(p2, p1);  // abs(p2 - p1)
+  const uint8x16_t a_p1_p0 = vabdq_u8(p1, p0);  // abs(p1 - p0)
+  const uint8x16_t a_q3_q2 = vabdq_u8(q3, q2);  // abs(q3 - q2)
+  const uint8x16_t a_q2_q1 = vabdq_u8(q2, q1);  // abs(q2 - q1)
+  const uint8x16_t a_q1_q0 = vabdq_u8(q1, q0);  // abs(q1 - q0)
+  const uint8x16_t max1 = vmaxq_u8(a_p3_p2, a_p2_p1);
+  const uint8x16_t max2 = vmaxq_u8(a_p1_p0, a_q3_q2);
+  const uint8x16_t max3 = vmaxq_u8(a_q2_q1, a_q1_q0);
+  const uint8x16_t max12 = vmaxq_u8(max1, max2);
+  const uint8x16_t max123 = vmaxq_u8(max12, max3);
+  const uint8x16_t mask2 = vcgeq_u8(ithresh_v, max123);
+  const uint8x16_t mask1 = NeedsFilter(p1, p0, q0, q1, thresh);
+  const uint8x16_t mask = vandq_u8(mask1, mask2);
+  return mask;
+}
+
+//  4-points filter
+
+static void ApplyFilter4(
+    const int8x16_t p1, const int8x16_t p0,
+    const int8x16_t q0, const int8x16_t q1,
+    const int8x16_t delta0,
+    uint8x16_t* const op1, uint8x16_t* const op0,
+    uint8x16_t* const oq0, uint8x16_t* const oq1) {
+  const int8x16_t kCst3 = vdupq_n_s8(0x03);
+  const int8x16_t kCst4 = vdupq_n_s8(0x04);
+  const int8x16_t delta1 = vqaddq_s8(delta0, kCst4);
+  const int8x16_t delta2 = vqaddq_s8(delta0, kCst3);
+  const int8x16_t a1 = vshrq_n_s8(delta1, 3);
+  const int8x16_t a2 = vshrq_n_s8(delta2, 3);
+  const int8x16_t a3 = vrshrq_n_s8(a1, 1);   // a3 = (a1 + 1) >> 1
+  *op0 = FlipSignBack(vqaddq_s8(p0, a2));  // clip(p0 + a2)
+  *oq0 = FlipSignBack(vqsubq_s8(q0, a1));  // clip(q0 - a1)
+  *op1 = FlipSignBack(vqaddq_s8(p1, a3));  // clip(p1 + a3)
+  *oq1 = FlipSignBack(vqsubq_s8(q1, a3));  // clip(q1 - a3)
+}
+
+static void DoFilter4(
+    const uint8x16_t p1, const uint8x16_t p0,
+    const uint8x16_t q0, const uint8x16_t q1,
+    const uint8x16_t mask, const uint8x16_t hev_mask,
+    uint8x16_t* const op1, uint8x16_t* const op0,
+    uint8x16_t* const oq0, uint8x16_t* const oq1) {
+  // This is a fused version of DoFilter2() calling ApplyFilter2 directly
+  const int8x16_t p1s = FlipSign(p1);
+  int8x16_t p0s = FlipSign(p0);
+  int8x16_t q0s = FlipSign(q0);
+  const int8x16_t q1s = FlipSign(q1);
+  const uint8x16_t simple_lf_mask = vandq_u8(mask, hev_mask);
+
+  // do_filter2 part (simple loopfilter on pixels with hev)
+  {
+    const int8x16_t delta = GetBaseDelta(p1s, p0s, q0s, q1s);
+    const int8x16_t simple_lf_delta =
+        vandq_s8(delta, vreinterpretq_s8_u8(simple_lf_mask));
+    ApplyFilter2NoFlip(p0s, q0s, simple_lf_delta, &p0s, &q0s);
+  }
+
+  // do_filter4 part (complex loopfilter on pixels without hev)
+  {
+    const int8x16_t delta0 = GetBaseDelta0(p0s, q0s);
+    // we use: (mask & hev_mask) ^ mask = mask & !hev_mask
+    const uint8x16_t complex_lf_mask = veorq_u8(simple_lf_mask, mask);
+    const int8x16_t complex_lf_delta =
+        vandq_s8(delta0, vreinterpretq_s8_u8(complex_lf_mask));
+    ApplyFilter4(p1s, p0s, q0s, q1s, complex_lf_delta, op1, op0, oq0, oq1);
+  }
+}
+
+//  6-points filter
+
+static void ApplyFilter6(
+    const int8x16_t p2, const int8x16_t p1, const int8x16_t p0,
+    const int8x16_t q0, const int8x16_t q1, const int8x16_t q2,
+    const int8x16_t delta,
+    uint8x16_t* const op2, uint8x16_t* const op1, uint8x16_t* const op0,
+    uint8x16_t* const oq0, uint8x16_t* const oq1, uint8x16_t* const oq2) {
+  const int16x8_t kCst63 = vdupq_n_s16(63);
+  const int8x8_t kCst27 = vdup_n_s8(27);
+  const int8x8_t kCst18 = vdup_n_s8(18);
+  const int8x8_t kCst9 = vdup_n_s8(9);
+  const int8x8_t delta_lo = vget_low_s8(delta);
+  const int8x8_t delta_hi = vget_high_s8(delta);
+  const int16x8_t s1_lo = vmlal_s8(kCst63, kCst27, delta_lo);  // 63 + 27 * a
+  const int16x8_t s1_hi = vmlal_s8(kCst63, kCst27, delta_hi);  // 63 + 27 * a
+  const int16x8_t s2_lo = vmlal_s8(kCst63, kCst18, delta_lo);  // 63 + 18 * a
+  const int16x8_t s2_hi = vmlal_s8(kCst63, kCst18, delta_hi);  // 63 + 18 * a
+  const int16x8_t s3_lo = vmlal_s8(kCst63, kCst9, delta_lo);   // 63 + 9 * a
+  const int16x8_t s3_hi = vmlal_s8(kCst63, kCst9, delta_hi);   // 63 + 9 * a
+  const int8x8_t a1_lo = vqshrn_n_s16(s1_lo, 7);
+  const int8x8_t a1_hi = vqshrn_n_s16(s1_hi, 7);
+  const int8x8_t a2_lo = vqshrn_n_s16(s2_lo, 7);
+  const int8x8_t a2_hi = vqshrn_n_s16(s2_hi, 7);
+  const int8x8_t a3_lo = vqshrn_n_s16(s3_lo, 7);
+  const int8x8_t a3_hi = vqshrn_n_s16(s3_hi, 7);
+  const int8x16_t a1 = vcombine_s8(a1_lo, a1_hi);
+  const int8x16_t a2 = vcombine_s8(a2_lo, a2_hi);
+  const int8x16_t a3 = vcombine_s8(a3_lo, a3_hi);
+
+  *op0 = FlipSignBack(vqaddq_s8(p0, a1));  // clip(p0 + a1)
+  *oq0 = FlipSignBack(vqsubq_s8(q0, a1));  // clip(q0 - q1)
+  *oq1 = FlipSignBack(vqsubq_s8(q1, a2));  // clip(q1 - a2)
+  *op1 = FlipSignBack(vqaddq_s8(p1, a2));  // clip(p1 + a2)
+  *oq2 = FlipSignBack(vqsubq_s8(q2, a3));  // clip(q2 - a3)
+  *op2 = FlipSignBack(vqaddq_s8(p2, a3));  // clip(p2 + a3)
+}
+
+static void DoFilter6(
+    const uint8x16_t p2, const uint8x16_t p1, const uint8x16_t p0,
+    const uint8x16_t q0, const uint8x16_t q1, const uint8x16_t q2,
+    const uint8x16_t mask, const uint8x16_t hev_mask,
+    uint8x16_t* const op2, uint8x16_t* const op1, uint8x16_t* const op0,
+    uint8x16_t* const oq0, uint8x16_t* const oq1, uint8x16_t* const oq2) {
+  // This is a fused version of DoFilter2() calling ApplyFilter2 directly
+  const int8x16_t p2s = FlipSign(p2);
+  const int8x16_t p1s = FlipSign(p1);
+  int8x16_t p0s = FlipSign(p0);
+  int8x16_t q0s = FlipSign(q0);
+  const int8x16_t q1s = FlipSign(q1);
+  const int8x16_t q2s = FlipSign(q2);
+  const uint8x16_t simple_lf_mask = vandq_u8(mask, hev_mask);
+  const int8x16_t delta0 = GetBaseDelta(p1s, p0s, q0s, q1s);
+
+  // do_filter2 part (simple loopfilter on pixels with hev)
+  {
+    const int8x16_t simple_lf_delta =
+        vandq_s8(delta0, vreinterpretq_s8_u8(simple_lf_mask));
+    ApplyFilter2NoFlip(p0s, q0s, simple_lf_delta, &p0s, &q0s);
+  }
+
+  // do_filter6 part (complex loopfilter on pixels without hev)
+  {
+    // we use: (mask & hev_mask) ^ mask = mask & !hev_mask
+    const uint8x16_t complex_lf_mask = veorq_u8(simple_lf_mask, mask);
+    const int8x16_t complex_lf_delta =
+        vandq_s8(delta0, vreinterpretq_s8_u8(complex_lf_mask));
+    ApplyFilter6(p2s, p1s, p0s, q0s, q1s, q2s, complex_lf_delta,
+                 op2, op1, op0, oq0, oq1, oq2);
+  }
+}
+
+// on macroblock edges
+
+static void VFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
+  Load16x8(p, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  {
+    const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3,
+                                         ithresh, thresh);
+    const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh);
+    uint8x16_t op2, op1, op0, oq0, oq1, oq2;
+    DoFilter6(p2, p1, p0, q0, q1, q2, mask, hev_mask,
+              &op2, &op1, &op0, &oq0, &oq1, &oq2);
+    Store16x2(op2, op1, p - 2 * stride, stride);
+    Store16x2(op0, oq0, p + 0 * stride, stride);
+    Store16x2(oq1, oq2, p + 2 * stride, stride);
+  }
+}
+
+static void HFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
+  Load8x16(p, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  {
+    const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3,
+                                         ithresh, thresh);
+    const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh);
+    uint8x16_t op2, op1, op0, oq0, oq1, oq2;
+    DoFilter6(p2, p1, p0, q0, q1, q2, mask, hev_mask,
+              &op2, &op1, &op0, &oq0, &oq1, &oq2);
+    Store2x16(op2, op1, p - 2, stride);
+    Store2x16(op0, oq0, p + 0, stride);
+    Store2x16(oq1, oq2, p + 2, stride);
+  }
+}
+
+// on three inner edges
+static void VFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
+  uint32_t k;
+  uint8x16_t p3, p2, p1, p0;
+  Load16x4(p + 2  * stride, stride, &p3, &p2, &p1, &p0);
+  for (k = 3; k != 0; --k) {
+    uint8x16_t q0, q1, q2, q3;
+    p += 4 * stride;
+    Load16x4(p + 2  * stride, stride, &q0, &q1, &q2, &q3);
+    {
+      const uint8x16_t mask =
+          NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, ithresh, thresh);
+      const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh);
+      // p3 and p2 are not just temporary variables here: they will be
+      // re-used for next span. And q2/q3 will become p1/p0 accordingly.
+      DoFilter4(p1, p0, q0, q1, mask, hev_mask, &p1, &p0, &p3, &p2);
+      Store16x4(p1, p0, p3, p2, p, stride);
+      p1 = q2;
+      p0 = q3;
+    }
+  }
+}
+
+#if !defined(WORK_AROUND_GCC)
+static void HFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
+  uint32_t k;
+  uint8x16_t p3, p2, p1, p0;
+  Load4x16(p + 2, stride, &p3, &p2, &p1, &p0);
+  for (k = 3; k != 0; --k) {
+    uint8x16_t q0, q1, q2, q3;
+    p += 4;
+    Load4x16(p + 2, stride, &q0, &q1, &q2, &q3);
+    {
+      const uint8x16_t mask =
+          NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3, ithresh, thresh);
+      const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh);
+      DoFilter4(p1, p0, q0, q1, mask, hev_mask, &p1, &p0, &p3, &p2);
+      Store4x16(p1, p0, p3, p2, p, stride);
+      p1 = q2;
+      p0 = q3;
+    }
+  }
+}
+#endif  // !WORK_AROUND_GCC
+
+// 8-pixels wide variant, for chroma filtering
+static void VFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
+  uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
+  Load8x8x2(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  {
+    const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3,
+                                         ithresh, thresh);
+    const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh);
+    uint8x16_t op2, op1, op0, oq0, oq1, oq2;
+    DoFilter6(p2, p1, p0, q0, q1, q2, mask, hev_mask,
+              &op2, &op1, &op0, &oq0, &oq1, &oq2);
+    Store8x2x2(op2, op1, u - 2 * stride, v - 2 * stride, stride);
+    Store8x2x2(op0, oq0, u + 0 * stride, v + 0 * stride, stride);
+    Store8x2x2(oq1, oq2, u + 2 * stride, v + 2 * stride, stride);
+  }
+}
+static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
+  u += 4 * stride;
+  v += 4 * stride;
+  Load8x8x2(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  {
+    const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3,
+                                         ithresh, thresh);
+    const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh);
+    uint8x16_t op1, op0, oq0, oq1;
+    DoFilter4(p1, p0, q0, q1, mask, hev_mask, &op1, &op0, &oq0, &oq1);
+    Store8x4x2(op1, op0, oq0, oq1, u, v, stride);
+  }
+}
+
+#if !defined(WORK_AROUND_GCC)
+static void HFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
+  uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
+  Load8x8x2T(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  {
+    const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3,
+                                         ithresh, thresh);
+    const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh);
+    uint8x16_t op2, op1, op0, oq0, oq1, oq2;
+    DoFilter6(p2, p1, p0, q0, q1, q2, mask, hev_mask,
+              &op2, &op1, &op0, &oq0, &oq1, &oq2);
+    Store6x8x2(op2, op1, op0, oq0, oq1, oq2, u, v, stride);
+  }
+}
+
+static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  uint8x16_t p3, p2, p1, p0, q0, q1, q2, q3;
+  u += 4;
+  v += 4;
+  Load8x8x2T(u, v, stride, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  {
+    const uint8x16_t mask = NeedsFilter2(p3, p2, p1, p0, q0, q1, q2, q3,
+                                         ithresh, thresh);
+    const uint8x16_t hev_mask = NeedsHev(p1, p0, q0, q1, hev_thresh);
+    uint8x16_t op1, op0, oq0, oq1;
+    DoFilter4(p1, p0, q0, q1, mask, hev_mask, &op1, &op0, &oq0, &oq1);
+    Store4x8x2(op1, op0, oq0, oq1, u, v, stride);
+  }
+}
+#endif  // !WORK_AROUND_GCC
+
+//-----------------------------------------------------------------------------
+// Inverse transforms (Paragraph 14.4)
+
+// Technically these are unsigned but vqdmulh is only available in signed.
+// vqdmulh returns high half (effectively >> 16) but also doubles the value,
+// changing the >> 16 to >> 15 and requiring an additional >> 1.
+// We use this to our advantage with kC2. The canonical value is 35468.
+// However, the high bit is set so treating it as signed will give incorrect
+// results. We avoid this by down shifting by 1 here to clear the highest bit.
+// Combined with the doubling effect of vqdmulh we get >> 16.
+// This can not be applied to kC1 because the lowest bit is set. Down shifting
+// the constant would reduce precision.
+
+// libwebp uses a trick to avoid some extra addition that libvpx does.
+// Instead of:
+// temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1) >> 16);
+// libwebp adds 1 << 16 to cospi8sqrt2minus1 (kC1). However, this causes the
+// same issue with kC1 and vqdmulh that we work around by down shifting kC2
 
+static const int16_t kC1 = 20091;
+static const int16_t kC2 = 17734;  // half of kC2, actually. See comment above.
+
+#if defined(WEBP_USE_INTRINSICS)
+static WEBP_INLINE void Transpose8x2(const int16x8_t in0, const int16x8_t in1,
+                                     int16x8x2_t* const out) {
+  // a0 a1 a2 a3 | b0 b1 b2 b3   => a0 b0 c0 d0 | a1 b1 c1 d1
+  // c0 c1 c2 c3 | d0 d1 d2 d3      a2 b2 c2 d2 | a3 b3 c3 d3
+  const int16x8x2_t tmp0 = vzipq_s16(in0, in1);   // a0 c0 a1 c1 a2 c2 ...
+                                                  // b0 d0 b1 d1 b2 d2 ...
+  *out = vzipq_s16(tmp0.val[0], tmp0.val[1]);
+}
+
+static WEBP_INLINE void TransformPass(int16x8x2_t* const rows) {
+  // {rows} = in0 | in4
+  //          in8 | in12
+  // B1 = in4 | in12
+  const int16x8_t B1 =
+      vcombine_s16(vget_high_s16(rows->val[0]), vget_high_s16(rows->val[1]));
+  // C0 = kC1 * in4 | kC1 * in12
+  // C1 = kC2 * in4 | kC2 * in12
+  const int16x8_t C0 = vsraq_n_s16(B1, vqdmulhq_n_s16(B1, kC1), 1);
+  const int16x8_t C1 = vqdmulhq_n_s16(B1, kC2);
+  const int16x4_t a = vqadd_s16(vget_low_s16(rows->val[0]),
+                                vget_low_s16(rows->val[1]));   // in0 + in8
+  const int16x4_t b = vqsub_s16(vget_low_s16(rows->val[0]),
+                                vget_low_s16(rows->val[1]));   // in0 - in8
+  // c = kC2 * in4 - kC1 * in12
+  // d = kC1 * in4 + kC2 * in12
+  const int16x4_t c = vqsub_s16(vget_low_s16(C1), vget_high_s16(C0));
+  const int16x4_t d = vqadd_s16(vget_low_s16(C0), vget_high_s16(C1));
+  const int16x8_t D0 = vcombine_s16(a, b);      // D0 = a | b
+  const int16x8_t D1 = vcombine_s16(d, c);      // D1 = d | c
+  const int16x8_t E0 = vqaddq_s16(D0, D1);      // a+d | b+c
+  const int16x8_t E_tmp = vqsubq_s16(D0, D1);   // a-d | b-c
+  const int16x8_t E1 = vcombine_s16(vget_high_s16(E_tmp), vget_low_s16(E_tmp));
+  Transpose8x2(E0, E1, rows);
+}
+
+static void TransformOne(const int16_t* in, uint8_t* dst) {
+  int16x8x2_t rows;
+  INIT_VECTOR2(rows, vld1q_s16(in + 0), vld1q_s16(in + 8));
+  TransformPass(&rows);
+  TransformPass(&rows);
+  Add4x4(rows.val[0], rows.val[1], dst);
+}
+
+#else
+
+static void TransformOne(const int16_t* in, uint8_t* dst) {
+  const int kBPS = BPS;
+  // kC1, kC2. Padded because vld1.16 loads 8 bytes
+  const int16_t constants[4] = { kC1, kC2, 0, 0 };
   /* Adapted from libvpx: vp8/common/arm/neon/shortidct4x4llm_neon.asm */
   __asm__ volatile (
     "vld1.16         {q1, q2}, [%[in]]           \n"
@@ -304,26 +1168,472 @@ static void TransformOneNEON(const int16_t *in, uint8_t *dst) {
   );
 }
 
-static void TransformTwoNEON(const int16_t* in, uint8_t* dst, int do_two) {
-  TransformOneNEON(in, dst);
+#endif    // WEBP_USE_INTRINSICS
+
+static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
+  TransformOne(in, dst);
   if (do_two) {
-    TransformOneNEON(in + 16, dst + 4);
+    TransformOne(in + 16, dst + 4);
   }
 }
 
-extern void VP8DspInitNEON(void);
+static void TransformDC(const int16_t* in, uint8_t* dst) {
+  const int16x8_t DC = vdupq_n_s16(in[0]);
+  Add4x4(DC, DC, dst);
+}
+
+//------------------------------------------------------------------------------
+
+#define STORE_WHT(dst, col, rows) do {                  \
+  *dst = vgetq_lane_s32(rows.val[0], col); (dst) += 16; \
+  *dst = vgetq_lane_s32(rows.val[1], col); (dst) += 16; \
+  *dst = vgetq_lane_s32(rows.val[2], col); (dst) += 16; \
+  *dst = vgetq_lane_s32(rows.val[3], col); (dst) += 16; \
+} while (0)
+
+static void TransformWHT(const int16_t* in, int16_t* out) {
+  int32x4x4_t tmp;
 
-void VP8DspInitNEON(void) {
-  VP8Transform = TransformTwoNEON;
+  {
+    // Load the source.
+    const int16x4_t in00_03 = vld1_s16(in + 0);
+    const int16x4_t in04_07 = vld1_s16(in + 4);
+    const int16x4_t in08_11 = vld1_s16(in + 8);
+    const int16x4_t in12_15 = vld1_s16(in + 12);
+    const int32x4_t a0 = vaddl_s16(in00_03, in12_15);  // in[0..3] + in[12..15]
+    const int32x4_t a1 = vaddl_s16(in04_07, in08_11);  // in[4..7] + in[8..11]
+    const int32x4_t a2 = vsubl_s16(in04_07, in08_11);  // in[4..7] - in[8..11]
+    const int32x4_t a3 = vsubl_s16(in00_03, in12_15);  // in[0..3] - in[12..15]
+    tmp.val[0] = vaddq_s32(a0, a1);
+    tmp.val[1] = vaddq_s32(a3, a2);
+    tmp.val[2] = vsubq_s32(a0, a1);
+    tmp.val[3] = vsubq_s32(a3, a2);
+    // Arrange the temporary results column-wise.
+    tmp = Transpose4x4(tmp);
+  }
+
+  {
+    const int32x4_t kCst3 = vdupq_n_s32(3);
+    const int32x4_t dc = vaddq_s32(tmp.val[0], kCst3);  // add rounder
+    const int32x4_t a0 = vaddq_s32(dc, tmp.val[3]);
+    const int32x4_t a1 = vaddq_s32(tmp.val[1], tmp.val[2]);
+    const int32x4_t a2 = vsubq_s32(tmp.val[1], tmp.val[2]);
+    const int32x4_t a3 = vsubq_s32(dc, tmp.val[3]);
+
+    tmp.val[0] = vaddq_s32(a0, a1);
+    tmp.val[1] = vaddq_s32(a3, a2);
+    tmp.val[2] = vsubq_s32(a0, a1);
+    tmp.val[3] = vsubq_s32(a3, a2);
+
+    // right shift the results by 3.
+    tmp.val[0] = vshrq_n_s32(tmp.val[0], 3);
+    tmp.val[1] = vshrq_n_s32(tmp.val[1], 3);
+    tmp.val[2] = vshrq_n_s32(tmp.val[2], 3);
+    tmp.val[3] = vshrq_n_s32(tmp.val[3], 3);
 
-  VP8SimpleVFilter16 = SimpleVFilter16NEON;
-  VP8SimpleHFilter16 = SimpleHFilter16NEON;
-  VP8SimpleVFilter16i = SimpleVFilter16iNEON;
-  VP8SimpleHFilter16i = SimpleHFilter16iNEON;
+    STORE_WHT(out, 0, tmp);
+    STORE_WHT(out, 1, tmp);
+    STORE_WHT(out, 2, tmp);
+    STORE_WHT(out, 3, tmp);
+  }
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
+#undef STORE_WHT
+
+//------------------------------------------------------------------------------
+
+#define MUL(a, b) (((a) * (b)) >> 16)
+static void TransformAC3(const int16_t* in, uint8_t* dst) {
+  static const int kC1_full = 20091 + (1 << 16);
+  static const int kC2_full = 35468;
+  const int16x4_t A = vld1_dup_s16(in);
+  const int16x4_t c4 = vdup_n_s16(MUL(in[4], kC2_full));
+  const int16x4_t d4 = vdup_n_s16(MUL(in[4], kC1_full));
+  const int c1 = MUL(in[1], kC2_full);
+  const int d1 = MUL(in[1], kC1_full);
+  const uint64_t cd = (uint64_t)( d1 & 0xffff) <<  0 |
+                      (uint64_t)( c1 & 0xffff) << 16 |
+                      (uint64_t)(-c1 & 0xffff) << 32 |
+                      (uint64_t)(-d1 & 0xffff) << 48;
+  const int16x4_t CD = vcreate_s16(cd);
+  const int16x4_t B = vqadd_s16(A, CD);
+  const int16x8_t m0_m1 = vcombine_s16(vqadd_s16(B, d4), vqadd_s16(B, c4));
+  const int16x8_t m2_m3 = vcombine_s16(vqsub_s16(B, c4), vqsub_s16(B, d4));
+  Add4x4(m0_m1, m2_m3, dst);
+}
+#undef MUL
+
+//------------------------------------------------------------------------------
+// 4x4
+
+static void DC4(uint8_t* dst) {    // DC
+  const uint8x8_t A = vld1_u8(dst - BPS);  // top row
+  const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+  const uint16x4_t p1 = vpadd_u16(p0, p0);
+  const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + 0 * BPS - 1));
+  const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + 1 * BPS - 1));
+  const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + 2 * BPS - 1));
+  const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + 3 * BPS - 1));
+  const uint16x8_t s0 = vaddq_u16(L0, L1);
+  const uint16x8_t s1 = vaddq_u16(L2, L3);
+  const uint16x8_t s01 = vaddq_u16(s0, s1);
+  const uint16x8_t sum = vaddq_u16(s01, vcombine_u16(p1, p1));
+  const uint8x8_t dc0 = vrshrn_n_u16(sum, 3);  // (sum + 4) >> 3
+  const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    vst1_lane_u32((uint32_t*)(dst + i * BPS), vreinterpret_u32_u8(dc), 0);
+  }
+}
+
+// TrueMotion (4x4 + 8x8)
+static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) {
+  const uint8x8_t TL = vld1_dup_u8(dst - BPS - 1);  // top-left pixel 'A[-1]'
+  const uint8x8_t T = vld1_u8(dst - BPS);  // top row 'A[0..3]'
+  const int16x8_t d = vreinterpretq_s16_u16(vsubl_u8(T, TL));  // A[c] - A[-1]
+  int y;
+  for (y = 0; y < size; y += 4) {
+    // left edge
+    const int16x8_t L0 = ConvertU8ToS16(vld1_dup_u8(dst + 0 * BPS - 1));
+    const int16x8_t L1 = ConvertU8ToS16(vld1_dup_u8(dst + 1 * BPS - 1));
+    const int16x8_t L2 = ConvertU8ToS16(vld1_dup_u8(dst + 2 * BPS - 1));
+    const int16x8_t L3 = ConvertU8ToS16(vld1_dup_u8(dst + 3 * BPS - 1));
+    const int16x8_t r0 = vaddq_s16(L0, d);  // L[r] + A[c] - A[-1]
+    const int16x8_t r1 = vaddq_s16(L1, d);
+    const int16x8_t r2 = vaddq_s16(L2, d);
+    const int16x8_t r3 = vaddq_s16(L3, d);
+    // Saturate and store the result.
+    const uint32x2_t r0_u32 = vreinterpret_u32_u8(vqmovun_s16(r0));
+    const uint32x2_t r1_u32 = vreinterpret_u32_u8(vqmovun_s16(r1));
+    const uint32x2_t r2_u32 = vreinterpret_u32_u8(vqmovun_s16(r2));
+    const uint32x2_t r3_u32 = vreinterpret_u32_u8(vqmovun_s16(r3));
+    if (size == 4) {
+      vst1_lane_u32((uint32_t*)(dst + 0 * BPS), r0_u32, 0);
+      vst1_lane_u32((uint32_t*)(dst + 1 * BPS), r1_u32, 0);
+      vst1_lane_u32((uint32_t*)(dst + 2 * BPS), r2_u32, 0);
+      vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3_u32, 0);
+    } else {
+      vst1_u32((uint32_t*)(dst + 0 * BPS), r0_u32);
+      vst1_u32((uint32_t*)(dst + 1 * BPS), r1_u32);
+      vst1_u32((uint32_t*)(dst + 2 * BPS), r2_u32);
+      vst1_u32((uint32_t*)(dst + 3 * BPS), r3_u32);
+    }
+    dst += 4 * BPS;
+  }
+}
+
+static void TM4(uint8_t* dst) { TrueMotion(dst, 4); }
+
+static void VE4(uint8_t* dst) {    // vertical
+  // NB: avoid vld1_u64 here as an alignment hint may be added -> SIGBUS.
+  const uint64x1_t A0 = vreinterpret_u64_u8(vld1_u8(dst - BPS - 1));  // top row
+  const uint64x1_t A1 = vshr_n_u64(A0, 8);
+  const uint64x1_t A2 = vshr_n_u64(A0, 16);
+  const uint8x8_t ABCDEFGH = vreinterpret_u8_u64(A0);
+  const uint8x8_t BCDEFGH0 = vreinterpret_u8_u64(A1);
+  const uint8x8_t CDEFGH00 = vreinterpret_u8_u64(A2);
+  const uint8x8_t b = vhadd_u8(ABCDEFGH, CDEFGH00);
+  const uint8x8_t avg = vrhadd_u8(b, BCDEFGH0);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    vst1_lane_u32((uint32_t*)(dst + i * BPS), vreinterpret_u32_u8(avg), 0);
+  }
+}
+
+static void RD4(uint8_t* dst) {   // Down-right
+  const uint8x8_t XABCD_u8 = vld1_u8(dst - BPS - 1);
+  const uint64x1_t XABCD = vreinterpret_u64_u8(XABCD_u8);
+  const uint64x1_t ____XABC = vshl_n_u64(XABCD, 32);
+  const uint32_t I = dst[-1 + 0 * BPS];
+  const uint32_t J = dst[-1 + 1 * BPS];
+  const uint32_t K = dst[-1 + 2 * BPS];
+  const uint32_t L = dst[-1 + 3 * BPS];
+  const uint64x1_t LKJI____ = vcreate_u64(L | (K << 8) | (J << 16) | (I << 24));
+  const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC);
+  const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8));
+  const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16));
+  const uint8_t D = vget_lane_u8(XABCD_u8, 4);
+  const uint8x8_t JIXABCD_ = vset_lane_u8(D, JIXABC__, 6);
+  const uint8x8_t LKJIXABC_u8 = vreinterpret_u8_u64(LKJIXABC);
+  const uint8x8_t avg1 = vhadd_u8(JIXABCD_, LKJIXABC_u8);
+  const uint8x8_t avg2 = vrhadd_u8(avg1, KJIXABC_);
+  const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
+  const uint32x2_t r3 = vreinterpret_u32_u8(avg2);
+  const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
+  const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
+  const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
+  vst1_lane_u32((uint32_t*)(dst + 0 * BPS), r0, 0);
+  vst1_lane_u32((uint32_t*)(dst + 1 * BPS), r1, 0);
+  vst1_lane_u32((uint32_t*)(dst + 2 * BPS), r2, 0);
+  vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3, 0);
+}
+
+static void LD4(uint8_t* dst) {    // Down-left
+  // Note using the same shift trick as VE4() is slower here.
+  const uint8x8_t ABCDEFGH = vld1_u8(dst - BPS + 0);
+  const uint8x8_t BCDEFGH0 = vld1_u8(dst - BPS + 1);
+  const uint8x8_t CDEFGH00 = vld1_u8(dst - BPS + 2);
+  const uint8x8_t CDEFGHH0 = vset_lane_u8(dst[-BPS + 7], CDEFGH00, 6);
+  const uint8x8_t avg1 = vhadd_u8(ABCDEFGH, CDEFGHH0);
+  const uint8x8_t avg2 = vrhadd_u8(avg1, BCDEFGH0);
+  const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
+  const uint32x2_t r0 = vreinterpret_u32_u8(avg2);
+  const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
+  const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
+  const uint32x2_t r3 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
+  vst1_lane_u32((uint32_t*)(dst + 0 * BPS), r0, 0);
+  vst1_lane_u32((uint32_t*)(dst + 1 * BPS), r1, 0);
+  vst1_lane_u32((uint32_t*)(dst + 2 * BPS), r2, 0);
+  vst1_lane_u32((uint32_t*)(dst + 3 * BPS), r3, 0);
+}
+
+//------------------------------------------------------------------------------
+// Chroma
+
+static void VE8uv(uint8_t* dst) {    // vertical
+  const uint8x8_t top = vld1_u8(dst - BPS);
+  int j;
+  for (j = 0; j < 8; ++j) {
+    vst1_u8(dst + j * BPS, top);
+  }
+}
+
+static void HE8uv(uint8_t* dst) {    // horizontal
+  int j;
+  for (j = 0; j < 8; ++j) {
+    const uint8x8_t left = vld1_dup_u8(dst - 1);
+    vst1_u8(dst, left);
+    dst += BPS;
+  }
+}
+
+static WEBP_INLINE void DC8(uint8_t* dst, int do_top, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_top) {
+    const uint8x8_t A = vld1_u8(dst - BPS);  // top row
+    const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_top = vcombine_u16(p2, p2);
+  }
+
+  if (do_left) {
+    const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + 0 * BPS - 1));
+    const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + 1 * BPS - 1));
+    const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + 2 * BPS - 1));
+    const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + 3 * BPS - 1));
+    const uint16x8_t L4 = vmovl_u8(vld1_u8(dst + 4 * BPS - 1));
+    const uint16x8_t L5 = vmovl_u8(vld1_u8(dst + 5 * BPS - 1));
+    const uint16x8_t L6 = vmovl_u8(vld1_u8(dst + 6 * BPS - 1));
+    const uint16x8_t L7 = vmovl_u8(vld1_u8(dst + 7 * BPS - 1));
+    const uint16x8_t s0 = vaddq_u16(L0, L1);
+    const uint16x8_t s1 = vaddq_u16(L2, L3);
+    const uint16x8_t s2 = vaddq_u16(L4, L5);
+    const uint16x8_t s3 = vaddq_u16(L6, L7);
+    const uint16x8_t s01 = vaddq_u16(s0, s1);
+    const uint16x8_t s23 = vaddq_u16(s2, s3);
+    sum_left = vaddq_u16(s01, s23);
+  }
+
+  if (do_top && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 4);
+  } else if (do_top) {
+    dc0 = vrshrn_n_u16(sum_top, 3);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 3);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 8; ++i) {
+      vst1_u32((uint32_t*)(dst + i * BPS), vreinterpret_u32_u8(dc));
+    }
+  }
+}
+
+static void DC8uv(uint8_t* dst) { DC8(dst, 1, 1); }
+static void DC8uvNoTop(uint8_t* dst) { DC8(dst, 0, 1); }
+static void DC8uvNoLeft(uint8_t* dst) { DC8(dst, 1, 0); }
+static void DC8uvNoTopLeft(uint8_t* dst) { DC8(dst, 0, 0); }
+
+static void TM8uv(uint8_t* dst) { TrueMotion(dst, 8); }
+
+//------------------------------------------------------------------------------
+// 16x16
+
+static void VE16(uint8_t* dst) {     // vertical
+  const uint8x16_t top = vld1q_u8(dst - BPS);
+  int j;
+  for (j = 0; j < 16; ++j) {
+    vst1q_u8(dst + j * BPS, top);
+  }
+}
+
+static void HE16(uint8_t* dst) {     // horizontal
+  int j;
+  for (j = 0; j < 16; ++j) {
+    const uint8x16_t left = vld1q_dup_u8(dst - 1);
+    vst1q_u8(dst, left);
+    dst += BPS;
+  }
+}
+
+static WEBP_INLINE void DC16(uint8_t* dst, int do_top, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_top) {
+    const uint8x16_t A = vld1q_u8(dst - BPS);  // top row
+    const uint16x8_t p0 = vpaddlq_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_top = vcombine_u16(p3, p3);
+  }
+
+  if (do_left) {
+    int i;
+    sum_left = vdupq_n_u16(0);
+    for (i = 0; i < 16; i += 8) {
+      const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + (i + 0) * BPS - 1));
+      const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + (i + 1) * BPS - 1));
+      const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + (i + 2) * BPS - 1));
+      const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + (i + 3) * BPS - 1));
+      const uint16x8_t L4 = vmovl_u8(vld1_u8(dst + (i + 4) * BPS - 1));
+      const uint16x8_t L5 = vmovl_u8(vld1_u8(dst + (i + 5) * BPS - 1));
+      const uint16x8_t L6 = vmovl_u8(vld1_u8(dst + (i + 6) * BPS - 1));
+      const uint16x8_t L7 = vmovl_u8(vld1_u8(dst + (i + 7) * BPS - 1));
+      const uint16x8_t s0 = vaddq_u16(L0, L1);
+      const uint16x8_t s1 = vaddq_u16(L2, L3);
+      const uint16x8_t s2 = vaddq_u16(L4, L5);
+      const uint16x8_t s3 = vaddq_u16(L6, L7);
+      const uint16x8_t s01 = vaddq_u16(s0, s1);
+      const uint16x8_t s23 = vaddq_u16(s2, s3);
+      const uint16x8_t sum = vaddq_u16(s01, s23);
+      sum_left = vaddq_u16(sum_left, sum);
+    }
+  }
+
+  if (do_top && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 5);
+  } else if (do_top) {
+    dc0 = vrshrn_n_u16(sum_top, 4);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 4);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 16; ++i) {
+      vst1q_u8(dst + i * BPS, dc);
+    }
+  }
+}
+
+static void DC16TopLeft(uint8_t* dst) { DC16(dst, 1, 1); }
+static void DC16NoTop(uint8_t* dst) { DC16(dst, 0, 1); }
+static void DC16NoLeft(uint8_t* dst) { DC16(dst, 1, 0); }
+static void DC16NoTopLeft(uint8_t* dst) { DC16(dst, 0, 0); }
+
+static void TM16(uint8_t* dst) {
+  const uint8x8_t TL = vld1_dup_u8(dst - BPS - 1);  // top-left pixel 'A[-1]'
+  const uint8x16_t T = vld1q_u8(dst - BPS);  // top row 'A[0..15]'
+  // A[c] - A[-1]
+  const int16x8_t d_lo = vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(T), TL));
+  const int16x8_t d_hi = vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(T), TL));
+  int y;
+  for (y = 0; y < 16; y += 4) {
+    // left edge
+    const int16x8_t L0 = ConvertU8ToS16(vld1_dup_u8(dst + 0 * BPS - 1));
+    const int16x8_t L1 = ConvertU8ToS16(vld1_dup_u8(dst + 1 * BPS - 1));
+    const int16x8_t L2 = ConvertU8ToS16(vld1_dup_u8(dst + 2 * BPS - 1));
+    const int16x8_t L3 = ConvertU8ToS16(vld1_dup_u8(dst + 3 * BPS - 1));
+    const int16x8_t r0_lo = vaddq_s16(L0, d_lo);  // L[r] + A[c] - A[-1]
+    const int16x8_t r1_lo = vaddq_s16(L1, d_lo);
+    const int16x8_t r2_lo = vaddq_s16(L2, d_lo);
+    const int16x8_t r3_lo = vaddq_s16(L3, d_lo);
+    const int16x8_t r0_hi = vaddq_s16(L0, d_hi);
+    const int16x8_t r1_hi = vaddq_s16(L1, d_hi);
+    const int16x8_t r2_hi = vaddq_s16(L2, d_hi);
+    const int16x8_t r3_hi = vaddq_s16(L3, d_hi);
+    // Saturate and store the result.
+    const uint8x16_t row0 = vcombine_u8(vqmovun_s16(r0_lo), vqmovun_s16(r0_hi));
+    const uint8x16_t row1 = vcombine_u8(vqmovun_s16(r1_lo), vqmovun_s16(r1_hi));
+    const uint8x16_t row2 = vcombine_u8(vqmovun_s16(r2_lo), vqmovun_s16(r2_hi));
+    const uint8x16_t row3 = vcombine_u8(vqmovun_s16(r3_lo), vqmovun_s16(r3_hi));
+    vst1q_u8(dst + 0 * BPS, row0);
+    vst1q_u8(dst + 1 * BPS, row1);
+    vst1q_u8(dst + 2 * BPS, row2);
+    vst1q_u8(dst + 3 * BPS, row3);
+    dst += 4 * BPS;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8DspInitNEON(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitNEON(void) {
+  VP8Transform = TransformTwo;
+  VP8TransformAC3 = TransformAC3;
+  VP8TransformDC = TransformDC;
+  VP8TransformWHT = TransformWHT;
+
+  VP8VFilter16 = VFilter16;
+  VP8VFilter16i = VFilter16i;
+  VP8HFilter16 = HFilter16;
+#if !defined(WORK_AROUND_GCC)
+  VP8HFilter16i = HFilter16i;
 #endif
+  VP8VFilter8 = VFilter8;
+  VP8VFilter8i = VFilter8i;
+#if !defined(WORK_AROUND_GCC)
+  VP8HFilter8 = HFilter8;
+  VP8HFilter8i = HFilter8i;
+#endif
+  VP8SimpleVFilter16 = SimpleVFilter16;
+  VP8SimpleHFilter16 = SimpleHFilter16;
+  VP8SimpleVFilter16i = SimpleVFilter16i;
+  VP8SimpleHFilter16i = SimpleHFilter16i;
+
+  VP8PredLuma4[0] = DC4;
+  VP8PredLuma4[1] = TM4;
+  VP8PredLuma4[2] = VE4;
+  VP8PredLuma4[4] = RD4;
+  VP8PredLuma4[6] = LD4;
+
+  VP8PredLuma16[0] = DC16TopLeft;
+  VP8PredLuma16[1] = TM16;
+  VP8PredLuma16[2] = VE16;
+  VP8PredLuma16[3] = HE16;
+  VP8PredLuma16[4] = DC16NoTop;
+  VP8PredLuma16[5] = DC16NoLeft;
+  VP8PredLuma16[6] = DC16NoTopLeft;
+
+  VP8PredChroma8[0] = DC8uv;
+  VP8PredChroma8[1] = TM8uv;
+  VP8PredChroma8[2] = VE8uv;
+  VP8PredChroma8[3] = HE8uv;
+  VP8PredChroma8[4] = DC8uvNoTop;
+  VP8PredChroma8[5] = DC8uvNoLeft;
+  VP8PredChroma8[6] = DC8uvNoTopLeft;
+}
+
+#else  // !WEBP_USE_NEON
+
+WEBP_DSP_INIT_STUB(VP8DspInitNEON)
 
-#endif   // WEBP_USE_NEON
+#endif  // WEBP_USE_NEON
diff --git a/drivers/webp/dsp/dec_sse2.c b/drivers/webp/dsp/dec_sse2.c
index 472b68ecb8..d4838b9210 100644
--- a/drivers/webp/dsp/dec_sse2.c
+++ b/drivers/webp/dsp/dec_sse2.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // SSE2 version of some decoding functions (idct, loop filtering).
@@ -14,17 +16,17 @@
 
 #if defined(WEBP_USE_SSE2)
 
+// The 3-coeff sparse transform in SSE2 is not really faster than the plain-C
+// one it seems => disable it by default. Uncomment the following to enable:
+// #define USE_TRANSFORM_AC3
+
 #include <emmintrin.h>
 #include "../dec/vp8i.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)
 
-static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
+static void Transform(const int16_t* in, uint8_t* dst, int do_two) {
   // This implementation makes use of 16-bit fixed point versions of two
   // multiply constants:
   //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@@ -50,19 +52,19 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
   // vectors will just contain random value we'll never use nor store.
   __m128i in0, in1, in2, in3;
   {
-    in0 = _mm_loadl_epi64((__m128i*)&in[0]);
-    in1 = _mm_loadl_epi64((__m128i*)&in[4]);
-    in2 = _mm_loadl_epi64((__m128i*)&in[8]);
-    in3 = _mm_loadl_epi64((__m128i*)&in[12]);
+    in0 = _mm_loadl_epi64((const __m128i*)&in[0]);
+    in1 = _mm_loadl_epi64((const __m128i*)&in[4]);
+    in2 = _mm_loadl_epi64((const __m128i*)&in[8]);
+    in3 = _mm_loadl_epi64((const __m128i*)&in[12]);
     // a00 a10 a20 a30   x x x x
     // a01 a11 a21 a31   x x x x
     // a02 a12 a22 a32   x x x x
     // a03 a13 a23 a33   x x x x
     if (do_two) {
-      const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]);
-      const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]);
-      const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]);
-      const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]);
+      const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]);
+      const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]);
+      const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]);
+      const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]);
       in0 = _mm_unpacklo_epi64(in0, inB0);
       in1 = _mm_unpacklo_epi64(in1, inB1);
       in2 = _mm_unpacklo_epi64(in2, inB2);
@@ -194,21 +196,21 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
 
   // Add inverse transform to 'dst' and store.
   {
-    const __m128i zero = _mm_set1_epi16(0);
+    const __m128i zero = _mm_setzero_si128();
     // Load the reference(s).
     __m128i dst0, dst1, dst2, dst3;
     if (do_two) {
       // Load eight bytes/pixels per line.
-      dst0 = _mm_loadl_epi64((__m128i*)&dst[0 * BPS]);
-      dst1 = _mm_loadl_epi64((__m128i*)&dst[1 * BPS]);
-      dst2 = _mm_loadl_epi64((__m128i*)&dst[2 * BPS]);
-      dst3 = _mm_loadl_epi64((__m128i*)&dst[3 * BPS]);
+      dst0 = _mm_loadl_epi64((__m128i*)(dst + 0 * BPS));
+      dst1 = _mm_loadl_epi64((__m128i*)(dst + 1 * BPS));
+      dst2 = _mm_loadl_epi64((__m128i*)(dst + 2 * BPS));
+      dst3 = _mm_loadl_epi64((__m128i*)(dst + 3 * BPS));
     } else {
       // Load four bytes/pixels per line.
-      dst0 = _mm_cvtsi32_si128(*(int*)&dst[0 * BPS]);
-      dst1 = _mm_cvtsi32_si128(*(int*)&dst[1 * BPS]);
-      dst2 = _mm_cvtsi32_si128(*(int*)&dst[2 * BPS]);
-      dst3 = _mm_cvtsi32_si128(*(int*)&dst[3 * BPS]);
+      dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));
+      dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));
+      dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
+      dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
     }
     // Convert to 16b.
     dst0 = _mm_unpacklo_epi8(dst0, zero);
@@ -228,20 +230,66 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
     // Store the results.
     if (do_two) {
       // Store eight bytes/pixels per line.
-      _mm_storel_epi64((__m128i*)&dst[0 * BPS], dst0);
-      _mm_storel_epi64((__m128i*)&dst[1 * BPS], dst1);
-      _mm_storel_epi64((__m128i*)&dst[2 * BPS], dst2);
-      _mm_storel_epi64((__m128i*)&dst[3 * BPS], dst3);
+      _mm_storel_epi64((__m128i*)(dst + 0 * BPS), dst0);
+      _mm_storel_epi64((__m128i*)(dst + 1 * BPS), dst1);
+      _mm_storel_epi64((__m128i*)(dst + 2 * BPS), dst2);
+      _mm_storel_epi64((__m128i*)(dst + 3 * BPS), dst3);
     } else {
       // Store four bytes/pixels per line.
-      *((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(dst0);
-      *((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(dst1);
-      *((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(dst2);
-      *((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(dst3);
+      *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
+      *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
+      *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
+      *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
     }
   }
 }
 
+#if defined(USE_TRANSFORM_AC3)
+#define MUL(a, b) (((a) * (b)) >> 16)
+static void TransformAC3(const int16_t* in, uint8_t* dst) {
+  static const int kC1 = 20091 + (1 << 16);
+  static const int kC2 = 35468;
+  const __m128i A = _mm_set1_epi16(in[0] + 4);
+  const __m128i c4 = _mm_set1_epi16(MUL(in[4], kC2));
+  const __m128i d4 = _mm_set1_epi16(MUL(in[4], kC1));
+  const int c1 = MUL(in[1], kC2);
+  const int d1 = MUL(in[1], kC1);
+  const __m128i CD = _mm_set_epi16(0, 0, 0, 0, -d1, -c1, c1, d1);
+  const __m128i B = _mm_adds_epi16(A, CD);
+  const __m128i m0 = _mm_adds_epi16(B, d4);
+  const __m128i m1 = _mm_adds_epi16(B, c4);
+  const __m128i m2 = _mm_subs_epi16(B, c4);
+  const __m128i m3 = _mm_subs_epi16(B, d4);
+  const __m128i zero = _mm_setzero_si128();
+  // Load the source pixels.
+  __m128i dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));
+  __m128i dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));
+  __m128i dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
+  __m128i dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
+  // Convert to 16b.
+  dst0 = _mm_unpacklo_epi8(dst0, zero);
+  dst1 = _mm_unpacklo_epi8(dst1, zero);
+  dst2 = _mm_unpacklo_epi8(dst2, zero);
+  dst3 = _mm_unpacklo_epi8(dst3, zero);
+  // Add the inverse transform.
+  dst0 = _mm_adds_epi16(dst0, _mm_srai_epi16(m0, 3));
+  dst1 = _mm_adds_epi16(dst1, _mm_srai_epi16(m1, 3));
+  dst2 = _mm_adds_epi16(dst2, _mm_srai_epi16(m2, 3));
+  dst3 = _mm_adds_epi16(dst3, _mm_srai_epi16(m3, 3));
+  // Unsigned saturate to 8b.
+  dst0 = _mm_packus_epi16(dst0, dst0);
+  dst1 = _mm_packus_epi16(dst1, dst1);
+  dst2 = _mm_packus_epi16(dst2, dst2);
+  dst3 = _mm_packus_epi16(dst3, dst3);
+  // Store the results.
+  *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
+  *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
+  *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
+  *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
+}
+#undef MUL
+#endif   // USE_TRANSFORM_AC3
+
 //------------------------------------------------------------------------------
 // Loop Filter (Paragraph 15)
 
@@ -250,20 +298,14 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
     _mm_subs_epu8((q), (p)),                                                   \
     _mm_subs_epu8((p), (q)))
 
-// Shift each byte of "a" by N bits while preserving by the sign bit.
-//
-// It first shifts the lower bytes of the words and then the upper bytes and
-// then merges the results together.
-#define SIGNED_SHIFT_N(a, N) {                                                 \
-  __m128i t = a;                                                               \
-  t = _mm_slli_epi16(t, 8);                                                    \
-  t = _mm_srai_epi16(t, N);                                                    \
-  t = _mm_srli_epi16(t, 8);                                                    \
-                                                                               \
-  a = _mm_srai_epi16(a, N + 8);                                                \
-  a = _mm_slli_epi16(a, 8);                                                    \
-                                                                               \
-  a = _mm_or_si128(t, a);                                                      \
+// Shift each byte of "x" by 3 bits while preserving by the sign bit.
+static WEBP_INLINE void SignedShift8b(__m128i* const x) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i lo_0 = _mm_unpacklo_epi8(zero, *x);
+  const __m128i hi_0 = _mm_unpackhi_epi8(zero, *x);
+  const __m128i lo_1 = _mm_srai_epi16(lo_0, 3 + 8);
+  const __m128i hi_1 = _mm_srai_epi16(hi_0, 3 + 8);
+  *x = _mm_packs_epi16(lo_1, hi_1);
 }
 
 #define FLIP_SIGN_BIT2(a, b) {                                                 \
@@ -276,103 +318,124 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
   FLIP_SIGN_BIT2(c, d);                                                        \
 }
 
-#define GET_NOTHEV(p1, p0, q0, q1, hev_thresh, not_hev) {                      \
-  const __m128i zero = _mm_setzero_si128();                                    \
-  const __m128i t1 = MM_ABS(p1, p0);                                           \
-  const __m128i t2 = MM_ABS(q1, q0);                                           \
-                                                                               \
-  const __m128i h = _mm_set1_epi8(hev_thresh);                                 \
-  const __m128i t3 = _mm_subs_epu8(t1, h);  /* abs(p1 - p0) - hev_tresh */     \
-  const __m128i t4 = _mm_subs_epu8(t2, h);  /* abs(q1 - q0) - hev_tresh */     \
-                                                                               \
-  not_hev = _mm_or_si128(t3, t4);                                              \
-  not_hev = _mm_cmpeq_epi8(not_hev, zero); /* not_hev <= t1 && not_hev <= t2 */\
-}
-
-#define GET_BASE_DELTA(p1, p0, q0, q1, o) {                                    \
-  const __m128i qp0 = _mm_subs_epi8(q0, p0);  /* q0 - p0 */                    \
-  o = _mm_subs_epi8(p1, q1);            /* p1 - q1 */                          \
-  o = _mm_adds_epi8(o, qp0);            /* p1 - q1 + 1 * (q0 - p0) */          \
-  o = _mm_adds_epi8(o, qp0);            /* p1 - q1 + 2 * (q0 - p0) */          \
-  o = _mm_adds_epi8(o, qp0);            /* p1 - q1 + 3 * (q0 - p0) */          \
-}
-
-#define DO_SIMPLE_FILTER(p0, q0, fl) {                                         \
-  const __m128i three = _mm_set1_epi8(3);                                      \
-  const __m128i four = _mm_set1_epi8(4);                                       \
-  __m128i v3 = _mm_adds_epi8(fl, three);                                       \
-  __m128i v4 = _mm_adds_epi8(fl, four);                                        \
-                                                                               \
-  /* Do +4 side */                                                             \
-  SIGNED_SHIFT_N(v4, 3);                /* v4 >> 3  */                         \
-  q0 = _mm_subs_epi8(q0, v4);           /* q0 -= v4 */                         \
-                                                                               \
-  /* Now do +3 side */                                                         \
-  SIGNED_SHIFT_N(v3, 3);                /* v3 >> 3  */                         \
-  p0 = _mm_adds_epi8(p0, v3);           /* p0 += v3 */                         \
+// input/output is uint8_t
+static WEBP_INLINE void GetNotHEV(const __m128i* const p1,
+                                  const __m128i* const p0,
+                                  const __m128i* const q0,
+                                  const __m128i* const q1,
+                                  int hev_thresh, __m128i* const not_hev) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t_1 = MM_ABS(*p1, *p0);
+  const __m128i t_2 = MM_ABS(*q1, *q0);
+
+  const __m128i h = _mm_set1_epi8(hev_thresh);
+  const __m128i t_max = _mm_max_epu8(t_1, t_2);
+
+  const __m128i t_max_h = _mm_subs_epu8(t_max, h);
+  *not_hev = _mm_cmpeq_epi8(t_max_h, zero);  // not_hev <= t1 && not_hev <= t2
 }
 
-// Updates values of 2 pixels at MB edge during complex filtering.
-// Update operations:
-// q = q - a and p = p + a; where a = [(a_hi >> 7), (a_lo >> 7)]
-#define UPDATE_2PIXELS(pi, qi, a_lo, a_hi) {                                   \
-  const __m128i a_lo7 = _mm_srai_epi16(a_lo, 7);                               \
-  const __m128i a_hi7 = _mm_srai_epi16(a_hi, 7);                               \
-  const __m128i a = _mm_packs_epi16(a_lo7, a_hi7);                             \
-  pi = _mm_adds_epi8(pi, a);                                                   \
-  qi = _mm_subs_epi8(qi, a);                                                   \
+// input pixels are int8_t
+static WEBP_INLINE void GetBaseDelta(const __m128i* const p1,
+                                     const __m128i* const p0,
+                                     const __m128i* const q0,
+                                     const __m128i* const q1,
+                                     __m128i* const delta) {
+  // beware of addition order, for saturation!
+  const __m128i p1_q1 = _mm_subs_epi8(*p1, *q1);   // p1 - q1
+  const __m128i q0_p0 = _mm_subs_epi8(*q0, *p0);   // q0 - p0
+  const __m128i s1 = _mm_adds_epi8(p1_q1, q0_p0);  // p1 - q1 + 1 * (q0 - p0)
+  const __m128i s2 = _mm_adds_epi8(q0_p0, s1);     // p1 - q1 + 2 * (q0 - p0)
+  const __m128i s3 = _mm_adds_epi8(q0_p0, s2);     // p1 - q1 + 3 * (q0 - p0)
+  *delta = s3;
 }
 
-static void NeedsFilter(const __m128i* p1, const __m128i* p0, const __m128i* q0,
-                        const __m128i* q1, int thresh, __m128i *mask) {
-  __m128i t1 = MM_ABS(*p1, *q1);        // abs(p1 - q1)
-  *mask = _mm_set1_epi8(0xFE);
-  t1 = _mm_and_si128(t1, *mask);        // set lsb of each byte to zero
-  t1 = _mm_srli_epi16(t1, 1);           // abs(p1 - q1) / 2
+// input and output are int8_t
+static WEBP_INLINE void DoSimpleFilter(__m128i* const p0, __m128i* const q0,
+                                       const __m128i* const fl) {
+  const __m128i k3 = _mm_set1_epi8(3);
+  const __m128i k4 = _mm_set1_epi8(4);
+  __m128i v3 = _mm_adds_epi8(*fl, k3);
+  __m128i v4 = _mm_adds_epi8(*fl, k4);
+
+  SignedShift8b(&v4);                  // v4 >> 3
+  SignedShift8b(&v3);                  // v3 >> 3
+  *q0 = _mm_subs_epi8(*q0, v4);        // q0 -= v4
+  *p0 = _mm_adds_epi8(*p0, v3);        // p0 += v3
+}
 
-  *mask = MM_ABS(*p0, *q0);             // abs(p0 - q0)
-  *mask = _mm_adds_epu8(*mask, *mask);  // abs(p0 - q0) * 2
-  *mask = _mm_adds_epu8(*mask, t1);     // abs(p0 - q0) * 2 + abs(p1 - q1) / 2
+// Updates values of 2 pixels at MB edge during complex filtering.
+// Update operations:
+// q = q - delta and p = p + delta; where delta = [(a_hi >> 7), (a_lo >> 7)]
+// Pixels 'pi' and 'qi' are int8_t on input, uint8_t on output (sign flip).
+static WEBP_INLINE void Update2Pixels(__m128i* const pi, __m128i* const qi,
+                                      const __m128i* const a0_lo,
+                                      const __m128i* const a0_hi) {
+  const __m128i a1_lo = _mm_srai_epi16(*a0_lo, 7);
+  const __m128i a1_hi = _mm_srai_epi16(*a0_hi, 7);
+  const __m128i delta = _mm_packs_epi16(a1_lo, a1_hi);
+  const __m128i sign_bit = _mm_set1_epi8(0x80);
+  *pi = _mm_adds_epi8(*pi, delta);
+  *qi = _mm_subs_epi8(*qi, delta);
+  FLIP_SIGN_BIT2(*pi, *qi);
+}
 
-  t1 = _mm_set1_epi8(thresh);
-  *mask = _mm_subs_epu8(*mask, t1);     // mask <= thresh
-  *mask = _mm_cmpeq_epi8(*mask, _mm_setzero_si128());
+// input pixels are uint8_t
+static WEBP_INLINE void NeedsFilter(const __m128i* const p1,
+                                    const __m128i* const p0,
+                                    const __m128i* const q0,
+                                    const __m128i* const q1,
+                                    int thresh, __m128i* const mask) {
+  const __m128i m_thresh = _mm_set1_epi8(thresh);
+  const __m128i t1 = MM_ABS(*p1, *q1);        // abs(p1 - q1)
+  const __m128i kFE = _mm_set1_epi8(0xFE);
+  const __m128i t2 = _mm_and_si128(t1, kFE);  // set lsb of each byte to zero
+  const __m128i t3 = _mm_srli_epi16(t2, 1);   // abs(p1 - q1) / 2
+
+  const __m128i t4 = MM_ABS(*p0, *q0);        // abs(p0 - q0)
+  const __m128i t5 = _mm_adds_epu8(t4, t4);   // abs(p0 - q0) * 2
+  const __m128i t6 = _mm_adds_epu8(t5, t3);   // abs(p0-q0)*2 + abs(p1-q1)/2
+
+  const __m128i t7 = _mm_subs_epu8(t6, m_thresh);  // mask <= m_thresh
+  *mask = _mm_cmpeq_epi8(t7, _mm_setzero_si128());
 }
 
 //------------------------------------------------------------------------------
 // Edge filtering functions
 
 // Applies filter on 2 pixels (p0 and q0)
-static WEBP_INLINE void DoFilter2(const __m128i* p1, __m128i* p0, __m128i* q0,
-                                  const __m128i* q1, int thresh) {
+static WEBP_INLINE void DoFilter2(__m128i* const p1, __m128i* const p0,
+                                  __m128i* const q0, __m128i* const q1,
+                                  int thresh) {
   __m128i a, mask;
   const __m128i sign_bit = _mm_set1_epi8(0x80);
+  // convert p1/q1 to int8_t (for GetBaseDelta)
   const __m128i p1s = _mm_xor_si128(*p1, sign_bit);
   const __m128i q1s = _mm_xor_si128(*q1, sign_bit);
 
   NeedsFilter(p1, p0, q0, q1, thresh, &mask);
 
-  // convert to signed values
   FLIP_SIGN_BIT2(*p0, *q0);
-
-  GET_BASE_DELTA(p1s, *p0, *q0, q1s, a);
+  GetBaseDelta(&p1s, p0, q0, &q1s, &a);
   a = _mm_and_si128(a, mask);     // mask filter values we don't care about
-  DO_SIMPLE_FILTER(*p0, *q0, a);
-
-  // unoffset
+  DoSimpleFilter(p0, q0, &a);
   FLIP_SIGN_BIT2(*p0, *q0);
 }
 
 // Applies filter on 4 pixels (p1, p0, q0 and q1)
-static WEBP_INLINE void DoFilter4(__m128i* p1, __m128i *p0,
-                                  __m128i* q0, __m128i* q1,
-                                  const __m128i* mask, int hev_thresh) {
+static WEBP_INLINE void DoFilter4(__m128i* const p1, __m128i* const p0,
+                                  __m128i* const q0, __m128i* const q1,
+                                  const __m128i* const mask, int hev_thresh) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bit = _mm_set1_epi8(0x80);
+  const __m128i k64 = _mm_set1_epi8(64);
+  const __m128i k3 = _mm_set1_epi8(3);
+  const __m128i k4 = _mm_set1_epi8(4);
   __m128i not_hev;
   __m128i t1, t2, t3;
-  const __m128i sign_bit = _mm_set1_epi8(0x80);
 
   // compute hev mask
-  GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev);
+  GetNotHEV(p1, p0, q0, q1, hev_thresh, &not_hev);
 
   // convert to signed values
   FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
@@ -385,135 +448,115 @@ static WEBP_INLINE void DoFilter4(__m128i* p1, __m128i *p0,
   t1 = _mm_adds_epi8(t1, t2);          // hev(p1 - q1) + 3 * (q0 - p0)
   t1 = _mm_and_si128(t1, *mask);       // mask filter values we don't care about
 
-  // Do +4 side
-  t2 = _mm_set1_epi8(4);
-  t2 = _mm_adds_epi8(t1, t2);        // 3 * (q0 - p0) + (p1 - q1) + 4
-  SIGNED_SHIFT_N(t2, 3);             // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3
-  t3 = t2;                           // save t2
-  *q0 = _mm_subs_epi8(*q0, t2);      // q0 -= t2
-
-  // Now do +3 side
-  t2 = _mm_set1_epi8(3);
-  t2 = _mm_adds_epi8(t1, t2);        // +3 instead of +4
-  SIGNED_SHIFT_N(t2, 3);             // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3
+  t2 = _mm_adds_epi8(t1, k3);        // 3 * (q0 - p0) + hev(p1 - q1) + 3
+  t3 = _mm_adds_epi8(t1, k4);        // 3 * (q0 - p0) + hev(p1 - q1) + 4
+  SignedShift8b(&t2);                // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3
+  SignedShift8b(&t3);                // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3
   *p0 = _mm_adds_epi8(*p0, t2);      // p0 += t2
+  *q0 = _mm_subs_epi8(*q0, t3);      // q0 -= t3
+  FLIP_SIGN_BIT2(*p0, *q0);
 
-  t2 = _mm_set1_epi8(1);
-  t3 = _mm_adds_epi8(t3, t2);
-  SIGNED_SHIFT_N(t3, 1);             // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 4
+  // this is equivalent to signed (a + 1) >> 1 calculation
+  t2 = _mm_add_epi8(t3, sign_bit);
+  t3 = _mm_avg_epu8(t2, zero);
+  t3 = _mm_sub_epi8(t3, k64);
 
   t3 = _mm_and_si128(not_hev, t3);   // if !hev
   *q1 = _mm_subs_epi8(*q1, t3);      // q1 -= t3
   *p1 = _mm_adds_epi8(*p1, t3);      // p1 += t3
-
-  // unoffset
-  FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
+  FLIP_SIGN_BIT2(*p1, *q1);
 }
 
 // Applies filter on 6 pixels (p2, p1, p0, q0, q1 and q2)
-static WEBP_INLINE void DoFilter6(__m128i *p2, __m128i* p1, __m128i *p0,
-                                  __m128i* q0, __m128i* q1, __m128i *q2,
-                                  const __m128i* mask, int hev_thresh) {
-  __m128i a, not_hev;
+static WEBP_INLINE void DoFilter6(__m128i* const p2, __m128i* const p1,
+                                  __m128i* const p0, __m128i* const q0,
+                                  __m128i* const q1, __m128i* const q2,
+                                  const __m128i* const mask, int hev_thresh) {
+  const __m128i zero = _mm_setzero_si128();
   const __m128i sign_bit = _mm_set1_epi8(0x80);
+  __m128i a, not_hev;
 
   // compute hev mask
-  GET_NOTHEV(*p1, *p0, *q0, *q1, hev_thresh, not_hev);
+  GetNotHEV(p1, p0, q0, q1, hev_thresh, &not_hev);
 
-  // convert to signed values
   FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
   FLIP_SIGN_BIT2(*p2, *q2);
-
-  GET_BASE_DELTA(*p1, *p0, *q0, *q1, a);
+  GetBaseDelta(p1, p0, q0, q1, &a);
 
   { // do simple filter on pixels with hev
     const __m128i m = _mm_andnot_si128(not_hev, *mask);
     const __m128i f = _mm_and_si128(a, m);
-    DO_SIMPLE_FILTER(*p0, *q0, f);
+    DoSimpleFilter(p0, q0, &f);
   }
+
   { // do strong filter on pixels with not hev
-    const __m128i zero = _mm_setzero_si128();
-    const __m128i nine = _mm_set1_epi16(0x0900);
-    const __m128i sixty_three = _mm_set1_epi16(63);
+    const __m128i k9 = _mm_set1_epi16(0x0900);
+    const __m128i k63 = _mm_set1_epi16(63);
 
     const __m128i m = _mm_and_si128(not_hev, *mask);
     const __m128i f = _mm_and_si128(a, m);
+
     const __m128i f_lo = _mm_unpacklo_epi8(zero, f);
     const __m128i f_hi = _mm_unpackhi_epi8(zero, f);
 
-    const __m128i f9_lo = _mm_mulhi_epi16(f_lo, nine);   // Filter (lo) * 9
-    const __m128i f9_hi = _mm_mulhi_epi16(f_hi, nine);   // Filter (hi) * 9
-    const __m128i f18_lo = _mm_add_epi16(f9_lo, f9_lo);  // Filter (lo) * 18
-    const __m128i f18_hi = _mm_add_epi16(f9_hi, f9_hi);  // Filter (hi) * 18
+    const __m128i f9_lo = _mm_mulhi_epi16(f_lo, k9);    // Filter (lo) * 9
+    const __m128i f9_hi = _mm_mulhi_epi16(f_hi, k9);    // Filter (hi) * 9
 
-    const __m128i a2_lo = _mm_add_epi16(f9_lo, sixty_three);  // Filter * 9 + 63
-    const __m128i a2_hi = _mm_add_epi16(f9_hi, sixty_three);  // Filter * 9 + 63
+    const __m128i a2_lo = _mm_add_epi16(f9_lo, k63);    // Filter * 9 + 63
+    const __m128i a2_hi = _mm_add_epi16(f9_hi, k63);    // Filter * 9 + 63
 
-    const __m128i a1_lo = _mm_add_epi16(f18_lo, sixty_three);  // F... * 18 + 63
-    const __m128i a1_hi = _mm_add_epi16(f18_hi, sixty_three);  // F... * 18 + 63
+    const __m128i a1_lo = _mm_add_epi16(a2_lo, f9_lo);  // Filter * 18 + 63
+    const __m128i a1_hi = _mm_add_epi16(a2_hi, f9_hi);  // Filter * 18 + 63
 
-    const __m128i a0_lo = _mm_add_epi16(f18_lo, a2_lo);  // Filter * 27 + 63
-    const __m128i a0_hi = _mm_add_epi16(f18_hi, a2_hi);  // Filter * 27 + 63
+    const __m128i a0_lo = _mm_add_epi16(a1_lo, f9_lo);  // Filter * 27 + 63
+    const __m128i a0_hi = _mm_add_epi16(a1_hi, f9_hi);  // Filter * 27 + 63
 
-    UPDATE_2PIXELS(*p2, *q2, a2_lo, a2_hi);
-    UPDATE_2PIXELS(*p1, *q1, a1_lo, a1_hi);
-    UPDATE_2PIXELS(*p0, *q0, a0_lo, a0_hi);
+    Update2Pixels(p2, q2, &a2_lo, &a2_hi);
+    Update2Pixels(p1, q1, &a1_lo, &a1_hi);
+    Update2Pixels(p0, q0, &a0_lo, &a0_hi);
   }
+}
 
-  // unoffset
-  FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
-  FLIP_SIGN_BIT2(*p2, *q2);
+// memcpy() is the safe way of moving potentially unaligned 32b memory.
+static WEBP_INLINE uint32_t MemToUint32(const uint8_t* const ptr) {
+  uint32_t A;
+  memcpy(&A, (const int*)ptr, sizeof(A));
+  return A;
 }
 
 // reads 8 rows across a vertical edge.
-//
-// TODO(somnath): Investigate _mm_shuffle* also see if it can be broken into
-// two Load4x4() to avoid code duplication.
-static WEBP_INLINE void Load8x4(const uint8_t* b, int stride,
-                                __m128i* p, __m128i* q) {
-  __m128i t1, t2;
-
-  // Load 0th, 1st, 4th and 5th rows
-  __m128i r0 =  _mm_cvtsi32_si128(*((int*)&b[0 * stride]));  // 03 02 01 00
-  __m128i r1 =  _mm_cvtsi32_si128(*((int*)&b[1 * stride]));  // 13 12 11 10
-  __m128i r4 =  _mm_cvtsi32_si128(*((int*)&b[4 * stride]));  // 43 42 41 40
-  __m128i r5 =  _mm_cvtsi32_si128(*((int*)&b[5 * stride]));  // 53 52 51 50
-
-  r0 = _mm_unpacklo_epi32(r0, r4);               // 43 42 41 40 03 02 01 00
-  r1 = _mm_unpacklo_epi32(r1, r5);               // 53 52 51 50 13 12 11 10
-
-  // t1 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
-  t1 = _mm_unpacklo_epi8(r0, r1);
-
-  // Load 2nd, 3rd, 6th and 7th rows
-  r0 =  _mm_cvtsi32_si128(*((int*)&b[2 * stride]));          // 23 22 21 22
-  r1 =  _mm_cvtsi32_si128(*((int*)&b[3 * stride]));          // 33 32 31 30
-  r4 =  _mm_cvtsi32_si128(*((int*)&b[6 * stride]));          // 63 62 61 60
-  r5 =  _mm_cvtsi32_si128(*((int*)&b[7 * stride]));          // 73 72 71 70
-
-  r0 = _mm_unpacklo_epi32(r0, r4);               // 63 62 61 60 23 22 21 20
-  r1 = _mm_unpacklo_epi32(r1, r5);               // 73 72 71 70 33 32 31 30
-
-  // t2 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
-  t2 = _mm_unpacklo_epi8(r0, r1);
-
-  // t1 = 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
-  // t2 = 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40
-  r0 = t1;
-  t1 = _mm_unpacklo_epi16(t1, t2);
-  t2 = _mm_unpackhi_epi16(r0, t2);
+static WEBP_INLINE void Load8x4(const uint8_t* const b, int stride,
+                                __m128i* const p, __m128i* const q) {
+  // A0 = 63 62 61 60 23 22 21 20 43 42 41 40 03 02 01 00
+  // A1 = 73 72 71 70 33 32 31 30 53 52 51 50 13 12 11 10
+  const __m128i A0 = _mm_set_epi32(
+      MemToUint32(&b[6 * stride]), MemToUint32(&b[2 * stride]),
+      MemToUint32(&b[4 * stride]), MemToUint32(&b[0 * stride]));
+  const __m128i A1 = _mm_set_epi32(
+      MemToUint32(&b[7 * stride]), MemToUint32(&b[3 * stride]),
+      MemToUint32(&b[5 * stride]), MemToUint32(&b[1 * stride]));
+
+  // B0 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
+  // B1 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
+  const __m128i B0 = _mm_unpacklo_epi8(A0, A1);
+  const __m128i B1 = _mm_unpackhi_epi8(A0, A1);
+
+  // C0 = 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
+  // C1 = 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40
+  const __m128i C0 = _mm_unpacklo_epi16(B0, B1);
+  const __m128i C1 = _mm_unpackhi_epi16(B0, B1);
 
   // *p = 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00
   // *q = 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
-  *p = _mm_unpacklo_epi32(t1, t2);
-  *q = _mm_unpackhi_epi32(t1, t2);
+  *p = _mm_unpacklo_epi32(C0, C1);
+  *q = _mm_unpackhi_epi32(C0, C1);
 }
 
-static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8,
+static WEBP_INLINE void Load16x4(const uint8_t* const r0,
+                                 const uint8_t* const r8,
                                  int stride,
-                                 __m128i* p1, __m128i* p0,
-                                 __m128i* q0, __m128i* q1) {
-  __m128i t1, t2;
+                                 __m128i* const p1, __m128i* const p0,
+                                 __m128i* const q0, __m128i* const q1) {
   // Assume the pixels around the edge (|) are numbered as follows
   //                00 01 | 02 03
   //                10 11 | 12 13
@@ -532,19 +575,21 @@ static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8,
   Load8x4(r0, stride, p1, q0);
   Load8x4(r8, stride, p0, q1);
 
-  t1 = *p1;
-  t2 = *q0;
-  // p1 = f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
-  // p0 = f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01
-  // q0 = f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
-  // q1 = f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
-  *p1 = _mm_unpacklo_epi64(t1, *p0);
-  *p0 = _mm_unpackhi_epi64(t1, *p0);
-  *q0 = _mm_unpacklo_epi64(t2, *q1);
-  *q1 = _mm_unpackhi_epi64(t2, *q1);
+  {
+    // p1 = f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
+    // p0 = f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01
+    // q0 = f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+    // q1 = f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
+    const __m128i t1 = *p1;
+    const __m128i t2 = *q0;
+    *p1 = _mm_unpacklo_epi64(t1, *p0);
+    *p0 = _mm_unpackhi_epi64(t1, *p0);
+    *q0 = _mm_unpacklo_epi64(t2, *q1);
+    *q1 = _mm_unpackhi_epi64(t2, *q1);
+  }
 }
 
-static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) {
+static WEBP_INLINE void Store4x4(__m128i* const x, uint8_t* dst, int stride) {
   int i;
   for (i = 0; i < 4; ++i, dst += stride) {
     *((int32_t*)dst) = _mm_cvtsi128_si32(*x);
@@ -553,48 +598,51 @@ static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) {
 }
 
 // Transpose back and store
-static WEBP_INLINE void Store16x4(uint8_t* r0, uint8_t* r8, int stride,
-                                  __m128i* p1, __m128i* p0,
-                                  __m128i* q0, __m128i* q1) {
-  __m128i t1;
+static WEBP_INLINE void Store16x4(const __m128i* const p1,
+                                  const __m128i* const p0,
+                                  const __m128i* const q0,
+                                  const __m128i* const q1,
+                                  uint8_t* r0, uint8_t* r8,
+                                  int stride) {
+  __m128i t1, p1_s, p0_s, q0_s, q1_s;
 
   // p0 = 71 70 61 60 51 50 41 40 31 30 21 20 11 10 01 00
   // p1 = f1 f0 e1 e0 d1 d0 c1 c0 b1 b0 a1 a0 91 90 81 80
   t1 = *p0;
-  *p0 = _mm_unpacklo_epi8(*p1, t1);
-  *p1 = _mm_unpackhi_epi8(*p1, t1);
+  p0_s = _mm_unpacklo_epi8(*p1, t1);
+  p1_s = _mm_unpackhi_epi8(*p1, t1);
 
   // q0 = 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02
   // q1 = f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82
   t1 = *q0;
-  *q0 = _mm_unpacklo_epi8(t1, *q1);
-  *q1 = _mm_unpackhi_epi8(t1, *q1);
+  q0_s = _mm_unpacklo_epi8(t1, *q1);
+  q1_s = _mm_unpackhi_epi8(t1, *q1);
 
   // p0 = 33 32 31 30 23 22 21 20 13 12 11 10 03 02 01 00
   // q0 = 73 72 71 70 63 62 61 60 53 52 51 50 43 42 41 40
-  t1 = *p0;
-  *p0 = _mm_unpacklo_epi16(t1, *q0);
-  *q0 = _mm_unpackhi_epi16(t1, *q0);
+  t1 = p0_s;
+  p0_s = _mm_unpacklo_epi16(t1, q0_s);
+  q0_s = _mm_unpackhi_epi16(t1, q0_s);
 
   // p1 = b3 b2 b1 b0 a3 a2 a1 a0 93 92 91 90 83 82 81 80
   // q1 = f3 f2 f1 f0 e3 e2 e1 e0 d3 d2 d1 d0 c3 c2 c1 c0
-  t1 = *p1;
-  *p1 = _mm_unpacklo_epi16(t1, *q1);
-  *q1 = _mm_unpackhi_epi16(t1, *q1);
+  t1 = p1_s;
+  p1_s = _mm_unpacklo_epi16(t1, q1_s);
+  q1_s = _mm_unpackhi_epi16(t1, q1_s);
 
-  Store4x4(p0, r0, stride);
+  Store4x4(&p0_s, r0, stride);
   r0 += 4 * stride;
-  Store4x4(q0, r0, stride);
+  Store4x4(&q0_s, r0, stride);
 
-  Store4x4(p1, r8, stride);
+  Store4x4(&p1_s, r8, stride);
   r8 += 4 * stride;
-  Store4x4(q1, r8, stride);
+  Store4x4(&q1_s, r8, stride);
 }
 
 //------------------------------------------------------------------------------
 // Simple In-loop filtering (Paragraph 15.2)
 
-static void SimpleVFilter16SSE2(uint8_t* p, int stride, int thresh) {
+static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
   // Load
   __m128i p1 = _mm_loadu_si128((__m128i*)&p[-2 * stride]);
   __m128i p0 = _mm_loadu_si128((__m128i*)&p[-stride]);
@@ -605,49 +653,49 @@ static void SimpleVFilter16SSE2(uint8_t* p, int stride, int thresh) {
 
   // Store
   _mm_storeu_si128((__m128i*)&p[-stride], p0);
-  _mm_storeu_si128((__m128i*)p, q0);
+  _mm_storeu_si128((__m128i*)&p[0], q0);
 }
 
-static void SimpleHFilter16SSE2(uint8_t* p, int stride, int thresh) {
+static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
   __m128i p1, p0, q0, q1;
 
   p -= 2;  // beginning of p1
 
-  Load16x4(p, p + 8 * stride,  stride, &p1, &p0, &q0, &q1);
+  Load16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1);
   DoFilter2(&p1, &p0, &q0, &q1, thresh);
-  Store16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1);
+  Store16x4(&p1, &p0, &q0, &q1, p, p + 8 * stride, stride);
 }
 
-static void SimpleVFilter16iSSE2(uint8_t* p, int stride, int thresh) {
+static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
   int k;
   for (k = 3; k > 0; --k) {
     p += 4 * stride;
-    SimpleVFilter16SSE2(p, stride, thresh);
+    SimpleVFilter16(p, stride, thresh);
   }
 }
 
-static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) {
+static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
   int k;
   for (k = 3; k > 0; --k) {
     p += 4;
-    SimpleHFilter16SSE2(p, stride, thresh);
+    SimpleHFilter16(p, stride, thresh);
   }
 }
 
 //------------------------------------------------------------------------------
 // Complex In-loop filtering (Paragraph 15.3)
 
-#define MAX_DIFF1(p3, p2, p1, p0, m) {                                         \
-  m = MM_ABS(p3, p2);                                                          \
+#define MAX_DIFF1(p3, p2, p1, p0, m) do {                                      \
+  m = MM_ABS(p1, p0);                                                          \
+  m = _mm_max_epu8(m, MM_ABS(p3, p2));                                         \
   m = _mm_max_epu8(m, MM_ABS(p2, p1));                                         \
-  m = _mm_max_epu8(m, MM_ABS(p1, p0));                                         \
-}
+} while (0)
 
-#define MAX_DIFF2(p3, p2, p1, p0, m) {                                         \
+#define MAX_DIFF2(p3, p2, p1, p0, m) do {                                      \
+  m = _mm_max_epu8(m, MM_ABS(p1, p0));                                         \
   m = _mm_max_epu8(m, MM_ABS(p3, p2));                                         \
   m = _mm_max_epu8(m, MM_ABS(p2, p1));                                         \
-  m = _mm_max_epu8(m, MM_ABS(p1, p0));                                         \
-}
+} while (0)
 
 #define LOAD_H_EDGES4(p, stride, e1, e2, e3, e4) {                             \
   e1 = _mm_loadu_si128((__m128i*)&(p)[0 * stride]);                            \
@@ -656,10 +704,11 @@ static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) {
   e4 = _mm_loadu_si128((__m128i*)&(p)[3 * stride]);                            \
 }
 
-#define LOADUV_H_EDGE(p, u, v, stride) {                                       \
-  p = _mm_loadl_epi64((__m128i*)&(u)[(stride)]);                               \
-  p = _mm_unpacklo_epi64(p, _mm_loadl_epi64((__m128i*)&(v)[(stride)]));        \
-}
+#define LOADUV_H_EDGE(p, u, v, stride) do {                                    \
+  const __m128i U = _mm_loadl_epi64((__m128i*)&(u)[(stride)]);                 \
+  const __m128i V = _mm_loadl_epi64((__m128i*)&(v)[(stride)]);                 \
+  p = _mm_unpacklo_epi64(U, V);                                                \
+} while (0)
 
 #define LOADUV_H_EDGES4(u, v, stride, e1, e2, e3, e4) {                        \
   LOADUV_H_EDGE(e1, u, v, 0 * stride);                                         \
@@ -674,18 +723,23 @@ static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) {
   _mm_storel_epi64((__m128i*)&v[(stride)], p);                                 \
 }
 
-#define COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask) {               \
-  __m128i fl_yes;                                                              \
-  const __m128i it = _mm_set1_epi8(ithresh);                                   \
-  mask = _mm_subs_epu8(mask, it);                                              \
-  mask = _mm_cmpeq_epi8(mask, _mm_setzero_si128());                            \
-  NeedsFilter(&p1, &p0, &q0, &q1, thresh, &fl_yes);                            \
-  mask = _mm_and_si128(mask, fl_yes);                                          \
+static WEBP_INLINE void ComplexMask(const __m128i* const p1,
+                                    const __m128i* const p0,
+                                    const __m128i* const q0,
+                                    const __m128i* const q1,
+                                    int thresh, int ithresh,
+                                    __m128i* const mask) {
+  const __m128i it = _mm_set1_epi8(ithresh);
+  const __m128i diff = _mm_subs_epu8(*mask, it);
+  const __m128i thresh_mask = _mm_cmpeq_epi8(diff, _mm_setzero_si128());
+  __m128i filter_mask;
+  NeedsFilter(p1, p0, q0, q1, thresh, &filter_mask);
+  *mask = _mm_and_si128(thresh_mask, filter_mask);
 }
 
 // on macroblock edges
-static void VFilter16SSE2(uint8_t* p, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void VFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
   __m128i t1;
   __m128i mask;
   __m128i p2, p1, p0, q0, q1, q2;
@@ -698,20 +752,20 @@ static void VFilter16SSE2(uint8_t* p, int stride,
   LOAD_H_EDGES4(p, stride, q0, q1, q2, t1);
   MAX_DIFF2(t1, q2, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
 
   // Store
   _mm_storeu_si128((__m128i*)&p[-3 * stride], p2);
   _mm_storeu_si128((__m128i*)&p[-2 * stride], p1);
   _mm_storeu_si128((__m128i*)&p[-1 * stride], p0);
-  _mm_storeu_si128((__m128i*)&p[0 * stride], q0);
-  _mm_storeu_si128((__m128i*)&p[1 * stride], q1);
-  _mm_storeu_si128((__m128i*)&p[2 * stride], q2);
+  _mm_storeu_si128((__m128i*)&p[+0 * stride], q0);
+  _mm_storeu_si128((__m128i*)&p[+1 * stride], q1);
+  _mm_storeu_si128((__m128i*)&p[+2 * stride], q2);
 }
 
-static void HFilter16SSE2(uint8_t* p, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void HFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i p3, p2, p1, p0, q0, q1, q2, q3;
 
@@ -722,71 +776,78 @@ static void HFilter16SSE2(uint8_t* p, int stride,
   Load16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3);  // q0, q1, q2, q3
   MAX_DIFF2(q3, q2, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
 
-  Store16x4(b, b + 8 * stride, stride, &p3, &p2, &p1, &p0);
-  Store16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3);
+  Store16x4(&p3, &p2, &p1, &p0, b, b + 8 * stride, stride);
+  Store16x4(&q0, &q1, &q2, &q3, p, p + 8 * stride, stride);
 }
 
 // on three inner edges
-static void VFilter16iSSE2(uint8_t* p, int stride,
-                           int thresh, int ithresh, int hev_thresh) {
+static void VFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
   int k;
-  __m128i mask;
-  __m128i t1, t2, p1, p0, q0, q1;
+  __m128i p3, p2, p1, p0;   // loop invariants
 
-  for (k = 3; k > 0; --k) {
-    // Load p3, p2, p1, p0
-    LOAD_H_EDGES4(p, stride, t2, t1, p1, p0);
-    MAX_DIFF1(t2, t1, p1, p0, mask);
+  LOAD_H_EDGES4(p, stride, p3, p2, p1, p0);  // prologue
 
+  for (k = 3; k > 0; --k) {
+    __m128i mask, tmp1, tmp2;
+    uint8_t* const b = p + 2 * stride;   // beginning of p1
     p += 4 * stride;
 
-    // Load q0, q1, q2, q3
-    LOAD_H_EDGES4(p, stride, q0, q1, t1, t2);
-    MAX_DIFF2(t2, t1, q1, q0, mask);
+    MAX_DIFF1(p3, p2, p1, p0, mask);   // compute partial mask
+    LOAD_H_EDGES4(p, stride, p3, p2, tmp1, tmp2);
+    MAX_DIFF2(p3, p2, tmp1, tmp2, mask);
 
-    COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
-    DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
+    // p3 and p2 are not just temporary variables here: they will be
+    // re-used for next span. And q2/q3 will become p1/p0 accordingly.
+    ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask);
+    DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh);
 
     // Store
-    _mm_storeu_si128((__m128i*)&p[-2 * stride], p1);
-    _mm_storeu_si128((__m128i*)&p[-1 * stride], p0);
-    _mm_storeu_si128((__m128i*)&p[0 * stride], q0);
-    _mm_storeu_si128((__m128i*)&p[1 * stride], q1);
+    _mm_storeu_si128((__m128i*)&b[0 * stride], p1);
+    _mm_storeu_si128((__m128i*)&b[1 * stride], p0);
+    _mm_storeu_si128((__m128i*)&b[2 * stride], p3);
+    _mm_storeu_si128((__m128i*)&b[3 * stride], p2);
+
+    // rotate samples
+    p1 = tmp1;
+    p0 = tmp2;
   }
 }
 
-static void HFilter16iSSE2(uint8_t* p, int stride,
-                           int thresh, int ithresh, int hev_thresh) {
+static void HFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
   int k;
-  uint8_t* b;
-  __m128i mask;
-  __m128i t1, t2, p1, p0, q0, q1;
+  __m128i p3, p2, p1, p0;   // loop invariants
+
+  Load16x4(p, p + 8 * stride, stride, &p3, &p2, &p1, &p0);  // prologue
 
   for (k = 3; k > 0; --k) {
-    b = p;
-    Load16x4(b, b + 8 * stride, stride, &t2, &t1, &p1, &p0);  // p3, p2, p1, p0
-    MAX_DIFF1(t2, t1, p1, p0, mask);
+    __m128i mask, tmp1, tmp2;
+    uint8_t* const b = p + 2;   // beginning of p1
 
-    b += 4;  // beginning of q0
-    Load16x4(b, b + 8 * stride, stride, &q0, &q1, &t1, &t2);  // q0, q1, q2, q3
-    MAX_DIFF2(t2, t1, q1, q0, mask);
+    p += 4;  // beginning of q0 (and next span)
 
-    COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
-    DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
+    MAX_DIFF1(p3, p2, p1, p0, mask);   // compute partial mask
+    Load16x4(p, p + 8 * stride, stride, &p3, &p2, &tmp1, &tmp2);
+    MAX_DIFF2(p3, p2, tmp1, tmp2, mask);
 
-    b -= 2;  // beginning of p1
-    Store16x4(b, b + 8 * stride, stride, &p1, &p0, &q0, &q1);
+    ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask);
+    DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh);
 
-    p += 4;
+    Store16x4(&p1, &p0, &p3, &p2, b, b + 8 * stride, stride);
+
+    // rotate samples
+    p1 = tmp1;
+    p0 = tmp2;
   }
 }
 
 // 8-pixels wide variant, for chroma filtering
-static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
-                         int thresh, int ithresh, int hev_thresh) {
+static void VFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, p2, p1, p0, q0, q1, q2;
 
@@ -798,7 +859,7 @@ static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
   LOADUV_H_EDGES4(u, v, stride, q0, q1, q2, t1);
   MAX_DIFF2(t1, q2, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
 
   // Store
@@ -810,8 +871,8 @@ static void VFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
   STOREUV(q2, u, v, 2 * stride);
 }
 
-static void HFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
-                         int thresh, int ithresh, int hev_thresh) {
+static void HFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i p3, p2, p1, p0, q0, q1, q2, q3;
 
@@ -823,15 +884,15 @@ static void HFilter8SSE2(uint8_t* u, uint8_t* v, int stride,
   Load16x4(u, v, stride, &q0, &q1, &q2, &q3);    // q0, q1, q2, q3
   MAX_DIFF2(q3, q2, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
 
-  Store16x4(tu, tv, stride, &p3, &p2, &p1, &p0);
-  Store16x4(u, v, stride, &q0, &q1, &q2, &q3);
+  Store16x4(&p3, &p2, &p1, &p0, tu, tv, stride);
+  Store16x4(&q0, &q1, &q2, &q3, u, v, stride);
 }
 
-static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, t2, p1, p0, q0, q1;
 
@@ -846,7 +907,7 @@ static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
   LOADUV_H_EDGES4(u, v, stride, q0, q1, t1, t2);
   MAX_DIFF2(t2, t1, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
 
   // Store
@@ -856,8 +917,8 @@ static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
   STOREUV(q1, u, v, 1 * stride);
 }
 
-static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
-                          int thresh, int ithresh, int hev_thresh) {
+static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
   __m128i mask;
   __m128i t1, t2, p1, p0, q0, q1;
   Load16x4(u, v, stride, &t2, &t1, &p1, &p0);   // p3, p2, p1, p0
@@ -868,36 +929,361 @@ static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
   Load16x4(u, v, stride, &q0, &q1, &t1, &t2);  // q0, q1, q2, q3
   MAX_DIFF2(t2, t1, q1, q0, mask);
 
-  COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+  ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
 
   u -= 2;  // beginning of p1
   v -= 2;
-  Store16x4(u, v, stride, &p1, &p0, &q0, &q1);
+  Store16x4(&p1, &p0, &q0, &q1, u, v, stride);
 }
 
-extern void VP8DspInitSSE2(void);
+//------------------------------------------------------------------------------
+// 4x4 predictions
+
+#define DST(x, y) dst[(x) + (y) * BPS]
+#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
+
+// We use the following 8b-arithmetic tricks:
+//     (a + 2 * b + c + 2) >> 2 = (AC + b + 1) >> 1
+//   where: AC = (a + c) >> 1 = [(a + c + 1) >> 1] - [(a^c) & 1]
+// and:
+//     (a + 2 * b + c + 2) >> 2 = (AB + BC + 1) >> 1 - (ab|bc)&lsb
+//   where: AC = (a + b + 1) >> 1,   BC = (b + c + 1) >> 1
+//   and ab = a ^ b, bc = b ^ c, lsb = (AC^BC)&1
+
+static void VE4(uint8_t* dst) {    // vertical
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS - 1));
+  const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i a = _mm_avg_epu8(ABCDEFGH, CDEFGH00);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one);
+  const __m128i b = _mm_subs_epu8(a, lsb);
+  const __m128i avg = _mm_avg_epu8(b, BCDEFGH0);
+  const uint32_t vals = _mm_cvtsi128_si32(avg);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    *(uint32_t*)(dst + i * BPS) = vals;
+  }
+}
+
+static void LD4(uint8_t* dst) {   // Down-Left
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS));
+  const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i CDEFGHH0 = _mm_insert_epi16(CDEFGH00, dst[-BPS + 7], 3);
+  const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, CDEFGHH0);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0);
+  *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(               abcdefg    );
+  *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1));
+  *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2));
+  *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3));
+}
+
+static void VR4(uint8_t* dst) {   // Vertical-Right
+  const __m128i one = _mm_set1_epi8(1);
+  const int I = dst[-1 + 0 * BPS];
+  const int J = dst[-1 + 1 * BPS];
+  const int K = dst[-1 + 2 * BPS];
+  const int X = dst[-1 - BPS];
+  const __m128i XABCD = _mm_loadl_epi64((__m128i*)(dst - BPS - 1));
+  const __m128i ABCD0 = _mm_srli_si128(XABCD, 1);
+  const __m128i abcd = _mm_avg_epu8(XABCD, ABCD0);
+  const __m128i _XABCD = _mm_slli_si128(XABCD, 1);
+  const __m128i IXABCD = _mm_insert_epi16(_XABCD, I | (X << 8), 0);
+  const __m128i avg1 = _mm_avg_epu8(IXABCD, ABCD0);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i efgh = _mm_avg_epu8(avg2, XABCD);
+  *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(               abcd    );
+  *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(               efgh    );
+  *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1));
+  *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1));
+
+  // these two are hard to implement in SSE2, so we keep the C-version:
+  DST(0, 2) = AVG3(J, I, X);
+  DST(0, 3) = AVG3(K, J, I);
+}
 
-void VP8DspInitSSE2(void) {
-  VP8Transform = TransformSSE2;
+static void VL4(uint8_t* dst) {   // Vertical-Left
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS));
+  const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH__ = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, BCDEFGH_);
+  const __m128i avg2 = _mm_avg_epu8(CDEFGH__, BCDEFGH_);
+  const __m128i avg3 = _mm_avg_epu8(avg1, avg2);
+  const __m128i lsb1 = _mm_and_si128(_mm_xor_si128(avg1, avg2), one);
+  const __m128i ab = _mm_xor_si128(ABCDEFGH, BCDEFGH_);
+  const __m128i bc = _mm_xor_si128(CDEFGH__, BCDEFGH_);
+  const __m128i abbc = _mm_or_si128(ab, bc);
+  const __m128i lsb2 = _mm_and_si128(abbc, lsb1);
+  const __m128i avg4 = _mm_subs_epu8(avg3, lsb2);
+  const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
+  *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(               avg1    );
+  *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(               avg4    );
+  *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1));
+  *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1));
+
+  // these two are hard to get and irregular
+  DST(3, 2) = (extra_out >> 0) & 0xff;
+  DST(3, 3) = (extra_out >> 8) & 0xff;
+}
+
+static void RD4(uint8_t* dst) {   // Down-right
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i XABCD = _mm_loadl_epi64((__m128i*)(dst - BPS - 1));
+  const __m128i ____XABCD = _mm_slli_si128(XABCD, 4);
+  const uint32_t I = dst[-1 + 0 * BPS];
+  const uint32_t J = dst[-1 + 1 * BPS];
+  const uint32_t K = dst[-1 + 2 * BPS];
+  const uint32_t L = dst[-1 + 3 * BPS];
+  const __m128i LKJI_____ =
+      _mm_cvtsi32_si128(L | (K << 8) | (J << 16) | (I << 24));
+  const __m128i LKJIXABCD = _mm_or_si128(LKJI_____, ____XABCD);
+  const __m128i KJIXABCD_ = _mm_srli_si128(LKJIXABCD, 1);
+  const __m128i JIXABCD__ = _mm_srli_si128(LKJIXABCD, 2);
+  const __m128i avg1 = _mm_avg_epu8(JIXABCD__, LKJIXABCD);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_);
+  *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(               abcdefg    );
+  *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1));
+  *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2));
+  *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3));
+}
+
+#undef DST
+#undef AVG3
+
+//------------------------------------------------------------------------------
+// Luma 16x16
+
+static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) {
+  const uint8_t* top = dst - BPS;
+  const __m128i zero = _mm_setzero_si128();
+  int y;
+  if (size == 4) {
+    const __m128i top_values = _mm_cvtsi32_si128(MemToUint32(top));
+    const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
+    for (y = 0; y < 4; ++y, dst += BPS) {
+      const int val = dst[-1] - top[-1];
+      const __m128i base = _mm_set1_epi16(val);
+      const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
+      *(int*)dst = _mm_cvtsi128_si32(out);
+    }
+  } else if (size == 8) {
+    const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
+    const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
+    for (y = 0; y < 8; ++y, dst += BPS) {
+      const int val = dst[-1] - top[-1];
+      const __m128i base = _mm_set1_epi16(val);
+      const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
+      _mm_storel_epi64((__m128i*)dst, out);
+    }
+  } else {
+    const __m128i top_values = _mm_loadu_si128((const __m128i*)top);
+    const __m128i top_base_0 = _mm_unpacklo_epi8(top_values, zero);
+    const __m128i top_base_1 = _mm_unpackhi_epi8(top_values, zero);
+    for (y = 0; y < 16; ++y, dst += BPS) {
+      const int val = dst[-1] - top[-1];
+      const __m128i base = _mm_set1_epi16(val);
+      const __m128i out_0 = _mm_add_epi16(base, top_base_0);
+      const __m128i out_1 = _mm_add_epi16(base, top_base_1);
+      const __m128i out = _mm_packus_epi16(out_0, out_1);
+      _mm_storeu_si128((__m128i*)dst, out);
+    }
+  }
+}
+
+static void TM4(uint8_t* dst)   { TrueMotion(dst, 4); }
+static void TM8uv(uint8_t* dst) { TrueMotion(dst, 8); }
+static void TM16(uint8_t* dst)  { TrueMotion(dst, 16); }
+
+static void VE16(uint8_t* dst) {
+  const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS));
+  int j;
+  for (j = 0; j < 16; ++j) {
+    _mm_storeu_si128((__m128i*)(dst + j * BPS), top);
+  }
+}
+
+static void HE16(uint8_t* dst) {     // horizontal
+  int j;
+  for (j = 16; j > 0; --j) {
+    const __m128i values = _mm_set1_epi8(dst[-1]);
+    _mm_storeu_si128((__m128i*)dst, values);
+    dst += BPS;
+  }
+}
+
+static WEBP_INLINE void Put16(uint8_t v, uint8_t* dst) {
+  int j;
+  const __m128i values = _mm_set1_epi8(v);
+  for (j = 0; j < 16; ++j) {
+    _mm_storeu_si128((__m128i*)(dst + j * BPS), values);
+  }
+}
+
+static void DC16(uint8_t* dst) {    // DC
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS));
+  const __m128i sad8x2 = _mm_sad_epu8(top, zero);
+  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
+  const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
+  int left = 0;
+  int j;
+  for (j = 0; j < 16; ++j) {
+    left += dst[-1 + j * BPS];
+  }
+  {
+    const int DC = _mm_cvtsi128_si32(sum) + left + 16;
+    Put16(DC >> 5, dst);
+  }
+}
+
+static void DC16NoTop(uint8_t* dst) {   // DC with top samples not available
+  int DC = 8;
+  int j;
+  for (j = 0; j < 16; ++j) {
+    DC += dst[-1 + j * BPS];
+  }
+  Put16(DC >> 4, dst);
+}
+
+static void DC16NoLeft(uint8_t* dst) {  // DC with left samples not available
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS));
+  const __m128i sad8x2 = _mm_sad_epu8(top, zero);
+  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
+  const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
+  const int DC = _mm_cvtsi128_si32(sum) + 8;
+  Put16(DC >> 4, dst);
+}
 
-  VP8VFilter16 = VFilter16SSE2;
-  VP8HFilter16 = HFilter16SSE2;
-  VP8VFilter8 = VFilter8SSE2;
-  VP8HFilter8 = HFilter8SSE2;
-  VP8VFilter16i = VFilter16iSSE2;
-  VP8HFilter16i = HFilter16iSSE2;
-  VP8VFilter8i = VFilter8iSSE2;
-  VP8HFilter8i = HFilter8iSSE2;
+static void DC16NoTopLeft(uint8_t* dst) {  // DC with no top and left samples
+  Put16(0x80, dst);
+}
+
+//------------------------------------------------------------------------------
+// Chroma
+
+static void VE8uv(uint8_t* dst) {    // vertical
+  int j;
+  const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS));
+  for (j = 0; j < 8; ++j) {
+    _mm_storel_epi64((__m128i*)(dst + j * BPS), top);
+  }
+}
+
+static void HE8uv(uint8_t* dst) {    // horizontal
+  int j;
+  for (j = 0; j < 8; ++j) {
+    const __m128i values = _mm_set1_epi8(dst[-1]);
+    _mm_storel_epi64((__m128i*)dst, values);
+    dst += BPS;
+  }
+}
+
+// helper for chroma-DC predictions
+static WEBP_INLINE void Put8x8uv(uint8_t v, uint8_t* dst) {
+  int j;
+  const __m128i values = _mm_set1_epi8(v);
+  for (j = 0; j < 8; ++j) {
+    _mm_storel_epi64((__m128i*)(dst + j * BPS), values);
+  }
+}
 
-  VP8SimpleVFilter16 = SimpleVFilter16SSE2;
-  VP8SimpleHFilter16 = SimpleHFilter16SSE2;
-  VP8SimpleVFilter16i = SimpleVFilter16iSSE2;
-  VP8SimpleHFilter16i = SimpleHFilter16iSSE2;
+static void DC8uv(uint8_t* dst) {     // DC
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS));
+  const __m128i sum = _mm_sad_epu8(top, zero);
+  int left = 0;
+  int j;
+  for (j = 0; j < 8; ++j) {
+    left += dst[-1 + j * BPS];
+  }
+  {
+    const int DC = _mm_cvtsi128_si32(sum) + left + 8;
+    Put8x8uv(DC >> 4, dst);
+  }
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
+static void DC8uvNoLeft(uint8_t* dst) {   // DC with no left samples
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS));
+  const __m128i sum = _mm_sad_epu8(top, zero);
+  const int DC = _mm_cvtsi128_si32(sum) + 4;
+  Put8x8uv(DC >> 3, dst);
+}
+
+static void DC8uvNoTop(uint8_t* dst) {  // DC with no top samples
+  int dc0 = 4;
+  int i;
+  for (i = 0; i < 8; ++i) {
+    dc0 += dst[-1 + i * BPS];
+  }
+  Put8x8uv(dc0 >> 3, dst);
+}
+
+static void DC8uvNoTopLeft(uint8_t* dst) {    // DC with nothing
+  Put8x8uv(0x80, dst);
+}
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8DspInitSSE2(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitSSE2(void) {
+  VP8Transform = Transform;
+#if defined(USE_TRANSFORM_AC3)
+  VP8TransformAC3 = TransformAC3;
 #endif
 
-#endif   // WEBP_USE_SSE2
+  VP8VFilter16 = VFilter16;
+  VP8HFilter16 = HFilter16;
+  VP8VFilter8 = VFilter8;
+  VP8HFilter8 = HFilter8;
+  VP8VFilter16i = VFilter16i;
+  VP8HFilter16i = HFilter16i;
+  VP8VFilter8i = VFilter8i;
+  VP8HFilter8i = HFilter8i;
+
+  VP8SimpleVFilter16 = SimpleVFilter16;
+  VP8SimpleHFilter16 = SimpleHFilter16;
+  VP8SimpleVFilter16i = SimpleVFilter16i;
+  VP8SimpleHFilter16i = SimpleHFilter16i;
+
+  VP8PredLuma4[1] = TM4;
+  VP8PredLuma4[2] = VE4;
+  VP8PredLuma4[4] = RD4;
+  VP8PredLuma4[5] = VR4;
+  VP8PredLuma4[6] = LD4;
+  VP8PredLuma4[7] = VL4;
+
+  VP8PredLuma16[0] = DC16;
+  VP8PredLuma16[1] = TM16;
+  VP8PredLuma16[2] = VE16;
+  VP8PredLuma16[3] = HE16;
+  VP8PredLuma16[4] = DC16NoTop;
+  VP8PredLuma16[5] = DC16NoLeft;
+  VP8PredLuma16[6] = DC16NoTopLeft;
+
+  VP8PredChroma8[0] = DC8uv;
+  VP8PredChroma8[1] = TM8uv;
+  VP8PredChroma8[2] = VE8uv;
+  VP8PredChroma8[3] = HE8uv;
+  VP8PredChroma8[4] = DC8uvNoTop;
+  VP8PredChroma8[5] = DC8uvNoLeft;
+  VP8PredChroma8[6] = DC8uvNoTopLeft;
+}
+
+#else  // !WEBP_USE_SSE2
+
+WEBP_DSP_INIT_STUB(VP8DspInitSSE2)
+
+#endif  // WEBP_USE_SSE2
diff --git a/drivers/webp/dsp/dsp.h b/drivers/webp/dsp/dsp.h
index fd686a8532..8395df40e4 100644
--- a/drivers/webp/dsp/dsp.h
+++ b/drivers/webp/dsp/dsp.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 //   Speed-critical functions.
@@ -12,44 +14,121 @@
 #ifndef WEBP_DSP_DSP_H_
 #define WEBP_DSP_DSP_H_
 
-#include "../types.h"
+#ifdef HAVE_CONFIG_H
+#include "../webp/config.h"
+#endif
+
+#include "../webp/types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
+#define BPS 32   // this is the common stride for enc/dec
+
 //------------------------------------------------------------------------------
 // CPU detection
 
-#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
+#if defined(__GNUC__)
+# define LOCAL_GCC_VERSION ((__GNUC__ << 8) | __GNUC_MINOR__)
+# define LOCAL_GCC_PREREQ(maj, min) \
+    (LOCAL_GCC_VERSION >= (((maj) << 8) | (min)))
+#else
+# define LOCAL_GCC_VERSION 0
+# define LOCAL_GCC_PREREQ(maj, min) 0
+#endif
+
+#ifndef __has_builtin
+# define __has_builtin(x) 0
+#endif
+
+#if defined(_MSC_VER) && _MSC_VER > 1310 && \
+    (defined(_M_X64) || defined(_M_IX86))
 #define WEBP_MSC_SSE2  // Visual C++ SSE2 targets
 #endif
 
-#if defined(__SSE2__) || defined(WEBP_MSC_SSE2)
+#if defined(_MSC_VER) && _MSC_VER >= 1500 && \
+    (defined(_M_X64) || defined(_M_IX86))
+#define WEBP_MSC_SSE41  // Visual C++ SSE4.1 targets
+#endif
+
+// WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp
+// files without intrinsics, allowing the corresponding Init() to be called.
+// Files containing intrinsics will need to be built targeting the instruction
+// set so should succeed on one of the earlier tests.
+#if defined(__SSE2__) || defined(WEBP_MSC_SSE2) || defined(WEBP_HAVE_SSE2)
 #define WEBP_USE_SSE2
 #endif
 
-#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__) && defined(__ARM_NEON__)
+#if defined(__SSE4_1__) || defined(WEBP_MSC_SSE41) || defined(WEBP_HAVE_SSE41)
+#define WEBP_USE_SSE41
+#endif
+
+#if defined(__AVX2__) || defined(WEBP_HAVE_AVX2)
+#define WEBP_USE_AVX2
+#endif
+
+#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__)
 #define WEBP_ANDROID_NEON  // Android targets that might support NEON
 #endif
 
-#if ( (defined(__ARM_NEON__) && !defined(__aarch64__)) || defined(WEBP_ANDROID_NEON)) && !defined(PSP2_ENABLED)
+// The intrinsics currently cause compiler errors with arm-nacl-gcc and the
+// inline assembly would need to be modified for use with Native Client.
+#if (defined(__ARM_NEON__) || defined(WEBP_ANDROID_NEON) || \
+     defined(__aarch64__)) && !defined(__native_client__)
+#define WEBP_USE_NEON
+#endif
+
+#if defined(_MSC_VER) && _MSC_VER >= 1700 && defined(_M_ARM)
 #define WEBP_USE_NEON
+#define WEBP_USE_INTRINSICS
+#endif
+
+#if defined(__mips__) && !defined(__mips64) && \
+    defined(__mips_isa_rev) && (__mips_isa_rev >= 1) && (__mips_isa_rev < 6)
+#define WEBP_USE_MIPS32
+#if (__mips_isa_rev >= 2)
+#define WEBP_USE_MIPS32_R2
+#if defined(__mips_dspr2) || (__mips_dsp_rev >= 2)
+#define WEBP_USE_MIPS_DSP_R2
+#endif
+#endif
+#endif
+
+// This macro prevents thread_sanitizer from reporting known concurrent writes.
+#define WEBP_TSAN_IGNORE_FUNCTION
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#undef WEBP_TSAN_IGNORE_FUNCTION
+#define WEBP_TSAN_IGNORE_FUNCTION __attribute__((no_sanitize_thread))
+#endif
 #endif
 
 typedef enum {
   kSSE2,
   kSSE3,
-  kNEON
+  kSSE4_1,
+  kAVX,
+  kAVX2,
+  kNEON,
+  kMIPS32,
+  kMIPSdspR2
 } CPUFeature;
 // returns true if the CPU supports the feature.
 typedef int (*VP8CPUInfo)(CPUFeature feature);
-extern VP8CPUInfo VP8GetCPUInfo;
+WEBP_EXTERN(VP8CPUInfo) VP8GetCPUInfo;
 
 //------------------------------------------------------------------------------
-// Encoding
+// Init stub generator
 
-int VP8GetAlpha(const int histo[]);
+// Defines an init function stub to ensure each module exposes a symbol,
+// avoiding a compiler warning.
+#define WEBP_DSP_INIT_STUB(func) \
+  extern void func(void); \
+  WEBP_TSAN_IGNORE_FUNCTION void func(void) {}
+
+//------------------------------------------------------------------------------
+// Encoding
 
 // Transforms
 // VP8Idct: Does one of two inverse transforms. If do_two is set, the transforms
@@ -60,7 +139,7 @@ typedef void (*VP8Fdct)(const uint8_t* src, const uint8_t* ref, int16_t* out);
 typedef void (*VP8WHT)(const int16_t* in, int16_t* out);
 extern VP8Idct VP8ITransform;
 extern VP8Fdct VP8FTransform;
-extern VP8WHT VP8ITransformWHT;
+extern VP8Fdct VP8FTransform2;   // performs two transforms at a time
 extern VP8WHT VP8FTransformWHT;
 // Predictions
 // *dst is the destination block. *top and *left can be NULL.
@@ -79,20 +158,63 @@ extern VP8WMetric VP8TDisto4x4, VP8TDisto16x16;
 
 typedef void (*VP8BlockCopy)(const uint8_t* src, uint8_t* dst);
 extern VP8BlockCopy VP8Copy4x4;
+extern VP8BlockCopy VP8Copy16x8;
 // Quantization
 struct VP8Matrix;   // forward declaration
 typedef int (*VP8QuantizeBlock)(int16_t in[16], int16_t out[16],
-                                int n, const struct VP8Matrix* const mtx);
+                                const struct VP8Matrix* const mtx);
+// Same as VP8QuantizeBlock, but quantizes two consecutive blocks.
+typedef int (*VP8Quantize2Blocks)(int16_t in[32], int16_t out[32],
+                                  const struct VP8Matrix* const mtx);
+
 extern VP8QuantizeBlock VP8EncQuantizeBlock;
+extern VP8Quantize2Blocks VP8EncQuantize2Blocks;
+
+// specific to 2nd transform:
+typedef int (*VP8QuantizeBlockWHT)(int16_t in[16], int16_t out[16],
+                                   const struct VP8Matrix* const mtx);
+extern VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
 
-// Compute susceptibility based on DCT-coeff histograms:
-// the higher, the "easier" the macroblock is to compress.
-typedef int (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred,
-                         int start_block, int end_block);
 extern const int VP8DspScan[16 + 4 + 4];
+
+// Collect histogram for susceptibility calculation.
+#define MAX_COEFF_THRESH   31   // size of histogram used by CollectHistogram.
+typedef struct {
+  // We only need to store max_value and last_non_zero, not the distribution.
+  int max_value;
+  int last_non_zero;
+} VP8Histogram;
+typedef void (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred,
+                          int start_block, int end_block,
+                          VP8Histogram* const histo);
 extern VP8CHisto VP8CollectHistogram;
+// General-purpose util function to help VP8CollectHistogram().
+void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1],
+                         VP8Histogram* const histo);
+
+// must be called before using any of the above
+void VP8EncDspInit(void);
+
+//------------------------------------------------------------------------------
+// cost functions (encoding)
+
+extern const uint16_t VP8EntropyCost[256];        // 8bit fixed-point log(p)
+// approximate cost per level:
+extern const uint16_t VP8LevelFixedCosts[2047 /*MAX_LEVEL*/ + 1];
+extern const uint8_t VP8EncBands[16 + 1];
+
+struct VP8Residual;
+typedef void (*VP8SetResidualCoeffsFunc)(const int16_t* const coeffs,
+                                         struct VP8Residual* const res);
+extern VP8SetResidualCoeffsFunc VP8SetResidualCoeffs;
 
-void VP8EncDspInit(void);   // must be called before using any of the above
+// Cost calculation function.
+typedef int (*VP8GetResidualCostFunc)(int ctx0,
+                                      const struct VP8Residual* const res);
+extern VP8GetResidualCostFunc VP8GetResidualCost;
+
+// must be called before anything using the above
+void VP8EncDspCostInit(void);
 
 //------------------------------------------------------------------------------
 // Decoding
@@ -101,17 +223,26 @@ typedef void (*VP8DecIdct)(const int16_t* coeffs, uint8_t* dst);
 // when doing two transforms, coeffs is actually int16_t[2][16].
 typedef void (*VP8DecIdct2)(const int16_t* coeffs, uint8_t* dst, int do_two);
 extern VP8DecIdct2 VP8Transform;
+extern VP8DecIdct VP8TransformAC3;
 extern VP8DecIdct VP8TransformUV;
 extern VP8DecIdct VP8TransformDC;
 extern VP8DecIdct VP8TransformDCUV;
-extern void (*VP8TransformWHT)(const int16_t* in, int16_t* out);
+extern VP8WHT VP8TransformWHT;
 
 // *dst is the destination block, with stride BPS. Boundary samples are
 // assumed accessible when needed.
 typedef void (*VP8PredFunc)(uint8_t* dst);
-extern const VP8PredFunc VP8PredLuma16[/* NUM_B_DC_MODES */];
-extern const VP8PredFunc VP8PredChroma8[/* NUM_B_DC_MODES */];
-extern const VP8PredFunc VP8PredLuma4[/* NUM_BMODES */];
+extern VP8PredFunc VP8PredLuma16[/* NUM_B_DC_MODES */];
+extern VP8PredFunc VP8PredChroma8[/* NUM_B_DC_MODES */];
+extern VP8PredFunc VP8PredLuma4[/* NUM_BMODES */];
+
+// clipping tables (for filtering)
+extern const int8_t* const VP8ksclip1;  // clips [-1020, 1020] to [-128, 127]
+extern const int8_t* const VP8ksclip2;  // clips [-112, 112] to [-16, 15]
+extern const uint8_t* const VP8kclip1;  // clips [-255,511] to [0,255]
+extern const uint8_t* const VP8kabs0;   // abs(x) for x in [-255,255]
+// must be called first
+void VP8InitClipTables(void);
 
 // simple filter (only for luma)
 typedef void (*VP8SimpleFilterFunc)(uint8_t* p, int stride, int thresh);
@@ -145,6 +276,8 @@ void VP8DspInit(void);
 
 #define FANCY_UPSAMPLING   // undefined to remove fancy upsampling support
 
+// Convert a pair of y/u/v lines together to the output rgb/a colorspace.
+// bottom_y can be NULL if only one line of output is needed (at top/bottom).
 typedef void (*WebPUpsampleLinePairFunc)(
     const uint8_t* top_y, const uint8_t* bottom_y,
     const uint8_t* top_u, const uint8_t* top_v,
@@ -156,18 +289,20 @@ typedef void (*WebPUpsampleLinePairFunc)(
 // Fancy upsampling functions to convert YUV to RGB(A) modes
 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 
-// Initializes SSE2 version of the fancy upsamplers.
-void WebPInitUpsamplersSSE2(void);
-
 #endif    // FANCY_UPSAMPLING
 
-// Point-sampling methods.
-typedef void (*WebPSampleLinePairFunc)(
-    const uint8_t* top_y, const uint8_t* bottom_y,
-    const uint8_t* u, const uint8_t* v,
-    uint8_t* top_dst, uint8_t* bottom_dst, int len);
+// Per-row point-sampling methods.
+typedef void (*WebPSamplerRowFunc)(const uint8_t* y,
+                                   const uint8_t* u, const uint8_t* v,
+                                   uint8_t* dst, int len);
+// Generic function to apply 'WebPSamplerRowFunc' to the whole plane:
+void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
+                             const uint8_t* u, const uint8_t* v, int uv_stride,
+                             uint8_t* dst, int dst_stride,
+                             int width, int height, WebPSamplerRowFunc func);
 
-extern const WebPSampleLinePairFunc WebPSamplers[/* MODE_LAST */];
+// Sampling functions to convert rows of YUV to RGB(A)
+extern WebPSamplerRowFunc WebPSamplers[/* MODE_LAST */];
 
 // General function for converting two lines of ARGB or RGBA.
 // 'alpha_is_last' should be true if 0xff000000 is stored in memory as
@@ -179,13 +314,84 @@ typedef void (*WebPYUV444Converter)(const uint8_t* y,
                                     const uint8_t* u, const uint8_t* v,
                                     uint8_t* dst, int len);
 
-extern const WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
+extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
 
-// Main function to be called
+// Must be called before using the WebPUpsamplers[] (and for premultiplied
+// colorspaces like rgbA, rgbA4444, etc)
 void WebPInitUpsamplers(void);
+// Must be called before using WebPSamplers[]
+void WebPInitSamplers(void);
+// Must be called before using WebPYUV444Converters[]
+void WebPInitYUV444Converters(void);
 
 //------------------------------------------------------------------------------
-// Pre-multiply planes with alpha values
+// ARGB -> YUV converters
+
+// Convert ARGB samples to luma Y.
+extern void (*WebPConvertARGBToY)(const uint32_t* argb, uint8_t* y, int width);
+// Convert ARGB samples to U/V with downsampling. do_store should be '1' for
+// even lines and '0' for odd ones. 'src_width' is the original width, not
+// the U/V one.
+extern void (*WebPConvertARGBToUV)(const uint32_t* argb, uint8_t* u, uint8_t* v,
+                                   int src_width, int do_store);
+
+// Convert a row of accumulated (four-values) of rgba32 toward U/V
+extern void (*WebPConvertRGBA32ToUV)(const uint16_t* rgb,
+                                     uint8_t* u, uint8_t* v, int width);
+
+// Convert RGB or BGR to Y
+extern void (*WebPConvertRGB24ToY)(const uint8_t* rgb, uint8_t* y, int width);
+extern void (*WebPConvertBGR24ToY)(const uint8_t* bgr, uint8_t* y, int width);
+
+// used for plain-C fallback.
+extern void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v,
+                                  int src_width, int do_store);
+extern void WebPConvertRGBA32ToUV_C(const uint16_t* rgb,
+                                    uint8_t* u, uint8_t* v, int width);
+
+// Must be called before using the above.
+void WebPInitConvertARGBToYUV(void);
+
+//------------------------------------------------------------------------------
+// Rescaler
+
+struct WebPRescaler;
+
+// Import a row of data and save its contribution in the rescaler.
+// 'channel' denotes the channel number to be imported. 'Expand' corresponds to
+// the wrk->x_expand case. Otherwise, 'Shrink' is to be used.
+typedef void (*WebPRescalerImportRowFunc)(struct WebPRescaler* const wrk,
+                                          const uint8_t* src);
+
+extern WebPRescalerImportRowFunc WebPRescalerImportRowExpand;
+extern WebPRescalerImportRowFunc WebPRescalerImportRowShrink;
+
+// Export one row (starting at x_out position) from rescaler.
+// 'Expand' corresponds to the wrk->y_expand case.
+// Otherwise 'Shrink' is to be used
+typedef void (*WebPRescalerExportRowFunc)(struct WebPRescaler* const wrk);
+extern WebPRescalerExportRowFunc WebPRescalerExportRowExpand;
+extern WebPRescalerExportRowFunc WebPRescalerExportRowShrink;
+
+// Plain-C implementation, as fall-back.
+extern void WebPRescalerImportRowExpandC(struct WebPRescaler* const wrk,
+                                         const uint8_t* src);
+extern void WebPRescalerImportRowShrinkC(struct WebPRescaler* const wrk,
+                                         const uint8_t* src);
+extern void WebPRescalerExportRowExpandC(struct WebPRescaler* const wrk);
+extern void WebPRescalerExportRowShrinkC(struct WebPRescaler* const wrk);
+
+// Main entry calls:
+extern void WebPRescalerImportRow(struct WebPRescaler* const wrk,
+                                  const uint8_t* src);
+// Export one row (starting at x_out position) from rescaler.
+extern void WebPRescalerExportRow(struct WebPRescaler* const wrk);
+
+// Must be called first before using the above.
+void WebPRescalerDspInit(void);
+
+//------------------------------------------------------------------------------
+// Utilities for processing transparent channel.
 
 // Apply alpha pre-multiply on an rgba, bgra or argb plane of size w * h.
 // alpha_first should be 0 for argb, 1 for rgba or bgra (where alpha is last).
@@ -196,14 +402,98 @@ extern void (*WebPApplyAlphaMultiply)(
 extern void (*WebPApplyAlphaMultiply4444)(
     uint8_t* rgba4444, int w, int h, int stride);
 
+// Dispatch the values from alpha[] plane to the ARGB destination 'dst'.
+// Returns true if alpha[] plane has non-trivial values different from 0xff.
+extern int (*WebPDispatchAlpha)(const uint8_t* alpha, int alpha_stride,
+                                int width, int height,
+                                uint8_t* dst, int dst_stride);
+
+// Transfer packed 8b alpha[] values to green channel in dst[], zero'ing the
+// A/R/B values. 'dst_stride' is the stride for dst[] in uint32_t units.
+extern void (*WebPDispatchAlphaToGreen)(const uint8_t* alpha, int alpha_stride,
+                                        int width, int height,
+                                        uint32_t* dst, int dst_stride);
+
+// Extract the alpha values from 32b values in argb[] and pack them into alpha[]
+// (this is the opposite of WebPDispatchAlpha).
+// Returns true if there's only trivial 0xff alpha values.
+extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride,
+                               int width, int height,
+                               uint8_t* alpha, int alpha_stride);
+
+// Pre-Multiply operation transforms x into x * A / 255  (where x=Y,R,G or B).
+// Un-Multiply operation transforms x into x * 255 / A.
+
+// Pre-Multiply or Un-Multiply (if 'inverse' is true) argb values in a row.
+extern void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse);
+
+// Same a WebPMultARGBRow(), but for several rows.
+void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows,
+                      int inverse);
+
+// Same for a row of single values, with side alpha values.
+extern void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha,
+                           int width, int inverse);
+
+// Same a WebPMultRow(), but for several 'num_rows' rows.
+void WebPMultRows(uint8_t* ptr, int stride,
+                  const uint8_t* alpha, int alpha_stride,
+                  int width, int num_rows, int inverse);
+
+// Plain-C versions, used as fallback by some implementations.
+void WebPMultRowC(uint8_t* const ptr, const uint8_t* const alpha,
+                  int width, int inverse);
+void WebPMultARGBRowC(uint32_t* const ptr, int width, int inverse);
+
 // To be called first before using the above.
-void WebPInitPremultiply(void);
+void WebPInitAlphaProcessing(void);
+
+// ARGB packing function: a/r/g/b input is rgba or bgra order.
+extern void (*VP8PackARGB)(const uint8_t* a, const uint8_t* r,
+                           const uint8_t* g, const uint8_t* b, int len,
+                           uint32_t* out);
+
+// RGB packing function. 'step' can be 3 or 4. r/g/b input is rgb or bgr order.
+extern void (*VP8PackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
+                          int len, int step, uint32_t* out);
 
-void WebPInitPremultiplySSE2(void);   // should not be called directly.
+// To be called first before using the above.
+void VP8EncDspARGBInit(void);
 
 //------------------------------------------------------------------------------
+// Filter functions
+
+typedef enum {     // Filter types.
+  WEBP_FILTER_NONE = 0,
+  WEBP_FILTER_HORIZONTAL,
+  WEBP_FILTER_VERTICAL,
+  WEBP_FILTER_GRADIENT,
+  WEBP_FILTER_LAST = WEBP_FILTER_GRADIENT + 1,  // end marker
+  WEBP_FILTER_BEST,    // meta-types
+  WEBP_FILTER_FAST
+} WEBP_FILTER_TYPE;
+
+typedef void (*WebPFilterFunc)(const uint8_t* in, int width, int height,
+                               int stride, uint8_t* out);
+typedef void (*WebPUnfilterFunc)(int width, int height, int stride,
+                                 int row, int num_rows, uint8_t* data);
+
+// Filter the given data using the given predictor.
+// 'in' corresponds to a 2-dimensional pixel array of size (stride * height)
+// in raster order.
+// 'stride' is number of bytes per scan line (with possible padding).
+// 'out' should be pre-allocated.
+extern WebPFilterFunc WebPFilters[WEBP_FILTER_LAST];
+
+// In-place reconstruct the original data from the given filtered data.
+// The reconstruction will be done for 'num_rows' rows starting from 'row'
+// (assuming rows upto 'row - 1' are already reconstructed).
+extern WebPUnfilterFunc WebPUnfilters[WEBP_FILTER_LAST];
+
+// To be called first before using the above.
+void VP8FiltersInit(void);
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/dsp/enc.c b/drivers/webp/dsp/enc.c
index 02234564be..95e63f89ab 100644
--- a/drivers/webp/dsp/enc.c
+++ b/drivers/webp/dsp/enc.c
@@ -1,47 +1,34 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Speed-critical encoding functions.
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
+#include <assert.h>
 #include <stdlib.h>  // for abs()
+
 #include "./dsp.h"
 #include "../enc/vp8enci.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+static WEBP_INLINE uint8_t clip_8b(int v) {
+  return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
+}
+
+static WEBP_INLINE int clip_max(int v, int max) {
+  return (v > max) ? max : v;
+}
 
 //------------------------------------------------------------------------------
 // Compute susceptibility based on DCT-coeff histograms:
 // the higher, the "easier" the macroblock is to compress.
 
-static int ClipAlpha(int alpha) {
-  return alpha < 0 ? 0 : alpha > 255 ? 255 : alpha;
-}
-
-int VP8GetAlpha(const int histo[MAX_COEFF_THRESH + 1]) {
-  int num = 0, den = 0, val = 0;
-  int k;
-  int alpha;
-  // note: changing this loop to avoid the numerous "k + 1" slows things down.
-  for (k = 0; k < MAX_COEFF_THRESH; ++k) {
-    if (histo[k + 1]) {
-      val += histo[k + 1];
-      num += val * (k + 1);
-      den += (k + 1) * (k + 1);
-    }
-  }
-  // we scale the value to a usable [0..255] range
-  alpha = den ? 10 * num / den - 5 : 0;
-  return ClipAlpha(alpha);
-}
-
 const int VP8DspScan[16 + 4 + 4] = {
   // Luma
   0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
@@ -53,27 +40,41 @@ const int VP8DspScan[16 + 4 + 4] = {
   8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
 };
 
-static int CollectHistogram(const uint8_t* ref, const uint8_t* pred,
-                            int start_block, int end_block) {
-  int histo[MAX_COEFF_THRESH + 1] = { 0 };
-  int16_t out[16];
-  int j, k;
+// general-purpose util function
+void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1],
+                         VP8Histogram* const histo) {
+  int max_value = 0, last_non_zero = 1;
+  int k;
+  for (k = 0; k <= MAX_COEFF_THRESH; ++k) {
+    const int value = distribution[k];
+    if (value > 0) {
+      if (value > max_value) max_value = value;
+      last_non_zero = k;
+    }
+  }
+  histo->max_value = max_value;
+  histo->last_non_zero = last_non_zero;
+}
+
+static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
+                             int start_block, int end_block,
+                             VP8Histogram* const histo) {
+  int j;
+  int distribution[MAX_COEFF_THRESH + 1] = { 0 };
   for (j = start_block; j < end_block; ++j) {
-    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
+    int k;
+    int16_t out[16];
 
-    // Convert coefficients to bin (within out[]).
-    for (k = 0; k < 16; ++k) {
-      const int v = abs(out[k]) >> 2;
-      out[k] = (v > MAX_COEFF_THRESH) ? MAX_COEFF_THRESH : v;
-    }
+    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
 
-    // Use bin to update histogram.
+    // Convert coefficients to bin.
     for (k = 0; k < 16; ++k) {
-      histo[out[k]]++;
+      const int v = abs(out[k]) >> 3;  // TODO(skal): add rounding?
+      const int clipped_value = clip_max(v, MAX_COEFF_THRESH);
+      ++distribution[clipped_value];
     }
   }
-
-  return VP8GetAlpha(histo);
+  VP8SetHistogramData(distribution, histo);
 }
 
 //------------------------------------------------------------------------------
@@ -85,19 +86,16 @@ static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
 // and make sure it's set to true _last_ (so as to be thread-safe)
 static volatile int tables_ok = 0;
 
-static void InitTables(void) {
+static WEBP_TSAN_IGNORE_FUNCTION void InitTables(void) {
   if (!tables_ok) {
     int i;
     for (i = -255; i <= 255 + 255; ++i) {
-      clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i;
+      clip1[255 + i] = clip_8b(i);
     }
     tables_ok = 1;
   }
 }
 
-static WEBP_INLINE uint8_t clip_8b(int v) {
-  return (!(v & ~0xff)) ? v : v < 0 ? 0 : 255;
-}
 
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)
@@ -154,84 +152,63 @@ static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
   int i;
   int tmp[16];
   for (i = 0; i < 4; ++i, src += BPS, ref += BPS) {
-    const int d0 = src[0] - ref[0];
+    const int d0 = src[0] - ref[0];   // 9bit dynamic range ([-255,255])
     const int d1 = src[1] - ref[1];
     const int d2 = src[2] - ref[2];
     const int d3 = src[3] - ref[3];
-    const int a0 = (d0 + d3) << 3;
-    const int a1 = (d1 + d2) << 3;
-    const int a2 = (d1 - d2) << 3;
-    const int a3 = (d0 - d3) << 3;
-    tmp[0 + i * 4] = (a0 + a1);
-    tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 14500) >> 12;
-    tmp[2 + i * 4] = (a0 - a1);
-    tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 +  7500) >> 12;
+    const int a0 = (d0 + d3);         // 10b                      [-510,510]
+    const int a1 = (d1 + d2);
+    const int a2 = (d1 - d2);
+    const int a3 = (d0 - d3);
+    tmp[0 + i * 4] = (a0 + a1) * 8;   // 14b                      [-8160,8160]
+    tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9;      // [-7536,7542]
+    tmp[2 + i * 4] = (a0 - a1) * 8;
+    tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 +  937) >> 9;
   }
   for (i = 0; i < 4; ++i) {
-    const int a0 = (tmp[0 + i] + tmp[12 + i]);
+    const int a0 = (tmp[0 + i] + tmp[12 + i]);  // 15b
     const int a1 = (tmp[4 + i] + tmp[ 8 + i]);
     const int a2 = (tmp[4 + i] - tmp[ 8 + i]);
     const int a3 = (tmp[0 + i] - tmp[12 + i]);
-    out[0 + i] = (a0 + a1 + 7) >> 4;
+    out[0 + i] = (a0 + a1 + 7) >> 4;            // 12b
     out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0);
     out[8 + i] = (a0 - a1 + 7) >> 4;
     out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16);
   }
 }
 
-static void ITransformWHT(const int16_t* in, int16_t* out) {
-  int tmp[16];
-  int i;
-  for (i = 0; i < 4; ++i) {
-    const int a0 = in[0 + i] + in[12 + i];
-    const int a1 = in[4 + i] + in[ 8 + i];
-    const int a2 = in[4 + i] - in[ 8 + i];
-    const int a3 = in[0 + i] - in[12 + i];
-    tmp[0  + i] = a0 + a1;
-    tmp[8  + i] = a0 - a1;
-    tmp[4  + i] = a3 + a2;
-    tmp[12 + i] = a3 - a2;
-  }
-  for (i = 0; i < 4; ++i) {
-    const int dc = tmp[0 + i * 4] + 3;    // w/ rounder
-    const int a0 = dc             + tmp[3 + i * 4];
-    const int a1 = tmp[1 + i * 4] + tmp[2 + i * 4];
-    const int a2 = tmp[1 + i * 4] - tmp[2 + i * 4];
-    const int a3 = dc             - tmp[3 + i * 4];
-    out[ 0] = (a0 + a1) >> 3;
-    out[16] = (a3 + a2) >> 3;
-    out[32] = (a0 - a1) >> 3;
-    out[48] = (a3 - a2) >> 3;
-    out += 64;
-  }
+static void FTransform2(const uint8_t* src, const uint8_t* ref, int16_t* out) {
+  VP8FTransform(src, ref, out);
+  VP8FTransform(src + 4, ref + 4, out + 16);
 }
 
 static void FTransformWHT(const int16_t* in, int16_t* out) {
-  int tmp[16];
+  // input is 12b signed
+  int32_t tmp[16];
   int i;
   for (i = 0; i < 4; ++i, in += 64) {
-    const int a0 = (in[0 * 16] + in[2 * 16]) << 2;
-    const int a1 = (in[1 * 16] + in[3 * 16]) << 2;
-    const int a2 = (in[1 * 16] - in[3 * 16]) << 2;
-    const int a3 = (in[0 * 16] - in[2 * 16]) << 2;
-    tmp[0 + i * 4] = (a0 + a1) + (a0 != 0);
+    const int a0 = (in[0 * 16] + in[2 * 16]);  // 13b
+    const int a1 = (in[1 * 16] + in[3 * 16]);
+    const int a2 = (in[1 * 16] - in[3 * 16]);
+    const int a3 = (in[0 * 16] - in[2 * 16]);
+    tmp[0 + i * 4] = a0 + a1;   // 14b
     tmp[1 + i * 4] = a3 + a2;
     tmp[2 + i * 4] = a3 - a2;
     tmp[3 + i * 4] = a0 - a1;
   }
   for (i = 0; i < 4; ++i) {
-    const int a0 = (tmp[0 + i] + tmp[8 + i]);
+    const int a0 = (tmp[0 + i] + tmp[8 + i]);  // 15b
     const int a1 = (tmp[4 + i] + tmp[12+ i]);
     const int a2 = (tmp[4 + i] - tmp[12+ i]);
     const int a3 = (tmp[0 + i] - tmp[8 + i]);
-    const int b0 = a0 + a1;
+    const int b0 = a0 + a1;    // 16b
     const int b1 = a3 + a2;
     const int b2 = a3 - a2;
     const int b3 = a0 - a1;
-    out[ 0 + i] = (b0 + (b0 > 0) + 3) >> 3;
-    out[ 4 + i] = (b1 + (b1 > 0) + 3) >> 3;
-    out[ 8 + i] = (b2 + (b2 > 0) + 3) >> 3;
-    out[12 + i] = (b3 + (b3 > 0) + 3) >> 3;
+    out[ 0 + i] = b0 >> 1;     // 15b
+    out[ 4 + i] = b1 >> 1;
+    out[ 8 + i] = b2 >> 1;
+    out[12 + i] = b3 >> 1;
   }
 }
 
@@ -241,8 +218,6 @@ static void FTransformWHT(const int16_t* in, int16_t* out) {
 //------------------------------------------------------------------------------
 // Intra predictions
 
-#define DST(x, y) dst[(x) + (y) * BPS]
-
 static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
   int j;
   for (j = 0; j < size; ++j) {
@@ -253,7 +228,7 @@ static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
 static WEBP_INLINE void VerticalPred(uint8_t* dst,
                                      const uint8_t* top, int size) {
   int j;
-  if (top) {
+  if (top != NULL) {
     for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
   } else {
     Fill(dst, 127, size);
@@ -262,7 +237,7 @@ static WEBP_INLINE void VerticalPred(uint8_t* dst,
 
 static WEBP_INLINE void HorizontalPred(uint8_t* dst,
                                        const uint8_t* left, int size) {
-  if (left) {
+  if (left != NULL) {
     int j;
     for (j = 0; j < size; ++j) {
       memset(dst + j * BPS, left[j], size);
@@ -275,8 +250,8 @@ static WEBP_INLINE void HorizontalPred(uint8_t* dst,
 static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
                                    const uint8_t* top, int size) {
   int y;
-  if (left) {
-    if (top) {
+  if (left != NULL) {
+    if (top != NULL) {
       const uint8_t* const clip = clip1 + 255 - left[-1];
       for (y = 0; y < size; ++y) {
         const uint8_t* const clip_table = clip + left[y];
@@ -294,7 +269,7 @@ static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
     // is equivalent to VE prediction where you just copy the top samples.
     // Note that if top samples are not available, the default value is
     // then 129, and not 127 as in the VerticalPred case.
-    if (top) {
+    if (top != NULL) {
       VerticalPred(dst, top, size);
     } else {
       Fill(dst, 129, size);
@@ -307,15 +282,15 @@ static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
                                int size, int round, int shift) {
   int DC = 0;
   int j;
-  if (top) {
+  if (top != NULL) {
     for (j = 0; j < size; ++j) DC += top[j];
-    if (left) {   // top and left present
+    if (left != NULL) {   // top and left present
       for (j = 0; j < size; ++j) DC += left[j];
     } else {      // top, but no left
       DC += DC;
     }
     DC = (DC + round) >> shift;
-  } else if (left) {   // left but no top
+  } else if (left != NULL) {   // left but no top
     for (j = 0; j < size; ++j) DC += left[j];
     DC += DC;
     DC = (DC + round) >> shift;
@@ -337,8 +312,8 @@ static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
   TrueMotion(C8TM8 + dst, left, top, 8);
   // V block
   dst += 8;
-  if (top) top += 8;
-  if (left) left += 16;
+  if (top != NULL) top += 8;
+  if (left != NULL) left += 16;
   DCMode(C8DC8 + dst, left, top, 8, 8, 4);
   VerticalPred(C8VE8 + dst, top, 8);
   HorizontalPred(C8HE8 + dst, left, 8);
@@ -359,6 +334,7 @@ static void Intra16Preds(uint8_t* dst,
 //------------------------------------------------------------------------------
 // luma 4x4 prediction
 
+#define DST(x, y) dst[(x) + (y) * BPS]
 #define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
 
@@ -589,30 +565,30 @@ static int TTransform(const uint8_t* in, const uint16_t* w) {
   int i;
   // horizontal pass
   for (i = 0; i < 4; ++i, in += BPS) {
-    const int a0 = (in[0] + in[2]) << 2;
-    const int a1 = (in[1] + in[3]) << 2;
-    const int a2 = (in[1] - in[3]) << 2;
-    const int a3 = (in[0] - in[2]) << 2;
-    tmp[0 + i * 4] = a0 + a1 + (a0 != 0);
+    const int a0 = in[0] + in[2];
+    const int a1 = in[1] + in[3];
+    const int a2 = in[1] - in[3];
+    const int a3 = in[0] - in[2];
+    tmp[0 + i * 4] = a0 + a1;
     tmp[1 + i * 4] = a3 + a2;
     tmp[2 + i * 4] = a3 - a2;
     tmp[3 + i * 4] = a0 - a1;
   }
   // vertical pass
   for (i = 0; i < 4; ++i, ++w) {
-    const int a0 = (tmp[0 + i] + tmp[8 + i]);
-    const int a1 = (tmp[4 + i] + tmp[12+ i]);
-    const int a2 = (tmp[4 + i] - tmp[12+ i]);
-    const int a3 = (tmp[0 + i] - tmp[8 + i]);
+    const int a0 = tmp[0 + i] + tmp[8 + i];
+    const int a1 = tmp[4 + i] + tmp[12+ i];
+    const int a2 = tmp[4 + i] - tmp[12+ i];
+    const int a3 = tmp[0 + i] - tmp[8 + i];
     const int b0 = a0 + a1;
     const int b1 = a3 + a2;
     const int b2 = a3 - a2;
     const int b3 = a0 - a1;
-    // abs((b + (b<0) + 3) >> 3) = (abs(b) + 3) >> 3
-    sum += w[ 0] * ((abs(b0) + 3) >> 3);
-    sum += w[ 4] * ((abs(b1) + 3) >> 3);
-    sum += w[ 8] * ((abs(b2) + 3) >> 3);
-    sum += w[12] * ((abs(b3) + 3) >> 3);
+
+    sum += w[ 0] * abs(b0);
+    sum += w[ 4] * abs(b1);
+    sum += w[ 8] * abs(b2);
+    sum += w[12] * abs(b3);
   }
   return sum;
 }
@@ -621,7 +597,7 @@ static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
                     const uint16_t* const w) {
   const int sum1 = TTransform(a, w);
   const int sum2 = TTransform(b, w);
-  return (abs(sum2 - sum1) + 8) >> 4;
+  return abs(sum2 - sum1) >> 5;
 }
 
 static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
@@ -646,21 +622,57 @@ static const uint8_t kZigzag[16] = {
 
 // Simple quantization
 static int QuantizeBlock(int16_t in[16], int16_t out[16],
-                         int n, const VP8Matrix* const mtx) {
+                         const VP8Matrix* const mtx) {
   int last = -1;
-  for (; n < 16; ++n) {
+  int n;
+  for (n = 0; n < 16; ++n) {
+    const int j = kZigzag[n];
+    const int sign = (in[j] < 0);
+    const uint32_t coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
+    if (coeff > mtx->zthresh_[j]) {
+      const uint32_t Q = mtx->q_[j];
+      const uint32_t iQ = mtx->iq_[j];
+      const uint32_t B = mtx->bias_[j];
+      int level = QUANTDIV(coeff, iQ, B);
+      if (level > MAX_LEVEL) level = MAX_LEVEL;
+      if (sign) level = -level;
+      in[j] = level * Q;
+      out[n] = level;
+      if (level) last = n;
+    } else {
+      out[n] = 0;
+      in[j] = 0;
+    }
+  }
+  return (last >= 0);
+}
+
+static int Quantize2Blocks(int16_t in[32], int16_t out[32],
+                           const VP8Matrix* const mtx) {
+  int nz;
+  nz  = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0;
+  nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1;
+  return nz;
+}
+
+static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
+                            const VP8Matrix* const mtx) {
+  int n, last = -1;
+  for (n = 0; n < 16; ++n) {
     const int j = kZigzag[n];
     const int sign = (in[j] < 0);
-    int coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
-    if (coeff > 2047) coeff = 2047;
+    const uint32_t coeff = sign ? -in[j] : in[j];
+    assert(mtx->sharpen_[j] == 0);
     if (coeff > mtx->zthresh_[j]) {
-      const int Q = mtx->q_[j];
-      const int iQ = mtx->iq_[j];
-      const int B = mtx->bias_[j];
-      out[n] = QUANTDIV(coeff, iQ, B);
-      if (sign) out[n] = -out[n];
-      in[j] = out[n] * Q;
-      if (out[n]) last = n;
+      const uint32_t Q = mtx->q_[j];
+      const uint32_t iQ = mtx->iq_[j];
+      const uint32_t B = mtx->bias_[j];
+      int level = QUANTDIV(coeff, iQ, B);
+      if (level > MAX_LEVEL) level = MAX_LEVEL;
+      if (sign) level = -level;
+      in[j] = level * Q;
+      out[n] = level;
+      if (level) last = n;
     } else {
       out[n] = 0;
       in[j] = 0;
@@ -672,16 +684,22 @@ static int QuantizeBlock(int16_t in[16], int16_t out[16],
 //------------------------------------------------------------------------------
 // Block copy
 
-static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int size) {
+static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int w, int h) {
   int y;
-  for (y = 0; y < size; ++y) {
-    memcpy(dst, src, size);
+  for (y = 0; y < h; ++y) {
+    memcpy(dst, src, w);
     src += BPS;
     dst += BPS;
   }
 }
 
-static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); }
+static void Copy4x4(const uint8_t* src, uint8_t* dst) {
+  Copy(src, dst, 4, 4);
+}
+
+static void Copy16x8(const uint8_t* src, uint8_t* dst) {
+  Copy(src, dst, 16, 8);
+}
 
 //------------------------------------------------------------------------------
 // Initialization
@@ -691,7 +709,7 @@ static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); }
 VP8CHisto VP8CollectHistogram;
 VP8Idct VP8ITransform;
 VP8Fdct VP8FTransform;
-VP8WHT VP8ITransformWHT;
+VP8Fdct VP8FTransform2;
 VP8WHT VP8FTransformWHT;
 VP8Intra4Preds VP8EncPredLuma4;
 VP8IntraPreds VP8EncPredLuma16;
@@ -703,18 +721,32 @@ VP8Metric VP8SSE4x4;
 VP8WMetric VP8TDisto4x4;
 VP8WMetric VP8TDisto16x16;
 VP8QuantizeBlock VP8EncQuantizeBlock;
+VP8Quantize2Blocks VP8EncQuantize2Blocks;
+VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
 VP8BlockCopy VP8Copy4x4;
+VP8BlockCopy VP8Copy16x8;
 
 extern void VP8EncDspInitSSE2(void);
+extern void VP8EncDspInitSSE41(void);
+extern void VP8EncDspInitAVX2(void);
+extern void VP8EncDspInitNEON(void);
+extern void VP8EncDspInitMIPS32(void);
+extern void VP8EncDspInitMIPSdspR2(void);
+
+static volatile VP8CPUInfo enc_last_cpuinfo_used =
+    (VP8CPUInfo)&enc_last_cpuinfo_used;
 
-void VP8EncDspInit(void) {
+WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInit(void) {
+  if (enc_last_cpuinfo_used == VP8GetCPUInfo) return;
+
+  VP8DspInit();  // common inverse transforms
   InitTables();
 
   // default C implementations
   VP8CollectHistogram = CollectHistogram;
   VP8ITransform = ITransform;
   VP8FTransform = FTransform;
-  VP8ITransformWHT = ITransformWHT;
+  VP8FTransform2 = FTransform2;
   VP8FTransformWHT = FTransformWHT;
   VP8EncPredLuma4 = Intra4Preds;
   VP8EncPredLuma16 = Intra16Preds;
@@ -726,18 +758,43 @@ void VP8EncDspInit(void) {
   VP8TDisto4x4 = Disto4x4;
   VP8TDisto16x16 = Disto16x16;
   VP8EncQuantizeBlock = QuantizeBlock;
+  VP8EncQuantize2Blocks = Quantize2Blocks;
+  VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
   VP8Copy4x4 = Copy4x4;
+  VP8Copy16x8 = Copy16x8;
 
   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
-  if (VP8GetCPUInfo) {
+  if (VP8GetCPUInfo != NULL) {
 #if defined(WEBP_USE_SSE2)
     if (VP8GetCPUInfo(kSSE2)) {
       VP8EncDspInitSSE2();
+#if defined(WEBP_USE_SSE41)
+      if (VP8GetCPUInfo(kSSE4_1)) {
+        VP8EncDspInitSSE41();
+      }
+#endif
+    }
+#endif
+#if defined(WEBP_USE_AVX2)
+    if (VP8GetCPUInfo(kAVX2)) {
+      VP8EncDspInitAVX2();
+    }
+#endif
+#if defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      VP8EncDspInitNEON();
+    }
+#endif
+#if defined(WEBP_USE_MIPS32)
+    if (VP8GetCPUInfo(kMIPS32)) {
+      VP8EncDspInitMIPS32();
+    }
+#endif
+#if defined(WEBP_USE_MIPS_DSP_R2)
+    if (VP8GetCPUInfo(kMIPSdspR2)) {
+      VP8EncDspInitMIPSdspR2();
     }
 #endif
   }
+  enc_last_cpuinfo_used = VP8GetCPUInfo;
 }
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/dsp/enc_sse2.c b/drivers/webp/dsp/enc_sse2.c
index b046761dc1..63d9cecd85 100644
--- a/drivers/webp/dsp/enc_sse2.c
+++ b/drivers/webp/dsp/enc_sse2.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // SSE2 version of speed-critical encoding functions.
@@ -15,64 +17,55 @@
 #include <stdlib.h>  // for abs()
 #include <emmintrin.h>
 
+#include "../enc/cost.h"
 #include "../enc/vp8enci.h"
+#include "../utils/utils.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
+//------------------------------------------------------------------------------
+// Quite useful macro for debugging. Left here for convenience.
+
+#if 0
+#include <stdio.h>
+static void PrintReg(const __m128i r, const char* const name, int size) {
+  int n;
+  union {
+    __m128i r;
+    uint8_t i8[16];
+    uint16_t i16[8];
+    uint32_t i32[4];
+    uint64_t i64[2];
+  } tmp;
+  tmp.r = r;
+  fprintf(stderr, "%s\t: ", name);
+  if (size == 8) {
+    for (n = 0; n < 16; ++n) fprintf(stderr, "%.2x ", tmp.i8[n]);
+  } else if (size == 16) {
+    for (n = 0; n < 8; ++n) fprintf(stderr, "%.4x ", tmp.i16[n]);
+  } else if (size == 32) {
+    for (n = 0; n < 4; ++n) fprintf(stderr, "%.8x ", tmp.i32[n]);
+  } else {
+    for (n = 0; n < 2; ++n) fprintf(stderr, "%.16lx ", tmp.i64[n]);
+  }
+  fprintf(stderr, "\n");
+}
 #endif
 
 //------------------------------------------------------------------------------
-// Compute susceptibility based on DCT-coeff histograms:
-// the higher, the "easier" the macroblock is to compress.
-
-static int CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
-                                int start_block, int end_block) {
-  int histo[MAX_COEFF_THRESH + 1] = { 0 };
-  int16_t out[16];
-  int j, k;
-  const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
-  for (j = start_block; j < end_block; ++j) {
-    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
-
-    // Convert coefficients to bin (within out[]).
-    {
-      // Load.
-      const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
-      const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
-      // sign(out) = out >> 15  (0x0000 if positive, 0xffff if negative)
-      const __m128i sign0 = _mm_srai_epi16(out0, 15);
-      const __m128i sign1 = _mm_srai_epi16(out1, 15);
-      // abs(out) = (out ^ sign) - sign
-      const __m128i xor0 = _mm_xor_si128(out0, sign0);
-      const __m128i xor1 = _mm_xor_si128(out1, sign1);
-      const __m128i abs0 = _mm_sub_epi16(xor0, sign0);
-      const __m128i abs1 = _mm_sub_epi16(xor1, sign1);
-      // v = abs(out) >> 2
-      const __m128i v0 = _mm_srai_epi16(abs0, 2);
-      const __m128i v1 = _mm_srai_epi16(abs1, 2);
-      // bin = min(v, MAX_COEFF_THRESH)
-      const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
-      const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
-      // Store.
-      _mm_storeu_si128((__m128i*)&out[0], bin0);
-      _mm_storeu_si128((__m128i*)&out[8], bin1);
-    }
+// util for unaligned loads.
 
-    // Use bin to update histogram.
-    for (k = 0; k < 16; ++k) {
-      histo[out[k]]++;
-    }
-  }
-
-  return VP8GetAlpha(histo);
+// memcpy() is the safe way of moving potentially unaligned 32b memory.
+static WEBP_INLINE uint32_t MemToUint32(const uint8_t* const ptr) {
+  uint32_t A;
+  memcpy(&A, (const int*)ptr, sizeof(A));
+  return A;
 }
 
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)
 
 // Does one or two inverse transforms.
-static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
-                           int do_two) {
+static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
+                       int do_two) {
   // This implementation makes use of 16-bit fixed point versions of two
   // multiply constants:
   //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
@@ -99,19 +92,19 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
   // use nor store.
   __m128i in0, in1, in2, in3;
   {
-    in0 = _mm_loadl_epi64((__m128i*)&in[0]);
-    in1 = _mm_loadl_epi64((__m128i*)&in[4]);
-    in2 = _mm_loadl_epi64((__m128i*)&in[8]);
-    in3 = _mm_loadl_epi64((__m128i*)&in[12]);
+    in0 = _mm_loadl_epi64((const __m128i*)&in[0]);
+    in1 = _mm_loadl_epi64((const __m128i*)&in[4]);
+    in2 = _mm_loadl_epi64((const __m128i*)&in[8]);
+    in3 = _mm_loadl_epi64((const __m128i*)&in[12]);
     // a00 a10 a20 a30   x x x x
     // a01 a11 a21 a31   x x x x
     // a02 a12 a22 a32   x x x x
     // a03 a13 a23 a33   x x x x
     if (do_two) {
-      const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]);
-      const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]);
-      const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]);
-      const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]);
+      const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]);
+      const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]);
+      const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]);
+      const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]);
       in0 = _mm_unpacklo_epi64(in0, inB0);
       in1 = _mm_unpacklo_epi64(in1, inB1);
       in2 = _mm_unpacklo_epi64(in2, inB2);
@@ -243,21 +236,21 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
 
   // Add inverse transform to 'ref' and store.
   {
-    const __m128i zero = _mm_set1_epi16(0);
+    const __m128i zero = _mm_setzero_si128();
     // Load the reference(s).
     __m128i ref0, ref1, ref2, ref3;
     if (do_two) {
       // Load eight bytes/pixels per line.
-      ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
-      ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
-      ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
-      ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
+      ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
+      ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
+      ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
+      ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
     } else {
       // Load four bytes/pixels per line.
-      ref0 = _mm_cvtsi32_si128(*(int*)&ref[0 * BPS]);
-      ref1 = _mm_cvtsi32_si128(*(int*)&ref[1 * BPS]);
-      ref2 = _mm_cvtsi32_si128(*(int*)&ref[2 * BPS]);
-      ref3 = _mm_cvtsi32_si128(*(int*)&ref[3 * BPS]);
+      ref0 = _mm_cvtsi32_si128(MemToUint32(&ref[0 * BPS]));
+      ref1 = _mm_cvtsi32_si128(MemToUint32(&ref[1 * BPS]));
+      ref2 = _mm_cvtsi32_si128(MemToUint32(&ref[2 * BPS]));
+      ref3 = _mm_cvtsi32_si128(MemToUint32(&ref[3 * BPS]));
     }
     // Convert to 16b.
     ref0 = _mm_unpacklo_epi8(ref0, zero);
@@ -291,200 +284,865 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
   }
 }
 
-static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
-                           int16_t* out) {
+static void FTransformPass1(const __m128i* const in01,
+                            const __m128i* const in23,
+                            __m128i* const out01,
+                            __m128i* const out32) {
+  const __m128i k937 = _mm_set1_epi32(937);
+  const __m128i k1812 = _mm_set1_epi32(1812);
+
+  const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8);
+  const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8);
+  const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352,
+                                            2217, 5352, 2217, 5352);
+  const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217,
+                                            -5352, 2217, -5352, 2217);
+
+  // *in01 = 00 01 10 11 02 03 12 13
+  // *in23 = 20 21 30 31 22 23 32 33
+  const __m128i shuf01_p = _mm_shufflehi_epi16(*in01, _MM_SHUFFLE(2, 3, 0, 1));
+  const __m128i shuf23_p = _mm_shufflehi_epi16(*in23, _MM_SHUFFLE(2, 3, 0, 1));
+  // 00 01 10 11 03 02 13 12
+  // 20 21 30 31 23 22 33 32
+  const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p);
+  const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p);
+  // 00 01 10 11 20 21 30 31
+  // 03 02 13 12 23 22 33 32
+  const __m128i a01 = _mm_add_epi16(s01, s32);
+  const __m128i a32 = _mm_sub_epi16(s01, s32);
+  // [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ]
+  // [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ]
+
+  const __m128i tmp0   = _mm_madd_epi16(a01, k88p);  // [ (a0 + a1) << 3, ... ]
+  const __m128i tmp2   = _mm_madd_epi16(a01, k88m);  // [ (a0 - a1) << 3, ... ]
+  const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p);
+  const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m);
+  const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812);
+  const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937);
+  const __m128i tmp1   = _mm_srai_epi32(tmp1_2, 9);
+  const __m128i tmp3   = _mm_srai_epi32(tmp3_2, 9);
+  const __m128i s03    = _mm_packs_epi32(tmp0, tmp2);
+  const __m128i s12    = _mm_packs_epi32(tmp1, tmp3);
+  const __m128i s_lo   = _mm_unpacklo_epi16(s03, s12);   // 0 1 0 1 0 1...
+  const __m128i s_hi   = _mm_unpackhi_epi16(s03, s12);   // 2 3 2 3 2 3
+  const __m128i v23    = _mm_unpackhi_epi32(s_lo, s_hi);
+  *out01 = _mm_unpacklo_epi32(s_lo, s_hi);
+  *out32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));  // 3 2 3 2 3 2..
+}
+
+static void FTransformPass2(const __m128i* const v01, const __m128i* const v32,
+                            int16_t* out) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i seven = _mm_set1_epi16(7);
-  const __m128i k7500 = _mm_set1_epi32(7500);
-  const __m128i k14500 = _mm_set1_epi32(14500);
-  const __m128i k51000 = _mm_set1_epi32(51000);
-  const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16));
   const __m128i k5352_2217 = _mm_set_epi16(5352,  2217, 5352,  2217,
                                            5352,  2217, 5352,  2217);
   const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352,
                                            2217, -5352, 2217, -5352);
+  const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16));
+  const __m128i k51000 = _mm_set1_epi32(51000);
+
+  // Same operations are done on the (0,3) and (1,2) pairs.
+  // a0 = v0 + v3
+  // a1 = v1 + v2
+  // a3 = v0 - v3
+  // a2 = v1 - v2
+  const __m128i a01 = _mm_add_epi16(*v01, *v32);
+  const __m128i a32 = _mm_sub_epi16(*v01, *v32);
+  const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
+  const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
+  const __m128i a01_plus_7 = _mm_add_epi16(a01, seven);
+
+  // d0 = (a0 + a1 + 7) >> 4;
+  // d2 = (a0 - a1 + 7) >> 4;
+  const __m128i c0 = _mm_add_epi16(a01_plus_7, a11);
+  const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11);
+  const __m128i d0 = _mm_srai_epi16(c0, 4);
+  const __m128i d2 = _mm_srai_epi16(c2, 4);
+
+  // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
+  // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
+  const __m128i b23 = _mm_unpacklo_epi16(a22, a32);
+  const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
+  const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
+  const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one);
+  const __m128i d3 = _mm_add_epi32(c3, k51000);
+  const __m128i e1 = _mm_srai_epi32(d1, 16);
+  const __m128i e3 = _mm_srai_epi32(d3, 16);
+  const __m128i f1 = _mm_packs_epi32(e1, e1);
+  const __m128i f3 = _mm_packs_epi32(e3, e3);
+  // f1 = f1 + (a3 != 0);
+  // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
+  // desired (0, 1), we add one earlier through k12000_plus_one.
+  // -> f1 = f1 + 1 - (a3 == 0)
+  const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
+
+  const __m128i d0_g1 = _mm_unpacklo_epi64(d0, g1);
+  const __m128i d2_f3 = _mm_unpacklo_epi64(d2, f3);
+  _mm_storeu_si128((__m128i*)&out[0], d0_g1);
+  _mm_storeu_si128((__m128i*)&out[8], d2_f3);
+}
+
+static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
+  const __m128i zero = _mm_setzero_si128();
 
+  // Load src and convert to 16b.
+  const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]);
+  const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]);
+  const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]);
+  const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]);
+  const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
+  const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
+  const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
+  const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
+  // Load ref and convert to 16b.
+  const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
+  const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
+  const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
+  const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
+  const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
+  const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
+  const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
+  const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
+  // Compute difference. -> 00 01 02 03 00 00 00 00
+  const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
+  const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
+  const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
+  const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
+
+  // Unpack and shuffle
+  // 00 01 02 03   0 0 0 0
+  // 10 11 12 13   0 0 0 0
+  // 20 21 22 23   0 0 0 0
+  // 30 31 32 33   0 0 0 0
+  const __m128i shuf01 = _mm_unpacklo_epi32(diff0, diff1);
+  const __m128i shuf23 = _mm_unpacklo_epi32(diff2, diff3);
   __m128i v01, v32;
 
-  // Difference between src and ref and initial transpose.
-  {
-    // Load src and convert to 16b.
-    const __m128i src0 = _mm_loadl_epi64((__m128i*)&src[0 * BPS]);
-    const __m128i src1 = _mm_loadl_epi64((__m128i*)&src[1 * BPS]);
-    const __m128i src2 = _mm_loadl_epi64((__m128i*)&src[2 * BPS]);
-    const __m128i src3 = _mm_loadl_epi64((__m128i*)&src[3 * BPS]);
-    const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
-    const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
-    const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
-    const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
-    // Load ref and convert to 16b.
-    const __m128i ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
-    const __m128i ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
-    const __m128i ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
-    const __m128i ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
-    const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
-    const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
-    const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
-    const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
-    // Compute difference.
-    const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
-    const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
-    const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
-    const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
-
-    // Transpose.
-    // 00 01 02 03   0 0 0 0
-    // 10 11 12 13   0 0 0 0
-    // 20 21 22 23   0 0 0 0
-    // 30 31 32 33   0 0 0 0
-    const __m128i transpose0_0 = _mm_unpacklo_epi16(diff0, diff1);
-    const __m128i transpose0_1 = _mm_unpacklo_epi16(diff2, diff3);
-    // 00 10 01 11   02 12 03 13
-    // 20 30 21 31   22 32 23 33
-    const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
-    v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
-    v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));
-    // a02 a12 a22 a32   a03 a13 a23 a33
-    // a00 a10 a20 a30   a01 a11 a21 a31
-    // a03 a13 a23 a33   a02 a12 a22 a32
-  }
-
-  // First pass and subsequent transpose.
-  {
-    // Same operations are done on the (0,3) and (1,2) pairs.
-    // b0 = (a0 + a3) << 3
-    // b1 = (a1 + a2) << 3
-    // b3 = (a0 - a3) << 3
-    // b2 = (a1 - a2) << 3
-    const __m128i a01 = _mm_add_epi16(v01, v32);
-    const __m128i a32 = _mm_sub_epi16(v01, v32);
-    const __m128i b01 = _mm_slli_epi16(a01, 3);
-    const __m128i b32 = _mm_slli_epi16(a32, 3);
-    const __m128i b11 = _mm_unpackhi_epi64(b01, b01);
-    const __m128i b22 = _mm_unpackhi_epi64(b32, b32);
-
-    // e0 = b0 + b1
-    // e2 = b0 - b1
-    const __m128i e0 = _mm_add_epi16(b01, b11);
-    const __m128i e2 = _mm_sub_epi16(b01, b11);
-    const __m128i e02 = _mm_unpacklo_epi64(e0, e2);
-
-    // e1 = (b3 * 5352 + b2 * 2217 + 14500) >> 12
-    // e3 = (b3 * 2217 - b2 * 5352 +  7500) >> 12
-    const __m128i b23 = _mm_unpacklo_epi16(b22, b32);
-    const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
-    const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
-    const __m128i d1 = _mm_add_epi32(c1, k14500);
-    const __m128i d3 = _mm_add_epi32(c3, k7500);
-    const __m128i e1 = _mm_srai_epi32(d1, 12);
-    const __m128i e3 = _mm_srai_epi32(d3, 12);
-    const __m128i e13 = _mm_packs_epi32(e1, e3);
-
-    // Transpose.
-    // 00 01 02 03  20 21 22 23
-    // 10 11 12 13  30 31 32 33
-    const __m128i transpose0_0 = _mm_unpacklo_epi16(e02, e13);
-    const __m128i transpose0_1 = _mm_unpackhi_epi16(e02, e13);
-    // 00 10 01 11   02 12 03 13
-    // 20 30 21 31   22 32 23 33
-    const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
-    v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
-    v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));
-    // 02 12 22 32   03 13 23 33
-    // 00 10 20 30   01 11 21 31
-    // 03 13 23 33   02 12 22 32
-  }
+  // First pass
+  FTransformPass1(&shuf01, &shuf23, &v01, &v32);
 
   // Second pass
+  FTransformPass2(&v01, &v32, out);
+}
+
+static void FTransform2(const uint8_t* src, const uint8_t* ref, int16_t* out) {
+  const __m128i zero = _mm_setzero_si128();
+
+  // Load src and convert to 16b.
+  const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]);
+  const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]);
+  const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]);
+  const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]);
+  const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
+  const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
+  const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
+  const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
+  // Load ref and convert to 16b.
+  const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
+  const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
+  const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
+  const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
+  const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
+  const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
+  const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
+  const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
+  // Compute difference. -> 00 01 02 03  00' 01' 02' 03'
+  const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
+  const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
+  const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
+  const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
+
+  // Unpack and shuffle
+  // 00 01 02 03   0 0 0 0
+  // 10 11 12 13   0 0 0 0
+  // 20 21 22 23   0 0 0 0
+  // 30 31 32 33   0 0 0 0
+  const __m128i shuf01l = _mm_unpacklo_epi32(diff0, diff1);
+  const __m128i shuf23l = _mm_unpacklo_epi32(diff2, diff3);
+  const __m128i shuf01h = _mm_unpackhi_epi32(diff0, diff1);
+  const __m128i shuf23h = _mm_unpackhi_epi32(diff2, diff3);
+  __m128i v01l, v32l;
+  __m128i v01h, v32h;
+
+  // First pass
+  FTransformPass1(&shuf01l, &shuf23l, &v01l, &v32l);
+  FTransformPass1(&shuf01h, &shuf23h, &v01h, &v32h);
+
+  // Second pass
+  FTransformPass2(&v01l, &v32l, out + 0);
+  FTransformPass2(&v01h, &v32h, out + 16);
+}
+
+static void FTransformWHTRow(const int16_t* const in, __m128i* const out) {
+  const __m128i kMult1 = _mm_set_epi16(0, 0, 0, 0, 1, 1, 1, 1);
+  const __m128i kMult2 = _mm_set_epi16(0, 0, 0, 0, -1, 1, -1, 1);
+  const __m128i src0 = _mm_loadl_epi64((__m128i*)&in[0 * 16]);
+  const __m128i src1 = _mm_loadl_epi64((__m128i*)&in[1 * 16]);
+  const __m128i src2 = _mm_loadl_epi64((__m128i*)&in[2 * 16]);
+  const __m128i src3 = _mm_loadl_epi64((__m128i*)&in[3 * 16]);
+  const __m128i A01 = _mm_unpacklo_epi16(src0, src1);  // A0 A1 | ...
+  const __m128i A23 = _mm_unpacklo_epi16(src2, src3);  // A2 A3 | ...
+  const __m128i B0 = _mm_adds_epi16(A01, A23);    // a0 | a1 | ...
+  const __m128i B1 = _mm_subs_epi16(A01, A23);    // a3 | a2 | ...
+  const __m128i C0 = _mm_unpacklo_epi32(B0, B1);  // a0 | a1 | a3 | a2
+  const __m128i C1 = _mm_unpacklo_epi32(B1, B0);  // a3 | a2 | a0 | a1
+  const __m128i D0 = _mm_madd_epi16(C0, kMult1);  // out0, out1
+  const __m128i D1 = _mm_madd_epi16(C1, kMult2);  // out2, out3
+  *out = _mm_unpacklo_epi64(D0, D1);
+}
+
+static void FTransformWHT(const int16_t* in, int16_t* out) {
+  __m128i row0, row1, row2, row3;
+  FTransformWHTRow(in + 0 * 64, &row0);
+  FTransformWHTRow(in + 1 * 64, &row1);
+  FTransformWHTRow(in + 2 * 64, &row2);
+  FTransformWHTRow(in + 3 * 64, &row3);
+
   {
-    // Same operations are done on the (0,3) and (1,2) pairs.
-    // a0 = v0 + v3
-    // a1 = v1 + v2
-    // a3 = v0 - v3
-    // a2 = v1 - v2
-    const __m128i a01 = _mm_add_epi16(v01, v32);
-    const __m128i a32 = _mm_sub_epi16(v01, v32);
-    const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
-    const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
-
-    // d0 = (a0 + a1 + 7) >> 4;
-    // d2 = (a0 - a1 + 7) >> 4;
-    const __m128i b0 = _mm_add_epi16(a01, a11);
-    const __m128i b2 = _mm_sub_epi16(a01, a11);
-    const __m128i c0 = _mm_add_epi16(b0, seven);
-    const __m128i c2 = _mm_add_epi16(b2, seven);
-    const __m128i d0 = _mm_srai_epi16(c0, 4);
-    const __m128i d2 = _mm_srai_epi16(c2, 4);
-
-    // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
-    // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
-    const __m128i b23 = _mm_unpacklo_epi16(a22, a32);
-    const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
-    const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
-    const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one);
-    const __m128i d3 = _mm_add_epi32(c3, k51000);
-    const __m128i e1 = _mm_srai_epi32(d1, 16);
-    const __m128i e3 = _mm_srai_epi32(d3, 16);
-    const __m128i f1 = _mm_packs_epi32(e1, e1);
-    const __m128i f3 = _mm_packs_epi32(e3, e3);
-    // f1 = f1 + (a3 != 0);
-    // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
-    // desired (0, 1), we add one earlier through k12000_plus_one.
-    const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
-
-    _mm_storel_epi64((__m128i*)&out[ 0], d0);
-    _mm_storel_epi64((__m128i*)&out[ 4], g1);
-    _mm_storel_epi64((__m128i*)&out[ 8], d2);
-    _mm_storel_epi64((__m128i*)&out[12], f3);
+    const __m128i a0 = _mm_add_epi32(row0, row2);
+    const __m128i a1 = _mm_add_epi32(row1, row3);
+    const __m128i a2 = _mm_sub_epi32(row1, row3);
+    const __m128i a3 = _mm_sub_epi32(row0, row2);
+    const __m128i b0 = _mm_srai_epi32(_mm_add_epi32(a0, a1), 1);
+    const __m128i b1 = _mm_srai_epi32(_mm_add_epi32(a3, a2), 1);
+    const __m128i b2 = _mm_srai_epi32(_mm_sub_epi32(a3, a2), 1);
+    const __m128i b3 = _mm_srai_epi32(_mm_sub_epi32(a0, a1), 1);
+    const __m128i out0 = _mm_packs_epi32(b0, b1);
+    const __m128i out1 = _mm_packs_epi32(b2, b3);
+    _mm_storeu_si128((__m128i*)&out[0], out0);
+    _mm_storeu_si128((__m128i*)&out[8], out1);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Compute susceptibility based on DCT-coeff histograms:
+// the higher, the "easier" the macroblock is to compress.
+
+static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
+                             int start_block, int end_block,
+                             VP8Histogram* const histo) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
+  int j;
+  int distribution[MAX_COEFF_THRESH + 1] = { 0 };
+  for (j = start_block; j < end_block; ++j) {
+    int16_t out[16];
+    int k;
+
+    FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
+
+    // Convert coefficients to bin (within out[]).
+    {
+      // Load.
+      const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
+      const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
+      const __m128i d0 = _mm_sub_epi16(zero, out0);
+      const __m128i d1 = _mm_sub_epi16(zero, out1);
+      const __m128i abs0 = _mm_max_epi16(out0, d0);   // abs(v), 16b
+      const __m128i abs1 = _mm_max_epi16(out1, d1);
+      // v = abs(out) >> 3
+      const __m128i v0 = _mm_srai_epi16(abs0, 3);
+      const __m128i v1 = _mm_srai_epi16(abs1, 3);
+      // bin = min(v, MAX_COEFF_THRESH)
+      const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
+      const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
+      // Store.
+      _mm_storeu_si128((__m128i*)&out[0], bin0);
+      _mm_storeu_si128((__m128i*)&out[8], bin1);
+    }
+
+    // Convert coefficients to bin.
+    for (k = 0; k < 16; ++k) {
+      ++distribution[out[k]];
+    }
   }
+  VP8SetHistogramData(distribution, histo);
+}
+
+//------------------------------------------------------------------------------
+// Intra predictions
+
+// helper for chroma-DC predictions
+static WEBP_INLINE void Put8x8uv(uint8_t v, uint8_t* dst) {
+  int j;
+  const __m128i values = _mm_set1_epi8(v);
+  for (j = 0; j < 8; ++j) {
+    _mm_storel_epi64((__m128i*)(dst + j * BPS), values);
+  }
+}
+
+static WEBP_INLINE void Put16(uint8_t v, uint8_t* dst) {
+  int j;
+  const __m128i values = _mm_set1_epi8(v);
+  for (j = 0; j < 16; ++j) {
+    _mm_store_si128((__m128i*)(dst + j * BPS), values);
+  }
+}
+
+static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
+  if (size == 4) {
+    int j;
+    for (j = 0; j < 4; ++j) {
+      memset(dst + j * BPS, value, 4);
+    }
+  } else if (size == 8) {
+    Put8x8uv(value, dst);
+  } else {
+    Put16(value, dst);
+  }
+}
+
+static WEBP_INLINE void VE8uv(uint8_t* dst, const uint8_t* top) {
+  int j;
+  const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
+  for (j = 0; j < 8; ++j) {
+    _mm_storel_epi64((__m128i*)(dst + j * BPS), top_values);
+  }
+}
+
+static WEBP_INLINE void VE16(uint8_t* dst, const uint8_t* top) {
+  const __m128i top_values = _mm_load_si128((const __m128i*)top);
+  int j;
+  for (j = 0; j < 16; ++j) {
+    _mm_store_si128((__m128i*)(dst + j * BPS), top_values);
+  }
+}
+
+static WEBP_INLINE void VerticalPred(uint8_t* dst,
+                                     const uint8_t* top, int size) {
+  if (top != NULL) {
+    if (size == 8) {
+      VE8uv(dst, top);
+    } else {
+      VE16(dst, top);
+    }
+  } else {
+    Fill(dst, 127, size);
+  }
+}
+
+static WEBP_INLINE void HE8uv(uint8_t* dst, const uint8_t* left) {
+  int j;
+  for (j = 0; j < 8; ++j) {
+    const __m128i values = _mm_set1_epi8(left[j]);
+    _mm_storel_epi64((__m128i*)dst, values);
+    dst += BPS;
+  }
+}
+
+static WEBP_INLINE void HE16(uint8_t* dst, const uint8_t* left) {
+  int j;
+  for (j = 0; j < 16; ++j) {
+    const __m128i values = _mm_set1_epi8(left[j]);
+    _mm_store_si128((__m128i*)dst, values);
+    dst += BPS;
+  }
+}
+
+static WEBP_INLINE void HorizontalPred(uint8_t* dst,
+                                       const uint8_t* left, int size) {
+  if (left != NULL) {
+    if (size == 8) {
+      HE8uv(dst, left);
+    } else {
+      HE16(dst, left);
+    }
+  } else {
+    Fill(dst, 129, size);
+  }
+}
+
+static WEBP_INLINE void TM(uint8_t* dst, const uint8_t* left,
+                           const uint8_t* top, int size) {
+  const __m128i zero = _mm_setzero_si128();
+  int y;
+  if (size == 8) {
+    const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
+    const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
+    for (y = 0; y < 8; ++y, dst += BPS) {
+      const int val = left[y] - left[-1];
+      const __m128i base = _mm_set1_epi16(val);
+      const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
+      _mm_storel_epi64((__m128i*)dst, out);
+    }
+  } else {
+    const __m128i top_values = _mm_load_si128((const __m128i*)top);
+    const __m128i top_base_0 = _mm_unpacklo_epi8(top_values, zero);
+    const __m128i top_base_1 = _mm_unpackhi_epi8(top_values, zero);
+    for (y = 0; y < 16; ++y, dst += BPS) {
+      const int val = left[y] - left[-1];
+      const __m128i base = _mm_set1_epi16(val);
+      const __m128i out_0 = _mm_add_epi16(base, top_base_0);
+      const __m128i out_1 = _mm_add_epi16(base, top_base_1);
+      const __m128i out = _mm_packus_epi16(out_0, out_1);
+      _mm_store_si128((__m128i*)dst, out);
+    }
+  }
+}
+
+static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
+                                   const uint8_t* top, int size) {
+  if (left != NULL) {
+    if (top != NULL) {
+      TM(dst, left, top, size);
+    } else {
+      HorizontalPred(dst, left, size);
+    }
+  } else {
+    // true motion without left samples (hence: with default 129 value)
+    // is equivalent to VE prediction where you just copy the top samples.
+    // Note that if top samples are not available, the default value is
+    // then 129, and not 127 as in the VerticalPred case.
+    if (top != NULL) {
+      VerticalPred(dst, top, size);
+    } else {
+      Fill(dst, 129, size);
+    }
+  }
+}
+
+static WEBP_INLINE void DC8uv(uint8_t* dst, const uint8_t* left,
+                              const uint8_t* top) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
+  const __m128i left_values = _mm_loadl_epi64((const __m128i*)left);
+  const __m128i sum_top = _mm_sad_epu8(top_values, zero);
+  const __m128i sum_left = _mm_sad_epu8(left_values, zero);
+  const int DC = _mm_cvtsi128_si32(sum_top) + _mm_cvtsi128_si32(sum_left) + 8;
+  Put8x8uv(DC >> 4, dst);
+}
+
+static WEBP_INLINE void DC8uvNoLeft(uint8_t* dst, const uint8_t* top) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
+  const __m128i sum = _mm_sad_epu8(top_values, zero);
+  const int DC = _mm_cvtsi128_si32(sum) + 4;
+  Put8x8uv(DC >> 3, dst);
+}
+
+static WEBP_INLINE void DC8uvNoTop(uint8_t* dst, const uint8_t* left) {
+  // 'left' is contiguous so we can reuse the top summation.
+  DC8uvNoLeft(dst, left);
+}
+
+static WEBP_INLINE void DC8uvNoTopLeft(uint8_t* dst) {
+  Put8x8uv(0x80, dst);
+}
+
+static WEBP_INLINE void DC8uvMode(uint8_t* dst, const uint8_t* left,
+                                  const uint8_t* top) {
+  if (top != NULL) {
+    if (left != NULL) {  // top and left present
+      DC8uv(dst, left, top);
+    } else {  // top, but no left
+      DC8uvNoLeft(dst, top);
+    }
+  } else if (left != NULL) {  // left but no top
+    DC8uvNoTop(dst, left);
+  } else {  // no top, no left, nothing.
+    DC8uvNoTopLeft(dst);
+  }
+}
+
+static WEBP_INLINE void DC16(uint8_t* dst, const uint8_t* left,
+                             const uint8_t* top) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top_row = _mm_load_si128((const __m128i*)top);
+  const __m128i left_row = _mm_load_si128((const __m128i*)left);
+  const __m128i sad8x2 = _mm_sad_epu8(top_row, zero);
+  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
+  const __m128i sum_top = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
+  const __m128i sad8x2_left = _mm_sad_epu8(left_row, zero);
+  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
+  const __m128i sum_left =
+      _mm_add_epi16(sad8x2_left, _mm_shuffle_epi32(sad8x2_left, 2));
+  const int DC = _mm_cvtsi128_si32(sum_top) + _mm_cvtsi128_si32(sum_left) + 16;
+  Put16(DC >> 5, dst);
+}
+
+static WEBP_INLINE void DC16NoLeft(uint8_t* dst, const uint8_t* top) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top_row = _mm_load_si128((const __m128i*)top);
+  const __m128i sad8x2 = _mm_sad_epu8(top_row, zero);
+  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
+  const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
+  const int DC = _mm_cvtsi128_si32(sum) + 8;
+  Put16(DC >> 4, dst);
+}
+
+static WEBP_INLINE void DC16NoTop(uint8_t* dst, const uint8_t* left) {
+  // 'left' is contiguous so we can reuse the top summation.
+  DC16NoLeft(dst, left);
+}
+
+static WEBP_INLINE void DC16NoTopLeft(uint8_t* dst) {
+  Put16(0x80, dst);
+}
+
+static WEBP_INLINE void DC16Mode(uint8_t* dst, const uint8_t* left,
+                                 const uint8_t* top) {
+  if (top != NULL) {
+    if (left != NULL) {  // top and left present
+      DC16(dst, left, top);
+    } else {  // top, but no left
+      DC16NoLeft(dst, top);
+    }
+  } else if (left != NULL) {  // left but no top
+    DC16NoTop(dst, left);
+  } else {  // no top, no left, nothing.
+    DC16NoTopLeft(dst);
+  }
+}
+
+//------------------------------------------------------------------------------
+// 4x4 predictions
+
+#define DST(x, y) dst[(x) + (y) * BPS]
+#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
+#define AVG2(a, b) (((a) + (b) + 1) >> 1)
+
+// We use the following 8b-arithmetic tricks:
+//     (a + 2 * b + c + 2) >> 2 = (AC + b + 1) >> 1
+//   where: AC = (a + c) >> 1 = [(a + c + 1) >> 1] - [(a^c) & 1]
+// and:
+//     (a + 2 * b + c + 2) >> 2 = (AB + BC + 1) >> 1 - (ab|bc)&lsb
+//   where: AC = (a + b + 1) >> 1,   BC = (b + c + 1) >> 1
+//   and ab = a ^ b, bc = b ^ c, lsb = (AC^BC)&1
+
+static WEBP_INLINE void VE4(uint8_t* dst, const uint8_t* top) {  // vertical
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(top - 1));
+  const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i a = _mm_avg_epu8(ABCDEFGH, CDEFGH00);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one);
+  const __m128i b = _mm_subs_epu8(a, lsb);
+  const __m128i avg = _mm_avg_epu8(b, BCDEFGH0);
+  const uint32_t vals = _mm_cvtsi128_si32(avg);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    *(uint32_t*)(dst + i * BPS) = vals;
+  }
+}
+
+static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) {  // horizontal
+  const int X = top[-1];
+  const int I = top[-2];
+  const int J = top[-3];
+  const int K = top[-4];
+  const int L = top[-5];
+  *(uint32_t*)(dst + 0 * BPS) = 0x01010101U * AVG3(X, I, J);
+  *(uint32_t*)(dst + 1 * BPS) = 0x01010101U * AVG3(I, J, K);
+  *(uint32_t*)(dst + 2 * BPS) = 0x01010101U * AVG3(J, K, L);
+  *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(K, L, L);
+}
+
+static WEBP_INLINE void DC4(uint8_t* dst, const uint8_t* top) {
+  uint32_t dc = 4;
+  int i;
+  for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
+  Fill(dst, dc >> 3, 4);
+}
+
+static WEBP_INLINE void LD4(uint8_t* dst, const uint8_t* top) {  // Down-Left
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top);
+  const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i CDEFGHH0 = _mm_insert_epi16(CDEFGH00, top[7], 3);
+  const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, CDEFGHH0);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0);
+  *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(               abcdefg    );
+  *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1));
+  *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2));
+  *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3));
+}
+
+static WEBP_INLINE void VR4(uint8_t* dst,
+                            const uint8_t* top) {  // Vertical-Right
+  const __m128i one = _mm_set1_epi8(1);
+  const int I = top[-2];
+  const int J = top[-3];
+  const int K = top[-4];
+  const int X = top[-1];
+  const __m128i XABCD = _mm_loadl_epi64((const __m128i*)(top - 1));
+  const __m128i ABCD0 = _mm_srli_si128(XABCD, 1);
+  const __m128i abcd = _mm_avg_epu8(XABCD, ABCD0);
+  const __m128i _XABCD = _mm_slli_si128(XABCD, 1);
+  const __m128i IXABCD = _mm_insert_epi16(_XABCD, I | (X << 8), 0);
+  const __m128i avg1 = _mm_avg_epu8(IXABCD, ABCD0);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i efgh = _mm_avg_epu8(avg2, XABCD);
+  *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(               abcd    );
+  *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(               efgh    );
+  *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1));
+  *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1));
+
+  // these two are hard to implement in SSE2, so we keep the C-version:
+  DST(0, 2) = AVG3(J, I, X);
+  DST(0, 3) = AVG3(K, J, I);
+}
+
+static WEBP_INLINE void VL4(uint8_t* dst,
+                            const uint8_t* top) {  // Vertical-Left
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top);
+  const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH__ = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, BCDEFGH_);
+  const __m128i avg2 = _mm_avg_epu8(CDEFGH__, BCDEFGH_);
+  const __m128i avg3 = _mm_avg_epu8(avg1, avg2);
+  const __m128i lsb1 = _mm_and_si128(_mm_xor_si128(avg1, avg2), one);
+  const __m128i ab = _mm_xor_si128(ABCDEFGH, BCDEFGH_);
+  const __m128i bc = _mm_xor_si128(CDEFGH__, BCDEFGH_);
+  const __m128i abbc = _mm_or_si128(ab, bc);
+  const __m128i lsb2 = _mm_and_si128(abbc, lsb1);
+  const __m128i avg4 = _mm_subs_epu8(avg3, lsb2);
+  const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
+  *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(               avg1    );
+  *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(               avg4    );
+  *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1));
+  *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1));
+
+  // these two are hard to get and irregular
+  DST(3, 2) = (extra_out >> 0) & 0xff;
+  DST(3, 3) = (extra_out >> 8) & 0xff;
+}
+
+static WEBP_INLINE void RD4(uint8_t* dst, const uint8_t* top) {  // Down-right
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i LKJIXABC = _mm_loadl_epi64((const __m128i*)(top - 5));
+  const __m128i LKJIXABCD = _mm_insert_epi16(LKJIXABC, top[3], 4);
+  const __m128i KJIXABCD_ = _mm_srli_si128(LKJIXABCD, 1);
+  const __m128i JIXABCD__ = _mm_srli_si128(LKJIXABCD, 2);
+  const __m128i avg1 = _mm_avg_epu8(JIXABCD__, LKJIXABCD);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_);
+  *(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(               abcdefg    );
+  *(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1));
+  *(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2));
+  *(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3));
+}
+
+static WEBP_INLINE void HU4(uint8_t* dst, const uint8_t* top) {
+  const int I = top[-2];
+  const int J = top[-3];
+  const int K = top[-4];
+  const int L = top[-5];
+  DST(0, 0) =             AVG2(I, J);
+  DST(2, 0) = DST(0, 1) = AVG2(J, K);
+  DST(2, 1) = DST(0, 2) = AVG2(K, L);
+  DST(1, 0) =             AVG3(I, J, K);
+  DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
+  DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
+  DST(3, 2) = DST(2, 2) =
+  DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
+}
+
+static WEBP_INLINE void HD4(uint8_t* dst, const uint8_t* top) {
+  const int X = top[-1];
+  const int I = top[-2];
+  const int J = top[-3];
+  const int K = top[-4];
+  const int L = top[-5];
+  const int A = top[0];
+  const int B = top[1];
+  const int C = top[2];
+
+  DST(0, 0) = DST(2, 1) = AVG2(I, X);
+  DST(0, 1) = DST(2, 2) = AVG2(J, I);
+  DST(0, 2) = DST(2, 3) = AVG2(K, J);
+  DST(0, 3)             = AVG2(L, K);
+
+  DST(3, 0)             = AVG3(A, B, C);
+  DST(2, 0)             = AVG3(X, A, B);
+  DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
+  DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
+  DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
+  DST(1, 3)             = AVG3(L, K, J);
+}
+
+static WEBP_INLINE void TM4(uint8_t* dst, const uint8_t* top) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top_values = _mm_cvtsi32_si128(MemToUint32(top));
+  const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
+  int y;
+  for (y = 0; y < 4; ++y, dst += BPS) {
+    const int val = top[-2 - y] - top[-1];
+    const __m128i base = _mm_set1_epi16(val);
+    const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
+    *(int*)dst = _mm_cvtsi128_si32(out);
+  }
+}
+
+#undef DST
+#undef AVG3
+#undef AVG2
+
+//------------------------------------------------------------------------------
+// luma 4x4 prediction
+
+// Left samples are top[-5 .. -2], top_left is top[-1], top are
+// located at top[0..3], and top right is top[4..7]
+static void Intra4Preds(uint8_t* dst, const uint8_t* top) {
+  DC4(I4DC4 + dst, top);
+  TM4(I4TM4 + dst, top);
+  VE4(I4VE4 + dst, top);
+  HE4(I4HE4 + dst, top);
+  RD4(I4RD4 + dst, top);
+  VR4(I4VR4 + dst, top);
+  LD4(I4LD4 + dst, top);
+  VL4(I4VL4 + dst, top);
+  HD4(I4HD4 + dst, top);
+  HU4(I4HU4 + dst, top);
+}
+
+//------------------------------------------------------------------------------
+// Chroma 8x8 prediction (paragraph 12.2)
+
+static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
+                             const uint8_t* top) {
+  // U block
+  DC8uvMode(C8DC8 + dst, left, top);
+  VerticalPred(C8VE8 + dst, top, 8);
+  HorizontalPred(C8HE8 + dst, left, 8);
+  TrueMotion(C8TM8 + dst, left, top, 8);
+  // V block
+  dst += 8;
+  if (top != NULL) top += 8;
+  if (left != NULL) left += 16;
+  DC8uvMode(C8DC8 + dst, left, top);
+  VerticalPred(C8VE8 + dst, top, 8);
+  HorizontalPred(C8HE8 + dst, left, 8);
+  TrueMotion(C8TM8 + dst, left, top, 8);
+}
+
+//------------------------------------------------------------------------------
+// luma 16x16 prediction (paragraph 12.3)
+
+static void Intra16Preds(uint8_t* dst,
+                         const uint8_t* left, const uint8_t* top) {
+  DC16Mode(I16DC16 + dst, left, top);
+  VerticalPred(I16VE16 + dst, top, 16);
+  HorizontalPred(I16HE16 + dst, left, 16);
+  TrueMotion(I16TM16 + dst, left, top, 16);
 }
 
 //------------------------------------------------------------------------------
 // Metric
 
-static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) {
-  const __m128i zero = _mm_set1_epi16(0);
+static WEBP_INLINE void SubtractAndAccumulate(const __m128i a, const __m128i b,
+                                              __m128i* const sum) {
+  // take abs(a-b) in 8b
+  const __m128i a_b = _mm_subs_epu8(a, b);
+  const __m128i b_a = _mm_subs_epu8(b, a);
+  const __m128i abs_a_b = _mm_or_si128(a_b, b_a);
+  // zero-extend to 16b
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i C0 = _mm_unpacklo_epi8(abs_a_b, zero);
+  const __m128i C1 = _mm_unpackhi_epi8(abs_a_b, zero);
+  // multiply with self
+  const __m128i sum1 = _mm_madd_epi16(C0, C0);
+  const __m128i sum2 = _mm_madd_epi16(C1, C1);
+  *sum = _mm_add_epi32(sum1, sum2);
+}
 
-  // Load values.
-  const __m128i a0 = _mm_loadl_epi64((__m128i*)&a[BPS * 0]);
-  const __m128i a1 = _mm_loadl_epi64((__m128i*)&a[BPS * 1]);
-  const __m128i a2 = _mm_loadl_epi64((__m128i*)&a[BPS * 2]);
-  const __m128i a3 = _mm_loadl_epi64((__m128i*)&a[BPS * 3]);
-  const __m128i b0 = _mm_loadl_epi64((__m128i*)&b[BPS * 0]);
-  const __m128i b1 = _mm_loadl_epi64((__m128i*)&b[BPS * 1]);
-  const __m128i b2 = _mm_loadl_epi64((__m128i*)&b[BPS * 2]);
-  const __m128i b3 = _mm_loadl_epi64((__m128i*)&b[BPS * 3]);
+static WEBP_INLINE int SSE_16xN(const uint8_t* a, const uint8_t* b,
+                                int num_pairs) {
+  __m128i sum = _mm_setzero_si128();
+  int32_t tmp[4];
+  int i;
+
+  for (i = 0; i < num_pairs; ++i) {
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[BPS * 0]);
+    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[BPS * 0]);
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[BPS * 1]);
+    const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[BPS * 1]);
+    __m128i sum1, sum2;
+    SubtractAndAccumulate(a0, b0, &sum1);
+    SubtractAndAccumulate(a1, b1, &sum2);
+    sum = _mm_add_epi32(sum, _mm_add_epi32(sum1, sum2));
+    a += 2 * BPS;
+    b += 2 * BPS;
+  }
+  _mm_storeu_si128((__m128i*)tmp, sum);
+  return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
+}
+
+static int SSE16x16(const uint8_t* a, const uint8_t* b) {
+  return SSE_16xN(a, b, 8);
+}
 
-  // Combine pair of lines and convert to 16b.
+static int SSE16x8(const uint8_t* a, const uint8_t* b) {
+  return SSE_16xN(a, b, 4);
+}
+
+#define LOAD_8x16b(ptr) \
+  _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i*)(ptr)), zero)
+
+static int SSE8x8(const uint8_t* a, const uint8_t* b) {
+  const __m128i zero = _mm_setzero_si128();
+  int num_pairs = 4;
+  __m128i sum = zero;
+  int32_t tmp[4];
+  while (num_pairs-- > 0) {
+    const __m128i a0 = LOAD_8x16b(&a[BPS * 0]);
+    const __m128i a1 = LOAD_8x16b(&a[BPS * 1]);
+    const __m128i b0 = LOAD_8x16b(&b[BPS * 0]);
+    const __m128i b1 = LOAD_8x16b(&b[BPS * 1]);
+    // subtract
+    const __m128i c0 = _mm_subs_epi16(a0, b0);
+    const __m128i c1 = _mm_subs_epi16(a1, b1);
+    // multiply/accumulate with self
+    const __m128i d0 = _mm_madd_epi16(c0, c0);
+    const __m128i d1 = _mm_madd_epi16(c1, c1);
+    // collect
+    const __m128i sum01 = _mm_add_epi32(d0, d1);
+    sum = _mm_add_epi32(sum, sum01);
+    a += 2 * BPS;
+    b += 2 * BPS;
+  }
+  _mm_storeu_si128((__m128i*)tmp, sum);
+  return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
+}
+#undef LOAD_8x16b
+
+static int SSE4x4(const uint8_t* a, const uint8_t* b) {
+  const __m128i zero = _mm_setzero_si128();
+
+  // Load values. Note that we read 8 pixels instead of 4,
+  // but the a/b buffers are over-allocated to that effect.
+  const __m128i a0 = _mm_loadl_epi64((const __m128i*)&a[BPS * 0]);
+  const __m128i a1 = _mm_loadl_epi64((const __m128i*)&a[BPS * 1]);
+  const __m128i a2 = _mm_loadl_epi64((const __m128i*)&a[BPS * 2]);
+  const __m128i a3 = _mm_loadl_epi64((const __m128i*)&a[BPS * 3]);
+  const __m128i b0 = _mm_loadl_epi64((const __m128i*)&b[BPS * 0]);
+  const __m128i b1 = _mm_loadl_epi64((const __m128i*)&b[BPS * 1]);
+  const __m128i b2 = _mm_loadl_epi64((const __m128i*)&b[BPS * 2]);
+  const __m128i b3 = _mm_loadl_epi64((const __m128i*)&b[BPS * 3]);
+  // Combine pair of lines.
   const __m128i a01 = _mm_unpacklo_epi32(a0, a1);
   const __m128i a23 = _mm_unpacklo_epi32(a2, a3);
   const __m128i b01 = _mm_unpacklo_epi32(b0, b1);
   const __m128i b23 = _mm_unpacklo_epi32(b2, b3);
+  // Convert to 16b.
   const __m128i a01s = _mm_unpacklo_epi8(a01, zero);
   const __m128i a23s = _mm_unpacklo_epi8(a23, zero);
   const __m128i b01s = _mm_unpacklo_epi8(b01, zero);
   const __m128i b23s = _mm_unpacklo_epi8(b23, zero);
+  // subtract, square and accumulate
+  const __m128i d0 = _mm_subs_epi16(a01s, b01s);
+  const __m128i d1 = _mm_subs_epi16(a23s, b23s);
+  const __m128i e0 = _mm_madd_epi16(d0, d0);
+  const __m128i e1 = _mm_madd_epi16(d1, d1);
+  const __m128i sum = _mm_add_epi32(e0, e1);
 
-  // Compute differences; (a-b)^2 = (abs(a-b))^2 = (sat8(a-b) + sat8(b-a))^2
-  // TODO(cduvivier): Dissassemble and figure out why this is fastest. We don't
-  //                  need absolute values, there is no need to do calculation
-  //                  in 8bit as we are already in 16bit, ... Yet this is what
-  //                  benchmarks the fastest!
-  const __m128i d0 = _mm_subs_epu8(a01s, b01s);
-  const __m128i d1 = _mm_subs_epu8(b01s, a01s);
-  const __m128i d2 = _mm_subs_epu8(a23s, b23s);
-  const __m128i d3 = _mm_subs_epu8(b23s, a23s);
-
-  // Square and add them all together.
-  const __m128i madd0 = _mm_madd_epi16(d0, d0);
-  const __m128i madd1 = _mm_madd_epi16(d1, d1);
-  const __m128i madd2 = _mm_madd_epi16(d2, d2);
-  const __m128i madd3 = _mm_madd_epi16(d3, d3);
-  const __m128i sum0 = _mm_add_epi32(madd0, madd1);
-  const __m128i sum1 = _mm_add_epi32(madd2, madd3);
-  const __m128i sum2 = _mm_add_epi32(sum0, sum1);
   int32_t tmp[4];
-  _mm_storeu_si128((__m128i*)tmp, sum2);
+  _mm_storeu_si128((__m128i*)tmp, sum);
   return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
 }
 
@@ -497,24 +1155,22 @@ static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) {
 // Hadamard transform
 // Returns the difference between the weighted sum of the absolute value of
 // transformed coefficients.
-static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
-                          const uint16_t* const w) {
+static int TTransform(const uint8_t* inA, const uint8_t* inB,
+                      const uint16_t* const w) {
   int32_t sum[4];
   __m128i tmp_0, tmp_1, tmp_2, tmp_3;
   const __m128i zero = _mm_setzero_si128();
-  const __m128i one = _mm_set1_epi16(1);
-  const __m128i three = _mm_set1_epi16(3);
 
-  // Load, combine and tranpose inputs.
+  // Load, combine and transpose inputs.
   {
-    const __m128i inA_0 = _mm_loadl_epi64((__m128i*)&inA[BPS * 0]);
-    const __m128i inA_1 = _mm_loadl_epi64((__m128i*)&inA[BPS * 1]);
-    const __m128i inA_2 = _mm_loadl_epi64((__m128i*)&inA[BPS * 2]);
-    const __m128i inA_3 = _mm_loadl_epi64((__m128i*)&inA[BPS * 3]);
-    const __m128i inB_0 = _mm_loadl_epi64((__m128i*)&inB[BPS * 0]);
-    const __m128i inB_1 = _mm_loadl_epi64((__m128i*)&inB[BPS * 1]);
-    const __m128i inB_2 = _mm_loadl_epi64((__m128i*)&inB[BPS * 2]);
-    const __m128i inB_3 = _mm_loadl_epi64((__m128i*)&inB[BPS * 3]);
+    const __m128i inA_0 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 0]);
+    const __m128i inA_1 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 1]);
+    const __m128i inA_2 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 2]);
+    const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]);
+    const __m128i inB_0 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 0]);
+    const __m128i inB_1 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 1]);
+    const __m128i inB_2 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 2]);
+    const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]);
 
     // Combine inA and inB (we'll do two transforms in parallel).
     const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0);
@@ -550,17 +1206,14 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
   // Horizontal pass and subsequent transpose.
   {
     // Calculate a and b (two 4x4 at once).
-    const __m128i a0 = _mm_slli_epi16(_mm_add_epi16(tmp_0, tmp_2), 2);
-    const __m128i a1 = _mm_slli_epi16(_mm_add_epi16(tmp_1, tmp_3), 2);
-    const __m128i a2 = _mm_slli_epi16(_mm_sub_epi16(tmp_1, tmp_3), 2);
-    const __m128i a3 = _mm_slli_epi16(_mm_sub_epi16(tmp_0, tmp_2), 2);
-    // b0_extra = (a0 != 0);
-    const __m128i b0_extra = _mm_andnot_si128(_mm_cmpeq_epi16 (a0, zero), one);
-    const __m128i b0_base = _mm_add_epi16(a0, a1);
+    const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
+    const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
+    const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
+    const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
+    const __m128i b0 = _mm_add_epi16(a0, a1);
     const __m128i b1 = _mm_add_epi16(a3, a2);
     const __m128i b2 = _mm_sub_epi16(a3, a2);
     const __m128i b3 = _mm_sub_epi16(a0, a1);
-    const __m128i b0 = _mm_add_epi16(b0_base, b0_extra);
     // a00 a01 a02 a03   b00 b01 b02 b03
     // a10 a11 a12 a13   b10 b11 b12 b13
     // a20 a21 a22 a23   b20 b21 b22 b23
@@ -598,8 +1251,8 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
     // Load all inputs.
     // TODO(cduvivier): Make variable declarations and allocations aligned so
     //                  we can use _mm_load_si128 instead of _mm_loadu_si128.
-    const __m128i w_0 = _mm_loadu_si128((__m128i*)&w[0]);
-    const __m128i w_8 = _mm_loadu_si128((__m128i*)&w[8]);
+    const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]);
+    const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]);
 
     // Calculate a and b (two 4x4 at once).
     const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
@@ -618,36 +1271,16 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
     __m128i B_b2 = _mm_unpackhi_epi64(b2, b3);
 
     {
-      // sign(b) = b >> 15  (0x0000 if positive, 0xffff if negative)
-      const __m128i sign_A_b0 = _mm_srai_epi16(A_b0, 15);
-      const __m128i sign_A_b2 = _mm_srai_epi16(A_b2, 15);
-      const __m128i sign_B_b0 = _mm_srai_epi16(B_b0, 15);
-      const __m128i sign_B_b2 = _mm_srai_epi16(B_b2, 15);
-
-      // b = abs(b) = (b ^ sign) - sign
-      A_b0 = _mm_xor_si128(A_b0, sign_A_b0);
-      A_b2 = _mm_xor_si128(A_b2, sign_A_b2);
-      B_b0 = _mm_xor_si128(B_b0, sign_B_b0);
-      B_b2 = _mm_xor_si128(B_b2, sign_B_b2);
-      A_b0 = _mm_sub_epi16(A_b0, sign_A_b0);
-      A_b2 = _mm_sub_epi16(A_b2, sign_A_b2);
-      B_b0 = _mm_sub_epi16(B_b0, sign_B_b0);
-      B_b2 = _mm_sub_epi16(B_b2, sign_B_b2);
+      const __m128i d0 = _mm_sub_epi16(zero, A_b0);
+      const __m128i d1 = _mm_sub_epi16(zero, A_b2);
+      const __m128i d2 = _mm_sub_epi16(zero, B_b0);
+      const __m128i d3 = _mm_sub_epi16(zero, B_b2);
+      A_b0 = _mm_max_epi16(A_b0, d0);   // abs(v), 16b
+      A_b2 = _mm_max_epi16(A_b2, d1);
+      B_b0 = _mm_max_epi16(B_b0, d2);
+      B_b2 = _mm_max_epi16(B_b2, d3);
     }
 
-    // b = abs(b) + 3
-    A_b0 = _mm_add_epi16(A_b0, three);
-    A_b2 = _mm_add_epi16(A_b2, three);
-    B_b0 = _mm_add_epi16(B_b0, three);
-    B_b2 = _mm_add_epi16(B_b2, three);
-
-    // abs((b + (b<0) + 3) >> 3) = (abs(b) + 3) >> 3
-    // b = (abs(b) + 3) >> 3
-    A_b0 = _mm_srai_epi16(A_b0, 3);
-    A_b2 = _mm_srai_epi16(A_b2, 3);
-    B_b0 = _mm_srai_epi16(B_b0, 3);
-    B_b2 = _mm_srai_epi16(B_b2, 3);
-
     // weighted sums
     A_b0 = _mm_madd_epi16(A_b0, w_0);
     A_b2 = _mm_madd_epi16(A_b2, w_8);
@@ -663,35 +1296,33 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
   return sum[0] + sum[1] + sum[2] + sum[3];
 }
 
-static int Disto4x4SSE2(const uint8_t* const a, const uint8_t* const b,
-                        const uint16_t* const w) {
-  const int diff_sum = TTransformSSE2(a, b, w);
-  return (abs(diff_sum) + 8) >> 4;
+static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
+                    const uint16_t* const w) {
+  const int diff_sum = TTransform(a, b, w);
+  return abs(diff_sum) >> 5;
 }
 
-static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b,
-                          const uint16_t* const w) {
+static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
+                      const uint16_t* const w) {
   int D = 0;
   int x, y;
   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
     for (x = 0; x < 16; x += 4) {
-      D += Disto4x4SSE2(a + x + y, b + x + y, w);
+      D += Disto4x4(a + x + y, b + x + y, w);
     }
   }
   return D;
 }
 
-
 //------------------------------------------------------------------------------
 // Quantization
 //
 
-// Simple quantization
-static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
-                             int n, const VP8Matrix* const mtx) {
-  const __m128i max_coeff_2047 = _mm_set1_epi16(2047);
-  const __m128i zero = _mm_set1_epi16(0);
-  __m128i sign0, sign8;
+static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16],
+                                       const uint16_t* const sharpen,
+                                       const VP8Matrix* const mtx) {
+  const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
+  const __m128i zero = _mm_setzero_si128();
   __m128i coeff0, coeff8;
   __m128i out0, out8;
   __m128i packed_out;
@@ -701,20 +1332,14 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   //                  we can use _mm_load_si128 instead of _mm_loadu_si128.
   __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
   __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
-  const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[0]);
-  const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[8]);
-  const __m128i iq0 = _mm_loadu_si128((__m128i*)&mtx->iq_[0]);
-  const __m128i iq8 = _mm_loadu_si128((__m128i*)&mtx->iq_[8]);
-  const __m128i bias0 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]);
-  const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]);
-  const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]);
-  const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]);
-  const __m128i zthresh0 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[0]);
-  const __m128i zthresh8 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[8]);
-
-  // sign(in) = in >> 15  (0x0000 if positive, 0xffff if negative)
-  sign0 = _mm_srai_epi16(in0, 15);
-  sign8 = _mm_srai_epi16(in8, 15);
+  const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]);
+  const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]);
+  const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]);
+  const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]);
+
+  // extract sign(in)  (0x0000 if positive, 0xffff if negative)
+  const __m128i sign0 = _mm_cmpgt_epi16(zero, in0);
+  const __m128i sign8 = _mm_cmpgt_epi16(zero, in8);
 
   // coeff = abs(in) = (in ^ sign) - sign
   coeff0 = _mm_xor_si128(in0, sign0);
@@ -723,43 +1348,47 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   coeff8 = _mm_sub_epi16(coeff8, sign8);
 
   // coeff = abs(in) + sharpen
-  coeff0 = _mm_add_epi16(coeff0, sharpen0);
-  coeff8 = _mm_add_epi16(coeff8, sharpen8);
-
-  // if (coeff > 2047) coeff = 2047
-  coeff0 = _mm_min_epi16(coeff0, max_coeff_2047);
-  coeff8 = _mm_min_epi16(coeff8, max_coeff_2047);
+  if (sharpen != NULL) {
+    const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]);
+    const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]);
+    coeff0 = _mm_add_epi16(coeff0, sharpen0);
+    coeff8 = _mm_add_epi16(coeff8, sharpen8);
+  }
 
-  // out = (coeff * iQ + B) >> QFIX;
+  // out = (coeff * iQ + B) >> QFIX
   {
     // doing calculations with 32b precision (QFIX=17)
     // out = (coeff * iQ)
-    __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
-    __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
-    __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
-    __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
+    const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
+    const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
+    const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
+    const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
     __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
     __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
     __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
     __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
-    // expand bias from 16b to 32b
-    __m128i bias_00 = _mm_unpacklo_epi16(bias0, zero);
-    __m128i bias_04 = _mm_unpackhi_epi16(bias0, zero);
-    __m128i bias_08 = _mm_unpacklo_epi16(bias8, zero);
-    __m128i bias_12 = _mm_unpackhi_epi16(bias8, zero);
     // out = (coeff * iQ + B)
+    const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]);
+    const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]);
+    const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]);
+    const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]);
     out_00 = _mm_add_epi32(out_00, bias_00);
     out_04 = _mm_add_epi32(out_04, bias_04);
     out_08 = _mm_add_epi32(out_08, bias_08);
     out_12 = _mm_add_epi32(out_12, bias_12);
-    // out = (coeff * iQ + B) >> QFIX;
+    // out = QUANTDIV(coeff, iQ, B, QFIX)
     out_00 = _mm_srai_epi32(out_00, QFIX);
     out_04 = _mm_srai_epi32(out_04, QFIX);
     out_08 = _mm_srai_epi32(out_08, QFIX);
     out_12 = _mm_srai_epi32(out_12, QFIX);
+
     // pack result as 16b
     out0 = _mm_packs_epi32(out_00, out_04);
     out8 = _mm_packs_epi32(out_08, out_12);
+
+    // if (coeff > 2047) coeff = 2047
+    out0 = _mm_min_epi16(out0, max_coeff_2047);
+    out8 = _mm_min_epi16(out8, max_coeff_2047);
   }
 
   // get sign back (if (sign[j]) out_n = -out_n)
@@ -772,17 +1401,8 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   in0 = _mm_mullo_epi16(out0, q0);
   in8 = _mm_mullo_epi16(out8, q8);
 
-  // if (coeff <= mtx->zthresh_) {in=0; out=0;}
-  {
-    __m128i cmp0 = _mm_cmpgt_epi16(coeff0, zthresh0);
-    __m128i cmp8 = _mm_cmpgt_epi16(coeff8, zthresh8);
-    in0 = _mm_and_si128(in0, cmp0);
-    in8 = _mm_and_si128(in8, cmp8);
-    _mm_storeu_si128((__m128i*)&in[0], in0);
-    _mm_storeu_si128((__m128i*)&in[8], in8);
-    out0 = _mm_and_si128(out0, cmp0);
-    out8 = _mm_and_si128(out8, cmp8);
-  }
+  _mm_storeu_si128((__m128i*)&in[0], in0);
+  _mm_storeu_si128((__m128i*)&in[8], in8);
 
   // zigzag the output before storing it.
   //
@@ -809,29 +1429,55 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   }
 
   // detect if all 'out' values are zeroes or not
-  {
-    int32_t tmp[4];
-    _mm_storeu_si128((__m128i*)tmp, packed_out);
-    if (n) {
-      tmp[0] &= ~0xff;
-    }
-    return (tmp[3] || tmp[2] || tmp[1] || tmp[0]);
-  }
+  return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
+}
+
+static int QuantizeBlock(int16_t in[16], int16_t out[16],
+                         const VP8Matrix* const mtx) {
+  return DoQuantizeBlock(in, out, &mtx->sharpen_[0], mtx);
+}
+
+static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
+                            const VP8Matrix* const mtx) {
+  return DoQuantizeBlock(in, out, NULL, mtx);
+}
+
+static int Quantize2Blocks(int16_t in[32], int16_t out[32],
+                           const VP8Matrix* const mtx) {
+  int nz;
+  const uint16_t* const sharpen = &mtx->sharpen_[0];
+  nz  = DoQuantizeBlock(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0;
+  nz |= DoQuantizeBlock(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1;
+  return nz;
 }
 
+//------------------------------------------------------------------------------
+// Entry point
+
 extern void VP8EncDspInitSSE2(void);
-void VP8EncDspInitSSE2(void) {
-  VP8CollectHistogram = CollectHistogramSSE2;
-  VP8EncQuantizeBlock = QuantizeBlockSSE2;
-  VP8ITransform = ITransformSSE2;
-  VP8FTransform = FTransformSSE2;
-  VP8SSE4x4 = SSE4x4SSE2;
-  VP8TDisto4x4 = Disto4x4SSE2;
-  VP8TDisto16x16 = Disto16x16SSE2;
-}
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
 
-#endif   // WEBP_USE_SSE2
+WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE2(void) {
+  VP8CollectHistogram = CollectHistogram;
+  VP8EncPredLuma16 = Intra16Preds;
+  VP8EncPredChroma8 = IntraChromaPreds;
+  VP8EncPredLuma4 = Intra4Preds;
+  VP8EncQuantizeBlock = QuantizeBlock;
+  VP8EncQuantize2Blocks = Quantize2Blocks;
+  VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
+  VP8ITransform = ITransform;
+  VP8FTransform = FTransform;
+  VP8FTransform2 = FTransform2;
+  VP8FTransformWHT = FTransformWHT;
+  VP8SSE16x16 = SSE16x16;
+  VP8SSE16x8 = SSE16x8;
+  VP8SSE8x8 = SSE8x8;
+  VP8SSE4x4 = SSE4x4;
+  VP8TDisto4x4 = Disto4x4;
+  VP8TDisto16x16 = Disto16x16;
+}
+
+#else  // !WEBP_USE_SSE2
+
+WEBP_DSP_INIT_STUB(VP8EncDspInitSSE2)
+
+#endif  // WEBP_USE_SSE2
diff --git a/drivers/webp/dsp/lossless.c b/drivers/webp/dsp/lossless.c
index 62a6b7b15a..5702eb3b17 100644
--- a/drivers/webp/dsp/lossless.c
+++ b/drivers/webp/dsp/lossless.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
 // -----------------------------------------------------------------------------
 //
 // Image transforms and color space conversion methods for lossless decoder.
@@ -11,170 +13,16 @@
 //          Jyrki Alakuijala (jyrki@google.com)
 //          Urvang Joshi (urvang@google.com)
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "./dsp.h"
 
 #include <math.h>
 #include <stdlib.h>
-#include "./lossless.h"
 #include "../dec/vp8li.h"
-#include "../dsp/yuv.h"
-#include "../dsp/dsp.h"
-#include "../enc/histogram.h"
+#include "../utils/endian_inl.h"
+#include "./lossless.h"
 
 #define MAX_DIFF_COST (1e30f)
 
-// lookup table for small values of log2(int)
-#define APPROX_LOG_MAX  4096
-#define LOG_2_RECIPROCAL 1.44269504088896338700465094007086
-#define LOG_LOOKUP_IDX_MAX 256
-static const float kLog2Table[LOG_LOOKUP_IDX_MAX] = {
-  0.0000000000000000f, 0.0000000000000000f,
-  1.0000000000000000f, 1.5849625007211560f,
-  2.0000000000000000f, 2.3219280948873621f,
-  2.5849625007211560f, 2.8073549220576041f,
-  3.0000000000000000f, 3.1699250014423121f,
-  3.3219280948873621f, 3.4594316186372973f,
-  3.5849625007211560f, 3.7004397181410921f,
-  3.8073549220576041f, 3.9068905956085187f,
-  4.0000000000000000f, 4.0874628412503390f,
-  4.1699250014423121f, 4.2479275134435852f,
-  4.3219280948873626f, 4.3923174227787606f,
-  4.4594316186372973f, 4.5235619560570130f,
-  4.5849625007211560f, 4.6438561897747243f,
-  4.7004397181410917f, 4.7548875021634682f,
-  4.8073549220576037f, 4.8579809951275718f,
-  4.9068905956085187f, 4.9541963103868749f,
-  5.0000000000000000f, 5.0443941193584533f,
-  5.0874628412503390f, 5.1292830169449663f,
-  5.1699250014423121f, 5.2094533656289501f,
-  5.2479275134435852f, 5.2854022188622487f,
-  5.3219280948873626f, 5.3575520046180837f,
-  5.3923174227787606f, 5.4262647547020979f,
-  5.4594316186372973f, 5.4918530963296747f,
-  5.5235619560570130f, 5.5545888516776376f,
-  5.5849625007211560f, 5.6147098441152083f,
-  5.6438561897747243f, 5.6724253419714951f,
-  5.7004397181410917f, 5.7279204545631987f,
-  5.7548875021634682f, 5.7813597135246599f,
-  5.8073549220576037f, 5.8328900141647412f,
-  5.8579809951275718f, 5.8826430493618415f,
-  5.9068905956085187f, 5.9307373375628866f,
-  5.9541963103868749f, 5.9772799234999167f,
-  6.0000000000000000f, 6.0223678130284543f,
-  6.0443941193584533f, 6.0660891904577720f,
-  6.0874628412503390f, 6.1085244567781691f,
-  6.1292830169449663f, 6.1497471195046822f,
-  6.1699250014423121f, 6.1898245588800175f,
-  6.2094533656289501f, 6.2288186904958804f,
-  6.2479275134435852f, 6.2667865406949010f,
-  6.2854022188622487f, 6.3037807481771030f,
-  6.3219280948873626f, 6.3398500028846243f,
-  6.3575520046180837f, 6.3750394313469245f,
-  6.3923174227787606f, 6.4093909361377017f,
-  6.4262647547020979f, 6.4429434958487279f,
-  6.4594316186372973f, 6.4757334309663976f,
-  6.4918530963296747f, 6.5077946401986963f,
-  6.5235619560570130f, 6.5391588111080309f,
-  6.5545888516776376f, 6.5698556083309478f,
-  6.5849625007211560f, 6.5999128421871278f,
-  6.6147098441152083f, 6.6293566200796094f,
-  6.6438561897747243f, 6.6582114827517946f,
-  6.6724253419714951f, 6.6865005271832185f,
-  6.7004397181410917f, 6.7142455176661224f,
-  6.7279204545631987f, 6.7414669864011464f,
-  6.7548875021634682f, 6.7681843247769259f,
-  6.7813597135246599f, 6.7944158663501061f,
-  6.8073549220576037f, 6.8201789624151878f,
-  6.8328900141647412f, 6.8454900509443747f,
-  6.8579809951275718f, 6.8703647195834047f,
-  6.8826430493618415f, 6.8948177633079437f,
-  6.9068905956085187f, 6.9188632372745946f,
-  6.9307373375628866f, 6.9425145053392398f,
-  6.9541963103868749f, 6.9657842846620869f,
-  6.9772799234999167f, 6.9886846867721654f,
-  7.0000000000000000f, 7.0112272554232539f,
-  7.0223678130284543f, 7.0334230015374501f,
-  7.0443941193584533f, 7.0552824355011898f,
-  7.0660891904577720f, 7.0768155970508308f,
-  7.0874628412503390f, 7.0980320829605263f,
-  7.1085244567781691f, 7.1189410727235076f,
-  7.1292830169449663f, 7.1395513523987936f,
-  7.1497471195046822f, 7.1598713367783890f,
-  7.1699250014423121f, 7.1799090900149344f,
-  7.1898245588800175f, 7.1996723448363644f,
-  7.2094533656289501f, 7.2191685204621611f,
-  7.2288186904958804f, 7.2384047393250785f,
-  7.2479275134435852f, 7.2573878426926521f,
-  7.2667865406949010f, 7.2761244052742375f,
-  7.2854022188622487f, 7.2946207488916270f,
-  7.3037807481771030f, 7.3128829552843557f,
-  7.3219280948873626f, 7.3309168781146167f,
-  7.3398500028846243f, 7.3487281542310771f,
-  7.3575520046180837f, 7.3663222142458160f,
-  7.3750394313469245f, 7.3837042924740519f,
-  7.3923174227787606f, 7.4008794362821843f,
-  7.4093909361377017f, 7.4178525148858982f,
-  7.4262647547020979f, 7.4346282276367245f,
-  7.4429434958487279f, 7.4512111118323289f,
-  7.4594316186372973f, 7.4676055500829976f,
-  7.4757334309663976f, 7.4838157772642563f,
-  7.4918530963296747f, 7.4998458870832056f,
-  7.5077946401986963f, 7.5156998382840427f,
-  7.5235619560570130f, 7.5313814605163118f,
-  7.5391588111080309f, 7.5468944598876364f,
-  7.5545888516776376f, 7.5622424242210728f,
-  7.5698556083309478f, 7.5774288280357486f,
-  7.5849625007211560f, 7.5924570372680806f,
-  7.5999128421871278f, 7.6073303137496104f,
-  7.6147098441152083f, 7.6220518194563764f,
-  7.6293566200796094f, 7.6366246205436487f,
-  7.6438561897747243f, 7.6510516911789281f,
-  7.6582114827517946f, 7.6653359171851764f,
-  7.6724253419714951f, 7.6794800995054464f,
-  7.6865005271832185f, 7.6934869574993252f,
-  7.7004397181410917f, 7.7073591320808825f,
-  7.7142455176661224f, 7.7210991887071855f,
-  7.7279204545631987f, 7.7347096202258383f,
-  7.7414669864011464f, 7.7481928495894605f,
-  7.7548875021634682f, 7.7615512324444795f,
-  7.7681843247769259f, 7.7747870596011736f,
-  7.7813597135246599f, 7.7879025593914317f,
-  7.7944158663501061f, 7.8008998999203047f,
-  7.8073549220576037f, 7.8137811912170374f,
-  7.8201789624151878f, 7.8265484872909150f,
-  7.8328900141647412f, 7.8392037880969436f,
-  7.8454900509443747f, 7.8517490414160571f,
-  7.8579809951275718f, 7.8641861446542797f,
-  7.8703647195834047f, 7.8765169465649993f,
-  7.8826430493618415f, 7.8887432488982591f,
-  7.8948177633079437f, 7.9008668079807486f,
-  7.9068905956085187f, 7.9128893362299619f,
-  7.9188632372745946f, 7.9248125036057812f,
-  7.9307373375628866f, 7.9366379390025709f,
-  7.9425145053392398f, 7.9483672315846778f,
-  7.9541963103868749f, 7.9600019320680805f,
-  7.9657842846620869f, 7.9715435539507719f,
-  7.9772799234999167f, 7.9829935746943103f,
-  7.9886846867721654f, 7.9943534368588577f
-};
-
-float VP8LFastLog2(int v) {
-  if (v < LOG_LOOKUP_IDX_MAX) {
-    return kLog2Table[v];
-  } else if (v < APPROX_LOG_MAX) {
-    int log_cnt = 0;
-    while (v >= LOG_LOOKUP_IDX_MAX) {
-      ++log_cnt;
-      v = v >> 1;
-    }
-    return kLog2Table[v] + (float)log_cnt;
-  } else {
-    return (float)(LOG_2_RECIPROCAL * log((double)v));
-  }
-}
-
 //------------------------------------------------------------------------------
 // Image transforms.
 
@@ -186,7 +34,7 @@ static WEBP_INLINE void AddPixelsEq(uint32_t* a, uint32_t b) {
 }
 
 static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) {
-  return (((a0 ^ a1) & 0xfefefefeL) >> 1) + (a0 & a1);
+  return (((a0 ^ a1) & 0xfefefefeu) >> 1) + (a0 & a1);
 }
 
 static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) {
@@ -221,7 +69,7 @@ static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1,
                                          (c1 >> 8) & 0xff,
                                          (c2 >> 8) & 0xff);
   const int b = AddSubtractComponentFull(c0 & 0xff, c1 & 0xff, c2 & 0xff);
-  return (a << 24) | (r << 16) | (g << 8) | b;
+  return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b;
 }
 
 static WEBP_INLINE int AddSubtractComponentHalf(int a, int b) {
@@ -235,22 +83,30 @@ static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
   const int r = AddSubtractComponentHalf((ave >> 16) & 0xff, (c2 >> 16) & 0xff);
   const int g = AddSubtractComponentHalf((ave >> 8) & 0xff, (c2 >> 8) & 0xff);
   const int b = AddSubtractComponentHalf((ave >> 0) & 0xff, (c2 >> 0) & 0xff);
-  return (a << 24) | (r << 16) | (g << 8) | b;
+  return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b;
 }
 
-static WEBP_INLINE int Sub3(int a, int b, int c) {
-  const int pa = b - c;
-  const int pb = a - c;
-  return abs(pa) - abs(pb);
+// gcc-4.9 on ARM generates incorrect code in Select() when Sub3() is inlined.
+#if defined(__arm__) && LOCAL_GCC_VERSION == 0x409
+# define LOCAL_INLINE __attribute__ ((noinline))
+#else
+# define LOCAL_INLINE WEBP_INLINE
+#endif
+
+static LOCAL_INLINE int Sub3(int a, int b, int c) {
+  const int pb = b - c;
+  const int pa = a - c;
+  return abs(pb) - abs(pa);
 }
 
+#undef LOCAL_INLINE
+
 static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
   const int pa_minus_pb =
       Sub3((a >> 24)       , (b >> 24)       , (c >> 24)       ) +
       Sub3((a >> 16) & 0xff, (b >> 16) & 0xff, (c >> 16) & 0xff) +
       Sub3((a >>  8) & 0xff, (b >>  8) & 0xff, (c >>  8) & 0xff) +
       Sub3((a      ) & 0xff, (b      ) & 0xff, (c      ) & 0xff);
-
   return (pa_minus_pb <= 0) ? a : b;
 }
 
@@ -317,208 +173,7 @@ static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
   return pred;
 }
 
-typedef uint32_t (*PredictorFunc)(uint32_t left, const uint32_t* const top);
-static const PredictorFunc kPredictors[16] = {
-  Predictor0, Predictor1, Predictor2, Predictor3,
-  Predictor4, Predictor5, Predictor6, Predictor7,
-  Predictor8, Predictor9, Predictor10, Predictor11,
-  Predictor12, Predictor13,
-  Predictor0, Predictor0    // <- padding security sentinels
-};
-
-// TODO(vikasa): Replace 256 etc with defines.
-static float PredictionCostSpatial(const int* counts,
-                                   int weight_0, double exp_val) {
-  const int significant_symbols = 16;
-  const double exp_decay_factor = 0.6;
-  double bits = weight_0 * counts[0];
-  int i;
-  for (i = 1; i < significant_symbols; ++i) {
-    bits += exp_val * (counts[i] + counts[256 - i]);
-    exp_val *= exp_decay_factor;
-  }
-  return (float)(-0.1 * bits);
-}
-
-// Compute the Shanon's entropy: Sum(p*log2(p))
-static float ShannonEntropy(const int* const array, int n) {
-  int i;
-  float retval = 0.f;
-  int sum = 0;
-  for (i = 0; i < n; ++i) {
-    if (array[i] != 0) {
-      sum += array[i];
-      retval -= VP8LFastSLog2(array[i]);
-    }
-  }
-  retval += VP8LFastSLog2(sum);
-  return retval;
-}
-
-static float PredictionCostSpatialHistogram(int accumulated[4][256],
-                                            int tile[4][256]) {
-  int i;
-  int k;
-  int combo[256];
-  double retval = 0;
-  for (i = 0; i < 4; ++i) {
-    const double exp_val = 0.94;
-    retval += PredictionCostSpatial(&tile[i][0], 1, exp_val);
-    retval += ShannonEntropy(&tile[i][0], 256);
-    for (k = 0; k < 256; ++k) {
-      combo[k] = accumulated[i][k] + tile[i][k];
-    }
-    retval += ShannonEntropy(&combo[0], 256);
-  }
-  return (float)retval;
-}
-
-static int GetBestPredictorForTile(int width, int height,
-                                   int tile_x, int tile_y, int bits,
-                                   int accumulated[4][256],
-                                   const uint32_t* const argb_scratch) {
-  const int kNumPredModes = 14;
-  const int col_start = tile_x << bits;
-  const int row_start = tile_y << bits;
-  const int tile_size = 1 << bits;
-  const int ymax = (tile_size <= height - row_start) ?
-      tile_size : height - row_start;
-  const int xmax = (tile_size <= width - col_start) ?
-      tile_size : width - col_start;
-  int histo[4][256];
-  float best_diff = MAX_DIFF_COST;
-  int best_mode = 0;
-
-  int mode;
-  for (mode = 0; mode < kNumPredModes; ++mode) {
-    const uint32_t* current_row = argb_scratch;
-    const PredictorFunc pred_func = kPredictors[mode];
-    float cur_diff;
-    int y;
-    memset(&histo[0][0], 0, sizeof(histo));
-    for (y = 0; y < ymax; ++y) {
-      int x;
-      const int row = row_start + y;
-      const uint32_t* const upper_row = current_row;
-      current_row = upper_row + width;
-      for (x = 0; x < xmax; ++x) {
-        const int col = col_start + x;
-        uint32_t predict;
-        uint32_t predict_diff;
-        if (row == 0) {
-          predict = (col == 0) ? ARGB_BLACK : current_row[col - 1];  // Left.
-        } else if (col == 0) {
-          predict = upper_row[col];  // Top.
-        } else {
-          predict = pred_func(current_row[col - 1], upper_row + col);
-        }
-        predict_diff = VP8LSubPixels(current_row[col], predict);
-        ++histo[0][predict_diff >> 24];
-        ++histo[1][((predict_diff >> 16) & 0xff)];
-        ++histo[2][((predict_diff >> 8) & 0xff)];
-        ++histo[3][(predict_diff & 0xff)];
-      }
-    }
-    cur_diff = PredictionCostSpatialHistogram(accumulated, histo);
-    if (cur_diff < best_diff) {
-      best_diff = cur_diff;
-      best_mode = mode;
-    }
-  }
-
-  return best_mode;
-}
-
-static void CopyTileWithPrediction(int width, int height,
-                                   int tile_x, int tile_y, int bits, int mode,
-                                   const uint32_t* const argb_scratch,
-                                   uint32_t* const argb) {
-  const int col_start = tile_x << bits;
-  const int row_start = tile_y << bits;
-  const int tile_size = 1 << bits;
-  const int ymax = (tile_size <= height - row_start) ?
-      tile_size : height - row_start;
-  const int xmax = (tile_size <= width - col_start) ?
-      tile_size : width - col_start;
-  const PredictorFunc pred_func = kPredictors[mode];
-  const uint32_t* current_row = argb_scratch;
-
-  int y;
-  for (y = 0; y < ymax; ++y) {
-    int x;
-    const int row = row_start + y;
-    const uint32_t* const upper_row = current_row;
-    current_row = upper_row + width;
-    for (x = 0; x < xmax; ++x) {
-      const int col = col_start + x;
-      const int pix = row * width + col;
-      uint32_t predict;
-      if (row == 0) {
-        predict = (col == 0) ? ARGB_BLACK : current_row[col - 1];  // Left.
-      } else if (col == 0) {
-        predict = upper_row[col];  // Top.
-      } else {
-        predict = pred_func(current_row[col - 1], upper_row + col);
-      }
-      argb[pix] = VP8LSubPixels(current_row[col], predict);
-    }
-  }
-}
-
-void VP8LResidualImage(int width, int height, int bits,
-                       uint32_t* const argb, uint32_t* const argb_scratch,
-                       uint32_t* const image) {
-  const int max_tile_size = 1 << bits;
-  const int tiles_per_row = VP8LSubSampleSize(width, bits);
-  const int tiles_per_col = VP8LSubSampleSize(height, bits);
-  uint32_t* const upper_row = argb_scratch;
-  uint32_t* const current_tile_rows = argb_scratch + width;
-  int tile_y;
-  int histo[4][256];
-  memset(histo, 0, sizeof(histo));
-  for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) {
-    const int tile_y_offset = tile_y * max_tile_size;
-    const int this_tile_height =
-        (tile_y < tiles_per_col - 1) ? max_tile_size : height - tile_y_offset;
-    int tile_x;
-    if (tile_y > 0) {
-      memcpy(upper_row, current_tile_rows + (max_tile_size - 1) * width,
-             width * sizeof(*upper_row));
-    }
-    memcpy(current_tile_rows, &argb[tile_y_offset * width],
-           this_tile_height * width * sizeof(*current_tile_rows));
-    for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) {
-      int pred;
-      int y;
-      const int tile_x_offset = tile_x * max_tile_size;
-      int all_x_max = tile_x_offset + max_tile_size;
-      if (all_x_max > width) {
-        all_x_max = width;
-      }
-      pred = GetBestPredictorForTile(width, height, tile_x, tile_y, bits, histo,
-                                     argb_scratch);
-      image[tile_y * tiles_per_row + tile_x] = 0xff000000u | (pred << 8);
-      CopyTileWithPrediction(width, height, tile_x, tile_y, bits, pred,
-                             argb_scratch, argb);
-      for (y = 0; y < max_tile_size; ++y) {
-        int ix;
-        int all_x;
-        int all_y = tile_y_offset + y;
-        if (all_y >= height) {
-          break;
-        }
-        ix = all_y * width + tile_x_offset;
-        for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
-          const uint32_t a = argb[ix];
-          ++histo[0][a >> 24];
-          ++histo[1][((a >> 16) & 0xff)];
-          ++histo[2][((a >> 8) & 0xff)];
-          ++histo[3][(a & 0xff)];
-        }
-      }
-    }
-  }
-}
+//------------------------------------------------------------------------------
 
 // Inverse prediction.
 static void PredictorInverseTransform(const VP8LTransform* const transform,
@@ -538,29 +193,36 @@ static void PredictorInverseTransform(const VP8LTransform* const transform,
 
   {
     int y = y_start;
-    const int mask = (1 << transform->bits_) - 1;
+    const int tile_width = 1 << transform->bits_;
+    const int mask = tile_width - 1;
+    const int safe_width = width & ~mask;
     const int tiles_per_row = VP8LSubSampleSize(width, transform->bits_);
     const uint32_t* pred_mode_base =
         transform->data_ + (y >> transform->bits_) * tiles_per_row;
 
     while (y < y_end) {
-      int x;
       const uint32_t pred2 = Predictor2(data[-1], data - width);
       const uint32_t* pred_mode_src = pred_mode_base;
-      PredictorFunc pred_func;
-
+      VP8LPredictorFunc pred_func;
+      int x = 1;
+      int t = 1;
       // First pixel follows the T (mode=2) mode.
       AddPixelsEq(data, pred2);
-
       // .. the rest:
-      pred_func = kPredictors[((*pred_mode_src++) >> 8) & 0xf];
-      for (x = 1; x < width; ++x) {
-        uint32_t pred;
-        if ((x & mask) == 0) {    // start of tile. Read predictor function.
-          pred_func = kPredictors[((*pred_mode_src++) >> 8) & 0xf];
+      while (x < safe_width) {
+        pred_func = VP8LPredictors[((*pred_mode_src++) >> 8) & 0xf];
+        for (; t < tile_width; ++t, ++x) {
+          const uint32_t pred = pred_func(data[x - 1], data + x - width);
+          AddPixelsEq(data + x, pred);
+        }
+        t = 0;
+      }
+      if (x < width) {
+        pred_func = VP8LPredictors[((*pred_mode_src++) >> 8) & 0xf];
+        for (; x < width; ++x) {
+          const uint32_t pred = pred_func(data[x - 1], data + x - width);
+          AddPixelsEq(data + x, pred);
         }
-        pred = pred_func(data[x - 1], data + x - width);
-        AddPixelsEq(data + x, pred);
       }
       data += width;
       ++y;
@@ -571,326 +233,47 @@ static void PredictorInverseTransform(const VP8LTransform* const transform,
   }
 }
 
-void VP8LSubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs) {
-  int i;
-  for (i = 0; i < num_pixs; ++i) {
-    const uint32_t argb = argb_data[i];
-    const uint32_t green = (argb >> 8) & 0xff;
-    const uint32_t new_r = (((argb >> 16) & 0xff) - green) & 0xff;
-    const uint32_t new_b = ((argb & 0xff) - green) & 0xff;
-    argb_data[i] = (argb & 0xff00ff00) | (new_r << 16) | new_b;
-  }
-}
-
 // Add green to blue and red channels (i.e. perform the inverse transform of
 // 'subtract green').
-static void AddGreenToBlueAndRed(const VP8LTransform* const transform,
-                                 int y_start, int y_end, uint32_t* data) {
-  const int width = transform->xsize_;
-  const uint32_t* const data_end = data + (y_end - y_start) * width;
-  while (data < data_end) {
-    const uint32_t argb = *data;
-    // "* 0001001u" is equivalent to "(green << 16) + green)"
+void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels) {
+  int i;
+  for (i = 0; i < num_pixels; ++i) {
+    const uint32_t argb = data[i];
     const uint32_t green = ((argb >> 8) & 0xff);
     uint32_t red_blue = (argb & 0x00ff00ffu);
     red_blue += (green << 16) | green;
     red_blue &= 0x00ff00ffu;
-    *data++ = (argb & 0xff00ff00u) | red_blue;
+    data[i] = (argb & 0xff00ff00u) | red_blue;
   }
 }
 
-typedef struct {
-  // Note: the members are uint8_t, so that any negative values are
-  // automatically converted to "mod 256" values.
-  uint8_t green_to_red_;
-  uint8_t green_to_blue_;
-  uint8_t red_to_blue_;
-} Multipliers;
-
-static WEBP_INLINE void MultipliersClear(Multipliers* m) {
-  m->green_to_red_ = 0;
-  m->green_to_blue_ = 0;
-  m->red_to_blue_ = 0;
-}
-
 static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred,
                                                 int8_t color) {
   return (uint32_t)((int)(color_pred) * color) >> 5;
 }
 
 static WEBP_INLINE void ColorCodeToMultipliers(uint32_t color_code,
-                                               Multipliers* const m) {
+                                               VP8LMultipliers* const m) {
   m->green_to_red_  = (color_code >>  0) & 0xff;
   m->green_to_blue_ = (color_code >>  8) & 0xff;
   m->red_to_blue_   = (color_code >> 16) & 0xff;
 }
 
-static WEBP_INLINE uint32_t MultipliersToColorCode(Multipliers* const m) {
-  return 0xff000000u |
-         ((uint32_t)(m->red_to_blue_) << 16) |
-         ((uint32_t)(m->green_to_blue_) << 8) |
-         m->green_to_red_;
-}
-
-static WEBP_INLINE uint32_t TransformColor(const Multipliers* const m,
-                                           uint32_t argb, int inverse) {
-  const uint32_t green = argb >> 8;
-  const uint32_t red = argb >> 16;
-  uint32_t new_red = red;
-  uint32_t new_blue = argb;
-
-  if (inverse) {
+void VP8LTransformColorInverse_C(const VP8LMultipliers* const m, uint32_t* data,
+                                 int num_pixels) {
+  int i;
+  for (i = 0; i < num_pixels; ++i) {
+    const uint32_t argb = data[i];
+    const uint32_t green = argb >> 8;
+    const uint32_t red = argb >> 16;
+    uint32_t new_red = red;
+    uint32_t new_blue = argb;
     new_red += ColorTransformDelta(m->green_to_red_, green);
     new_red &= 0xff;
     new_blue += ColorTransformDelta(m->green_to_blue_, green);
     new_blue += ColorTransformDelta(m->red_to_blue_, new_red);
     new_blue &= 0xff;
-  } else {
-    new_red -= ColorTransformDelta(m->green_to_red_, green);
-    new_red &= 0xff;
-    new_blue -= ColorTransformDelta(m->green_to_blue_, green);
-    new_blue -= ColorTransformDelta(m->red_to_blue_, red);
-    new_blue &= 0xff;
-  }
-  return (argb & 0xff00ff00u) | (new_red << 16) | (new_blue);
-}
-
-static WEBP_INLINE int SkipRepeatedPixels(const uint32_t* const argb,
-                                          int ix, int xsize) {
-  const uint32_t v = argb[ix];
-  if (ix >= xsize + 3) {
-    if (v == argb[ix - xsize] &&
-        argb[ix - 1] == argb[ix - xsize - 1] &&
-        argb[ix - 2] == argb[ix - xsize - 2] &&
-        argb[ix - 3] == argb[ix - xsize - 3]) {
-      return 1;
-    }
-    return v == argb[ix - 3] && v == argb[ix - 2] && v == argb[ix - 1];
-  } else if (ix >= 3) {
-    return v == argb[ix - 3] && v == argb[ix - 2] && v == argb[ix - 1];
-  }
-  return 0;
-}
-
-static float PredictionCostCrossColor(const int accumulated[256],
-                                      const int counts[256]) {
-  // Favor low entropy, locally and globally.
-  int i;
-  int combo[256];
-  for (i = 0; i < 256; ++i) {
-    combo[i] = accumulated[i] + counts[i];
-  }
-  return ShannonEntropy(combo, 256) +
-         ShannonEntropy(counts, 256) +
-         PredictionCostSpatial(counts, 3, 2.4);  // Favor small absolute values.
-}
-
-static Multipliers GetBestColorTransformForTile(
-    int tile_x, int tile_y, int bits,
-    Multipliers prevX,
-    Multipliers prevY,
-    int step, int xsize, int ysize,
-    int* accumulated_red_histo,
-    int* accumulated_blue_histo,
-    const uint32_t* const argb) {
-  float best_diff = MAX_DIFF_COST;
-  float cur_diff;
-  const int halfstep = step / 2;
-  const int max_tile_size = 1 << bits;
-  const int tile_y_offset = tile_y * max_tile_size;
-  const int tile_x_offset = tile_x * max_tile_size;
-  int green_to_red;
-  int green_to_blue;
-  int red_to_blue;
-  int all_x_max = tile_x_offset + max_tile_size;
-  int all_y_max = tile_y_offset + max_tile_size;
-  Multipliers best_tx;
-  MultipliersClear(&best_tx);
-  if (all_x_max > xsize) {
-    all_x_max = xsize;
-  }
-  if (all_y_max > ysize) {
-    all_y_max = ysize;
-  }
-  for (green_to_red = -64; green_to_red <= 64; green_to_red += halfstep) {
-    int histo[256] = { 0 };
-    int all_y;
-    Multipliers tx;
-    MultipliersClear(&tx);
-    tx.green_to_red_ = green_to_red & 0xff;
-
-    for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
-      uint32_t predict;
-      int ix = all_y * xsize + tile_x_offset;
-      int all_x;
-      for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
-        if (SkipRepeatedPixels(argb, ix, xsize)) {
-          continue;
-        }
-        predict = TransformColor(&tx, argb[ix], 0);
-        ++histo[(predict >> 16) & 0xff];  // red.
-      }
-    }
-    cur_diff = PredictionCostCrossColor(&accumulated_red_histo[0], &histo[0]);
-    if (tx.green_to_red_ == prevX.green_to_red_) {
-      cur_diff -= 3;  // favor keeping the areas locally similar
-    }
-    if (tx.green_to_red_ == prevY.green_to_red_) {
-      cur_diff -= 3;  // favor keeping the areas locally similar
-    }
-    if (tx.green_to_red_ == 0) {
-      cur_diff -= 3;
-    }
-    if (cur_diff < best_diff) {
-      best_diff = cur_diff;
-      best_tx = tx;
-    }
-  }
-  best_diff = MAX_DIFF_COST;
-  green_to_red = best_tx.green_to_red_;
-  for (green_to_blue = -32; green_to_blue <= 32; green_to_blue += step) {
-    for (red_to_blue = -32; red_to_blue <= 32; red_to_blue += step) {
-      int all_y;
-      int histo[256] = { 0 };
-      Multipliers tx;
-      tx.green_to_red_ = green_to_red;
-      tx.green_to_blue_ = green_to_blue;
-      tx.red_to_blue_ = red_to_blue;
-      for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
-        uint32_t predict;
-        int all_x;
-        int ix = all_y * xsize + tile_x_offset;
-        for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
-          if (SkipRepeatedPixels(argb, ix, xsize)) {
-            continue;
-          }
-          predict = TransformColor(&tx, argb[ix], 0);
-          ++histo[predict & 0xff];  // blue.
-        }
-      }
-      cur_diff =
-        PredictionCostCrossColor(&accumulated_blue_histo[0], &histo[0]);
-      if (tx.green_to_blue_ == prevX.green_to_blue_) {
-        cur_diff -= 3;  // favor keeping the areas locally similar
-      }
-      if (tx.green_to_blue_ == prevY.green_to_blue_) {
-        cur_diff -= 3;  // favor keeping the areas locally similar
-      }
-      if (tx.red_to_blue_ == prevX.red_to_blue_) {
-        cur_diff -= 3;  // favor keeping the areas locally similar
-      }
-      if (tx.red_to_blue_ == prevY.red_to_blue_) {
-        cur_diff -= 3;  // favor keeping the areas locally similar
-      }
-      if (tx.green_to_blue_ == 0) {
-        cur_diff -= 3;
-      }
-      if (tx.red_to_blue_ == 0) {
-        cur_diff -= 3;
-      }
-      if (cur_diff < best_diff) {
-        best_diff = cur_diff;
-        best_tx = tx;
-      }
-    }
-  }
-  return best_tx;
-}
-
-static void CopyTileWithColorTransform(int xsize, int ysize,
-                                       int tile_x, int tile_y, int bits,
-                                       Multipliers color_transform,
-                                       uint32_t* const argb) {
-  int y;
-  int xscan = 1 << bits;
-  int yscan = 1 << bits;
-  tile_x <<= bits;
-  tile_y <<= bits;
-  if (xscan > xsize - tile_x) {
-    xscan = xsize - tile_x;
-  }
-  if (yscan > ysize - tile_y) {
-    yscan = ysize - tile_y;
-  }
-  yscan += tile_y;
-  for (y = tile_y; y < yscan; ++y) {
-    int ix = y * xsize + tile_x;
-    const int end_ix = ix + xscan;
-    for (; ix < end_ix; ++ix) {
-      argb[ix] = TransformColor(&color_transform, argb[ix], 0);
-    }
-  }
-}
-
-void VP8LColorSpaceTransform(int width, int height, int bits, int step,
-                             uint32_t* const argb, uint32_t* image) {
-  const int max_tile_size = 1 << bits;
-  int tile_xsize = VP8LSubSampleSize(width, bits);
-  int tile_ysize = VP8LSubSampleSize(height, bits);
-  int accumulated_red_histo[256] = { 0 };
-  int accumulated_blue_histo[256] = { 0 };
-  int tile_y;
-  int tile_x;
-  Multipliers prevX;
-  Multipliers prevY;
-  MultipliersClear(&prevY);
-  MultipliersClear(&prevX);
-  for (tile_y = 0; tile_y < tile_ysize; ++tile_y) {
-    for (tile_x = 0; tile_x < tile_xsize; ++tile_x) {
-      Multipliers color_transform;
-      int all_x_max;
-      int y;
-      const int tile_y_offset = tile_y * max_tile_size;
-      const int tile_x_offset = tile_x * max_tile_size;
-      if (tile_y != 0) {
-        ColorCodeToMultipliers(image[tile_y * tile_xsize + tile_x - 1], &prevX);
-        ColorCodeToMultipliers(image[(tile_y - 1) * tile_xsize + tile_x],
-                               &prevY);
-      } else if (tile_x != 0) {
-        ColorCodeToMultipliers(image[tile_y * tile_xsize + tile_x - 1], &prevX);
-      }
-      color_transform =
-          GetBestColorTransformForTile(tile_x, tile_y, bits,
-                                       prevX, prevY,
-                                       step, width, height,
-                                       &accumulated_red_histo[0],
-                                       &accumulated_blue_histo[0],
-                                       argb);
-      image[tile_y * tile_xsize + tile_x] =
-          MultipliersToColorCode(&color_transform);
-      CopyTileWithColorTransform(width, height, tile_x, tile_y, bits,
-                                 color_transform, argb);
-
-      // Gather accumulated histogram data.
-      all_x_max = tile_x_offset + max_tile_size;
-      if (all_x_max > width) {
-        all_x_max = width;
-      }
-      for (y = 0; y < max_tile_size; ++y) {
-        int ix;
-        int all_x;
-        int all_y = tile_y_offset + y;
-        if (all_y >= height) {
-          break;
-        }
-        ix = all_y * width + tile_x_offset;
-        for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
-          if (ix >= 2 &&
-              argb[ix] == argb[ix - 2] &&
-              argb[ix] == argb[ix - 1]) {
-            continue;  // repeated pixels are handled by backward references
-          }
-          if (ix >= width + 2 &&
-              argb[ix - 2] == argb[ix - width - 2] &&
-              argb[ix - 1] == argb[ix - width - 1] &&
-              argb[ix] == argb[ix - width]) {
-            continue;  // repeated pixels are handled by backward references
-          }
-          ++accumulated_red_histo[(argb[ix] >> 16) & 0xff];
-          ++accumulated_blue_histo[argb[ix] & 0xff];
-        }
-      }
-    }
+    data[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue);
   }
 }
 
@@ -898,7 +281,10 @@ void VP8LColorSpaceTransform(int width, int height, int bits, int step,
 static void ColorSpaceInverseTransform(const VP8LTransform* const transform,
                                        int y_start, int y_end, uint32_t* data) {
   const int width = transform->xsize_;
-  const int mask = (1 << transform->bits_) - 1;
+  const int tile_width = 1 << transform->bits_;
+  const int mask = tile_width - 1;
+  const int safe_width = width & ~mask;
+  const int remaining_width = width - safe_width;
   const int tiles_per_row = VP8LSubSampleSize(width, transform->bits_);
   int y = y_start;
   const uint32_t* pred_row =
@@ -906,68 +292,89 @@ static void ColorSpaceInverseTransform(const VP8LTransform* const transform,
 
   while (y < y_end) {
     const uint32_t* pred = pred_row;
-    Multipliers m = { 0, 0, 0 };
-    int x;
-
-    for (x = 0; x < width; ++x) {
-      if ((x & mask) == 0) ColorCodeToMultipliers(*pred++, &m);
-      data[x] = TransformColor(&m, data[x], 1);
+    VP8LMultipliers m = { 0, 0, 0 };
+    const uint32_t* const data_safe_end = data + safe_width;
+    const uint32_t* const data_end = data + width;
+    while (data < data_safe_end) {
+      ColorCodeToMultipliers(*pred++, &m);
+      VP8LTransformColorInverse(&m, data, tile_width);
+      data += tile_width;
+    }
+    if (data < data_end) {  // Left-overs using C-version.
+      ColorCodeToMultipliers(*pred++, &m);
+      VP8LTransformColorInverse(&m, data, remaining_width);
+      data += remaining_width;
     }
-    data += width;
     ++y;
-    if ((y & mask) == 0) pred_row += tiles_per_row;;
+    if ((y & mask) == 0) pred_row += tiles_per_row;
   }
 }
 
 // Separate out pixels packed together using pixel-bundling.
-static void ColorIndexInverseTransform(
-    const VP8LTransform* const transform,
-    int y_start, int y_end, const uint32_t* src, uint32_t* dst) {
-  int y;
-  const int bits_per_pixel = 8 >> transform->bits_;
-  const int width = transform->xsize_;
-  const uint32_t* const color_map = transform->data_;
-  if (bits_per_pixel < 8) {
-    const int pixels_per_byte = 1 << transform->bits_;
-    const int count_mask = pixels_per_byte - 1;
-    const uint32_t bit_mask = (1 << bits_per_pixel) - 1;
-    for (y = y_start; y < y_end; ++y) {
-      uint32_t packed_pixels = 0;
-      int x;
-      for (x = 0; x < width; ++x) {
-        // We need to load fresh 'packed_pixels' once every 'pixels_per_byte'
-        // increments of x. Fortunately, pixels_per_byte is a power of 2, so
-        // can just use a mask for that, instead of decrementing a counter.
-        if ((x & count_mask) == 0) packed_pixels = ((*src++) >> 8) & 0xff;
-        *dst++ = color_map[packed_pixels & bit_mask];
-        packed_pixels >>= bits_per_pixel;
-      }
-    }
-  } else {
-    for (y = y_start; y < y_end; ++y) {
-      int x;
-      for (x = 0; x < width; ++x) {
-        *dst++ = color_map[((*src++) >> 8) & 0xff];
-      }
-    }
-  }
-}
+// We define two methods for ARGB data (uint32_t) and alpha-only data (uint8_t).
+#define COLOR_INDEX_INVERSE(FUNC_NAME, F_NAME, STATIC_DECL, TYPE, BIT_SUFFIX,  \
+                            GET_INDEX, GET_VALUE)                              \
+static void F_NAME(const TYPE* src, const uint32_t* const color_map,           \
+                   TYPE* dst, int y_start, int y_end, int width) {             \
+  int y;                                                                       \
+  for (y = y_start; y < y_end; ++y) {                                          \
+    int x;                                                                     \
+    for (x = 0; x < width; ++x) {                                              \
+      *dst++ = GET_VALUE(color_map[GET_INDEX(*src++)]);                        \
+    }                                                                          \
+  }                                                                            \
+}                                                                              \
+STATIC_DECL void FUNC_NAME(const VP8LTransform* const transform,               \
+                           int y_start, int y_end, const TYPE* src,            \
+                           TYPE* dst) {                                        \
+  int y;                                                                       \
+  const int bits_per_pixel = 8 >> transform->bits_;                            \
+  const int width = transform->xsize_;                                         \
+  const uint32_t* const color_map = transform->data_;                          \
+  if (bits_per_pixel < 8) {                                                    \
+    const int pixels_per_byte = 1 << transform->bits_;                         \
+    const int count_mask = pixels_per_byte - 1;                                \
+    const uint32_t bit_mask = (1 << bits_per_pixel) - 1;                       \
+    for (y = y_start; y < y_end; ++y) {                                        \
+      uint32_t packed_pixels = 0;                                              \
+      int x;                                                                   \
+      for (x = 0; x < width; ++x) {                                            \
+        /* We need to load fresh 'packed_pixels' once every                */  \
+        /* 'pixels_per_byte' increments of x. Fortunately, pixels_per_byte */  \
+        /* is a power of 2, so can just use a mask for that, instead of    */  \
+        /* decrementing a counter.                                         */  \
+        if ((x & count_mask) == 0) packed_pixels = GET_INDEX(*src++);          \
+        *dst++ = GET_VALUE(color_map[packed_pixels & bit_mask]);               \
+        packed_pixels >>= bits_per_pixel;                                      \
+      }                                                                        \
+    }                                                                          \
+  } else {                                                                     \
+    VP8LMapColor##BIT_SUFFIX(src, color_map, dst, y_start, y_end, width);      \
+  }                                                                            \
+}
+
+COLOR_INDEX_INVERSE(ColorIndexInverseTransform, MapARGB, static, uint32_t, 32b,
+                    VP8GetARGBIndex, VP8GetARGBValue)
+COLOR_INDEX_INVERSE(VP8LColorIndexInverseTransformAlpha, MapAlpha, , uint8_t,
+                    8b, VP8GetAlphaIndex, VP8GetAlphaValue)
+
+#undef COLOR_INDEX_INVERSE
 
 void VP8LInverseTransform(const VP8LTransform* const transform,
                           int row_start, int row_end,
                           const uint32_t* const in, uint32_t* const out) {
+  const int width = transform->xsize_;
   assert(row_start < row_end);
   assert(row_end <= transform->ysize_);
   switch (transform->type_) {
     case SUBTRACT_GREEN:
-      AddGreenToBlueAndRed(transform, row_start, row_end, out);
+      VP8LAddGreenToBlueAndRed(out, (row_end - row_start) * width);
       break;
     case PREDICTOR_TRANSFORM:
       PredictorInverseTransform(transform, row_start, row_end, out);
       if (row_end != transform->ysize_) {
         // The last predicted row in this iteration will be the top-pred row
         // for the first row in next iteration.
-        const int width = transform->xsize_;
         memcpy(out - width, out + (row_end - row_start - 1) * width,
                width * sizeof(*out));
       }
@@ -982,7 +389,7 @@ void VP8LInverseTransform(const VP8LTransform* const transform,
         // Also, note that this is the only transform that applies on
         // the effective width of VP8LSubSampleSize(xsize_, bits_). All other
         // transforms work on effective width of xsize_.
-        const int out_stride = (row_end - row_start) * transform->xsize_;
+        const int out_stride = (row_end - row_start) * width;
         const int in_stride = (row_end - row_start) *
             VP8LSubSampleSize(transform->xsize_, transform->bits_);
         uint32_t* const src = out + out_stride - in_stride;
@@ -1006,8 +413,8 @@ static int is_big_endian(void) {
   return (tmp.b[0] != 1);
 }
 
-static void ConvertBGRAToRGB(const uint32_t* src,
-                             int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToRGB_C(const uint32_t* src,
+                            int num_pixels, uint8_t* dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
@@ -1017,8 +424,8 @@ static void ConvertBGRAToRGB(const uint32_t* src,
   }
 }
 
-static void ConvertBGRAToRGBA(const uint32_t* src,
-                              int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToRGBA_C(const uint32_t* src,
+                             int num_pixels, uint8_t* dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
@@ -1029,28 +436,42 @@ static void ConvertBGRAToRGBA(const uint32_t* src,
   }
 }
 
-static void ConvertBGRAToRGBA4444(const uint32_t* src,
-                                  int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src,
+                                 int num_pixels, uint8_t* dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
-    *dst++ = ((argb >> 16) & 0xf0) | ((argb >> 12) & 0xf);
-    *dst++ = ((argb >>  0) & 0xf0) | ((argb >> 28) & 0xf);
+    const uint8_t rg = ((argb >> 16) & 0xf0) | ((argb >> 12) & 0xf);
+    const uint8_t ba = ((argb >>  0) & 0xf0) | ((argb >> 28) & 0xf);
+#ifdef WEBP_SWAP_16BIT_CSP
+    *dst++ = ba;
+    *dst++ = rg;
+#else
+    *dst++ = rg;
+    *dst++ = ba;
+#endif
   }
 }
 
-static void ConvertBGRAToRGB565(const uint32_t* src,
-                                int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToRGB565_C(const uint32_t* src,
+                               int num_pixels, uint8_t* dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
-    *dst++ = ((argb >> 16) & 0xf8) | ((argb >> 13) & 0x7);
-    *dst++ = ((argb >>  5) & 0xe0) | ((argb >>  3) & 0x1f);
+    const uint8_t rg = ((argb >> 16) & 0xf8) | ((argb >> 13) & 0x7);
+    const uint8_t gb = ((argb >>  5) & 0xe0) | ((argb >>  3) & 0x1f);
+#ifdef WEBP_SWAP_16BIT_CSP
+    *dst++ = gb;
+    *dst++ = rg;
+#else
+    *dst++ = rg;
+    *dst++ = gb;
+#endif
   }
 }
 
-static void ConvertBGRAToBGR(const uint32_t* src,
-                             int num_pixels, uint8_t* dst) {
+void VP8LConvertBGRAToBGR_C(const uint32_t* src,
+                            int num_pixels, uint8_t* dst) {
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
@@ -1065,21 +486,24 @@ static void CopyOrSwap(const uint32_t* src, int num_pixels, uint8_t* dst,
   if (is_big_endian() == swap_on_big_endian) {
     const uint32_t* const src_end = src + num_pixels;
     while (src < src_end) {
-      uint32_t argb = *src++;
-#if !defined(__BIG_ENDIAN__) && (defined(__i386__) || defined(__x86_64__))
-      __asm__ volatile("bswap %0" : "=r"(argb) : "0"(argb));
-      *(uint32_t*)dst = argb;
-      dst += sizeof(argb);
-#elif !defined(__BIG_ENDIAN__) && defined(_MSC_VER)
-      argb = _byteswap_ulong(argb);
-      *(uint32_t*)dst = argb;
-      dst += sizeof(argb);
-#else
-      *dst++ = (argb >> 24) & 0xff;
-      *dst++ = (argb >> 16) & 0xff;
-      *dst++ = (argb >>  8) & 0xff;
-      *dst++ = (argb >>  0) & 0xff;
+      const uint32_t argb = *src++;
+
+#if !defined(WORDS_BIGENDIAN)
+#if !defined(WEBP_REFERENCE_IMPLEMENTATION)
+      *(uint32_t*)dst = BSwap32(argb);
+#else  // WEBP_REFERENCE_IMPLEMENTATION
+      dst[0] = (argb >> 24) & 0xff;
+      dst[1] = (argb >> 16) & 0xff;
+      dst[2] = (argb >>  8) & 0xff;
+      dst[3] = (argb >>  0) & 0xff;
 #endif
+#else  // WORDS_BIGENDIAN
+      dst[0] = (argb >>  0) & 0xff;
+      dst[1] = (argb >>  8) & 0xff;
+      dst[2] = (argb >> 16) & 0xff;
+      dst[3] = (argb >> 24) & 0xff;
+#endif
+      dst += sizeof(argb);
     }
   } else {
     memcpy(dst, src, num_pixels * sizeof(*src));
@@ -1090,17 +514,17 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
                          WEBP_CSP_MODE out_colorspace, uint8_t* const rgba) {
   switch (out_colorspace) {
     case MODE_RGB:
-      ConvertBGRAToRGB(in_data, num_pixels, rgba);
+      VP8LConvertBGRAToRGB(in_data, num_pixels, rgba);
       break;
     case MODE_RGBA:
-      ConvertBGRAToRGBA(in_data, num_pixels, rgba);
+      VP8LConvertBGRAToRGBA(in_data, num_pixels, rgba);
       break;
     case MODE_rgbA:
-      ConvertBGRAToRGBA(in_data, num_pixels, rgba);
+      VP8LConvertBGRAToRGBA(in_data, num_pixels, rgba);
       WebPApplyAlphaMultiply(rgba, 0, num_pixels, 1, 0);
       break;
     case MODE_BGR:
-      ConvertBGRAToBGR(in_data, num_pixels, rgba);
+      VP8LConvertBGRAToBGR(in_data, num_pixels, rgba);
       break;
     case MODE_BGRA:
       CopyOrSwap(in_data, num_pixels, rgba, 1);
@@ -1117,14 +541,14 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
       WebPApplyAlphaMultiply(rgba, 1, num_pixels, 1, 0);
       break;
     case MODE_RGBA_4444:
-      ConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
+      VP8LConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
       break;
     case MODE_rgbA_4444:
-      ConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
+      VP8LConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
       WebPApplyAlphaMultiply4444(rgba, num_pixels, 1, 0);
       break;
     case MODE_RGB_565:
-      ConvertBGRAToRGB565(in_data, num_pixels, rgba);
+      VP8LConvertBGRAToRGB565(in_data, num_pixels, rgba);
       break;
     default:
       assert(0);          // Code flow should not reach here.
@@ -1133,6 +557,79 @@ void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
+VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed;
+VP8LPredictorFunc VP8LPredictors[16];
+
+VP8LTransformColorFunc VP8LTransformColorInverse;
+
+VP8LConvertFunc VP8LConvertBGRAToRGB;
+VP8LConvertFunc VP8LConvertBGRAToRGBA;
+VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
+VP8LConvertFunc VP8LConvertBGRAToRGB565;
+VP8LConvertFunc VP8LConvertBGRAToBGR;
+
+VP8LMapARGBFunc VP8LMapColor32b;
+VP8LMapAlphaFunc VP8LMapColor8b;
+
+extern void VP8LDspInitSSE2(void);
+extern void VP8LDspInitNEON(void);
+extern void VP8LDspInitMIPSdspR2(void);
+
+static volatile VP8CPUInfo lossless_last_cpuinfo_used =
+    (VP8CPUInfo)&lossless_last_cpuinfo_used;
+
+WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInit(void) {
+  if (lossless_last_cpuinfo_used == VP8GetCPUInfo) return;
+
+  VP8LPredictors[0] = Predictor0;
+  VP8LPredictors[1] = Predictor1;
+  VP8LPredictors[2] = Predictor2;
+  VP8LPredictors[3] = Predictor3;
+  VP8LPredictors[4] = Predictor4;
+  VP8LPredictors[5] = Predictor5;
+  VP8LPredictors[6] = Predictor6;
+  VP8LPredictors[7] = Predictor7;
+  VP8LPredictors[8] = Predictor8;
+  VP8LPredictors[9] = Predictor9;
+  VP8LPredictors[10] = Predictor10;
+  VP8LPredictors[11] = Predictor11;
+  VP8LPredictors[12] = Predictor12;
+  VP8LPredictors[13] = Predictor13;
+  VP8LPredictors[14] = Predictor0;     // <- padding security sentinels
+  VP8LPredictors[15] = Predictor0;
+
+  VP8LAddGreenToBlueAndRed = VP8LAddGreenToBlueAndRed_C;
+
+  VP8LTransformColorInverse = VP8LTransformColorInverse_C;
+
+  VP8LConvertBGRAToRGB = VP8LConvertBGRAToRGB_C;
+  VP8LConvertBGRAToRGBA = VP8LConvertBGRAToRGBA_C;
+  VP8LConvertBGRAToRGBA4444 = VP8LConvertBGRAToRGBA4444_C;
+  VP8LConvertBGRAToRGB565 = VP8LConvertBGRAToRGB565_C;
+  VP8LConvertBGRAToBGR = VP8LConvertBGRAToBGR_C;
+
+  VP8LMapColor32b = MapARGB;
+  VP8LMapColor8b = MapAlpha;
+
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      VP8LDspInitSSE2();
+    }
+#endif
+#if defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      VP8LDspInitNEON();
+    }
+#endif
+#if defined(WEBP_USE_MIPS_DSP_R2)
+    if (VP8GetCPUInfo(kMIPSdspR2)) {
+      VP8LDspInitMIPSdspR2();
+    }
 #endif
+  }
+  lossless_last_cpuinfo_used = VP8GetCPUInfo;
+}
+
+//------------------------------------------------------------------------------
diff --git a/drivers/webp/dsp/lossless.h b/drivers/webp/dsp/lossless.h
index 7c7d5555ed..ee6771333f 100644
--- a/drivers/webp/dsp/lossless.h
+++ b/drivers/webp/dsp/lossless.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
 // -----------------------------------------------------------------------------
 //
 // Image transforms and color space conversion methods for lossless decoder.
@@ -13,15 +15,42 @@
 #ifndef WEBP_DSP_LOSSLESS_H_
 #define WEBP_DSP_LOSSLESS_H_
 
-#include "../types.h"
-#include "../decode.h"
+#include "../webp/types.h"
+#include "../webp/decode.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#include "../enc/histogram.h"
+#include "../utils/utils.h"
+
+#ifdef __cplusplus
 extern "C" {
 #endif
 
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+#include "../enc/delta_palettization.h"
+#endif  // WEBP_EXPERIMENTAL_FEATURES
+
+// Not a trivial literal symbol.
+#define VP8L_NON_TRIVIAL_SYM (0xffffffff)
+
 //------------------------------------------------------------------------------
-// Image transforms.
+// Decoding
+
+typedef uint32_t (*VP8LPredictorFunc)(uint32_t left, const uint32_t* const top);
+extern VP8LPredictorFunc VP8LPredictors[16];
+
+typedef void (*VP8LProcessBlueAndRedFunc)(uint32_t* argb_data, int num_pixels);
+extern VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed;
+
+typedef struct {
+  // Note: the members are uint8_t, so that any negative values are
+  // automatically converted to "mod 256" values.
+  uint8_t green_to_red_;
+  uint8_t green_to_blue_;
+  uint8_t red_to_blue_;
+} VP8LMultipliers;
+typedef void (*VP8LTransformColorFunc)(const VP8LMultipliers* const m,
+                                       uint32_t* argb_data, int num_pixels);
+extern VP8LTransformColorFunc VP8LTransformColorInverse;
 
 struct VP8LTransform;  // Defined in dec/vp8li.h.
 
@@ -33,23 +62,110 @@ void VP8LInverseTransform(const struct VP8LTransform* const transform,
                           int row_start, int row_end,
                           const uint32_t* const in, uint32_t* const out);
 
-// Subtracts green from blue and red channels.
-void VP8LSubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs);
-
-void VP8LResidualImage(int width, int height, int bits,
-                       uint32_t* const argb, uint32_t* const argb_scratch,
-                       uint32_t* const image);
-
-void VP8LColorSpaceTransform(int width, int height, int bits, int step,
-                             uint32_t* const argb, uint32_t* image);
-
-//------------------------------------------------------------------------------
 // Color space conversion.
+typedef void (*VP8LConvertFunc)(const uint32_t* src, int num_pixels,
+                                uint8_t* dst);
+extern VP8LConvertFunc VP8LConvertBGRAToRGB;
+extern VP8LConvertFunc VP8LConvertBGRAToRGBA;
+extern VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
+extern VP8LConvertFunc VP8LConvertBGRAToRGB565;
+extern VP8LConvertFunc VP8LConvertBGRAToBGR;
 
 // Converts from BGRA to other color spaces.
 void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
                          WEBP_CSP_MODE out_colorspace, uint8_t* const rgba);
 
+// color mapping related functions.
+static WEBP_INLINE uint32_t VP8GetARGBIndex(uint32_t idx) {
+  return (idx >> 8) & 0xff;
+}
+
+static WEBP_INLINE uint8_t VP8GetAlphaIndex(uint8_t idx) {
+  return idx;
+}
+
+static WEBP_INLINE uint32_t VP8GetARGBValue(uint32_t val) {
+  return val;
+}
+
+static WEBP_INLINE uint8_t VP8GetAlphaValue(uint32_t val) {
+  return (val >> 8) & 0xff;
+}
+
+typedef void (*VP8LMapARGBFunc)(const uint32_t* src,
+                                const uint32_t* const color_map,
+                                uint32_t* dst, int y_start,
+                                int y_end, int width);
+typedef void (*VP8LMapAlphaFunc)(const uint8_t* src,
+                                 const uint32_t* const color_map,
+                                 uint8_t* dst, int y_start,
+                                 int y_end, int width);
+
+extern VP8LMapARGBFunc VP8LMapColor32b;
+extern VP8LMapAlphaFunc VP8LMapColor8b;
+
+// Similar to the static method ColorIndexInverseTransform() that is part of
+// lossless.c, but used only for alpha decoding. It takes uint8_t (rather than
+// uint32_t) arguments for 'src' and 'dst'.
+void VP8LColorIndexInverseTransformAlpha(
+    const struct VP8LTransform* const transform, int y_start, int y_end,
+    const uint8_t* src, uint8_t* dst);
+
+// Expose some C-only fallback functions
+void VP8LTransformColorInverse_C(const VP8LMultipliers* const m,
+                                 uint32_t* data, int num_pixels);
+
+void VP8LConvertBGRAToRGB_C(const uint32_t* src, int num_pixels, uint8_t* dst);
+void VP8LConvertBGRAToRGBA_C(const uint32_t* src, int num_pixels, uint8_t* dst);
+void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src,
+                                 int num_pixels, uint8_t* dst);
+void VP8LConvertBGRAToRGB565_C(const uint32_t* src,
+                               int num_pixels, uint8_t* dst);
+void VP8LConvertBGRAToBGR_C(const uint32_t* src, int num_pixels, uint8_t* dst);
+void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels);
+
+// Must be called before calling any of the above methods.
+void VP8LDspInit(void);
+
+//------------------------------------------------------------------------------
+// Encoding
+
+extern VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
+extern VP8LTransformColorFunc VP8LTransformColor;
+typedef void (*VP8LCollectColorBlueTransformsFunc)(
+    const uint32_t* argb, int stride,
+    int tile_width, int tile_height,
+    int green_to_blue, int red_to_blue, int histo[]);
+extern VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms;
+
+typedef void (*VP8LCollectColorRedTransformsFunc)(
+    const uint32_t* argb, int stride,
+    int tile_width, int tile_height,
+    int green_to_red, int histo[]);
+extern VP8LCollectColorRedTransformsFunc VP8LCollectColorRedTransforms;
+
+// Expose some C-only fallback functions
+void VP8LTransformColor_C(const VP8LMultipliers* const m,
+                          uint32_t* data, int num_pixels);
+void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels);
+void VP8LCollectColorRedTransforms_C(const uint32_t* argb, int stride,
+                                     int tile_width, int tile_height,
+                                     int green_to_red, int histo[]);
+void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride,
+                                      int tile_width, int tile_height,
+                                      int green_to_blue, int red_to_blue,
+                                      int histo[]);
+
+//------------------------------------------------------------------------------
+// Image transforms.
+
+void VP8LResidualImage(int width, int height, int bits, int low_effort,
+                       uint32_t* const argb, uint32_t* const argb_scratch,
+                       uint32_t* const image);
+
+void VP8LColorSpaceTransform(int width, int height, int bits, int quality,
+                             uint32_t* const argb, uint32_t* image);
+
 //------------------------------------------------------------------------------
 // Misc methods.
 
@@ -59,10 +175,136 @@ static WEBP_INLINE uint32_t VP8LSubSampleSize(uint32_t size,
   return (size + (1 << sampling_bits) - 1) >> sampling_bits;
 }
 
-// Faster logarithm for integers, with the property of log2(0) == 0.
-float VP8LFastLog2(int v);
+// -----------------------------------------------------------------------------
+// Faster logarithm for integers. Small values use a look-up table.
+#define LOG_LOOKUP_IDX_MAX 256
+extern const float kLog2Table[LOG_LOOKUP_IDX_MAX];
+extern const float kSLog2Table[LOG_LOOKUP_IDX_MAX];
+typedef float (*VP8LFastLog2SlowFunc)(uint32_t v);
+
+extern VP8LFastLog2SlowFunc VP8LFastLog2Slow;
+extern VP8LFastLog2SlowFunc VP8LFastSLog2Slow;
+
+static WEBP_INLINE float VP8LFastLog2(uint32_t v) {
+  return (v < LOG_LOOKUP_IDX_MAX) ? kLog2Table[v] : VP8LFastLog2Slow(v);
+}
 // Fast calculation of v * log2(v) for integer input.
-static WEBP_INLINE float VP8LFastSLog2(int v) { return VP8LFastLog2(v) * v; }
+static WEBP_INLINE float VP8LFastSLog2(uint32_t v) {
+  return (v < LOG_LOOKUP_IDX_MAX) ? kSLog2Table[v] : VP8LFastSLog2Slow(v);
+}
+
+// -----------------------------------------------------------------------------
+// Huffman-cost related functions.
+
+typedef double (*VP8LCostFunc)(const uint32_t* population, int length);
+typedef double (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y,
+                                       int length);
+
+extern VP8LCostFunc VP8LExtraCost;
+extern VP8LCostCombinedFunc VP8LExtraCostCombined;
+
+typedef struct {        // small struct to hold counters
+  int counts[2];        // index: 0=zero steak, 1=non-zero streak
+  int streaks[2][2];    // [zero/non-zero][streak<3 / streak>=3]
+} VP8LStreaks;
+
+typedef VP8LStreaks (*VP8LCostCountFunc)(const uint32_t* population,
+                                         int length);
+typedef VP8LStreaks (*VP8LCostCombinedCountFunc)(const uint32_t* X,
+                                                 const uint32_t* Y, int length);
+
+extern VP8LCostCountFunc VP8LHuffmanCostCount;
+extern VP8LCostCombinedCountFunc VP8LHuffmanCostCombinedCount;
+
+// Get the symbol entropy for the distribution 'population'.
+// Set 'trivial_sym', if there's only one symbol present in the distribution.
+double VP8LPopulationCost(const uint32_t* const population, int length,
+                          uint32_t* const trivial_sym);
+
+// Get the combined symbol entropy for the distributions 'X' and 'Y'.
+double VP8LGetCombinedEntropy(const uint32_t* const X,
+                              const uint32_t* const Y, int length);
+
+double VP8LBitsEntropy(const uint32_t* const array, int n,
+                       uint32_t* const trivial_symbol);
+
+// 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);
+
+typedef void (*VP8LHistogramAddFunc)(const VP8LHistogram* const a,
+                                     const VP8LHistogram* const b,
+                                     VP8LHistogram* const out);
+extern VP8LHistogramAddFunc VP8LHistogramAdd;
+
+// -----------------------------------------------------------------------------
+// PrefixEncode()
+
+static WEBP_INLINE int VP8LBitsLog2Ceiling(uint32_t n) {
+  const int log_floor = BitsLog2Floor(n);
+  if (n == (n & ~(n - 1)))  // zero or a power of two.
+    return log_floor;
+  else
+    return log_floor + 1;
+}
+
+// Splitting of distance and length codes into prefixes and
+// extra bits. The prefixes are encoded with an entropy code
+// while the extra bits are stored just as normal bits.
+static WEBP_INLINE void VP8LPrefixEncodeBitsNoLUT(int distance, int* const code,
+                                                  int* const extra_bits) {
+  const int highest_bit = BitsLog2Floor(--distance);
+  const int second_highest_bit = (distance >> (highest_bit - 1)) & 1;
+  *extra_bits = highest_bit - 1;
+  *code = 2 * highest_bit + second_highest_bit;
+}
+
+static WEBP_INLINE void VP8LPrefixEncodeNoLUT(int distance, int* const code,
+                                              int* const extra_bits,
+                                              int* const extra_bits_value) {
+  const int highest_bit = BitsLog2Floor(--distance);
+  const int second_highest_bit = (distance >> (highest_bit - 1)) & 1;
+  *extra_bits = highest_bit - 1;
+  *extra_bits_value = distance & ((1 << *extra_bits) - 1);
+  *code = 2 * highest_bit + second_highest_bit;
+}
+
+#define PREFIX_LOOKUP_IDX_MAX   512
+typedef struct {
+  int8_t code_;
+  int8_t extra_bits_;
+} VP8LPrefixCode;
+
+// These tables are derived using VP8LPrefixEncodeNoLUT.
+extern const VP8LPrefixCode kPrefixEncodeCode[PREFIX_LOOKUP_IDX_MAX];
+extern const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX];
+static WEBP_INLINE void VP8LPrefixEncodeBits(int distance, int* const code,
+                                             int* const extra_bits) {
+  if (distance < PREFIX_LOOKUP_IDX_MAX) {
+    const VP8LPrefixCode prefix_code = kPrefixEncodeCode[distance];
+    *code = prefix_code.code_;
+    *extra_bits = prefix_code.extra_bits_;
+  } else {
+    VP8LPrefixEncodeBitsNoLUT(distance, code, extra_bits);
+  }
+}
+
+static WEBP_INLINE void VP8LPrefixEncode(int distance, int* const code,
+                                         int* const extra_bits,
+                                         int* const extra_bits_value) {
+  if (distance < PREFIX_LOOKUP_IDX_MAX) {
+    const VP8LPrefixCode prefix_code = kPrefixEncodeCode[distance];
+    *code = prefix_code.code_;
+    *extra_bits = prefix_code.extra_bits_;
+    *extra_bits_value = kPrefixEncodeExtraBitsValue[distance];
+  } else {
+    VP8LPrefixEncodeNoLUT(distance, code, extra_bits, extra_bits_value);
+  }
+}
 
 // In-place difference of each component with mod 256.
 static WEBP_INLINE uint32_t VP8LSubPixels(uint32_t a, uint32_t b) {
@@ -73,9 +315,15 @@ static WEBP_INLINE uint32_t VP8LSubPixels(uint32_t a, uint32_t b) {
   return (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu);
 }
 
+void VP8LBundleColorMap(const uint8_t* const row, int width,
+                        int xbits, uint32_t* const dst);
+
+// Must be called before calling any of the above methods.
+void VP8LEncDspInit(void);
+
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/dsp/upsampling.c b/drivers/webp/dsp/upsampling.c
index 4855eb1432..651274fcee 100644
--- a/drivers/webp/dsp/upsampling.c
+++ b/drivers/webp/dsp/upsampling.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // YUV to RGB upsampling functions.
@@ -12,9 +14,7 @@
 #include "./dsp.h"
 #include "./yuv.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include <assert.h>
 
 //------------------------------------------------------------------------------
 // Fancy upsampler
@@ -32,7 +32,7 @@ WebPUpsampleLinePairFunc WebPUpsamplers[MODE_LAST];
 //  ([3*a +   b + 9*c + 3*d      a + 3*b + 3*c + 9*d]   [8 8]) / 16
 
 // We process u and v together stashed into 32bit (16bit each).
-#define LOAD_UV(u,v) ((u) | ((v) << 16))
+#define LOAD_UV(u, v) ((u) | ((v) << 16))
 
 #define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                                  \
 static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
@@ -43,11 +43,12 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
   const int last_pixel_pair = (len - 1) >> 1;                                  \
   uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]);   /* top-left sample */        \
   uint32_t l_uv  = LOAD_UV(cur_u[0], cur_v[0]);   /* left-sample */            \
-  if (top_y) {                                                                 \
+  assert(top_y != NULL);                                                       \
+  {                                                                            \
     const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2;                \
     FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst);                          \
   }                                                                            \
-  if (bottom_y) {                                                              \
+  if (bottom_y != NULL) {                                                      \
     const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2;                \
     FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst);                    \
   }                                                                            \
@@ -58,7 +59,7 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
     const uint32_t avg = tl_uv + t_uv + l_uv + uv + 0x00080008u;               \
     const uint32_t diag_12 = (avg + 2 * (t_uv + l_uv)) >> 3;                   \
     const uint32_t diag_03 = (avg + 2 * (tl_uv + uv)) >> 3;                    \
-    if (top_y) {                                                               \
+    {                                                                          \
       const uint32_t uv0 = (diag_12 + tl_uv) >> 1;                             \
       const uint32_t uv1 = (diag_03 + t_uv) >> 1;                              \
       FUNC(top_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16),                          \
@@ -66,7 +67,7 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
       FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16),                          \
            top_dst + (2 * x - 0) * XSTEP);                                     \
     }                                                                          \
-    if (bottom_y) {                                                            \
+    if (bottom_y != NULL) {                                                    \
       const uint32_t uv0 = (diag_03 + l_uv) >> 1;                              \
       const uint32_t uv1 = (diag_12 + uv) >> 1;                                \
       FUNC(bottom_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16),                       \
@@ -78,12 +79,12 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
     l_uv = uv;                                                                 \
   }                                                                            \
   if (!(len & 1)) {                                                            \
-    if (top_y) {                                                               \
+    {                                                                          \
       const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2;              \
       FUNC(top_y[len - 1], uv0 & 0xff, (uv0 >> 16),                            \
            top_dst + (len - 1) * XSTEP);                                       \
     }                                                                          \
-    if (bottom_y) {                                                            \
+    if (bottom_y != NULL) {                                                    \
       const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2;              \
       FUNC(bottom_y[len - 1], uv0 & 0xff, (uv0 >> 16),                         \
            bottom_dst + (len - 1) * XSTEP);                                    \
@@ -106,57 +107,6 @@ UPSAMPLE_FUNC(UpsampleRgb565LinePair,  VP8YuvToRgb565,  2)
 #endif  // FANCY_UPSAMPLING
 
 //------------------------------------------------------------------------------
-// simple point-sampling
-
-#define SAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                                    \
-static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
-                      const uint8_t* u, const uint8_t* v,                      \
-                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
-  int i;                                                                       \
-  for (i = 0; i < len - 1; i += 2) {                                           \
-    FUNC(top_y[0], u[0], v[0], top_dst);                                       \
-    FUNC(top_y[1], u[0], v[0], top_dst + XSTEP);                               \
-    FUNC(bottom_y[0], u[0], v[0], bottom_dst);                                 \
-    FUNC(bottom_y[1], u[0], v[0], bottom_dst + XSTEP);                         \
-    top_y += 2;                                                                \
-    bottom_y += 2;                                                             \
-    u++;                                                                       \
-    v++;                                                                       \
-    top_dst += 2 * XSTEP;                                                      \
-    bottom_dst += 2 * XSTEP;                                                   \
-  }                                                                            \
-  if (i == len - 1) {    /* last one */                                        \
-    FUNC(top_y[0], u[0], v[0], top_dst);                                       \
-    FUNC(bottom_y[0], u[0], v[0], bottom_dst);                                 \
-  }                                                                            \
-}
-
-// All variants implemented.
-SAMPLE_FUNC(SampleRgbLinePair,      VP8YuvToRgb,  3)
-SAMPLE_FUNC(SampleBgrLinePair,      VP8YuvToBgr,  3)
-SAMPLE_FUNC(SampleRgbaLinePair,     VP8YuvToRgba, 4)
-SAMPLE_FUNC(SampleBgraLinePair,     VP8YuvToBgra, 4)
-SAMPLE_FUNC(SampleArgbLinePair,     VP8YuvToArgb, 4)
-SAMPLE_FUNC(SampleRgba4444LinePair, VP8YuvToRgba4444, 2)
-SAMPLE_FUNC(SampleRgb565LinePair,   VP8YuvToRgb565, 2)
-
-#undef SAMPLE_FUNC
-
-const WebPSampleLinePairFunc WebPSamplers[MODE_LAST] = {
-  SampleRgbLinePair,       // MODE_RGB
-  SampleRgbaLinePair,      // MODE_RGBA
-  SampleBgrLinePair,       // MODE_BGR
-  SampleBgraLinePair,      // MODE_BGRA
-  SampleArgbLinePair,      // MODE_ARGB
-  SampleRgba4444LinePair,  // MODE_RGBA_4444
-  SampleRgb565LinePair,    // MODE_RGB_565
-  SampleRgbaLinePair,      // MODE_rgbA
-  SampleBgraLinePair,      // MODE_bgrA
-  SampleArgbLinePair,      // MODE_Argb
-  SampleRgba4444LinePair   // MODE_rgbA_4444
-};
-
-//------------------------------------------------------------------------------
 
 #if !defined(FANCY_UPSAMPLING)
 #define DUAL_SAMPLE_FUNC(FUNC_NAME, FUNC)                                      \
@@ -166,7 +116,8 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y,              \
                       uint8_t* top_dst, uint8_t* bot_dst, int len) {           \
   const int half_len = len >> 1;                                               \
   int x;                                                                       \
-  if (top_dst != NULL) {                                                       \
+  assert(top_dst != NULL);                                                     \
+  {                                                                            \
     for (x = 0; x < half_len; ++x) {                                           \
       FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x + 0);         \
       FUNC(top_y[2 * x + 1], top_u[x], top_v[x], top_dst + 8 * x + 4);         \
@@ -202,116 +153,75 @@ WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last) {
 // YUV444 converter
 
 #define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP)                                    \
-static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
-                      uint8_t* dst, int len) {                                 \
+extern void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
+                      uint8_t* dst, int len);                                  \
+void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,           \
+               uint8_t* dst, int len) {                                        \
   int i;                                                                       \
   for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]);           \
 }
 
-YUV444_FUNC(Yuv444ToRgb,      VP8YuvToRgb,  3)
-YUV444_FUNC(Yuv444ToBgr,      VP8YuvToBgr,  3)
-YUV444_FUNC(Yuv444ToRgba,     VP8YuvToRgba, 4)
-YUV444_FUNC(Yuv444ToBgra,     VP8YuvToBgra, 4)
-YUV444_FUNC(Yuv444ToArgb,     VP8YuvToArgb, 4)
-YUV444_FUNC(Yuv444ToRgba4444, VP8YuvToRgba4444, 2)
-YUV444_FUNC(Yuv444ToRgb565,   VP8YuvToRgb565, 2)
+YUV444_FUNC(WebPYuv444ToRgbC,      VP8YuvToRgb,  3)
+YUV444_FUNC(WebPYuv444ToBgrC,      VP8YuvToBgr,  3)
+YUV444_FUNC(WebPYuv444ToRgbaC,     VP8YuvToRgba, 4)
+YUV444_FUNC(WebPYuv444ToBgraC,     VP8YuvToBgra, 4)
+YUV444_FUNC(WebPYuv444ToArgbC,     VP8YuvToArgb, 4)
+YUV444_FUNC(WebPYuv444ToRgba4444C, VP8YuvToRgba4444, 2)
+YUV444_FUNC(WebPYuv444ToRgb565C,   VP8YuvToRgb565, 2)
 
 #undef YUV444_FUNC
 
-const WebPYUV444Converter WebPYUV444Converters[MODE_LAST] = {
-  Yuv444ToRgb,       // MODE_RGB
-  Yuv444ToRgba,      // MODE_RGBA
-  Yuv444ToBgr,       // MODE_BGR
-  Yuv444ToBgra,      // MODE_BGRA
-  Yuv444ToArgb,      // MODE_ARGB
-  Yuv444ToRgba4444,  // MODE_RGBA_4444
-  Yuv444ToRgb565,    // MODE_RGB_565
-  Yuv444ToRgba,      // MODE_rgbA
-  Yuv444ToBgra,      // MODE_bgrA
-  Yuv444ToArgb,      // MODE_Argb
-  Yuv444ToRgba4444   // MODE_rgbA_4444
-};
-
-//------------------------------------------------------------------------------
-// Premultiplied modes
-
-// non dithered-modes
-
-// (x * a * 32897) >> 23 is bit-wise equivalent to (int)(x * a / 255.)
-// for all 8bit x or a. For bit-wise equivalence to (int)(x * a / 255. + .5),
-// one can use instead: (x * a * 65793 + (1 << 23)) >> 24
-#if 1     // (int)(x * a / 255.)
-#define MULTIPLIER(a)   ((a) * 32897UL)
-#define PREMULTIPLY(x, m) (((x) * (m)) >> 23)
-#else     // (int)(x * a / 255. + .5)
-#define MULTIPLIER(a) ((a) * 65793UL)
-#define PREMULTIPLY(x, m) (((x) * (m) + (1UL << 23)) >> 24)
-#endif
-
-static void ApplyAlphaMultiply(uint8_t* rgba, int alpha_first,
-                               int w, int h, int stride) {
-  while (h-- > 0) {
-    uint8_t* const rgb = rgba + (alpha_first ? 1 : 0);
-    const uint8_t* const alpha = rgba + (alpha_first ? 0 : 3);
-    int i;
-    for (i = 0; i < w; ++i) {
-      const uint32_t a = alpha[4 * i];
-      if (a != 0xff) {
-        const uint32_t mult = MULTIPLIER(a);
-        rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult);
-        rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult);
-        rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult);
-      }
-    }
-    rgba += stride;
-  }
-}
-#undef MULTIPLIER
-#undef PREMULTIPLY
-
-// rgbA4444
+WebPYUV444Converter WebPYUV444Converters[MODE_LAST];
 
-#define MULTIPLIER(a)  ((a) * 0x1111)    // 0x1111 ~= (1 << 16) / 15
+extern void WebPInitYUV444ConvertersMIPSdspR2(void);
+extern void WebPInitYUV444ConvertersSSE2(void);
 
-static WEBP_INLINE uint8_t dither_hi(uint8_t x) {
-  return (x & 0xf0) | (x >> 4);
-}
+static volatile VP8CPUInfo upsampling_last_cpuinfo_used1 =
+    (VP8CPUInfo)&upsampling_last_cpuinfo_used1;
 
-static WEBP_INLINE uint8_t dither_lo(uint8_t x) {
-  return (x & 0x0f) | (x << 4);
-}
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444Converters(void) {
+  if (upsampling_last_cpuinfo_used1 == VP8GetCPUInfo) return;
 
-static WEBP_INLINE uint8_t multiply(uint8_t x, uint32_t m) {
-  return (x * m) >> 16;
-}
+  WebPYUV444Converters[MODE_RGB]       = WebPYuv444ToRgbC;
+  WebPYUV444Converters[MODE_RGBA]      = WebPYuv444ToRgbaC;
+  WebPYUV444Converters[MODE_BGR]       = WebPYuv444ToBgrC;
+  WebPYUV444Converters[MODE_BGRA]      = WebPYuv444ToBgraC;
+  WebPYUV444Converters[MODE_ARGB]      = WebPYuv444ToArgbC;
+  WebPYUV444Converters[MODE_RGBA_4444] = WebPYuv444ToRgba4444C;
+  WebPYUV444Converters[MODE_RGB_565]   = WebPYuv444ToRgb565C;
+  WebPYUV444Converters[MODE_rgbA]      = WebPYuv444ToRgbaC;
+  WebPYUV444Converters[MODE_bgrA]      = WebPYuv444ToBgraC;
+  WebPYUV444Converters[MODE_Argb]      = WebPYuv444ToArgbC;
+  WebPYUV444Converters[MODE_rgbA_4444] = WebPYuv444ToRgba4444C;
 
-static void ApplyAlphaMultiply4444(uint8_t* rgba4444,
-                                   int w, int h, int stride) {
-  while (h-- > 0) {
-    int i;
-    for (i = 0; i < w; ++i) {
-      const uint8_t a = (rgba4444[2 * i + 1] & 0x0f);
-      const uint32_t mult = MULTIPLIER(a);
-      const uint8_t r = multiply(dither_hi(rgba4444[2 * i + 0]), mult);
-      const uint8_t g = multiply(dither_lo(rgba4444[2 * i + 0]), mult);
-      const uint8_t b = multiply(dither_hi(rgba4444[2 * i + 1]), mult);
-      rgba4444[2 * i + 0] = (r & 0xf0) | ((g >> 4) & 0x0f);
-      rgba4444[2 * i + 1] = (b & 0xf0) | a;
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitYUV444ConvertersSSE2();
+    }
+#endif
+#if defined(WEBP_USE_MIPS_DSP_R2)
+    if (VP8GetCPUInfo(kMIPSdspR2)) {
+      WebPInitYUV444ConvertersMIPSdspR2();
     }
-    rgba4444 += stride;
+#endif
   }
+  upsampling_last_cpuinfo_used1 = VP8GetCPUInfo;
 }
-#undef MULTIPLIER
-
-void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int)
-    = ApplyAlphaMultiply;
-void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int)
-    = ApplyAlphaMultiply4444;
 
 //------------------------------------------------------------------------------
-// Main call
+// Main calls
+
+extern void WebPInitUpsamplersSSE2(void);
+extern void WebPInitUpsamplersNEON(void);
+extern void WebPInitUpsamplersMIPSdspR2(void);
+
+static volatile VP8CPUInfo upsampling_last_cpuinfo_used2 =
+    (VP8CPUInfo)&upsampling_last_cpuinfo_used2;
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplers(void) {
+  if (upsampling_last_cpuinfo_used2 == VP8GetCPUInfo) return;
 
-void WebPInitUpsamplers(void) {
 #ifdef FANCY_UPSAMPLING
   WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair;
   WebPUpsamplers[MODE_RGBA]      = UpsampleRgbaLinePair;
@@ -320,38 +230,31 @@ void WebPInitUpsamplers(void) {
   WebPUpsamplers[MODE_ARGB]      = UpsampleArgbLinePair;
   WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
   WebPUpsamplers[MODE_RGB_565]   = UpsampleRgb565LinePair;
-
-  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
-  if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
-    if (VP8GetCPUInfo(kSSE2)) {
-      WebPInitUpsamplersSSE2();
-    }
-#endif
-  }
-#endif  // FANCY_UPSAMPLING
-}
-
-void WebPInitPremultiply(void) {
-  WebPApplyAlphaMultiply = ApplyAlphaMultiply;
-  WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply4444;
-
-#ifdef FANCY_UPSAMPLING
   WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair;
   WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair;
   WebPUpsamplers[MODE_Argb]      = UpsampleArgbLinePair;
   WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
 
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
   if (VP8GetCPUInfo != NULL) {
 #if defined(WEBP_USE_SSE2)
     if (VP8GetCPUInfo(kSSE2)) {
-      WebPInitPremultiplySSE2();
+      WebPInitUpsamplersSSE2();
+    }
+#endif
+#if defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      WebPInitUpsamplersNEON();
+    }
+#endif
+#if defined(WEBP_USE_MIPS_DSP_R2)
+    if (VP8GetCPUInfo(kMIPSdspR2)) {
+      WebPInitUpsamplersMIPSdspR2();
     }
 #endif
   }
 #endif  // FANCY_UPSAMPLING
+  upsampling_last_cpuinfo_used2 = VP8GetCPUInfo;
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
+//------------------------------------------------------------------------------
diff --git a/drivers/webp/dsp/upsampling_sse2.c b/drivers/webp/dsp/upsampling_sse2.c
index 8cb275a02b..b85808e271 100644
--- a/drivers/webp/dsp/upsampling_sse2.c
+++ b/drivers/webp/dsp/upsampling_sse2.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // SSE2 version of YUV to RGB upsampling functions.
@@ -18,10 +20,6 @@
 #include <string.h>
 #include "./yuv.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 #ifdef FANCY_UPSAMPLING
 
 // We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows
@@ -49,23 +47,23 @@ extern "C" {
   (out) = _mm_sub_epi8(tmp0, tmp4);    /* (k + in + 1) / 2 - lsb_correction */ \
 } while (0)
 
-// pack and store two alterning pixel rows
+// pack and store two alternating pixel rows
 #define PACK_AND_STORE(a, b, da, db, out) do {                                 \
-  const __m128i ta = _mm_avg_epu8(a, da);  /* (9a + 3b + 3c +  d + 8) / 16 */  \
-  const __m128i tb = _mm_avg_epu8(b, db);  /* (3a + 9b +  c + 3d + 8) / 16 */  \
-  const __m128i t1 = _mm_unpacklo_epi8(ta, tb);                                \
-  const __m128i t2 = _mm_unpackhi_epi8(ta, tb);                                \
-  _mm_store_si128(((__m128i*)(out)) + 0, t1);                                  \
-  _mm_store_si128(((__m128i*)(out)) + 1, t2);                                  \
+  const __m128i t_a = _mm_avg_epu8(a, da);  /* (9a + 3b + 3c +  d + 8) / 16 */ \
+  const __m128i t_b = _mm_avg_epu8(b, db);  /* (3a + 9b +  c + 3d + 8) / 16 */ \
+  const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b);                             \
+  const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b);                             \
+  _mm_store_si128(((__m128i*)(out)) + 0, t_1);                                 \
+  _mm_store_si128(((__m128i*)(out)) + 1, t_2);                                 \
 } while (0)
 
 // Loads 17 pixels each from rows r1 and r2 and generates 32 pixels.
 #define UPSAMPLE_32PIXELS(r1, r2, out) {                                       \
   const __m128i one = _mm_set1_epi8(1);                                        \
-  const __m128i a = _mm_loadu_si128((__m128i*)&(r1)[0]);                       \
-  const __m128i b = _mm_loadu_si128((__m128i*)&(r1)[1]);                       \
-  const __m128i c = _mm_loadu_si128((__m128i*)&(r2)[0]);                       \
-  const __m128i d = _mm_loadu_si128((__m128i*)&(r2)[1]);                       \
+  const __m128i a = _mm_loadu_si128((const __m128i*)&(r1)[0]);                 \
+  const __m128i b = _mm_loadu_si128((const __m128i*)&(r1)[1]);                 \
+  const __m128i c = _mm_loadu_si128((const __m128i*)&(r2)[0]);                 \
+  const __m128i d = _mm_loadu_si128((const __m128i*)&(r2)[1]);                 \
                                                                                \
   const __m128i s = _mm_avg_epu8(a, d);        /* s = (a + d + 1) / 2 */       \
   const __m128i t = _mm_avg_epu8(b, c);        /* t = (b + c + 1) / 2 */       \
@@ -85,8 +83,8 @@ extern "C" {
   GET_M(ad, s, diag2);                  /* diag2 = (3a + b + c + 3d) / 8 */    \
                                                                                \
   /* pack the alternate pixels */                                              \
-  PACK_AND_STORE(a, b, diag1, diag2, &(out)[0 * 32]);                          \
-  PACK_AND_STORE(c, d, diag2, diag1, &(out)[2 * 32]);                          \
+  PACK_AND_STORE(a, b, diag1, diag2, out +      0);  /* store top */           \
+  PACK_AND_STORE(c, d, diag2, diag1, out + 2 * 32);  /* store bottom */        \
 }
 
 // Turn the macro into a function for reducing code-size when non-critical
@@ -106,104 +104,140 @@ static void Upsample32Pixels(const uint8_t r1[], const uint8_t r2[],
   Upsample32Pixels(r1, r2, out);                                               \
 }
 
-#define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, uv,                          \
+#define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y,                              \
                     top_dst, bottom_dst, cur_x, num_pixels) {                  \
   int n;                                                                       \
-  if (top_y) {                                                                 \
-    for (n = 0; n < (num_pixels); ++n) {                                       \
-      FUNC(top_y[(cur_x) + n], (uv)[n], (uv)[32 + n],                          \
-           top_dst + ((cur_x) + n) * XSTEP);                                   \
-    }                                                                          \
+  for (n = 0; n < (num_pixels); ++n) {                                         \
+    FUNC(top_y[(cur_x) + n], r_u[n], r_v[n],                                   \
+         top_dst + ((cur_x) + n) * XSTEP);                                     \
   }                                                                            \
-  if (bottom_y) {                                                              \
+  if (bottom_y != NULL) {                                                      \
     for (n = 0; n < (num_pixels); ++n) {                                       \
-      FUNC(bottom_y[(cur_x) + n], (uv)[64 + n], (uv)[64 + 32 + n],             \
+      FUNC(bottom_y[(cur_x) + n], r_u[64 + n], r_v[64 + n],                    \
            bottom_dst + ((cur_x) + n) * XSTEP);                                \
     }                                                                          \
   }                                                                            \
 }
 
+#define CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y,                           \
+                       top_dst, bottom_dst, cur_x) do {                        \
+  FUNC##32(top_y + (cur_x), r_u, r_v, top_dst + (cur_x) * XSTEP);              \
+  if (bottom_y != NULL) {                                                      \
+    FUNC##32(bottom_y + (cur_x), r_u + 64, r_v + 64,                           \
+             bottom_dst + (cur_x) * XSTEP);                                    \
+  }                                                                            \
+} while (0)
+
 #define SSE2_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                             \
 static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
                       const uint8_t* top_u, const uint8_t* top_v,              \
                       const uint8_t* cur_u, const uint8_t* cur_v,              \
                       uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
-  int b;                                                                       \
-  /* 16 byte aligned array to cache reconstructed u and v */                   \
+  int uv_pos, pos;                                                             \
+  /* 16byte-aligned array to cache reconstructed u and v */                    \
   uint8_t uv_buf[4 * 32 + 15];                                                 \
-  uint8_t* const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);            \
-  const int uv_len = (len + 1) >> 1;                                           \
-  /* 17 pixels must be read-able for each block */                             \
-  const int num_blocks = (uv_len - 1) >> 4;                                    \
-  const int leftover = uv_len - num_blocks * 16;                               \
-  const int last_pos = 1 + 32 * num_blocks;                                    \
-                                                                               \
-  const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1;                         \
-  const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;                         \
+  uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);             \
+  uint8_t* const r_v = r_u + 32;                                               \
                                                                                \
-  assert(len > 0);                                                             \
-  /* Treat the first pixel in regular way */                                   \
-  if (top_y) {                                                                 \
-    const int u0 = (top_u[0] + u_diag) >> 1;                                   \
-    const int v0 = (top_v[0] + v_diag) >> 1;                                   \
-    FUNC(top_y[0], u0, v0, top_dst);                                           \
+  assert(top_y != NULL);                                                       \
+  {   /* Treat the first pixel in regular way */                               \
+    const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1;                       \
+    const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;                       \
+    const int u0_t = (top_u[0] + u_diag) >> 1;                                 \
+    const int v0_t = (top_v[0] + v_diag) >> 1;                                 \
+    FUNC(top_y[0], u0_t, v0_t, top_dst);                                       \
+    if (bottom_y != NULL) {                                                    \
+      const int u0_b = (cur_u[0] + u_diag) >> 1;                               \
+      const int v0_b = (cur_v[0] + v_diag) >> 1;                               \
+      FUNC(bottom_y[0], u0_b, v0_b, bottom_dst);                               \
+    }                                                                          \
   }                                                                            \
-  if (bottom_y) {                                                              \
-    const int u0 = (cur_u[0] + u_diag) >> 1;                                   \
-    const int v0 = (cur_v[0] + v_diag) >> 1;                                   \
-    FUNC(bottom_y[0], u0, v0, bottom_dst);                                     \
+  /* For UPSAMPLE_32PIXELS, 17 u/v values must be read-able for each block */  \
+  for (pos = 1, uv_pos = 0; pos + 32 + 1 <= len; pos += 32, uv_pos += 16) {    \
+    UPSAMPLE_32PIXELS(top_u + uv_pos, cur_u + uv_pos, r_u);                    \
+    UPSAMPLE_32PIXELS(top_v + uv_pos, cur_v + uv_pos, r_v);                    \
+    CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst, pos);    \
   }                                                                            \
-                                                                               \
-  for (b = 0; b < num_blocks; ++b) {                                           \
-    UPSAMPLE_32PIXELS(top_u, cur_u, r_uv + 0 * 32);                            \
-    UPSAMPLE_32PIXELS(top_v, cur_v, r_uv + 1 * 32);                            \
-    CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst,       \
-                32 * b + 1, 32)                                                \
-    top_u += 16;                                                               \
-    cur_u += 16;                                                               \
-    top_v += 16;                                                               \
-    cur_v += 16;                                                               \
+  if (len > 1) {                                                               \
+    const int left_over = ((len + 1) >> 1) - (pos >> 1);                       \
+    assert(left_over > 0);                                                     \
+    UPSAMPLE_LAST_BLOCK(top_u + uv_pos, cur_u + uv_pos, left_over, r_u);       \
+    UPSAMPLE_LAST_BLOCK(top_v + uv_pos, cur_v + uv_pos, left_over, r_v);       \
+    CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst,             \
+                pos, len - pos);                                               \
   }                                                                            \
-                                                                               \
-  UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv + 0 * 32);                  \
-  UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 1 * 32);                  \
-  CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst,         \
-              last_pos, len - last_pos);                                       \
 }
 
 // SSE2 variants of the fancy upsampler.
-SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePairSSE2,  VP8YuvToRgb,  3)
-SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePairSSE2,  VP8YuvToBgr,  3)
-SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePairSSE2, VP8YuvToRgba, 4)
-SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4)
+SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePair,  VP8YuvToRgb,  3)
+SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePair,  VP8YuvToBgr,  3)
+SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePair, VP8YuvToRgba, 4)
+SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePair, VP8YuvToBgra, 4)
 
 #undef GET_M
 #undef PACK_AND_STORE
 #undef UPSAMPLE_32PIXELS
 #undef UPSAMPLE_LAST_BLOCK
 #undef CONVERT2RGB
+#undef CONVERT2RGB_32
 #undef SSE2_UPSAMPLE_FUNC
 
 //------------------------------------------------------------------------------
+// Entry point
 
 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 
-void WebPInitUpsamplersSSE2(void) {
-  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePairSSE2;
-  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2;
-  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePairSSE2;
-  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2;
-}
+extern void WebPInitUpsamplersSSE2(void);
 
-void WebPInitPremultiplySSE2(void) {
-  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2;
-  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2;
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersSSE2(void) {
+  VP8YUVInitSSE2();
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair;
+  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair;
+  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
 }
 
 #endif  // FANCY_UPSAMPLING
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
+//------------------------------------------------------------------------------
+
+extern WebPYUV444Converter WebPYUV444Converters[/* MODE_LAST */];
+extern void WebPInitYUV444ConvertersSSE2(void);
+
+#define YUV444_FUNC(FUNC_NAME, CALL, XSTEP) \
+extern void WebP##FUNC_NAME##C(const uint8_t* y, const uint8_t* u,             \
+                               const uint8_t* v, uint8_t* dst, int len);       \
+static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
+                      uint8_t* dst, int len) {                                 \
+  int i;                                                                       \
+  const int max_len = len & ~31;                                               \
+  for (i = 0; i < max_len; i += 32) CALL(y + i, u + i, v + i, dst + i * XSTEP);\
+  if (i < len) {  /* C-fallback */                                             \
+    WebP##FUNC_NAME##C(y + i, u + i, v + i, dst + i * XSTEP, len - i);         \
+  }                                                                            \
+}
 
-#endif   // WEBP_USE_SSE2
+YUV444_FUNC(Yuv444ToRgba, VP8YuvToRgba32, 4);
+YUV444_FUNC(Yuv444ToBgra, VP8YuvToBgra32, 4);
+YUV444_FUNC(Yuv444ToRgb, VP8YuvToRgb32, 3);
+YUV444_FUNC(Yuv444ToBgr, VP8YuvToBgr32, 3);
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersSSE2(void) {
+  VP8YUVInitSSE2();
+  WebPYUV444Converters[MODE_RGBA] = Yuv444ToRgba;
+  WebPYUV444Converters[MODE_BGRA] = Yuv444ToBgra;
+  WebPYUV444Converters[MODE_RGB]  = Yuv444ToRgb;
+  WebPYUV444Converters[MODE_BGR]  = Yuv444ToBgr;
+}
+
+#else
+
+WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersSSE2)
+
+#endif  // WEBP_USE_SSE2
+
+#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_SSE2))
+WEBP_DSP_INIT_STUB(WebPInitUpsamplersSSE2)
+#endif
diff --git a/drivers/webp/dsp/yuv.c b/drivers/webp/dsp/yuv.c
index 7f05f9a3aa..f50a253168 100644
--- a/drivers/webp/dsp/yuv.c
+++ b/drivers/webp/dsp/yuv.c
@@ -1,26 +1,19 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
-// YUV->RGB conversion function
+// YUV->RGB conversion functions
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include "./yuv.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-enum { YUV_HALF = 1 << (YUV_FIX - 1) };
-
-int16_t VP8kVToR[256], VP8kUToB[256];
-int32_t VP8kVToG[256], VP8kUToG[256];
-uint8_t VP8kClip[YUV_RANGE_MAX - YUV_RANGE_MIN];
-uint8_t VP8kClip4Bits[YUV_RANGE_MAX - YUV_RANGE_MIN];
+#if defined(WEBP_YUV_USE_TABLE)
 
 static int done = 0;
 
@@ -28,11 +21,17 @@ static WEBP_INLINE uint8_t clip(int v, int max_value) {
   return v < 0 ? 0 : v > max_value ? max_value : v;
 }
 
-void VP8YUVInit(void) {
+int16_t VP8kVToR[256], VP8kUToB[256];
+int32_t VP8kVToG[256], VP8kUToG[256];
+uint8_t VP8kClip[YUV_RANGE_MAX - YUV_RANGE_MIN];
+uint8_t VP8kClip4Bits[YUV_RANGE_MAX - YUV_RANGE_MIN];
+
+WEBP_TSAN_IGNORE_FUNCTION void VP8YUVInit(void) {
   int i;
   if (done) {
     return;
   }
+#ifndef USE_YUVj
   for (i = 0; i < 256; ++i) {
     VP8kVToR[i] = (89858 * (i - 128) + YUV_HALF) >> YUV_FIX;
     VP8kUToG[i] = -22014 * (i - 128) + YUV_HALF;
@@ -44,9 +43,238 @@ void VP8YUVInit(void) {
     VP8kClip[i - YUV_RANGE_MIN] = clip(k, 255);
     VP8kClip4Bits[i - YUV_RANGE_MIN] = clip((k + 8) >> 4, 15);
   }
+#else
+  for (i = 0; i < 256; ++i) {
+    VP8kVToR[i] = (91881 * (i - 128) + YUV_HALF) >> YUV_FIX;
+    VP8kUToG[i] = -22554 * (i - 128) + YUV_HALF;
+    VP8kVToG[i] = -46802 * (i - 128);
+    VP8kUToB[i] = (116130 * (i - 128) + YUV_HALF) >> YUV_FIX;
+  }
+  for (i = YUV_RANGE_MIN; i < YUV_RANGE_MAX; ++i) {
+    const int k = i;
+    VP8kClip[i - YUV_RANGE_MIN] = clip(k, 255);
+    VP8kClip4Bits[i - YUV_RANGE_MIN] = clip((k + 8) >> 4, 15);
+  }
+#endif
+
   done = 1;
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
+#else
+
+WEBP_TSAN_IGNORE_FUNCTION void VP8YUVInit(void) {}
+
+#endif  // WEBP_YUV_USE_TABLE
+
+//-----------------------------------------------------------------------------
+// Plain-C version
+
+#define ROW_FUNC(FUNC_NAME, FUNC, XSTEP)                                       \
+static void FUNC_NAME(const uint8_t* y,                                        \
+                      const uint8_t* u, const uint8_t* v,                      \
+                      uint8_t* dst, int len) {                                 \
+  const uint8_t* const end = dst + (len & ~1) * XSTEP;                         \
+  while (dst != end) {                                                         \
+    FUNC(y[0], u[0], v[0], dst);                                               \
+    FUNC(y[1], u[0], v[0], dst + XSTEP);                                       \
+    y += 2;                                                                    \
+    ++u;                                                                       \
+    ++v;                                                                       \
+    dst += 2 * XSTEP;                                                          \
+  }                                                                            \
+  if (len & 1) {                                                               \
+    FUNC(y[0], u[0], v[0], dst);                                               \
+  }                                                                            \
+}                                                                              \
+
+// All variants implemented.
+ROW_FUNC(YuvToRgbRow,      VP8YuvToRgb,  3)
+ROW_FUNC(YuvToBgrRow,      VP8YuvToBgr,  3)
+ROW_FUNC(YuvToRgbaRow,     VP8YuvToRgba, 4)
+ROW_FUNC(YuvToBgraRow,     VP8YuvToBgra, 4)
+ROW_FUNC(YuvToArgbRow,     VP8YuvToArgb, 4)
+ROW_FUNC(YuvToRgba4444Row, VP8YuvToRgba4444, 2)
+ROW_FUNC(YuvToRgb565Row,   VP8YuvToRgb565, 2)
+
+#undef ROW_FUNC
+
+// Main call for processing a plane with a WebPSamplerRowFunc function:
+void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
+                             const uint8_t* u, const uint8_t* v, int uv_stride,
+                             uint8_t* dst, int dst_stride,
+                             int width, int height, WebPSamplerRowFunc func) {
+  int j;
+  for (j = 0; j < height; ++j) {
+    func(y, u, v, dst, width);
+    y += y_stride;
+    if (j & 1) {
+      u += uv_stride;
+      v += uv_stride;
+    }
+    dst += dst_stride;
+  }
+}
+
+//-----------------------------------------------------------------------------
+// Main call
+
+WebPSamplerRowFunc WebPSamplers[MODE_LAST];
+
+extern void WebPInitSamplersSSE2(void);
+extern void WebPInitSamplersMIPS32(void);
+extern void WebPInitSamplersMIPSdspR2(void);
+
+static volatile VP8CPUInfo yuv_last_cpuinfo_used =
+    (VP8CPUInfo)&yuv_last_cpuinfo_used;
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplers(void) {
+  if (yuv_last_cpuinfo_used == VP8GetCPUInfo) return;
+
+  WebPSamplers[MODE_RGB]       = YuvToRgbRow;
+  WebPSamplers[MODE_RGBA]      = YuvToRgbaRow;
+  WebPSamplers[MODE_BGR]       = YuvToBgrRow;
+  WebPSamplers[MODE_BGRA]      = YuvToBgraRow;
+  WebPSamplers[MODE_ARGB]      = YuvToArgbRow;
+  WebPSamplers[MODE_RGBA_4444] = YuvToRgba4444Row;
+  WebPSamplers[MODE_RGB_565]   = YuvToRgb565Row;
+  WebPSamplers[MODE_rgbA]      = YuvToRgbaRow;
+  WebPSamplers[MODE_bgrA]      = YuvToBgraRow;
+  WebPSamplers[MODE_Argb]      = YuvToArgbRow;
+  WebPSamplers[MODE_rgbA_4444] = YuvToRgba4444Row;
+
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitSamplersSSE2();
+    }
+#endif  // WEBP_USE_SSE2
+#if defined(WEBP_USE_MIPS32)
+    if (VP8GetCPUInfo(kMIPS32)) {
+      WebPInitSamplersMIPS32();
+    }
+#endif  // WEBP_USE_MIPS32
+#if defined(WEBP_USE_MIPS_DSP_R2)
+    if (VP8GetCPUInfo(kMIPSdspR2)) {
+      WebPInitSamplersMIPSdspR2();
+    }
+#endif  // WEBP_USE_MIPS_DSP_R2
+  }
+  yuv_last_cpuinfo_used = VP8GetCPUInfo;
+}
+
+//-----------------------------------------------------------------------------
+// ARGB -> YUV converters
+
+static void ConvertARGBToY(const uint32_t* argb, uint8_t* y, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32_t p = argb[i];
+    y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >>  0) & 0xff,
+                     YUV_HALF);
+  }
+}
+
+void WebPConvertARGBToUV_C(const uint32_t* argb, uint8_t* u, uint8_t* v,
+                           int src_width, int do_store) {
+  // No rounding. Last pixel is dealt with separately.
+  const int uv_width = src_width >> 1;
+  int i;
+  for (i = 0; i < uv_width; ++i) {
+    const uint32_t v0 = argb[2 * i + 0];
+    const uint32_t v1 = argb[2 * i + 1];
+    // VP8RGBToU/V expects four accumulated pixels. Hence we need to
+    // scale r/g/b value by a factor 2. We just shift v0/v1 one bit less.
+    const int r = ((v0 >> 15) & 0x1fe) + ((v1 >> 15) & 0x1fe);
+    const int g = ((v0 >>  7) & 0x1fe) + ((v1 >>  7) & 0x1fe);
+    const int b = ((v0 <<  1) & 0x1fe) + ((v1 <<  1) & 0x1fe);
+    const int tmp_u = VP8RGBToU(r, g, b, YUV_HALF << 2);
+    const int tmp_v = VP8RGBToV(r, g, b, YUV_HALF << 2);
+    if (do_store) {
+      u[i] = tmp_u;
+      v[i] = tmp_v;
+    } else {
+      // Approximated average-of-four. But it's an acceptable diff.
+      u[i] = (u[i] + tmp_u + 1) >> 1;
+      v[i] = (v[i] + tmp_v + 1) >> 1;
+    }
+  }
+  if (src_width & 1) {       // last pixel
+    const uint32_t v0 = argb[2 * i + 0];
+    const int r = (v0 >> 14) & 0x3fc;
+    const int g = (v0 >>  6) & 0x3fc;
+    const int b = (v0 <<  2) & 0x3fc;
+    const int tmp_u = VP8RGBToU(r, g, b, YUV_HALF << 2);
+    const int tmp_v = VP8RGBToV(r, g, b, YUV_HALF << 2);
+    if (do_store) {
+      u[i] = tmp_u;
+      v[i] = tmp_v;
+    } else {
+      u[i] = (u[i] + tmp_u + 1) >> 1;
+      v[i] = (v[i] + tmp_v + 1) >> 1;
+    }
+  }
+}
+
+//-----------------------------------------------------------------------------
+
+static void ConvertRGB24ToY(const uint8_t* rgb, uint8_t* y, int width) {
+  int i;
+  for (i = 0; i < width; ++i, rgb += 3) {
+    y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF);
+  }
+}
+
+static void ConvertBGR24ToY(const uint8_t* bgr, uint8_t* y, int width) {
+  int i;
+  for (i = 0; i < width; ++i, bgr += 3) {
+    y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
+  }
+}
+
+void WebPConvertRGBA32ToUV_C(const uint16_t* rgb,
+                             uint8_t* u, uint8_t* v, int width) {
+  int i;
+  for (i = 0; i < width; i += 1, rgb += 4) {
+    const int r = rgb[0], g = rgb[1], b = rgb[2];
+    u[i] = VP8RGBToU(r, g, b, YUV_HALF << 2);
+    v[i] = VP8RGBToV(r, g, b, YUV_HALF << 2);
+  }
+}
+
+//-----------------------------------------------------------------------------
+
+void (*WebPConvertRGB24ToY)(const uint8_t* rgb, uint8_t* y, int width);
+void (*WebPConvertBGR24ToY)(const uint8_t* bgr, uint8_t* y, int width);
+void (*WebPConvertRGBA32ToUV)(const uint16_t* rgb,
+                              uint8_t* u, uint8_t* v, int width);
+
+void (*WebPConvertARGBToY)(const uint32_t* argb, uint8_t* y, int width);
+void (*WebPConvertARGBToUV)(const uint32_t* argb, uint8_t* u, uint8_t* v,
+                            int src_width, int do_store);
+
+static volatile VP8CPUInfo rgba_to_yuv_last_cpuinfo_used =
+    (VP8CPUInfo)&rgba_to_yuv_last_cpuinfo_used;
+
+extern void WebPInitConvertARGBToYUVSSE2(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) {
+  if (rgba_to_yuv_last_cpuinfo_used == VP8GetCPUInfo) return;
+
+  WebPConvertARGBToY = ConvertARGBToY;
+  WebPConvertARGBToUV = WebPConvertARGBToUV_C;
+
+  WebPConvertRGB24ToY = ConvertRGB24ToY;
+  WebPConvertBGR24ToY = ConvertBGR24ToY;
+
+  WebPConvertRGBA32ToUV = WebPConvertRGBA32ToUV_C;
+
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitConvertARGBToYUVSSE2();
+    }
+#endif  // WEBP_USE_SSE2
+  }
+  rgba_to_yuv_last_cpuinfo_used = VP8GetCPUInfo;
+}
diff --git a/drivers/webp/dsp/yuv.h b/drivers/webp/dsp/yuv.h
index a569109c54..af435a5b3e 100644
--- a/drivers/webp/dsp/yuv.h
+++ b/drivers/webp/dsp/yuv.h
@@ -1,36 +1,165 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // inline YUV<->RGB conversion function
 //
+// The exact naming is Y'CbCr, following the ITU-R BT.601 standard.
+// More information at: http://en.wikipedia.org/wiki/YCbCr
+// Y = 0.2569 * R + 0.5044 * G + 0.0979 * B + 16
+// U = -0.1483 * R - 0.2911 * G + 0.4394 * B + 128
+// V = 0.4394 * R - 0.3679 * G - 0.0715 * B + 128
+// We use 16bit fixed point operations for RGB->YUV conversion (YUV_FIX).
+//
+// For the Y'CbCr to RGB conversion, the BT.601 specification reads:
+//   R = 1.164 * (Y-16) + 1.596 * (V-128)
+//   G = 1.164 * (Y-16) - 0.813 * (V-128) - 0.391 * (U-128)
+//   B = 1.164 * (Y-16)                   + 2.018 * (U-128)
+// where Y is in the [16,235] range, and U/V in the [16,240] range.
+// In the table-lookup version (WEBP_YUV_USE_TABLE), the common factor
+// "1.164 * (Y-16)" can be handled as an offset in the VP8kClip[] table.
+// So in this case the formulae should read:
+//   R = 1.164 * [Y + 1.371 * (V-128)                  ] - 18.624
+//   G = 1.164 * [Y - 0.698 * (V-128) - 0.336 * (U-128)] - 18.624
+//   B = 1.164 * [Y                   + 1.733 * (U-128)] - 18.624
+// once factorized.
+// For YUV->RGB conversion, only 14bit fixed precision is used (YUV_FIX2).
+// That's the maximum possible for a convenient ARM implementation.
+//
 // Author: Skal (pascal.massimino@gmail.com)
 
 #ifndef WEBP_DSP_YUV_H_
 #define WEBP_DSP_YUV_H_
 
+#include "./dsp.h"
 #include "../dec/decode_vp8.h"
 
+// Define the following to use the LUT-based code:
+// #define WEBP_YUV_USE_TABLE
+
+#if defined(WEBP_EXPERIMENTAL_FEATURES)
+// Do NOT activate this feature for real compression. This is only experimental!
+// This flag is for comparison purpose against JPEG's "YUVj" natural colorspace.
+// This colorspace is close to Rec.601's Y'CbCr model with the notable
+// difference of allowing larger range for luma/chroma.
+// See http://en.wikipedia.org/wiki/YCbCr#JPEG_conversion paragraph, and its
+// difference with http://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion
+// #define USE_YUVj
+#endif
+
 //------------------------------------------------------------------------------
 // YUV -> RGB conversion
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-enum { YUV_FIX = 16,                // fixed-point precision
-       YUV_RANGE_MIN = -227,        // min value of r/g/b output
-       YUV_RANGE_MAX = 256 + 226    // max value of r/g/b output
+enum {
+  YUV_FIX = 16,                    // fixed-point precision for RGB->YUV
+  YUV_HALF = 1 << (YUV_FIX - 1),
+  YUV_MASK = (256 << YUV_FIX) - 1,
+  YUV_RANGE_MIN = -227,            // min value of r/g/b output
+  YUV_RANGE_MAX = 256 + 226,       // max value of r/g/b output
+
+  YUV_FIX2 = 14,                   // fixed-point precision for YUV->RGB
+  YUV_HALF2 = 1 << (YUV_FIX2 - 1),
+  YUV_MASK2 = (256 << YUV_FIX2) - 1
 };
+
+// These constants are 14b fixed-point version of ITU-R BT.601 constants.
+#define kYScale 19077    // 1.164 = 255 / 219
+#define kVToR   26149    // 1.596 = 255 / 112 * 0.701
+#define kUToG   6419     // 0.391 = 255 / 112 * 0.886 * 0.114 / 0.587
+#define kVToG   13320    // 0.813 = 255 / 112 * 0.701 * 0.299 / 0.587
+#define kUToB   33050    // 2.018 = 255 / 112 * 0.886
+#define kRCst (-kYScale * 16 - kVToR * 128 + YUV_HALF2)
+#define kGCst (-kYScale * 16 + kUToG * 128 + kVToG * 128 + YUV_HALF2)
+#define kBCst (-kYScale * 16 - kUToB * 128 + YUV_HALF2)
+
+//------------------------------------------------------------------------------
+
+#if !defined(WEBP_YUV_USE_TABLE)
+
+// slower on x86 by ~7-8%, but bit-exact with the SSE2 version
+
+static WEBP_INLINE int VP8Clip8(int v) {
+  return ((v & ~YUV_MASK2) == 0) ? (v >> YUV_FIX2) : (v < 0) ? 0 : 255;
+}
+
+static WEBP_INLINE int VP8YUVToR(int y, int v) {
+  return VP8Clip8(kYScale * y + kVToR * v + kRCst);
+}
+
+static WEBP_INLINE int VP8YUVToG(int y, int u, int v) {
+  return VP8Clip8(kYScale * y - kUToG * u - kVToG * v + kGCst);
+}
+
+static WEBP_INLINE int VP8YUVToB(int y, int u) {
+  return VP8Clip8(kYScale * y + kUToB * u + kBCst);
+}
+
+static WEBP_INLINE void VP8YuvToRgb(int y, int u, int v,
+                                    uint8_t* const rgb) {
+  rgb[0] = VP8YUVToR(y, v);
+  rgb[1] = VP8YUVToG(y, u, v);
+  rgb[2] = VP8YUVToB(y, u);
+}
+
+static WEBP_INLINE void VP8YuvToBgr(int y, int u, int v,
+                                    uint8_t* const bgr) {
+  bgr[0] = VP8YUVToB(y, u);
+  bgr[1] = VP8YUVToG(y, u, v);
+  bgr[2] = VP8YUVToR(y, v);
+}
+
+static WEBP_INLINE void VP8YuvToRgb565(int y, int u, int v,
+                                       uint8_t* const rgb) {
+  const int r = VP8YUVToR(y, v);      // 5 usable bits
+  const int g = VP8YUVToG(y, u, v);   // 6 usable bits
+  const int b = VP8YUVToB(y, u);      // 5 usable bits
+  const int rg = (r & 0xf8) | (g >> 5);
+  const int gb = ((g << 3) & 0xe0) | (b >> 3);
+#ifdef WEBP_SWAP_16BIT_CSP
+  rgb[0] = gb;
+  rgb[1] = rg;
+#else
+  rgb[0] = rg;
+  rgb[1] = gb;
+#endif
+}
+
+static WEBP_INLINE void VP8YuvToRgba4444(int y, int u, int v,
+                                         uint8_t* const argb) {
+  const int r = VP8YUVToR(y, v);        // 4 usable bits
+  const int g = VP8YUVToG(y, u, v);     // 4 usable bits
+  const int b = VP8YUVToB(y, u);        // 4 usable bits
+  const int rg = (r & 0xf0) | (g >> 4);
+  const int ba = (b & 0xf0) | 0x0f;     // overwrite the lower 4 bits
+#ifdef WEBP_SWAP_16BIT_CSP
+  argb[0] = ba;
+  argb[1] = rg;
+#else
+  argb[0] = rg;
+  argb[1] = ba;
+#endif
+}
+
+#else
+
+// Table-based version, not totally equivalent to the SSE2 version.
+// Rounding diff is only +/-1 though.
+
 extern int16_t VP8kVToR[256], VP8kUToB[256];
 extern int32_t VP8kVToG[256], VP8kUToG[256];
 extern uint8_t VP8kClip[YUV_RANGE_MAX - YUV_RANGE_MIN];
 extern uint8_t VP8kClip4Bits[YUV_RANGE_MAX - YUV_RANGE_MIN];
 
-static WEBP_INLINE void VP8YuvToRgb(uint8_t y, uint8_t u, uint8_t v,
+static WEBP_INLINE void VP8YuvToRgb(int y, int u, int v,
                                     uint8_t* const rgb) {
   const int r_off = VP8kVToR[v];
   const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
@@ -40,42 +169,60 @@ static WEBP_INLINE void VP8YuvToRgb(uint8_t y, uint8_t u, uint8_t v,
   rgb[2] = VP8kClip[y + b_off - YUV_RANGE_MIN];
 }
 
-static WEBP_INLINE void VP8YuvToRgb565(uint8_t y, uint8_t u, uint8_t v,
-                                       uint8_t* const rgb) {
+static WEBP_INLINE void VP8YuvToBgr(int y, int u, int v,
+                                    uint8_t* const bgr) {
   const int r_off = VP8kVToR[v];
   const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
   const int b_off = VP8kUToB[u];
-  rgb[0] = ((VP8kClip[y + r_off - YUV_RANGE_MIN] & 0xf8) |
-            (VP8kClip[y + g_off - YUV_RANGE_MIN] >> 5));
-  rgb[1] = (((VP8kClip[y + g_off - YUV_RANGE_MIN] << 3) & 0xe0) |
-            (VP8kClip[y + b_off - YUV_RANGE_MIN] >> 3));
-}
-
-static WEBP_INLINE void VP8YuvToArgb(uint8_t y, uint8_t u, uint8_t v,
-                                     uint8_t* const argb) {
-  argb[0] = 0xff;
-  VP8YuvToRgb(y, u, v, argb + 1);
+  bgr[0] = VP8kClip[y + b_off - YUV_RANGE_MIN];
+  bgr[1] = VP8kClip[y + g_off - YUV_RANGE_MIN];
+  bgr[2] = VP8kClip[y + r_off - YUV_RANGE_MIN];
 }
 
-static WEBP_INLINE void VP8YuvToRgba4444(uint8_t y, uint8_t u, uint8_t v,
-                                         uint8_t* const argb) {
+static WEBP_INLINE void VP8YuvToRgb565(int y, int u, int v,
+                                       uint8_t* const rgb) {
   const int r_off = VP8kVToR[v];
   const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
   const int b_off = VP8kUToB[u];
-  // Don't update alpha (last 4 bits of argb[1])
-  argb[0] = ((VP8kClip4Bits[y + r_off - YUV_RANGE_MIN] << 4) |
-             VP8kClip4Bits[y + g_off - YUV_RANGE_MIN]);
-  argb[1] = 0x0f | (VP8kClip4Bits[y + b_off - YUV_RANGE_MIN] << 4);
+  const int rg = ((VP8kClip[y + r_off - YUV_RANGE_MIN] & 0xf8) |
+                  (VP8kClip[y + g_off - YUV_RANGE_MIN] >> 5));
+  const int gb = (((VP8kClip[y + g_off - YUV_RANGE_MIN] << 3) & 0xe0) |
+                   (VP8kClip[y + b_off - YUV_RANGE_MIN] >> 3));
+#ifdef WEBP_SWAP_16BIT_CSP
+  rgb[0] = gb;
+  rgb[1] = rg;
+#else
+  rgb[0] = rg;
+  rgb[1] = gb;
+#endif
 }
 
-static WEBP_INLINE void VP8YuvToBgr(uint8_t y, uint8_t u, uint8_t v,
-                                    uint8_t* const bgr) {
+static WEBP_INLINE void VP8YuvToRgba4444(int y, int u, int v,
+                                         uint8_t* const argb) {
   const int r_off = VP8kVToR[v];
   const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
   const int b_off = VP8kUToB[u];
-  bgr[0] = VP8kClip[y + b_off - YUV_RANGE_MIN];
-  bgr[1] = VP8kClip[y + g_off - YUV_RANGE_MIN];
-  bgr[2] = VP8kClip[y + r_off - YUV_RANGE_MIN];
+  const int rg = ((VP8kClip4Bits[y + r_off - YUV_RANGE_MIN] << 4) |
+                   VP8kClip4Bits[y + g_off - YUV_RANGE_MIN]);
+  const int ba = (VP8kClip4Bits[y + b_off - YUV_RANGE_MIN] << 4) | 0x0f;
+#ifdef WEBP_SWAP_16BIT_CSP
+  argb[0] = ba;
+  argb[1] = rg;
+#else
+  argb[0] = rg;
+  argb[1] = ba;
+#endif
+}
+
+#endif  // WEBP_YUV_USE_TABLE
+
+//-----------------------------------------------------------------------------
+// Alpha handling variants
+
+static WEBP_INLINE void VP8YuvToArgb(uint8_t y, uint8_t u, uint8_t v,
+                                     uint8_t* const argb) {
+  argb[0] = 0xff;
+  VP8YuvToRgb(y, u, v, argb + 1);
 }
 
 static WEBP_INLINE void VP8YuvToBgra(uint8_t y, uint8_t u, uint8_t v,
@@ -93,35 +240,79 @@ static WEBP_INLINE void VP8YuvToRgba(uint8_t y, uint8_t u, uint8_t v,
 // Must be called before everything, to initialize the tables.
 void VP8YUVInit(void);
 
+//-----------------------------------------------------------------------------
+// SSE2 extra functions (mostly for upsampling_sse2.c)
+
+#if defined(WEBP_USE_SSE2)
+
+// When the following is defined, tables are initialized statically, adding ~12k
+// to the binary size. Otherwise, they are initialized at run-time (small cost).
+#define WEBP_YUV_USE_SSE2_TABLES
+
+// Process 32 pixels and store the result (24b or 32b per pixel) in *dst.
+void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                    uint8_t* dst);
+void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                   uint8_t* dst);
+void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                    uint8_t* dst);
+void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                   uint8_t* dst);
+
+// Must be called to initialize tables before using the functions.
+void VP8YUVInitSSE2(void);
+
+#endif    // WEBP_USE_SSE2
+
 //------------------------------------------------------------------------------
 // RGB -> YUV conversion
-// The exact naming is Y'CbCr, following the ITU-R BT.601 standard.
-// More information at: http://en.wikipedia.org/wiki/YCbCr
-// Y = 0.2569 * R + 0.5044 * G + 0.0979 * B + 16
-// U = -0.1483 * R - 0.2911 * G + 0.4394 * B + 128
-// V = 0.4394 * R - 0.3679 * G - 0.0715 * B + 128
-// We use 16bit fixed point operations.
 
-static WEBP_INLINE int VP8ClipUV(int v) {
-   v = (v + (257 << (YUV_FIX + 2 - 1))) >> (YUV_FIX + 2);
-   return ((v & ~0xff) == 0) ? v : (v < 0) ? 0 : 255;
+// Stub functions that can be called with various rounding values:
+static WEBP_INLINE int VP8ClipUV(int uv, int rounding) {
+  uv = (uv + rounding + (128 << (YUV_FIX + 2))) >> (YUV_FIX + 2);
+  return ((uv & ~0xff) == 0) ? uv : (uv < 0) ? 0 : 255;
 }
 
-static WEBP_INLINE int VP8RGBToY(int r, int g, int b) {
-  const int kRound = (1 << (YUV_FIX - 1)) + (16 << YUV_FIX);
+#ifndef USE_YUVj
+
+static WEBP_INLINE int VP8RGBToY(int r, int g, int b, int rounding) {
   const int luma = 16839 * r + 33059 * g + 6420 * b;
-  return (luma + kRound) >> YUV_FIX;  // no need to clip
+  return (luma + rounding + (16 << YUV_FIX)) >> YUV_FIX;  // no need to clip
 }
 
-static WEBP_INLINE int VP8RGBToU(int r, int g, int b) {
-  return VP8ClipUV(-9719 * r - 19081 * g + 28800 * b);
+static WEBP_INLINE int VP8RGBToU(int r, int g, int b, int rounding) {
+  const int u = -9719 * r - 19081 * g + 28800 * b;
+  return VP8ClipUV(u, rounding);
 }
 
-static WEBP_INLINE int VP8RGBToV(int r, int g, int b) {
-  return VP8ClipUV(+28800 * r - 24116 * g - 4684 * b);
+static WEBP_INLINE int VP8RGBToV(int r, int g, int b, int rounding) {
+  const int v = +28800 * r - 24116 * g - 4684 * b;
+  return VP8ClipUV(v, rounding);
+}
+
+#else
+
+// This JPEG-YUV colorspace, only for comparison!
+// These are also 16bit precision coefficients from Rec.601, but with full
+// [0..255] output range.
+static WEBP_INLINE int VP8RGBToY(int r, int g, int b, int rounding) {
+  const int luma = 19595 * r + 38470 * g + 7471 * b;
+  return (luma + rounding) >> YUV_FIX;  // no need to clip
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
+static WEBP_INLINE int VP8RGBToU(int r, int g, int b, int rounding) {
+  const int u = -11058 * r - 21710 * g + 32768 * b;
+  return VP8ClipUV(u, rounding);
+}
+
+static WEBP_INLINE int VP8RGBToV(int r, int g, int b, int rounding) {
+  const int v = 32768 * r - 27439 * g - 5329 * b;
+  return VP8ClipUV(v, rounding);
+}
+
+#endif    // USE_YUVj
+
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/enc/alpha.c b/drivers/webp/enc/alpha.c
index e554eb7f30..fad6346e43 100644
--- a/drivers/webp/enc/alpha.c
+++ b/drivers/webp/enc/alpha.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
@@ -13,13 +15,11 @@
 #include <stdlib.h>
 
 #include "./vp8enci.h"
+#include "../dsp/dsp.h"
 #include "../utils/filters.h"
 #include "../utils/quant_levels.h"
-#include "../format_constants.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "../utils/utils.h"
+#include "../webp/format_constants.h"
 
 // -----------------------------------------------------------------------------
 // Encodes the given alpha data via specified compression method 'method'.
@@ -36,7 +36,7 @@ extern "C" {
 //
 // 'output' corresponds to the buffer containing compressed alpha data.
 //          This buffer is allocated by this method and caller should call
-//          free(*output) when done.
+//          WebPSafeFree(*output) when done.
 // 'output_size' corresponds to size of this compressed alpha buffer.
 //
 // Returns 1 on successfully encoding the alpha and
@@ -48,12 +48,11 @@ extern "C" {
 
 static int EncodeLossless(const uint8_t* const data, int width, int height,
                           int effort_level,  // in [0..6] range
-                          VP8BitWriter* const bw,
+                          VP8LBitWriter* const bw,
                           WebPAuxStats* const stats) {
   int ok = 0;
   WebPConfig config;
   WebPPicture picture;
-  VP8LBitWriter tmp_bw;
 
   WebPPictureInit(&picture);
   picture.width = width;
@@ -63,53 +62,51 @@ static int EncodeLossless(const uint8_t* const data, int width, int height,
   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) | 0xff000000u;
-      }
-      src += width;
-      dst += picture.argb_stride;
-    }
-  }
+  WebPDispatchAlphaToGreen(data, width, picture.width, picture.height,
+                           picture.argb, picture.argb_stride);
 
   WebPConfigInit(&config);
   config.lossless = 1;
   config.method = effort_level;  // impact is very small
-  // Set moderate default quality setting for alpha. Higher qualities (80 and
-  // above) could be very slow.
-  config.quality = 10.f + 15.f * effort_level;
-  if (config.quality > 100.f) config.quality = 100.f;
+  // 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);
+  ok = (VP8LEncodeStream(&config, &picture, bw) == VP8_ENC_OK);
   WebPPictureFree(&picture);
-  if (ok) {
-    const uint8_t* const data = VP8LBitWriterFinish(&tmp_bw);
-    const size_t data_size = VP8LBitWriterNumBytes(&tmp_bw);
-    VP8BitWriterAppend(bw, data, data_size);
+  ok = ok && !bw->error_;
+  if (!ok) {
+    VP8LBitWriterWipeOut(bw);
+    return 0;
   }
-  VP8LBitWriterDestroy(&tmp_bw);
-  return ok && !bw->error_;
+  return 1;
 }
 
 // -----------------------------------------------------------------------------
 
+// Small struct to hold the result of a filter mode compression attempt.
+typedef struct {
+  size_t score;
+  VP8BitWriter bw;
+  WebPAuxStats stats;
+} FilterTrial;
+
+// This function always returns an initialized 'bw' object, even upon error.
 static int EncodeAlphaInternal(const uint8_t* const data, int width, int height,
                                int method, int filter, int reduce_levels,
                                int effort_level,  // in [0..6] range
                                uint8_t* const tmp_alpha,
-                               VP8BitWriter* const bw,
-                               WebPAuxStats* const stats) {
+                               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;
+  const uint8_t* output = NULL;
+  size_t output_size = 0;
+  VP8LBitWriter tmp_bw;
 
   assert((uint64_t)data_size == (uint64_t)width * height);  // as per spec
   assert(filter >= 0 && filter < WEBP_FILTER_LAST);
@@ -118,43 +115,163 @@ static int EncodeAlphaInternal(const uint8_t* const data, int width, int height,
   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(bw, expected_size);
-  VP8BitWriterAppend(bw, &header, ALPHA_HEADER_LEN);
-
   filter_func = WebPFilters[filter];
-  if (filter_func) {
-    filter_func(data, width, height, 1, width, tmp_alpha);
+  if (filter_func != NULL) {
+    filter_func(data, width, height, width, tmp_alpha);
     alpha_src = tmp_alpha;
   }  else {
     alpha_src = data;
   }
 
+  if (method != ALPHA_NO_COMPRESSION) {
+    ok = VP8LBitWriterInit(&tmp_bw, data_size >> 3);
+    ok = ok && EncodeLossless(alpha_src, width, height, effort_level,
+                              &tmp_bw, &result->stats);
+    if (ok) {
+      output = VP8LBitWriterFinish(&tmp_bw);
+      output_size = VP8LBitWriterNumBytes(&tmp_bw);
+      if (output_size > data_size) {
+        // compressed size is larger than source! Revert to uncompressed mode.
+        method = ALPHA_NO_COMPRESSION;
+        VP8LBitWriterWipeOut(&tmp_bw);
+      }
+    } else {
+      VP8LBitWriterWipeOut(&tmp_bw);
+      return 0;
+    }
+  }
+
   if (method == ALPHA_NO_COMPRESSION) {
-    ok = VP8BitWriterAppend(bw, alpha_src, width * height);
-    ok = ok && !bw->error_;
-  } else {
-    ok = EncodeLossless(alpha_src, width, height, effort_level, bw, stats);
-    VP8BitWriterFinish(bw);
+    output = alpha_src;
+    output_size = data_size;
+    ok = 1;
+  }
+
+  // Emit final result.
+  header = method | (filter << 2);
+  if (reduce_levels) header |= ALPHA_PREPROCESSED_LEVELS << 4;
+
+  VP8BitWriterInit(&result->bw, ALPHA_HEADER_LEN + output_size);
+  ok = ok && VP8BitWriterAppend(&result->bw, &header, ALPHA_HEADER_LEN);
+  ok = ok && VP8BitWriterAppend(&result->bw, output, output_size);
+
+  if (method != ALPHA_NO_COMPRESSION) {
+    VP8LBitWriterWipeOut(&tmp_bw);
   }
+  ok = ok && !result->bw.error_;
+  result->score = VP8BitWriterSize(&result->bw);
   return ok;
 }
 
 // -----------------------------------------------------------------------------
 
-// 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
+        : WebPEstimateBestFilter(alpha, width, height, width);
+    bit_map |= 1 << filter;
+    // For large number of colors, try FILTER_NONE in addition to the best
+    // filter as well.
+    if (try_filter_none || num_colors > kMaxColorsForFilterNone) {
+      bit_map |= FILTER_TRY_NONE;
+    }
+  } else if (filter == WEBP_FILTER_NONE) {
+    bit_map = FILTER_TRY_NONE;
+  } else {  // WEBP_FILTER_BEST -> try all
+    bit_map = FILTER_TRY_ALL;
+  }
+  return bit_map;
+}
+
+static void InitFilterTrial(FilterTrial* const score) {
+  score->score = (size_t)~0U;
+  VP8BitWriterInit(&score->bw, 0);
+}
+
+static int ApplyFiltersAndEncode(const uint8_t* alpha, int width, int height,
+                                 size_t data_size, int method, int filter,
+                                 int reduce_levels, int effort_level,
+                                 uint8_t** const output,
+                                 size_t* const output_size,
+                                 WebPAuxStats* const stats) {
+  int ok = 1;
+  FilterTrial best;
+  uint32_t try_map =
+      GetFilterMap(alpha, width, height, filter, effort_level);
+  InitFilterTrial(&best);
+
+  if (try_map != FILTER_TRY_NONE) {
+    uint8_t* filtered_alpha =  (uint8_t*)WebPSafeMalloc(1ULL, data_size);
+    if (filtered_alpha == NULL) return 0;
+
+    for (filter = WEBP_FILTER_NONE; ok && try_map; ++filter, try_map >>= 1) {
+      if (try_map & 1) {
+        FilterTrial trial;
+        ok = EncodeAlphaInternal(alpha, width, height, method, filter,
+                                 reduce_levels, effort_level, filtered_alpha,
+                                 &trial);
+        if (ok && trial.score < best.score) {
+          VP8BitWriterWipeOut(&best.bw);
+          best = trial;
+        } else {
+          VP8BitWriterWipeOut(&trial.bw);
+        }
+      }
+    }
+    WebPSafeFree(filtered_alpha);
+  } else {
+    ok = EncodeAlphaInternal(alpha, width, height, method, WEBP_FILTER_NONE,
+                             reduce_levels, effort_level, NULL, &best);
+  }
+  if (ok) {
+    if (stats != NULL) {
+      stats->lossless_features = best.stats.lossless_features;
+      stats->histogram_bits = best.stats.histogram_bits;
+      stats->transform_bits = best.stats.transform_bits;
+      stats->cache_bits = best.stats.cache_bits;
+      stats->palette_size = best.stats.palette_size;
+      stats->lossless_size = best.stats.lossless_size;
+      stats->lossless_hdr_size = best.stats.lossless_hdr_size;
+      stats->lossless_data_size = best.stats.lossless_data_size;
+    }
+    *output_size = VP8BitWriterSize(&best.bw);
+    *output = VP8BitWriterBuf(&best.bw);
+  } else {
+    VP8BitWriterWipeOut(&best.bw);
+  }
+  return ok;
 }
 
 static int EncodeAlpha(VP8Encoder* const enc,
@@ -187,13 +304,18 @@ static int EncodeAlpha(VP8Encoder* const enc,
     return 0;
   }
 
-  quant_alpha = (uint8_t*)malloc(data_size);
+  if (method == ALPHA_NO_COMPRESSION) {
+    // Don't filter, as filtering will make no impact on compressed size.
+    filter = WEBP_FILTER_NONE;
+  }
+
+  quant_alpha = (uint8_t*)WebPSafeMalloc(1ULL, data_size);
   if (quant_alpha == NULL) {
     return 0;
   }
 
   // Extract alpha data (width x height) from raw_data (stride x height).
-  CopyPlane(pic->a, pic->a_stride, quant_alpha, width, width, height);
+  WebPCopyPlane(pic->a, pic->a_stride, quant_alpha, width, width, height);
 
   if (reduce_levels) {  // No Quantization required for 'quality = 100'.
     // 16 alpha levels gives quite a low MSE w.r.t original alpha plane hence
@@ -205,126 +327,99 @@ static int EncodeAlpha(VP8Encoder* const enc,
   }
 
   if (ok) {
-    VP8BitWriter bw;
-    int test_filter;
-    uint8_t* filtered_alpha = NULL;
-
-    // We always test WEBP_FILTER_NONE first.
-    ok = EncodeAlphaInternal(quant_alpha, width, height,
-                             method, WEBP_FILTER_NONE, reduce_levels,
-                             effort_level, NULL, &bw, pic->stats);
-    if (!ok) {
-      VP8BitWriterWipeOut(&bw);
-      goto End;
-    }
-
-    if (filter == WEBP_FILTER_FAST) {  // Quick estimate of a second candidate?
-      filter = EstimateBestFilter(quant_alpha, width, height, width);
-    }
-    // Stop?
-    if (filter == WEBP_FILTER_NONE) {
-      goto Ok;
-    }
-
-    filtered_alpha = (uint8_t*)malloc(data_size);
-    ok = (filtered_alpha != NULL);
-    if (!ok) {
-      goto End;
+    VP8FiltersInit();
+    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;
     }
-
-    // Try the other mode(s).
-    {
-      WebPAuxStats best_stats;
-      size_t best_score = VP8BitWriterSize(&bw);
-
-      memset(&best_stats, 0, sizeof(best_stats));  // prevent spurious warning
-      if (pic->stats != NULL) best_stats = *pic->stats;
-      for (test_filter = WEBP_FILTER_HORIZONTAL;
-           ok && (test_filter <= WEBP_FILTER_GRADIENT);
-           ++test_filter) {
-        VP8BitWriter tmp_bw;
-        if (filter != WEBP_FILTER_BEST && test_filter != filter) {
-          continue;
-        }
-        ok = EncodeAlphaInternal(quant_alpha, width, height,
-                                 method, test_filter, reduce_levels,
-                                 effort_level, filtered_alpha, &tmp_bw,
-                                 pic->stats);
-        if (ok) {
-          const size_t score = VP8BitWriterSize(&tmp_bw);
-          if (score < best_score) {
-            // swap bitwriter objects.
-            VP8BitWriter tmp = tmp_bw;
-            tmp_bw = bw;
-            bw = tmp;
-            best_score = score;
-            if (pic->stats != NULL) best_stats = *pic->stats;
-          }
-        } else {
-          VP8BitWriterWipeOut(&bw);
-        }
-        VP8BitWriterWipeOut(&tmp_bw);
-      }
-      if (pic->stats != NULL) *pic->stats = best_stats;
-    }
- Ok:
-    if (ok) {
-      *output_size = VP8BitWriterSize(&bw);
-      *output = VP8BitWriterBuf(&bw);
-      if (pic->stats != NULL) {         // need stats?
-        pic->stats->coded_size += (int)(*output_size);
-        enc->sse_[3] = sse;
-      }
-    }
-    free(filtered_alpha);
   }
- End:
-  free(quant_alpha);
+
+  WebPSafeFree(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.
+    WebPSafeFree(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) {
+  WebPInitAlphaProcessing();
   enc->has_alpha_ = WebPPictureHasTransparency(enc->pic_);
   enc->alpha_data_ = NULL;
   enc->alpha_data_size_ = 0;
+  if (enc->thread_level_ > 0) {
+    WebPWorker* const worker = &enc->alpha_worker_;
+    WebPGetWorkerInterface()->Init(worker);
+    worker->data1 = enc;
+    worker->data2 = NULL;
+    worker->hook = (WebPWorkerHook)CompressAlphaJob;
+  }
 }
 
-int VP8EncFinishAlpha(VP8Encoder* const enc) {
+int VP8EncStartAlpha(VP8Encoder* const enc) {
   if (enc->has_alpha_) {
-    const WebPConfig* config = enc->config_;
-    uint8_t* tmp_data = NULL;
-    size_t tmp_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, &tmp_data, &tmp_size)) {
-      return 0;
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      // Makes sure worker is good to go.
+      if (!WebPGetWorkerInterface()->Reset(worker)) {
+        return 0;
+      }
+      WebPGetWorkerInterface()->Launch(worker);
+      return 1;
+    } else {
+      return CompressAlphaJob(enc, NULL);   // just do the job right away
     }
-    if (tmp_size != (uint32_t)tmp_size) {  // Sanity check.
-      free(tmp_data);
-      return 0;
+  }
+  return 1;
+}
+
+int VP8EncFinishAlpha(VP8Encoder* const enc) {
+  if (enc->has_alpha_) {
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      if (!WebPGetWorkerInterface()->Sync(worker)) return 0;  // error
     }
-    enc->alpha_data_size_ = (uint32_t)tmp_size;
-    enc->alpha_data_ = tmp_data;
   }
   return WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
 }
 
-void VP8EncDeleteAlpha(VP8Encoder* const enc) {
-  free(enc->alpha_data_);
+int VP8EncDeleteAlpha(VP8Encoder* const enc) {
+  int ok = 1;
+  if (enc->thread_level_ > 0) {
+    WebPWorker* const worker = &enc->alpha_worker_;
+    // finish anything left in flight
+    ok = WebPGetWorkerInterface()->Sync(worker);
+    // still need to end the worker, even if !ok
+    WebPGetWorkerInterface()->End(worker);
+  }
+  WebPSafeFree(enc->alpha_data_);
   enc->alpha_data_ = NULL;
   enc->alpha_data_size_ = 0;
   enc->has_alpha_ = 0;
+  return ok;
 }
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/analysis.c b/drivers/webp/enc/analysis.c
index 22cfb492e7..b55128fd48 100644
--- a/drivers/webp/enc/analysis.c
+++ b/drivers/webp/enc/analysis.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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
@@ -17,16 +19,8 @@
 #include "./cost.h"
 #include "../utils/utils.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 #define MAX_ITERS_K_MEANS  6
 
-static int ClipAlpha(int alpha) {
-  return alpha < 0 ? 0 : alpha > 255 ? 255 : alpha;
-}
-
 //------------------------------------------------------------------------------
 // Smooth the segment map by replacing isolated block by the majority of its
 // neighbours.
@@ -36,7 +30,7 @@ static void SmoothSegmentMap(VP8Encoder* const enc) {
   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));
+  uint8_t* const tmp = (uint8_t*)WebPSafeMalloc(w * h, sizeof(*tmp));
   assert((uint64_t)(w * h) == (uint64_t)w * h);   // no overflow, as per spec
 
   if (tmp == NULL) return;
@@ -57,6 +51,7 @@ static void SmoothSegmentMap(VP8Encoder* const enc) {
       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;
@@ -68,54 +63,14 @@ static void SmoothSegmentMap(VP8Encoder* const enc) {
       mb->segment_ = tmp[x + y * w];
     }
   }
-  free(tmp);
+  WebPSafeFree(tmp);
 }
 
 //------------------------------------------------------------------------------
-// Finalize Segment probability based on the coding tree
-
-static int GetProba(int a, int b) {
-  int proba;
-  const int total = a + b;
-  if (total == 0) return 255;  // that's the default probability.
-  proba = (255 * a + total / 2) / total;
-  return 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) {
-    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;
-  }
-}
+// set segment susceptibility alpha_ / beta_
 
 static WEBP_INLINE int clip(int v, int m, int M) {
-  return v < m ? m : v > M ? M : v;
+  return (v < m) ? m : (v > M) ? M : v;
 }
 
 static void SetSegmentAlphas(VP8Encoder* const enc,
@@ -142,28 +97,77 @@ static void SetSegmentAlphas(VP8Encoder* const enc,
 }
 
 //------------------------------------------------------------------------------
+// Compute susceptibility based on DCT-coeff histograms:
+// the higher, the "easier" the macroblock is to compress.
+
+#define MAX_ALPHA 255                // 8b of precision for susceptibilities.
+#define ALPHA_SCALE (2 * MAX_ALPHA)  // scaling factor for alpha.
+#define DEFAULT_ALPHA (-1)
+#define IS_BETTER_ALPHA(alpha, best_alpha) ((alpha) > (best_alpha))
+
+static int FinalAlphaValue(int alpha) {
+  alpha = MAX_ALPHA - alpha;
+  return clip(alpha, 0, MAX_ALPHA);
+}
+
+static int GetAlpha(const VP8Histogram* const histo) {
+  // 'alpha' will later be clipped to [0..MAX_ALPHA] range, clamping outer
+  // values which happen to be mostly noise. This leaves the maximum precision
+  // for handling the useful small values which contribute most.
+  const int max_value = histo->max_value;
+  const int last_non_zero = histo->last_non_zero;
+  const int alpha =
+      (max_value > 1) ? ALPHA_SCALE * last_non_zero / max_value : 0;
+  return alpha;
+}
+
+static void InitHistogram(VP8Histogram* const histo) {
+  histo->max_value = 0;
+  histo->last_non_zero = 1;
+}
+
+static void MergeHistograms(const VP8Histogram* const in,
+                            VP8Histogram* const out) {
+  if (in->max_value > out->max_value) {
+    out->max_value = in->max_value;
+  }
+  if (in->last_non_zero > out->last_non_zero) {
+    out->last_non_zero = in->last_non_zero;
+  }
+}
+
+//------------------------------------------------------------------------------
 // Simplified k-Means, to assign Nb segments based on alpha-histogram
 
-static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
-  const int nb = enc->segment_hdr_.num_segments_;
+static void AssignSegments(VP8Encoder* const enc,
+                           const int alphas[MAX_ALPHA + 1]) {
+  // 'num_segments_' is previously validated and <= NUM_MB_SEGMENTS, but an
+  // explicit check is needed to avoid spurious warning about 'n + 1' exceeding
+  // array bounds of 'centers' with some compilers (noticed with gcc-4.9).
+  const int nb = (enc->segment_hdr_.num_segments_ < NUM_MB_SEGMENTS) ?
+                 enc->segment_hdr_.num_segments_ : NUM_MB_SEGMENTS;
   int centers[NUM_MB_SEGMENTS];
   int weighted_average = 0;
-  int map[256];
+  int map[MAX_ALPHA + 1];
   int a, n, k;
-  int min_a = 0, max_a = 255, range_a;
+  int min_a = 0, max_a = MAX_ALPHA, range_a;
   // 'int' type is ok for histo, and won't overflow
   int accum[NUM_MB_SEGMENTS], dist_accum[NUM_MB_SEGMENTS];
 
+  assert(nb >= 1);
+  assert(nb <= NUM_MB_SEGMENTS);
+
   // bracket the input
-  for (n = 0; n < 256 && alphas[n] == 0; ++n) {}
+  for (n = 0; n <= MAX_ALPHA && alphas[n] == 0; ++n) {}
   min_a = n;
-  for (n = 255; n > min_a && alphas[n] == 0; --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 (n = 1, k = 0; n < 2 * nb; n += 2) {
-    centers[k++] = min_a + (n * range_a) / (2 * nb);
+  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
@@ -178,7 +182,7 @@ static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
     n = 0;    // track the nearest center for current 'a'
     for (a = min_a; a <= max_a; ++a) {
       if (alphas[a]) {
-        while (n < nb - 1 && abs(a - centers[n + 1]) < abs(a - centers[n])) {
+        while (n + 1 < nb && abs(a - centers[n + 1]) < abs(a - centers[n])) {
           n++;
         }
         map[a] = n;
@@ -210,7 +214,7 @@ static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
     VP8MBInfo* const mb = &enc->mb_info_[n];
     const int alpha = mb->alpha_;
     mb->segment_ = map[alpha];
-    mb->alpha_ = centers[map[alpha]];     // just for the record.
+    mb->alpha_ = centers[map[alpha]];  // for the record.
   }
 
   if (nb > 1) {
@@ -218,7 +222,6 @@ static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
     if (smooth) SmoothSegmentMap(enc);
   }
 
-  SetSegmentProbas(enc);                             // Assign final proba
   SetSegmentAlphas(enc, centers, weighted_average);  // pick some alphas.
 }
 
@@ -227,24 +230,30 @@ static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
 // 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,
-// we don't need to test all the possible modes during the analysis phase.
+// Number of modes to inspect for alpha_ evaluation. We don't need to test all
+// the possible modes during the analysis phase: we risk falling into a local
+// optimum, or be subject to boundary effect
 #define MAX_INTRA16_MODE 2
 #define MAX_INTRA4_MODE  2
 #define MAX_UV_MODE      2
 
 static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
-  const int max_mode = (it->enc_->method_ >= 3) ? MAX_INTRA16_MODE : 4;
+  const int max_mode = MAX_INTRA16_MODE;
   int mode;
-  int best_alpha = -1;
+  int best_alpha = DEFAULT_ALPHA;
   int best_mode = 0;
 
   VP8MakeLuma16Preds(it);
   for (mode = 0; mode < max_mode; ++mode) {
-    const int alpha = VP8CollectHistogram(it->yuv_in_ + Y_OFF,
-                                          it->yuv_p_ + VP8I16ModeOffsets[mode],
-                                          0, 16);
-    if (alpha > best_alpha) {
+    VP8Histogram histo;
+    int alpha;
+
+    InitHistogram(&histo);
+    VP8CollectHistogram(it->yuv_in_ + Y_OFF_ENC,
+                        it->yuv_p_ + VP8I16ModeOffsets[mode],
+                        0, 16, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
       best_alpha = alpha;
       best_mode = mode;
     }
@@ -256,46 +265,62 @@ static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
 static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
                                    int best_alpha) {
   uint8_t modes[16];
-  const int max_mode = (it->enc_->method_ >= 3) ? MAX_INTRA4_MODE : NUM_BMODES;
-  int i4_alpha = 0;
+  const int max_mode = MAX_INTRA4_MODE;
+  int i4_alpha;
+  VP8Histogram total_histo;
+  int cur_histo = 0;
+  InitHistogram(&total_histo);
+
   VP8IteratorStartI4(it);
   do {
     int mode;
-    int best_mode_alpha = -1;
-    const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
+    int best_mode_alpha = DEFAULT_ALPHA;
+    VP8Histogram histos[2];
+    const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC + VP8Scan[it->i4_];
 
     VP8MakeIntra4Preds(it);
     for (mode = 0; mode < max_mode; ++mode) {
-      const int alpha = VP8CollectHistogram(src,
-                                            it->yuv_p_ + VP8I4ModeOffsets[mode],
-                                            0, 1);
-      if (alpha > best_mode_alpha) {
+      int alpha;
+
+      InitHistogram(&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.
       }
     }
-    i4_alpha += best_mode_alpha;
+    // 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));
+  } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF_ENC));
 
-  if (i4_alpha > best_alpha) {
+  i4_alpha = GetAlpha(&total_histo);
+  if (IS_BETTER_ALPHA(i4_alpha, best_alpha)) {
     VP8SetIntra4Mode(it, modes);
-    best_alpha = ClipAlpha(i4_alpha);
+    best_alpha = i4_alpha;
   }
   return best_alpha;
 }
 
 static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
-  int best_alpha = -1;
+  int best_alpha = DEFAULT_ALPHA;
   int best_mode = 0;
-  const int max_mode = (it->enc_->method_ >= 3) ? MAX_UV_MODE : 4;
+  const int max_mode = MAX_UV_MODE;
   int mode;
+
   VP8MakeChroma8Preds(it);
   for (mode = 0; mode < max_mode; ++mode) {
-    const int alpha = VP8CollectHistogram(it->yuv_in_ + U_OFF,
-                                          it->yuv_p_ + VP8UVModeOffsets[mode],
-                                          16, 16 + 4 + 4);
-    if (alpha > best_alpha) {
+    VP8Histogram histo;
+    int alpha;
+    InitHistogram(&histo);
+    VP8CollectHistogram(it->yuv_in_ + U_OFF_ENC,
+                        it->yuv_p_ + VP8UVModeOffsets[mode],
+                        16, 16 + 4 + 4, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
       best_alpha = alpha;
       best_mode = mode;
     }
@@ -305,7 +330,8 @@ static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
 }
 
 static void MBAnalyze(VP8EncIterator* const it,
-                      int alphas[256], int* const uv_alpha) {
+                      int alphas[MAX_ALPHA + 1],
+                      int* const alpha, int* const uv_alpha) {
   const VP8Encoder* const enc = it->enc_;
   int best_alpha, best_uv_alpha;
 
@@ -314,7 +340,7 @@ static void MBAnalyze(VP8EncIterator* const it,
   VP8SetSegment(it, 0);      // default segment, spec-wise.
 
   best_alpha = MBAnalyzeBestIntra16Mode(it);
-  if (enc->method_ != 3) {
+  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.
@@ -324,10 +350,22 @@ static void MBAnalyze(VP8EncIterator* const it,
   best_uv_alpha = MBAnalyzeBestUVMode(it);
 
   // Final susceptibility mix
-  best_alpha = (best_alpha + best_uv_alpha + 1) / 2;
+  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;
-  it->mb_->alpha_ = best_alpha;   // Informative only.
+}
+
+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;
 }
 
 //------------------------------------------------------------------------------
@@ -340,25 +378,124 @@ static void MBAnalyze(VP8EncIterator* const it,
 // and decide intra4/intra16, but that's usually almost always a bad choice at
 // this stage.
 
-int VP8EncAnalyze(VP8Encoder* const enc) {
-  int ok = 1;
-  int alphas[256] = { 0 };
-  VP8EncIterator it;
-
-  VP8IteratorInit(enc, &it);
+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;
-  do {
-    VP8IteratorImport(&it);
-    MBAnalyze(&it, alphas, &enc->uv_alpha_);
-    ok = VP8IteratorProgress(&it, 20);
-    // Let's pretend we have perfect lossless reconstruction.
-  } while (ok && VP8IteratorNext(&it, it.yuv_in_));
-  enc->uv_alpha_ /= enc->mb_w_ * enc->mb_h_;
-  if (ok) AssignSegments(enc, alphas);
+  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 + WEBP_ALIGN_CST];
+    uint8_t* const scratch = (uint8_t*)WEBP_ALIGN(tmp);
+    do {
+      // Let's pretend we have perfect lossless reconstruction.
+      VP8IteratorImport(it, scratch);
+      MBAnalyze(it, job->alphas, &job->alpha, &job->uv_alpha);
+      ok = VP8IteratorProgress(it, job->delta_progress);
+    } while (ok && VP8IteratorNext(it));
+  }
   return ok;
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
+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) {
+  WebPGetWorkerInterface()->Init(&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
+    const WebPWorkerInterface* const worker_interface =
+        WebPGetWorkerInterface();
+    SegmentJob main_job;
+    if (do_mt) {
+      SegmentJob side_job;
+      // Note the use of '&' instead of '&&' because we must call the functions
+      // no matter what.
+      InitSegmentJob(enc, &main_job, 0, split_row);
+      InitSegmentJob(enc, &side_job, split_row, last_row);
+      // we don't need to call Reset() on main_job.worker, since we're calling
+      // WebPWorkerExecute() on it
+      ok &= worker_interface->Reset(&side_job.worker);
+      // launch the two jobs in parallel
+      if (ok) {
+        worker_interface->Launch(&side_job.worker);
+        worker_interface->Execute(&main_job.worker);
+        ok &= worker_interface->Sync(&side_job.worker);
+        ok &= worker_interface->Sync(&main_job.worker);
+      }
+      worker_interface->End(&side_job.worker);
+      if (ok) MergeJobs(&side_job, &main_job);  // merge results together
+    } else {
+      // Even for single-thread case, we use the generic Worker tools.
+      InitSegmentJob(enc, &main_job, 0, last_row);
+      worker_interface->Execute(&main_job.worker);
+      ok &= worker_interface->Sync(&main_job.worker);
+    }
+    worker_interface->End(&main_job.worker);
+    if (ok) {
+      enc->alpha_ = main_job.alpha / total_mb;
+      enc->uv_alpha_ = main_job.uv_alpha / total_mb;
+      AssignSegments(enc, main_job.alphas);
+    }
+  } else {   // Use only one default segment.
+    ResetAllMBInfo(enc);
+  }
+  return ok;
+}
+
diff --git a/drivers/webp/enc/backward_references.c b/drivers/webp/enc/backward_references.c
index b8c8ece806..049125e521 100644
--- a/drivers/webp/enc/backward_references.c
+++ b/drivers/webp/enc/backward_references.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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)
@@ -10,7 +12,6 @@
 
 #include <assert.h>
 #include <math.h>
-#include <stdio.h>
 
 #include "./backward_references.h"
 #include "./histogram.h"
@@ -20,9 +21,9 @@
 
 #define VALUES_IN_BYTE 256
 
-#define HASH_BITS 18
-#define HASH_SIZE (1 << HASH_BITS)
-#define HASH_MULTIPLIER (0xc6a4a7935bd1e995ULL)
+#define MIN_BLOCK_SIZE 256  // minimum block size for backward references
+
+#define MAX_ENTROPY    (1e30f)
 
 // 1M window (4M bytes) minus 120 special codes for short distances.
 #define WINDOW_SIZE ((1 << 20) - 120)
@@ -31,14 +32,6 @@
 #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] = {
@@ -65,145 +58,275 @@ static int DistanceToPlaneCode(int xsize, int dist) {
 
 static WEBP_INLINE int FindMatchLength(const uint32_t* const array1,
                                        const uint32_t* const array2,
-                                       const int max_limit) {
+                                       int best_len_match,
+                                       int max_limit) {
+#if !defined(__x86_64__)
+  // TODO(vrabaud): Compare on other architectures.
   int match_len = 0;
+  // Before 'expensive' linear match, check if the two arrays match at the
+  // current best length index.
+  if (array1[best_len_match] != array2[best_len_match]) return 0;
   while (match_len < max_limit && array1[match_len] == array2[match_len]) {
     ++match_len;
   }
   return match_len;
+#else
+  const uint32_t* array1_32 = array1;
+  const uint32_t* array2_32 = array2;
+  // max value is aligned to (uint64_t*) array1
+  const uint32_t* const array1_32_max = array1 + (max_limit & ~1);
+
+  // Before 'expensive' linear match, check if the two arrays match at the
+  // current best length index.
+  if (array1[best_len_match] != array2[best_len_match]) return 0;
+
+  // TODO(vrabaud): add __predict_true on bound checking?
+  while (array1_32 < array1_32_max) {
+    if (*(uint64_t*)array1_32 == *(uint64_t*)array2_32) {
+      array1_32 += 2;
+      array2_32 += 2;
+    } else {
+      // if the uint32_t pointed to are the same, then the following ones have
+      // to be different
+      return (array1_32 - array1) + (*array1_32 == *array2_32);
+    }
+  }
+
+  // Deal with the potential last uint32_t.
+  if ((max_limit & 1) && (*array1_32 != *array2_32)) return max_limit - 1;
+  return max_limit;
+#endif
 }
 
 // -----------------------------------------------------------------------------
 //  VP8LBackwardRefs
 
-void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs) {
-  if (refs != NULL) {
-    refs->refs = NULL;
-    refs->size = 0;
-    refs->max_size = 0;
+struct PixOrCopyBlock {
+  PixOrCopyBlock* next_;   // next block (or NULL)
+  PixOrCopy* start_;       // data start
+  int size_;               // currently used size
+};
+
+static void ClearBackwardRefs(VP8LBackwardRefs* const refs) {
+  assert(refs != NULL);
+  if (refs->tail_ != NULL) {
+    *refs->tail_ = refs->free_blocks_;  // recycle all blocks at once
   }
+  refs->free_blocks_ = refs->refs_;
+  refs->tail_ = &refs->refs_;
+  refs->last_block_ = NULL;
+  refs->refs_ = NULL;
 }
 
-void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs) {
-  if (refs != NULL) {
-    free(refs->refs);
-    VP8LInitBackwardRefs(refs);
+void VP8LBackwardRefsClear(VP8LBackwardRefs* const refs) {
+  assert(refs != NULL);
+  ClearBackwardRefs(refs);
+  while (refs->free_blocks_ != NULL) {
+    PixOrCopyBlock* const next = refs->free_blocks_->next_;
+    WebPSafeFree(refs->free_blocks_);
+    refs->free_blocks_ = next;
   }
 }
 
-int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size) {
+void VP8LBackwardRefsInit(VP8LBackwardRefs* const refs, int block_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;
+  memset(refs, 0, sizeof(*refs));
+  refs->tail_ = &refs->refs_;
+  refs->block_size_ =
+      (block_size < MIN_BLOCK_SIZE) ? MIN_BLOCK_SIZE : block_size;
+}
+
+VP8LRefsCursor VP8LRefsCursorInit(const VP8LBackwardRefs* const refs) {
+  VP8LRefsCursor c;
+  c.cur_block_ = refs->refs_;
+  if (refs->refs_ != NULL) {
+    c.cur_pos = c.cur_block_->start_;
+    c.last_pos_ = c.cur_pos + c.cur_block_->size_;
+  } else {
+    c.cur_pos = NULL;
+    c.last_pos_ = NULL;
+  }
+  return c;
+}
+
+void VP8LRefsCursorNextBlock(VP8LRefsCursor* const c) {
+  PixOrCopyBlock* const b = c->cur_block_->next_;
+  c->cur_pos = (b == NULL) ? NULL : b->start_;
+  c->last_pos_ = (b == NULL) ? NULL : b->start_ + b->size_;
+  c->cur_block_ = b;
+}
+
+// Create a new block, either from the free list or allocated
+static PixOrCopyBlock* BackwardRefsNewBlock(VP8LBackwardRefs* const refs) {
+  PixOrCopyBlock* b = refs->free_blocks_;
+  if (b == NULL) {   // allocate new memory chunk
+    const size_t total_size =
+        sizeof(*b) + refs->block_size_ * sizeof(*b->start_);
+    b = (PixOrCopyBlock*)WebPSafeMalloc(1ULL, total_size);
+    if (b == NULL) {
+      refs->error_ |= 1;
+      return NULL;
+    }
+    b->start_ = (PixOrCopy*)((uint8_t*)b + sizeof(*b));  // not always aligned
+  } else {  // recycle from free-list
+    refs->free_blocks_ = b->next_;
+  }
+  *refs->tail_ = b;
+  refs->tail_ = &b->next_;
+  refs->last_block_ = b;
+  b->next_ = NULL;
+  b->size_ = 0;
+  return b;
+}
+
+static WEBP_INLINE void BackwardRefsCursorAdd(VP8LBackwardRefs* const refs,
+                                              const PixOrCopy v) {
+  PixOrCopyBlock* b = refs->last_block_;
+  if (b == NULL || b->size_ == refs->block_size_) {
+    b = BackwardRefsNewBlock(refs);
+    if (b == NULL) return;   // refs->error_ is set
+  }
+  b->start_[b->size_++] = v;
+}
+
+int VP8LBackwardRefsCopy(const VP8LBackwardRefs* const src,
+                         VP8LBackwardRefs* const dst) {
+  const PixOrCopyBlock* b = src->refs_;
+  ClearBackwardRefs(dst);
+  assert(src->block_size_ == dst->block_size_);
+  while (b != NULL) {
+    PixOrCopyBlock* const new_b = BackwardRefsNewBlock(dst);
+    if (new_b == NULL) return 0;   // dst->error_ is set
+    memcpy(new_b->start_, b->start_, b->size_ * sizeof(*b->start_));
+    new_b->size_ = b->size_;
+    b = b->next_;
+  }
   return 1;
 }
 
 // -----------------------------------------------------------------------------
 // Hash chains
 
-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) {
+// initialize as empty
+static void HashChainReset(VP8LHashChain* const p) {
   int i;
-  p->chain_ = (int*)WebPSafeMalloc((uint64_t)size, sizeof(*p->chain_));
-  if (p->chain_ == NULL) {
-    return 0;
-  }
-  for (i = 0; i < size; ++i) {
+  assert(p != NULL);
+  for (i = 0; i < p->size_; ++i) {
     p->chain_[i] = -1;
   }
   for (i = 0; i < HASH_SIZE; ++i) {
     p->hash_to_first_index_[i] = -1;
   }
+}
+
+int VP8LHashChainInit(VP8LHashChain* const p, int size) {
+  assert(p->size_ == 0);
+  assert(p->chain_ == NULL);
+  assert(size > 0);
+  p->chain_ = (int*)WebPSafeMalloc(size, sizeof(*p->chain_));
+  if (p->chain_ == NULL) return 0;
+  p->size_ = size;
+  HashChainReset(p);
   return 1;
 }
 
-static void HashChainDelete(HashChain* const p) {
-  if (p != NULL) {
-    free(p->chain_);
-    free(p);
-  }
+void VP8LHashChainClear(VP8LHashChain* const p) {
+  assert(p != NULL);
+  WebPSafeFree(p->chain_);
+  p->size_ = 0;
+  p->chain_ = NULL;
+}
+
+// -----------------------------------------------------------------------------
+
+#define HASH_MULTIPLIER_HI (0xc6a4a793U)
+#define HASH_MULTIPLIER_LO (0x5bd1e996U)
+
+static WEBP_INLINE uint32_t GetPixPairHash64(const uint32_t* const argb) {
+  uint32_t key;
+  key  = argb[1] * HASH_MULTIPLIER_HI;
+  key += argb[0] * HASH_MULTIPLIER_LO;
+  key = key >> (32 - HASH_BITS);
+  return key;
 }
 
 // Insertion of two pixels at a time.
-static void HashChainInsert(HashChain* const p,
+static void HashChainInsert(VP8LHashChain* const p,
                             const uint32_t* const argb, int pos) {
-  const uint64_t hash_code = GetPixPairHash64(argb);
+  const uint32_t hash_code = GetPixPairHash64(argb);
   p->chain_[pos] = p->hash_to_first_index_[hash_code];
   p->hash_to_first_index_[hash_code] = pos;
 }
 
-static int HashChainFindCopy(const HashChain* const p,
-                             int quality, int index, int xsize,
-                             const uint32_t* const argb, int maxlen,
+// Returns the maximum number of hash chain lookups to do for a
+// given compression quality. Return value in range [6, 86].
+static int GetMaxItersForQuality(int quality, int low_effort) {
+  return (low_effort ? 6 : 8) + (quality * quality) / 128;
+}
+
+static int GetWindowSizeForHashChain(int quality, int xsize) {
+  const int max_window_size = (quality > 75) ? WINDOW_SIZE
+                            : (quality > 50) ? (xsize << 8)
+                            : (quality > 25) ? (xsize << 6)
+                            : (xsize << 4);
+  assert(xsize > 0);
+  return (max_window_size > WINDOW_SIZE) ? WINDOW_SIZE : max_window_size;
+}
+
+static WEBP_INLINE int MaxFindCopyLength(int len) {
+  return (len < MAX_LENGTH) ? len : MAX_LENGTH;
+}
+
+static void HashChainFindOffset(const VP8LHashChain* const p, int base_position,
+                                const uint32_t* const argb, int len,
+                                int window_size, int* const distance_ptr) {
+  const uint32_t* const argb_start = argb + base_position;
+  const int min_pos =
+      (base_position > window_size) ? base_position - window_size : 0;
+  int pos;
+  assert(len <= MAX_LENGTH);
+  for (pos = p->hash_to_first_index_[GetPixPairHash64(argb_start)];
+       pos >= min_pos;
+       pos = p->chain_[pos]) {
+    const int curr_length =
+        FindMatchLength(argb + pos, argb_start, len - 1, len);
+    if (curr_length == len) break;
+  }
+  *distance_ptr = base_position - pos;
+}
+
+static int HashChainFindCopy(const VP8LHashChain* const p,
+                             int base_position,
+                             const uint32_t* const argb, int max_len,
+                             int window_size, int iter_max,
                              int* const distance_ptr,
                              int* const length_ptr) {
-  const uint64_t hash_code = GetPixPairHash64(&argb[index]);
-  int prev_length = 0;
-  int64_t best_val = 0;
+  const uint32_t* const argb_start = argb + base_position;
+  int iter = iter_max;
   int best_length = 0;
   int best_distance = 0;
-  const uint32_t* const argb_start = argb + index;
-  const int iter_min_mult = (quality < 50) ? 2 : (quality < 75) ? 4 : 8;
-  const int iter_min = -quality * iter_min_mult;
-  int iter_cnt = 10 + (quality >> 1);
-  const int min_pos = (index > WINDOW_SIZE) ? index - WINDOW_SIZE : 0;
+  const int min_pos =
+      (base_position > window_size) ? base_position - window_size : 0;
   int pos;
-
-  assert(xsize > 0);
-  for (pos = p->hash_to_first_index_[hash_code];
+  int length_max = 256;
+  if (max_len < length_max) {
+    length_max = max_len;
+  }
+  for (pos = p->hash_to_first_index_[GetPixPairHash64(argb_start)];
        pos >= min_pos;
        pos = p->chain_[pos]) {
-    int64_t val;
     int curr_length;
-    if (iter_cnt < 0) {
-      if (iter_cnt < iter_min || best_val >= 0xff0000) {
-        break;
-      }
-    }
-    --iter_cnt;
-    if (best_length != 0 &&
-        argb[pos + best_length - 1] != argb_start[best_length - 1]) {
-      continue;
-    }
-    curr_length = FindMatchLength(argb + pos, argb_start, maxlen);
-    if (curr_length < prev_length) {
-      continue;
+    int distance;
+    if (--iter < 0) {
+      break;
     }
-    val = 65536 * curr_length;
-    // Favoring 2d locality here gives savings for certain images.
-    if (index - pos < 9 * xsize) {
-      const int y = (index - pos) / xsize;
-      int x = (index - pos) % xsize;
-      if (x > xsize / 2) {
-        x = xsize - x;
-      }
-      if (x <= 7 && x >= -8) {
-        val -= y * y + x * x;
-      } else {
-        val -= 9 * 9 + 9 * 9;
-      }
-    } else {
-      val -= 9 * 9 + 9 * 9;
-    }
-    if (best_val < val) {
-      prev_length = curr_length;
-      best_val = val;
+
+    curr_length = FindMatchLength(argb + pos, argb_start, best_length, max_len);
+    if (best_length < curr_length) {
+      distance = base_position - pos;
       best_length = curr_length;
-      best_distance = index - pos;
-      if (curr_length >= MAX_LENGTH) {
-        break;
-      }
-      if ((best_distance == 1 || best_distance == xsize) &&
-          best_length >= 128) {
+      best_distance = distance;
+      if (curr_length >= length_max) {
         break;
       }
     }
@@ -213,140 +336,153 @@ static int HashChainFindCopy(const HashChain* const p,
   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);
+static WEBP_INLINE void AddSingleLiteral(uint32_t pixel, int use_color_cache,
+                                         VP8LColorCache* const hashers,
+                                         VP8LBackwardRefs* const refs) {
+  PixOrCopy v;
+  if (use_color_cache) {
+    const uint32_t key = VP8LColorCacheGetIndex(hashers, pixel);
+    if (VP8LColorCacheLookup(hashers, key) == pixel) {
+      v = PixOrCopyCreateCacheIdx(key);
+    } else {
+      v = PixOrCopyCreateLiteral(pixel);
+      VP8LColorCacheSet(hashers, key, pixel);
+    }
+  } else {
+    v = PixOrCopyCreateLiteral(pixel);
   }
-  refs->size = size;
+  BackwardRefsCursorAdd(refs, v);
 }
 
-static void BackwardReferencesRle(int xsize, int ysize,
-                                  const uint32_t* const argb,
-                                  VP8LBackwardRefs* const refs) {
+static int BackwardReferencesRle(int xsize, int ysize,
+                                 const uint32_t* const argb,
+                                 int cache_bits, VP8LBackwardRefs* const refs) {
   const int pix_count = xsize * ysize;
-  int 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;
+  int i, k;
+  const int use_color_cache = (cache_bits > 0);
+  VP8LColorCache hashers;
+
+  if (use_color_cache && !VP8LColorCacheInit(&hashers, cache_bits)) {
+    return 0;
+  }
+  ClearBackwardRefs(refs);
+  // Add first pixel as literal.
+  AddSingleLiteral(argb[0], use_color_cache, &hashers, refs);
+  i = 1;
+  while (i < pix_count) {
+    const int max_len = MaxFindCopyLength(pix_count - i);
+    const int kMinLength = 4;
+    const int rle_len = FindMatchLength(argb + i, argb + i - 1, 0, max_len);
+    const int prev_row_len = (i < xsize) ? 0 :
+        FindMatchLength(argb + i, argb + i - xsize, 0, max_len);
+    if (rle_len >= prev_row_len && rle_len >= kMinLength) {
+      BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(1, rle_len));
+      // We don't need to update the color cache here since it is always the
+      // same pixel being copied, and that does not change the color cache
+      // state.
+      i += rle_len;
+    } else if (prev_row_len >= kMinLength) {
+      BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(xsize, prev_row_len));
+      if (use_color_cache) {
+        for (k = 0; k < prev_row_len; ++k) {
+          VP8LColorCacheInsert(&hashers, argb[i + k]);
+        }
+      }
+      i += prev_row_len;
     } else {
-      PushBackCopy(refs, match_len);
-      match_len = 0;
-      refs->refs[refs->size++] = PixOrCopyCreateLiteral(argb[i]);
+      AddSingleLiteral(argb[i], use_color_cache, &hashers, refs);
+      i++;
     }
   }
-  PushBackCopy(refs, match_len);
+  if (use_color_cache) VP8LColorCacheClear(&hashers);
+  return !refs->error_;
 }
 
-static int BackwardReferencesHashChain(int xsize, int ysize,
-                                       const uint32_t* const argb,
-                                       int cache_bits, int quality,
-                                       VP8LBackwardRefs* const refs) {
+static int BackwardReferencesLz77(int xsize, int ysize,
+                                  const uint32_t* const argb, int cache_bits,
+                                  int quality, int low_effort,
+                                  VP8LHashChain* const hash_chain,
+                                  VP8LBackwardRefs* const refs) {
   int i;
   int ok = 0;
   int cc_init = 0;
   const int use_color_cache = (cache_bits > 0);
   const int pix_count = xsize * ysize;
-  HashChain* const hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
   VP8LColorCache hashers;
+  int iter_max = GetMaxItersForQuality(quality, low_effort);
+  const int window_size = GetWindowSizeForHashChain(quality, xsize);
+  int min_matches = 32;
 
-  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;
-  for (i = 0; i < pix_count; ) {
+  ClearBackwardRefs(refs);
+  HashChainReset(hash_chain);
+  for (i = 0; i < pix_count - 2; ) {
     // 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 maxlen = pix_count - i;
-      if (maxlen > MAX_LENGTH) {
-        maxlen = MAX_LENGTH;
-      }
-      HashChainFindCopy(hash_chain, quality, i, xsize, argb, maxlen,
-                        &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.
+    const int max_len = MaxFindCopyLength(pix_count - i);
+    HashChainFindCopy(hash_chain, i, argb, max_len, window_size,
+                      iter_max, &offset, &len);
+    if (len > MIN_LENGTH || (len == MIN_LENGTH && offset <= 512)) {
       int offset2 = 0;
       int len2 = 0;
       int k;
+      min_matches = 8;
       HashChainInsert(hash_chain, &argb[i], i);
-      if (i < pix_count - 2) {  // FindCopy(i+1,..) reads [i + 1] and [i + 2].
-        int maxlen = pix_count - (i + 1);
-        if (maxlen > MAX_LENGTH) {
-          maxlen = MAX_LENGTH;
-        }
-        HashChainFindCopy(hash_chain, quality,
-                          i + 1, xsize, argb, maxlen, &offset2, &len2);
+      if ((len < (max_len >> 2)) && !low_effort) {
+        // Evaluate Alternative#2: Insert the pixel at 'i' as literal, and code
+        // the pixels starting at 'i + 1' using backward reference.
+        HashChainFindCopy(hash_chain, i + 1, argb, max_len - 1,
+                          window_size, iter_max, &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 {
-            refs->refs[refs->size] = PixOrCopyCreateLiteral(pixel);
-          }
-          ++refs->size;
-          if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
+          AddSingleLiteral(argb[i], use_color_cache, &hashers, refs);
           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);
+      BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(offset, len));
       if (use_color_cache) {
         for (k = 0; k < len; ++k) {
           VP8LColorCacheInsert(&hashers, argb[i + k]);
         }
       }
       // Add to the hash_chain (but cannot add the last pixel).
-      {
+      if (offset >= 3 && offset != xsize) {
         const int last = (len < pix_count - 1 - i) ? len : pix_count - 1 - i;
-        for (k = 1; k < last; ++k) {
+        for (k = 2; k < last - 8; k += 2) {
+          HashChainInsert(hash_chain, &argb[i + k], i + k);
+        }
+        for (; 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 {
-        refs->refs[refs->size] = PixOrCopyCreateLiteral(pixel);
-      }
-      ++refs->size;
-      if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
-      if (i + 1 < pix_count) {
+      AddSingleLiteral(argb[i], use_color_cache, &hashers, refs);
+      HashChainInsert(hash_chain, &argb[i], i);
+      ++i;
+      --min_matches;
+      if (min_matches <= 0) {
+        AddSingleLiteral(argb[i], use_color_cache, &hashers, refs);
         HashChainInsert(hash_chain, &argb[i], i);
+        ++i;
       }
-      ++i;
     }
   }
-  ok = 1;
-Error:
+  while (i < pix_count) {
+    // Handle the last pixel(s).
+    AddSingleLiteral(argb[i], use_color_cache, &hashers, refs);
+    ++i;
+  }
+
+  ok = !refs->error_;
+ Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  HashChainDelete(hash_chain);
   return ok;
 }
 
@@ -355,18 +491,19 @@ Error:
 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];
+  double* literal_;
 } CostModel;
 
 static int BackwardReferencesTraceBackwards(
-    int xsize, int ysize, int recursive_cost_model,
-    const uint32_t* const argb, int cache_bits, VP8LBackwardRefs* const refs);
+    int xsize, int ysize, const uint32_t* const argb, int quality,
+    int cache_bits, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs* const refs);
 
 static void ConvertPopulationCountTableToBitEstimates(
-    int num_symbols, const int population_counts[], double output[]) {
-  int sum = 0;
+    int num_symbols, const uint32_t population_counts[], double output[]) {
+  uint32_t sum = 0;
   int nonzeros = 0;
   int i;
   for (i = 0; i < num_symbols; ++i) {
@@ -385,42 +522,29 @@ static void ConvertPopulationCountTableToBitEstimates(
   }
 }
 
-static int CostModelBuild(CostModel* const m, int xsize, int ysize,
-                          int recursion_level, const uint32_t* const argb,
-                          int cache_bits) {
+static int CostModelBuild(CostModel* const m, int cache_bits,
+                          VP8LBackwardRefs* const refs) {
   int ok = 0;
-  VP8LHistogram histo;
-  VP8LBackwardRefs refs;
-  const int quality = 100;
+  VP8LHistogram* const histo = VP8LAllocateHistogram(cache_bits);
+  if (histo == NULL) goto Error;
 
-  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize)) goto Error;
+  VP8LHistogramCreate(histo, refs, cache_bits);
 
-  if (recursion_level > 0) {
-    if (!BackwardReferencesTraceBackwards(xsize, ysize, recursion_level - 1,
-                                          argb, 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_);
+      VP8LHistogramNumCodes(histo->palette_code_bits_),
+      histo->literal_, m->literal_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.red_, m->red_);
+      VALUES_IN_BYTE, histo->red_, m->red_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.blue_, m->blue_);
+      VALUES_IN_BYTE, histo->blue_, m->blue_);
   ConvertPopulationCountTableToBitEstimates(
-      VALUES_IN_BYTE, histo.alpha_, m->alpha_);
+      VALUES_IN_BYTE, histo->alpha_, m->alpha_);
   ConvertPopulationCountTableToBitEstimates(
-      NUM_DISTANCE_CODES, histo.distance_, m->distance_);
+      NUM_DISTANCE_CODES, histo->distance_, m->distance_);
   ok = 1;
 
  Error:
-  VP8LClearBackwardRefs(&refs);
+  VP8LFreeHistogram(histo);
   return ok;
 }
 
@@ -438,203 +562,211 @@ static WEBP_INLINE double GetCacheCost(const CostModel* const m, uint32_t idx) {
 
 static WEBP_INLINE double GetLengthCost(const CostModel* const m,
                                         uint32_t length) {
-  int code, extra_bits_count, extra_bits_value;
-  PrefixEncode(length, &code, &extra_bits_count, &extra_bits_value);
-  return m->literal_[VALUES_IN_BYTE + code] + extra_bits_count;
+  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_count, extra_bits_value;
-  PrefixEncode(distance, &code, &extra_bits_count, &extra_bits_value);
-  return m->distance_[code] + extra_bits_count;
+  int code, extra_bits;
+  VP8LPrefixEncodeBits(distance, &code, &extra_bits);
+  return m->distance_[code] + extra_bits;
+}
+
+static void AddSingleLiteralWithCostModel(
+    const uint32_t* const argb, VP8LHashChain* const hash_chain,
+    VP8LColorCache* const hashers, const CostModel* const cost_model, int idx,
+    int is_last, int use_color_cache, double prev_cost, float* const cost,
+    uint16_t* const dist_array) {
+  double cost_val = prev_cost;
+  const uint32_t color = argb[0];
+  if (!is_last) {
+    HashChainInsert(hash_chain, argb, idx);
+  }
+  if (use_color_cache && VP8LColorCacheContains(hashers, color)) {
+    const double mul0 = 0.68;
+    const int ix = VP8LColorCacheGetIndex(hashers, color);
+    cost_val += GetCacheCost(cost_model, ix) * mul0;
+  } else {
+    const double mul1 = 0.82;
+    if (use_color_cache) VP8LColorCacheInsert(hashers, color);
+    cost_val += GetLiteralCost(cost_model, color) * mul1;
+  }
+  if (cost[idx] > cost_val) {
+    cost[idx] = (float)cost_val;
+    dist_array[idx] = 1;  // only one is inserted.
+  }
 }
 
 static int BackwardReferencesHashChainDistanceOnly(
-    int xsize, int ysize, int recursive_cost_model, const uint32_t* const argb,
-    int cache_bits, uint32_t* const dist_array) {
+    int xsize, int ysize, const uint32_t* const argb,
+    int quality, int cache_bits, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs* const refs, uint16_t* const dist_array) {
   int i;
   int ok = 0;
   int cc_init = 0;
-  const int quality = 100;
   const int pix_count = xsize * ysize;
   const int use_color_cache = (cache_bits > 0);
-  double* const cost =
-      (double*)WebPSafeMalloc((uint64_t)pix_count, sizeof(*cost));
-  CostModel* cost_model = (CostModel*)malloc(sizeof(*cost_model));
-  HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+  float* const cost =
+      (float*)WebPSafeMalloc(pix_count, sizeof(*cost));
+  const size_t literal_array_size = sizeof(double) *
+      (NUM_LITERAL_CODES + NUM_LENGTH_CODES +
+       ((cache_bits > 0) ? (1 << cache_bits) : 0));
+  const size_t cost_model_size = sizeof(CostModel) + literal_array_size;
+  CostModel* const cost_model =
+      (CostModel*)WebPSafeMalloc(1ULL, cost_model_size);
   VP8LColorCache hashers;
-  const double mul0 = (recursive_cost_model != 0) ? 1.0 : 0.68;
-  const double mul1 = (recursive_cost_model != 0) ? 1.0 : 0.82;
-
-  if (cost == NULL || cost_model == NULL || hash_chain == NULL) goto Error;
+  const int skip_length = 32 + quality;
+  const int skip_min_distance_code = 2;
+  int iter_max = GetMaxItersForQuality(quality, 0);
+  const int window_size = GetWindowSizeForHashChain(quality, xsize);
 
-  if (!HashChainInit(hash_chain, pix_count)) goto Error;
+  if (cost == NULL || cost_model == NULL) goto Error;
 
+  cost_model->literal_ = (double*)(cost_model + 1);
   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,
-                      cache_bits)) {
+  if (!CostModelBuild(cost_model, cache_bits, refs)) {
     goto Error;
   }
 
-  for (i = 0; i < pix_count; ++i) cost[i] = 1e100;
+  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;
-  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 maxlen = shortmax ? 2 : MAX_LENGTH;
-        if (maxlen > pix_count - i) {
-          maxlen = pix_count - i;
+  HashChainReset(hash_chain);
+  // Add first pixel as literal.
+  AddSingleLiteralWithCostModel(argb + 0, hash_chain, &hashers, cost_model, 0,
+                                0, use_color_cache, 0.0, cost, dist_array);
+  for (i = 1; i < pix_count - 1; ++i) {
+    int offset = 0;
+    int len = 0;
+    double prev_cost = cost[i - 1];
+    const int max_len = MaxFindCopyLength(pix_count - i);
+    HashChainFindCopy(hash_chain, i, argb, max_len, window_size,
+                      iter_max, &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;
         }
-        HashChainFindCopy(hash_chain, quality, i, xsize, argb, maxlen,
-                          &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] = 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 >= skip_length && code <= skip_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]);
           }
         }
-        // This if is for speedup only. It roughly doubles the speed, and
-        // makes compression worse by .1 %.
-        if (len >= 128 && code < 2) {
-          // 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 < pix_count - 1 - i) ? len
-                                                       : pix_count - 1 - i;
-            for (k = 0; k < last; ++k) {
-              HashChainInsert(hash_chain, &argb[i + k], 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;
         }
+        // 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 {
-        cost_val += GetLiteralCost(cost_model, argb[i]) * mul1;
-      }
-      if (cost[i] > cost_val) {
-        cost[i] = cost_val;
-        dist_array[i] = 1;  // only one is inserted.
+      if (len != MIN_LENGTH) {
+        int code_min_length;
+        double cost_total;
+        HashChainFindOffset(hash_chain, i, argb, MIN_LENGTH, window_size,
+                            &offset);
+        code_min_length = DistanceToPlaneCode(xsize, offset);
+        cost_total = prev_cost +
+            GetDistanceCost(cost_model, code_min_length) +
+            GetLengthCost(cost_model, 1);
+        if (cost[i + 1] > cost_total) {
+          cost[i + 1] = (float)cost_total;
+          dist_array[i + 1] = 2;
+        }
       }
-      if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
     }
+    AddSingleLiteralWithCostModel(argb + i, hash_chain, &hashers, cost_model, i,
+                                  0, use_color_cache, prev_cost, cost,
+                                  dist_array);
  next_symbol: ;
   }
-  // Last pixel still to do, it can only be a single step if not reached
-  // through cheaper means already.
-  ok = 1;
-Error:
+  // Handle the last pixel.
+  if (i == (pix_count - 1)) {
+    AddSingleLiteralWithCostModel(argb + i, hash_chain, &hashers, cost_model, i,
+                                  1, use_color_cache, cost[pix_count - 2], cost,
+                                  dist_array);
+  }
+  ok = !refs->error_;
+ Error:
   if (cc_init) VP8LColorCacheClear(&hashers);
-  HashChainDelete(hash_chain);
-  free(cost_model);
-  free(cost);
+  WebPSafeFree(cost_model);
+  WebPSafeFree(cost);
   return ok;
 }
 
-static int TraceBackwards(const uint32_t* const dist_array,
-                          int dist_array_size,
-                          uint32_t** const chosen_path,
-                          int* const chosen_path_size) {
-  int i;
-  // Count how many.
-  int count = 0;
-  for (i = dist_array_size - 1; i >= 0; ) {
-    int k = dist_array[i];
-    assert(k >= 1);
-    ++count;
-    i -= k;
-  }
-  // Allocate.
-  *chosen_path_size = count;
-  *chosen_path =
-      (uint32_t*)WebPSafeMalloc((uint64_t)count, sizeof(**chosen_path));
-  if (*chosen_path == NULL) return 0;
-
-  // Write in reverse order.
-  for (i = dist_array_size - 1; i >= 0; ) {
-    int k = dist_array[i];
-    assert(k >= 1);
-    (*chosen_path)[--count] = k;
-    i -= k;
+// We pack the path at the end of *dist_array and return
+// a pointer to this part of the array. Example:
+// dist_array = [1x2xx3x2] => packed [1x2x1232], chosen_path = [1232]
+static void TraceBackwards(uint16_t* const dist_array,
+                           int dist_array_size,
+                           uint16_t** const chosen_path,
+                           int* const chosen_path_size) {
+  uint16_t* path = dist_array + dist_array_size;
+  uint16_t* cur = dist_array + dist_array_size - 1;
+  while (cur >= dist_array) {
+    const int k = *cur;
+    --path;
+    *path = k;
+    cur -= k;
   }
-  return 1;
+  *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 cache_bits,
-    const uint32_t* const chosen_path, int chosen_path_size,
+    int xsize, int ysize, const uint32_t* const argb,
+    int quality, int cache_bits,
+    const uint16_t* const chosen_path, int chosen_path_size,
+    VP8LHashChain* const hash_chain,
     VP8LBackwardRefs* const refs) {
-  const int quality = 100;
   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 i = 0;
   int ok = 0;
   int cc_init = 0;
-  HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+  const int window_size = GetWindowSizeForHashChain(quality, xsize);
   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;
-  for (ix = 0; ix < chosen_path_size; ++ix, ++size) {
+  ClearBackwardRefs(refs);
+  HashChainReset(hash_chain);
+  for (ix = 0; ix < chosen_path_size; ++ix) {
     int offset = 0;
-    int len = 0;
-    int maxlen = chosen_path[ix];
-    if (maxlen != 1) {
-      HashChainFindCopy(hash_chain, quality,
-                        i, xsize, argb, maxlen, &offset, &len);
-      assert(len == maxlen);
-      refs->refs[size] = PixOrCopyCreateCopy(offset, len);
+    const int len = chosen_path[ix];
+    if (len != 1) {
+      int k;
+      HashChainFindOffset(hash_chain, i, argb, len, window_size, &offset);
+      BackwardRefsCursorAdd(refs, PixOrCopyCreateCopy(offset, len));
       if (use_color_cache) {
         for (k = 0; k < len; ++k) {
           VP8LColorCacheInsert(&hashers, argb[i + k]);
@@ -648,227 +780,330 @@ static int BackwardReferencesHashChainFollowChosenPath(
       }
       i += len;
     } else {
+      PixOrCopy v;
       if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
         // push pixel as a color cache index
         const int idx = VP8LColorCacheGetIndex(&hashers, argb[i]);
-        refs->refs[size] = PixOrCopyCreateCacheIdx(idx);
+        v = PixOrCopyCreateCacheIdx(idx);
       } else {
-        refs->refs[size] = PixOrCopyCreateLiteral(argb[i]);
+        if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
+        v = PixOrCopyCreateLiteral(argb[i]);
       }
-      if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
+      BackwardRefsCursorAdd(refs, v);
       if (i + 1 < pix_count) {
         HashChainInsert(hash_chain, &argb[i], i);
       }
       ++i;
     }
   }
-  assert(size <= refs->max_size);
-  refs->size = size;
-  ok = 1;
-Error:
+  ok = !refs->error_;
+ 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 cache_bits,
+                                            int quality, int cache_bits,
+                                            VP8LHashChain* const hash_chain,
                                             VP8LBackwardRefs* const refs) {
   int ok = 0;
   const int dist_array_size = xsize * ysize;
-  uint32_t* chosen_path = NULL;
+  uint16_t* chosen_path = NULL;
   int chosen_path_size = 0;
-  uint32_t* dist_array =
-      (uint32_t*)WebPSafeMalloc((uint64_t)dist_array_size, sizeof(*dist_array));
+  uint16_t* dist_array =
+      (uint16_t*)WebPSafeMalloc(dist_array_size, sizeof(*dist_array));
 
   if (dist_array == NULL) goto Error;
 
   if (!BackwardReferencesHashChainDistanceOnly(
-      xsize, ysize, recursive_cost_model, argb, cache_bits, dist_array)) {
-    goto Error;
-  }
-  if (!TraceBackwards(dist_array, dist_array_size,
-                      &chosen_path, &chosen_path_size)) {
+      xsize, ysize, argb, quality, cache_bits, hash_chain,
+      refs, dist_array)) {
     goto Error;
   }
-  free(dist_array);   // no need to retain this memory any longer
-  dist_array = NULL;
+  TraceBackwards(dist_array, dist_array_size, &chosen_path, &chosen_path_size);
   if (!BackwardReferencesHashChainFollowChosenPath(
-      xsize, ysize, argb, cache_bits, chosen_path, chosen_path_size, refs)) {
+      xsize, ysize, argb, quality, cache_bits, chosen_path, chosen_path_size,
+      hash_chain, refs)) {
     goto Error;
   }
   ok = 1;
  Error:
-  free(chosen_path);
-  free(dist_array);
+  WebPSafeFree(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 VP8LBackwardRefs* const refs) {
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  while (VP8LRefsCursorOk(&c)) {
+    if (PixOrCopyIsCopy(c.cur_pos)) {
+      const int dist = c.cur_pos->argb_or_distance;
       const int transformed_dist = DistanceToPlaneCode(xsize, dist);
-      refs->refs[i].argb_or_distance = transformed_dist;
+      c.cur_pos->argb_or_distance = transformed_dist;
     }
+    VP8LRefsCursorNext(&c);
   }
 }
 
-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);
+// Returns entropy for the given cache bits.
+static double ComputeCacheEntropy(const uint32_t* argb,
+                                  const VP8LBackwardRefs* const refs,
+                                  int cache_bits) {
+  const int use_color_cache = (cache_bits > 0);
+  int cc_init = 0;
+  double entropy = MAX_ENTROPY;
+  const double kSmallPenaltyForLargeCache = 4.0;
+  VP8LColorCache hashers;
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  VP8LHistogram* histo = VP8LAllocateHistogram(cache_bits);
+  if (histo == NULL) goto Error;
 
-  {
-    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);
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
   }
-
-  // Choose appropriate backward reference.
-  if (lz77_is_useful) {
-    // TraceBackwards is costly. Run it for higher qualities.
-    const int try_lz77_trace_backwards = (quality >= 75);
-    *best = refs_lz77;   // default guess: lz77 is better
-    VP8LClearBackwardRefs(&refs_rle);
-    if (try_lz77_trace_backwards) {
-      const int recursion_level = (num_pix < 320 * 200) ? 1 : 0;
-      VP8LBackwardRefs refs_trace;
-      if (!VP8LBackwardRefsAlloc(&refs_trace, num_pix)) {
-        goto End;
-      }
-      if (BackwardReferencesTraceBackwards(
-          width, height, recursion_level, argb, cache_bits, &refs_trace)) {
-        VP8LClearBackwardRefs(&refs_lz77);
-        *best = refs_trace;
-      }
+  if (!use_color_cache) {
+    while (VP8LRefsCursorOk(&c)) {
+      VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos);
+      VP8LRefsCursorNext(&c);
     }
   } else {
-    VP8LClearBackwardRefs(&refs_lz77);
-    *best = refs_rle;
+    while (VP8LRefsCursorOk(&c)) {
+      const PixOrCopy* const v = c.cur_pos;
+      if (PixOrCopyIsLiteral(v)) {
+        const uint32_t pix = *argb++;
+        const uint32_t key = VP8LColorCacheGetIndex(&hashers, pix);
+        if (VP8LColorCacheLookup(&hashers, key) == pix) {
+          ++histo->literal_[NUM_LITERAL_CODES + NUM_LENGTH_CODES + key];
+        } else {
+          VP8LColorCacheSet(&hashers, key, pix);
+          ++histo->blue_[pix & 0xff];
+          ++histo->literal_[(pix >> 8) & 0xff];
+          ++histo->red_[(pix >> 16) & 0xff];
+          ++histo->alpha_[pix >> 24];
+        }
+      } else {
+        int len = PixOrCopyLength(v);
+        int code, extra_bits;
+        VP8LPrefixEncodeBits(len, &code, &extra_bits);
+        ++histo->literal_[NUM_LITERAL_CODES + code];
+        VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
+        ++histo->distance_[code];
+        do {
+          VP8LColorCacheInsert(&hashers, *argb++);
+        } while(--len != 0);
+      }
+      VP8LRefsCursorNext(&c);
+    }
   }
+  entropy = VP8LHistogramEstimateBits(histo) +
+      kSmallPenaltyForLargeCache * cache_bits;
+ Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  VP8LFreeHistogram(histo);
+  return entropy;
+}
 
-  if (use_2d_locality) BackwardReferences2DLocality(width, best);
-
-  ok = 1;
-
- End:
-  if (!ok) {
-    VP8LClearBackwardRefs(best);
+// Evaluate optimal cache bits for the local color cache.
+// The input *best_cache_bits sets the maximum cache bits to use (passing 0
+// implies disabling the local color cache). The local color cache is also
+// disabled for the lower (<= 25) quality.
+// Returns 0 in case of memory error.
+static int CalculateBestCacheSize(const uint32_t* const argb,
+                                  int xsize, int ysize, int quality,
+                                  VP8LHashChain* const hash_chain,
+                                  VP8LBackwardRefs* const refs,
+                                  int* const lz77_computed,
+                                  int* const best_cache_bits) {
+  int eval_low = 1;
+  int eval_high = 1;
+  double entropy_low = MAX_ENTROPY;
+  double entropy_high = MAX_ENTROPY;
+  const double cost_mul = 5e-4;
+  int cache_bits_low = 0;
+  int cache_bits_high = (quality <= 25) ? 0 : *best_cache_bits;
+
+  assert(cache_bits_high <= MAX_COLOR_CACHE_BITS);
+
+  *lz77_computed = 0;
+  if (cache_bits_high == 0) {
+    *best_cache_bits = 0;
+    // Local color cache is disabled.
+    return 1;
   }
-  return ok;
+  if (!BackwardReferencesLz77(xsize, ysize, argb, cache_bits_low, quality, 0,
+                              hash_chain, refs)) {
+    return 0;
+  }
+  // Do a binary search to find the optimal entropy for cache_bits.
+  while (eval_low || eval_high) {
+    if (eval_low) {
+      entropy_low = ComputeCacheEntropy(argb, refs, cache_bits_low);
+      entropy_low += entropy_low * cache_bits_low * cost_mul;
+      eval_low = 0;
+    }
+    if (eval_high) {
+      entropy_high = ComputeCacheEntropy(argb, refs, cache_bits_high);
+      entropy_high += entropy_high * cache_bits_high * cost_mul;
+      eval_high = 0;
+    }
+    if (entropy_high < entropy_low) {
+      const int prev_cache_bits_low = cache_bits_low;
+      *best_cache_bits = cache_bits_high;
+      cache_bits_low = (cache_bits_low + cache_bits_high) / 2;
+      if (cache_bits_low != prev_cache_bits_low) eval_low = 1;
+    } else {
+      *best_cache_bits = cache_bits_low;
+      cache_bits_high = (cache_bits_low + cache_bits_high) / 2;
+      if (cache_bits_high != cache_bits_low) eval_high = 1;
+    }
+  }
+  *lz77_computed = 1;
+  return 1;
 }
 
-// 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) {
+// Update (in-place) backward references for specified cache_bits.
+static int BackwardRefsWithLocalCache(const uint32_t* const argb,
+                                      int cache_bits,
+                                      VP8LBackwardRefs* const refs) {
   int pixel_index = 0;
-  int i;
-  uint32_t k;
   VP8LColorCache hashers;
-  const int use_color_cache = (cache_bits > 0);
-  int cc_init = 0;
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  if (!VP8LColorCacheInit(&hashers, cache_bits)) return 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];
+  while (VP8LRefsCursorOk(&c)) {
+    PixOrCopy* const v = c.cur_pos;
     if (PixOrCopyIsLiteral(v)) {
-      if (use_color_cache &&
-          VP8LColorCacheContains(&hashers, argb[pixel_index])) {
-        // push pixel as a cache index
-        const int ix = VP8LColorCacheGetIndex(&hashers, argb[pixel_index]);
-        const PixOrCopy token = PixOrCopyCreateCacheIdx(ix);
-        VP8LHistogramAddSinglePixOrCopy(histo, &token);
+      const uint32_t argb_literal = v->argb_or_distance;
+      if (VP8LColorCacheContains(&hashers, argb_literal)) {
+        const int ix = VP8LColorCacheGetIndex(&hashers, argb_literal);
+        *v = PixOrCopyCreateCacheIdx(ix);
       } else {
-        VP8LHistogramAddSinglePixOrCopy(histo, v);
+        VP8LColorCacheInsert(&hashers, argb_literal);
       }
+      ++pixel_index;
     } else {
-      VP8LHistogramAddSinglePixOrCopy(histo, v);
-    }
-    if (use_color_cache) {
-      for (k = 0; k < PixOrCopyLength(v); ++k) {
-        VP8LColorCacheInsert(&hashers, argb[pixel_index + k]);
+      // refs was created without local cache, so it can not have cache indexes.
+      int k;
+      assert(PixOrCopyIsCopy(v));
+      for (k = 0; k < v->len; ++k) {
+        VP8LColorCacheInsert(&hashers, argb[pixel_index++]);
       }
     }
-    pixel_index += PixOrCopyLength(v);
+    VP8LRefsCursorNext(&c);
   }
-  assert(pixel_index == xsize * ysize);
-  (void)xsize;  // xsize is not used in non-debug compilations otherwise.
-  (void)ysize;  // ysize is not used in non-debug compilations otherwise.
-  if (cc_init) VP8LColorCacheClear(&hashers);
+  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)) {
+static VP8LBackwardRefs* GetBackwardReferencesLowEffort(
+    int width, int height, const uint32_t* const argb, int quality,
+    int* const cache_bits, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs refs_array[2]) {
+  VP8LBackwardRefs* refs_lz77 = &refs_array[0];
+  *cache_bits = 0;
+  if (!BackwardReferencesLz77(width, height, argb, 0, quality,
+                              1 /* Low effort. */, hash_chain, refs_lz77)) {
+    return NULL;
+  }
+  BackwardReferences2DLocality(width, refs_lz77);
+  return refs_lz77;
+}
+
+static VP8LBackwardRefs* GetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int* const cache_bits, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs refs_array[2]) {
+  int lz77_is_useful;
+  int lz77_computed;
+  double bit_cost_lz77, bit_cost_rle;
+  VP8LBackwardRefs* best = NULL;
+  VP8LBackwardRefs* refs_lz77 = &refs_array[0];
+  VP8LBackwardRefs* refs_rle = &refs_array[1];
+  VP8LHistogram* histo = NULL;
+
+  if (!CalculateBestCacheSize(argb, width, height, quality, hash_chain,
+                              refs_lz77, &lz77_computed, cache_bits)) {
     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;
+
+  if (lz77_computed) {
+    // Transform refs_lz77 for the optimized cache_bits.
+    if (*cache_bits > 0) {
+      if (!BackwardRefsWithLocalCache(argb, *cache_bits, refs_lz77)) {
+        goto Error;
+      }
+    }
+  } else {
+    if (!BackwardReferencesLz77(width, height, argb, *cache_bits, quality,
+                                0 /* Low effort. */, hash_chain, refs_lz77)) {
+      goto Error;
     }
   }
-  ok = 1;
+
+  if (!BackwardReferencesRle(width, height, argb, *cache_bits, refs_rle)) {
+    goto Error;
+  }
+
+  histo = VP8LAllocateHistogram(*cache_bits);
+  if (histo == NULL) goto Error;
+
+  {
+    // 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);
+  }
+
+  // Choose appropriate backward reference.
+  if (lz77_is_useful) {
+    // TraceBackwards is costly. Don't execute it at lower quality.
+    const int try_lz77_trace_backwards = (quality >= 25);
+    best = refs_lz77;   // default guess: lz77 is better
+    if (try_lz77_trace_backwards) {
+      VP8LBackwardRefs* const refs_trace = refs_rle;
+      if (!VP8LBackwardRefsCopy(refs_lz77, refs_trace)) {
+        best = NULL;
+        goto Error;
+      }
+      if (BackwardReferencesTraceBackwards(width, height, argb, quality,
+                                           *cache_bits, hash_chain,
+                                           refs_trace)) {
+        double bit_cost_trace;
+        // Evaluate LZ77 coding.
+        VP8LHistogramCreate(histo, refs_trace, *cache_bits);
+        bit_cost_trace = VP8LHistogramEstimateBits(histo);
+        if (bit_cost_trace < bit_cost_lz77) {
+          best = refs_trace;
+        }
+      }
+    }
+  } else {
+    best = refs_rle;
+  }
+
+  BackwardReferences2DLocality(width, best);
+
  Error:
-  VP8LClearBackwardRefs(&refs);
-  return ok;
+  VP8LFreeHistogram(histo);
+  return best;
+}
+
+VP8LBackwardRefs* VP8LGetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int low_effort, int* const cache_bits, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs refs_array[2]) {
+  if (low_effort) {
+    return GetBackwardReferencesLowEffort(width, height, argb, quality,
+                                          cache_bits, hash_chain, refs_array);
+  } else {
+    return GetBackwardReferences(width, height, argb, quality, cache_bits,
+                                 hash_chain, refs_array);
+  }
 }
diff --git a/drivers/webp/enc/backward_references.h b/drivers/webp/enc/backward_references.h
index 8006a56ba1..daa084d846 100644
--- a/drivers/webp/enc/backward_references.h
+++ b/drivers/webp/enc/backward_references.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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)
@@ -13,82 +15,15 @@
 
 #include <assert.h>
 #include <stdlib.h>
-#include "../types.h"
-#include "../format_constants.h"
+#include "../webp/types.h"
+#include "../webp/format_constants.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#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))
-
-// -----------------------------------------------------------------------------
-// PrefixEncode()
-
-// use GNU builtins where available.
-#if defined(__GNUC__) && \
-    ((__GNUC__ == 3 && __GNUC_MINOR__ >= 4) || __GNUC__ >= 4)
-static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
-  return n == 0 ? -1 : 31 ^ __builtin_clz(n);
-}
-#elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
-#include <intrin.h>
-#pragma intrinsic(_BitScanReverse)
-
-static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
-  unsigned long first_set_bit;
-  return _BitScanReverse(&first_set_bit, n) ? first_set_bit : -1;
-}
-#else
-static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
-  int log = 0;
-  uint32_t value = n;
-  int i;
-
-  if (value == 0) return -1;
-  for (i = 4; i >= 0; --i) {
-    const int shift = (1 << i);
-    const uint32_t x = value >> shift;
-    if (x != 0) {
-      value = x;
-      log += shift;
-    }
-  }
-  return log;
-}
-#endif
-
-static WEBP_INLINE int VP8LBitsLog2Ceiling(uint32_t n) {
-  const int floor = BitsLog2Floor(n);
-  if (n == (n & ~(n - 1)))  // zero or a power of two.
-    return floor;
-  else
-    return floor + 1;
-}
-
-// Splitting of distance and length codes into prefixes and
-// extra bits. The prefixes are encoded with an entropy code
-// while the extra bits are stored just as normal bits.
-static WEBP_INLINE void PrefixEncode(int distance, int* const code,
-                                     int* const extra_bits_count,
-                                     int* const extra_bits_value) {
-  // Collect the two most significant bits where the highest bit is 1.
-  const int highest_bit = BitsLog2Floor(--distance);
-  // & 0x3f is to make behavior well defined when highest_bit
-  // does not exist or is the least significant bit.
-  const int second_highest_bit =
-      (distance >> ((highest_bit - 1) & 0x3f)) & 1;
-  *extra_bits_count = (highest_bit > 0) ? (highest_bit - 1) : 0;
-  *extra_bits_value = distance & ((1 << *extra_bits_count) - 1);
-  *code = (highest_bit > 0) ? (2 * highest_bit + second_highest_bit)
-                            : (highest_bit == 0) ? 1 : 0;
-}
+// The maximum allowed limit is 11.
+#define MAX_COLOR_CACHE_BITS 10
 
 // -----------------------------------------------------------------------------
 // PixOrCopy
@@ -173,39 +108,94 @@ static WEBP_INLINE uint32_t PixOrCopyDistance(const PixOrCopy* const p) {
 }
 
 // -----------------------------------------------------------------------------
-// VP8LBackwardRefs
+// VP8LHashChain
+
+#define HASH_BITS 18
+#define HASH_SIZE (1 << HASH_BITS)
+
+typedef struct VP8LHashChain VP8LHashChain;
+struct VP8LHashChain {
+  // 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_;
+  // This is the maximum size of the hash_chain that can be constructed.
+  // Typically this is the pixel count (width x height) for a given image.
+  int size_;
+};
 
-typedef struct {
-  PixOrCopy* refs;
-  int size;      // currently used
-  int max_size;  // maximum capacity
-} VP8LBackwardRefs;
+// Must be called first, to set size.
+int VP8LHashChainInit(VP8LHashChain* const p, int size);
+void VP8LHashChainClear(VP8LHashChain* const p);  // release memory
 
-// Initialize the object. Must be called first. 'refs' can be NULL.
-void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs);
+// -----------------------------------------------------------------------------
+// VP8LBackwardRefs (block-based backward-references storage)
+
+// maximum number of reference blocks the image will be segmented into
+#define MAX_REFS_BLOCK_PER_IMAGE 16
+
+typedef struct PixOrCopyBlock PixOrCopyBlock;   // forward declaration
+typedef struct VP8LBackwardRefs VP8LBackwardRefs;
+
+// Container for blocks chain
+struct VP8LBackwardRefs {
+  int block_size_;               // common block-size
+  int error_;                    // set to true if some memory error occurred
+  PixOrCopyBlock* refs_;         // list of currently used blocks
+  PixOrCopyBlock** tail_;        // for list recycling
+  PixOrCopyBlock* free_blocks_;  // free-list
+  PixOrCopyBlock* last_block_;   // used for adding new refs (internal)
+};
 
-// Release memory and re-initialize the object. 'refs' can be NULL.
-void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs);
+// Initialize the object. 'block_size' is the common block size to store
+// references (typically, width * height / MAX_REFS_BLOCK_PER_IMAGE).
+void VP8LBackwardRefsInit(VP8LBackwardRefs* const refs, int block_size);
+// Release memory for backward references.
+void VP8LBackwardRefsClear(VP8LBackwardRefs* const refs);
+// Copies the 'src' backward refs to the 'dst'. Returns 0 in case of error.
+int VP8LBackwardRefsCopy(const VP8LBackwardRefs* const src,
+                         VP8LBackwardRefs* const dst);
 
-// Allocate 'max_size' references. Returns false in case of memory error.
-int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size);
+// Cursor for iterating on references content
+typedef struct {
+  // public:
+  PixOrCopy* cur_pos;           // current position
+  // private:
+  PixOrCopyBlock* cur_block_;   // current block in the refs list
+  const PixOrCopy* last_pos_;   // sentinel for switching to next block
+} VP8LRefsCursor;
+
+// Returns a cursor positioned at the beginning of the references list.
+VP8LRefsCursor VP8LRefsCursorInit(const VP8LBackwardRefs* const refs);
+// Returns true if cursor is pointing at a valid position.
+static WEBP_INLINE int VP8LRefsCursorOk(const VP8LRefsCursor* const c) {
+  return (c->cur_pos != NULL);
+}
+// Move to next block of references. Internal, not to be called directly.
+void VP8LRefsCursorNextBlock(VP8LRefsCursor* const c);
+// Move to next position, or NULL. Should not be called if !VP8LRefsCursorOk().
+static WEBP_INLINE void VP8LRefsCursorNext(VP8LRefsCursor* const c) {
+  assert(c != NULL);
+  assert(VP8LRefsCursorOk(c));
+  if (++c->cur_pos == c->last_pos_) VP8LRefsCursorNextBlock(c);
+}
 
 // -----------------------------------------------------------------------------
 // Main entry points
 
 // 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);
-
-#if defined(__cplusplus) || defined(c_plusplus)
+// The input cache_bits to 'VP8LGetBackwardReferences' sets the maximum cache
+// bits to use (passing 0 implies disabling the local color cache).
+// The optimal cache bits is evaluated and set for the *cache_bits parameter.
+// The return value is the pointer to the best of the two backward refs viz,
+// refs[0] or refs[1].
+VP8LBackwardRefs* VP8LGetBackwardReferences(
+    int width, int height, const uint32_t* const argb, int quality,
+    int low_effort, int* const cache_bits, VP8LHashChain* const hash_chain,
+    VP8LBackwardRefs refs[2]);
+
+#ifdef __cplusplus
 }
 #endif
 
diff --git a/drivers/webp/enc/config.c b/drivers/webp/enc/config.c
index 4136f6c227..f9f7961d58 100644
--- a/drivers/webp/enc/config.c
+++ b/drivers/webp/enc/config.c
@@ -1,19 +1,17 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 "../encode.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "../webp/encode.h"
 
 //------------------------------------------------------------------------------
 // WebPConfig
@@ -31,9 +29,9 @@ int WebPConfigInitInternal(WebPConfig* config,
   config->target_PSNR = 0.;
   config->method = 4;
   config->sns_strength = 50;
-  config->filter_strength = 20;   // default: light filtering
+  config->filter_strength = 60;   // mid-filtering
   config->filter_sharpness = 0;
-  config->filter_type = 0;        // default: simple
+  config->filter_type = 1;        // default: strong (so U/V is filtered too)
   config->partitions = 0;
   config->segments = 4;
   config->pass = 1;
@@ -45,7 +43,15 @@ int WebPConfigInitInternal(WebPConfig* config,
   config->alpha_filtering = 1;
   config->alpha_quality = 100;
   config->lossless = 0;
+  config->exact = 0;
   config->image_hint = WEBP_HINT_DEFAULT;
+  config->emulate_jpeg_size = 0;
+  config->thread_level = 0;
+  config->low_memory = 0;
+  config->near_lossless = 100;
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  config->delta_palettization = 0;
+#endif // WEBP_EXPERIMENTAL_FEATURES
 
   // TODO(skal): tune.
   switch (preset) {
@@ -53,11 +59,13 @@ int WebPConfigInitInternal(WebPConfig* config,
       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;
@@ -67,10 +75,12 @@ int WebPConfigInitInternal(WebPConfig* config,
     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:
@@ -106,7 +116,7 @@ int WebPValidateConfig(const WebPConfig* config) {
     return 0;
   if (config->show_compressed < 0 || config->show_compressed > 1)
     return 0;
-  if (config->preprocessing < 0 || config->preprocessing > 1)
+  if (config->preprocessing < 0 || config->preprocessing > 7)
     return 0;
   if (config->partitions < 0 || config->partitions > 3)
     return 0;
@@ -120,13 +130,44 @@ int WebPValidateConfig(const WebPConfig* config) {
     return 0;
   if (config->lossless < 0 || config->lossless > 1)
     return 0;
+  if (config->near_lossless < 0 || config->near_lossless > 100)
+    return 0;
   if (config->image_hint >= WEBP_HINT_LAST)
     return 0;
+  if (config->emulate_jpeg_size < 0 || config->emulate_jpeg_size > 1)
+    return 0;
+  if (config->thread_level < 0 || config->thread_level > 1)
+    return 0;
+  if (config->low_memory < 0 || config->low_memory > 1)
+    return 0;
+  if (config->exact < 0 || config->exact > 1)
+    return 0;
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  if (config->delta_palettization < 0 || config->delta_palettization > 1)
+    return 0;
+#endif  // WEBP_EXPERIMENTAL_FEATURES
   return 1;
 }
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
+#define MAX_LEVEL 9
+
+// Mapping between -z level and -m / -q parameter settings.
+static const struct {
+  uint8_t method_;
+  uint8_t quality_;
+} kLosslessPresets[MAX_LEVEL + 1] = {
+  { 0,  0 }, { 1, 20 }, { 2, 25 }, { 3, 30 }, { 3, 50 },
+  { 4, 50 }, { 4, 75 }, { 4, 90 }, { 5, 90 }, { 6, 100 }
+};
+
+int WebPConfigLosslessPreset(WebPConfig* config, int level) {
+  if (config == NULL || level < 0 || level > MAX_LEVEL) return 0;
+  config->lossless = 1;
+  config->method = kLosslessPresets[level].method_;
+  config->quality = kLosslessPresets[level].quality_;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
diff --git a/drivers/webp/enc/cost.c b/drivers/webp/enc/cost.c
index 92e0cc713c..ae7fe01388 100644
--- a/drivers/webp/enc/cost.c
+++ b/drivers/webp/enc/cost.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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
@@ -11,42 +13,6 @@
 
 #include "./cost.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-//------------------------------------------------------------------------------
-// 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
 
@@ -73,267 +39,6 @@ const uint16_t VP8LevelCodes[MAX_VARIABLE_LEVEL][2] = {
   {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[2048] = {
-     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];
@@ -352,19 +57,21 @@ static int VariableLevelCost(int level, const uint8_t probas[NUM_PROBAS]) {
 //------------------------------------------------------------------------------
 // Pre-calc level costs once for all
 
-void VP8CalculateLevelCosts(VP8Proba* const proba) {
+void VP8CalculateLevelCosts(VP8EncProba* const proba) {
   int ctype, band, ctx;
 
   if (!proba->dirty_) return;  // nothing to do.
 
   for (ctype = 0; ctype < NUM_TYPES; ++ctype) {
+    int n;
     for (band = 0; band < NUM_BANDS; ++band) {
-      for(ctx = 0; ctx < NUM_CTX; ++ctx) {
+      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]);
+        const int cost0 = (ctx > 0) ? VP8BitCost(1, p[0]) : 0;
+        const int cost_base = VP8BitCost(1, p[1]) + cost0;
         int v;
-        table[0] = VP8BitCost(0, p[1]);
+        table[0] = VP8BitCost(0, p[1]) + cost0;
         for (v = 1; v <= MAX_VARIABLE_LEVEL; ++v) {
           table[v] = cost_base + VariableLevelCost(v, p);
         }
@@ -372,6 +79,12 @@ void VP8CalculateLevelCosts(VP8Proba* const proba) {
         // actually constant.
       }
     }
+    for (n = 0; n < 16; ++n) {    // replicate bands. We don't need to sentinel.
+      for (ctx = 0; ctx < NUM_CTX; ++ctx) {
+        proba->remapped_costs_[ctype][n][ctx] =
+            proba->level_cost_[ctype][VP8EncBands[n]][ctx];
+      }
+    }
   }
   proba->dirty_ = 0;
 }
@@ -385,110 +98,257 @@ 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] = {
-  { {  251, 1362, 1934, 2085, 2314, 2230, 1839, 1988, 2437, 2348 },
-    {  403,  680, 1507, 1519, 2060, 2005, 1992, 1914, 1924, 1733 },
-    {  353, 1121,  973, 1895, 2060, 1787, 1671, 1516, 2012, 1868 },
-    {  770,  852, 1581,  632, 1393, 1780, 1823, 1936, 1074, 1218 },
-    {  510, 1270, 1467, 1319,  847, 1279, 1792, 2094, 1080, 1353 },
-    {  488, 1322,  918, 1573, 1300,  883, 1814, 1752, 1756, 1502 },
-    {  425,  992, 1820, 1514, 1843, 2440,  937, 1771, 1924, 1129 },
-    {  363, 1248, 1257, 1970, 2194, 2385, 1569,  953, 1951, 1601 },
-    {  723, 1257, 1631,  964,  963, 1508, 1697, 1824,  671, 1418 },
-    {  635, 1038, 1573,  930, 1673, 1413, 1410, 1687, 1410,  749 } },
-  { {  451,  613, 1345, 1702, 1870, 1716, 1728, 1766, 2190, 2310 },
-    {  678,  453, 1171, 1443, 1925, 1831, 2045, 1781, 1887, 1602 },
-    {  711,  666,  674, 1718, 1910, 1493, 1775, 1193, 2325, 2325 },
-    {  883,  854, 1583,  542, 1800, 1878, 1664, 2149, 1207, 1087 },
-    {  669,  994, 1248, 1122,  949, 1179, 1376, 1729, 1070, 1244 },
-    {  715, 1026,  715, 1350, 1430,  930, 1717, 1296, 1479, 1479 },
-    {  544,  841, 1656, 1450, 2094, 3883, 1010, 1759, 2076,  809 },
-    {  610,  855,  957, 1553, 2067, 1561, 1704,  824, 2066, 1226 },
-    {  833,  960, 1416,  819, 1277, 1619, 1501, 1617,  757, 1182 },
-    {  711,  964, 1252,  879, 1441, 1828, 1508, 1636, 1594,  734 } },
-  { {  605,  764,  734, 1713, 1747, 1192, 1819, 1353, 1877, 2392 },
-    {  866,  641,  586, 1622, 2072, 1431, 1888, 1346, 2189, 1764 },
-    {  901,  851,  456, 2165, 2281, 1405, 1739, 1193, 2183, 2443 },
-    {  770, 1045,  952, 1078, 1342, 1191, 1436, 1063, 1303,  995 },
-    {  901, 1086,  727, 1170,  884, 1105, 1267, 1401, 1739, 1337 },
-    {  951, 1162,  595, 1488, 1388,  703, 1790, 1366, 2057, 1724 },
-    {  534,  986, 1273, 1987, 3273, 1485, 1024, 1399, 1583,  866 },
-    {  699, 1182,  695, 1978, 1726, 1986, 1326,  714, 1750, 1672 },
-    {  951, 1217, 1209,  920, 1062, 1441, 1548,  999,  952,  932 },
-    {  733, 1284,  784, 1256, 1557, 1098, 1257, 1357, 1414,  908 } },
-  { {  316, 1075, 1653, 1220, 2145, 2051, 1730, 2131, 1884, 1790 },
-    {  745,  516, 1404,  894, 1599, 2375, 2013, 2105, 1475, 1381 },
-    {  516,  729, 1088, 1319, 1637, 3426, 1636, 1275, 1531, 1453 },
-    {  894,  943, 2138,  468, 1704, 2259, 2069, 1763, 1266, 1158 },
-    {  605, 1025, 1235,  871, 1170, 1767, 1493, 1500, 1104, 1258 },
-    {  739,  826, 1207, 1151, 1412,  846, 1305, 2726, 1014, 1569 },
-    {  558,  825, 1820, 1398, 3344, 1556, 1218, 1550, 1228,  878 },
-    {  429,  951, 1089, 1816, 3861, 3861, 1556,  969, 1568, 1828 },
-    {  883,  961, 1752,  769, 1468, 1810, 2081, 2346,  613, 1298 },
-    {  803,  895, 1372,  641, 1303, 1708, 1686, 1700, 1306, 1033 } },
-  { {  439, 1267, 1270, 1579,  963, 1193, 1723, 1729, 1198, 1993 },
-    {  705,  725, 1029, 1153, 1176, 1103, 1821, 1567, 1259, 1574 },
-    {  723,  859,  802, 1253,  972, 1202, 1407, 1665, 1520, 1674 },
-    {  894,  960, 1254,  887, 1052, 1607, 1344, 1349,  865, 1150 },
-    {  833, 1312, 1337, 1205,  572, 1288, 1414, 1529, 1088, 1430 },
-    {  842, 1279, 1068, 1861,  862,  688, 1861, 1630, 1039, 1381 },
-    {  766,  938, 1279, 1546, 3338, 1550, 1031, 1542, 1288,  640 },
-    {  715, 1090,  835, 1609, 1100, 1100, 1603, 1019, 1102, 1617 },
-    {  894, 1813, 1500, 1188,  789, 1194, 1491, 1919,  617, 1333 },
-    {  610, 1076, 1644, 1281, 1283,  975, 1179, 1688, 1434,  889 } },
-  { {  544,  971, 1146, 1849, 1221,  740, 1857, 1621, 1683, 2430 },
-    {  723,  705,  961, 1371, 1426,  821, 2081, 2079, 1839, 1380 },
-    {  783,  857,  703, 2145, 1419,  814, 1791, 1310, 1609, 2206 },
-    {  997, 1000, 1153,  792, 1229, 1162, 1810, 1418,  942,  979 },
-    {  901, 1226,  883, 1289,  793,  715, 1904, 1649, 1319, 3108 },
-    {  979, 1478,  782, 2216, 1454,  455, 3092, 1591, 1997, 1664 },
-    {  663, 1110, 1504, 1114, 1522, 3311,  676, 1522, 1530, 1024 },
-    {  605, 1138, 1153, 1314, 1569, 1315, 1157,  804, 1574, 1320 },
-    {  770, 1216, 1218, 1227,  869, 1384, 1232, 1375,  834, 1239 },
-    {  775, 1007,  843, 1216, 1225, 1074, 2527, 1479, 1149,  975 } },
-  { {  477,  817, 1309, 1439, 1708, 1454, 1159, 1241, 1945, 1672 },
-    {  577,  796, 1112, 1271, 1618, 1458, 1087, 1345, 1831, 1265 },
-    {  663,  776,  753, 1940, 1690, 1690, 1227, 1097, 3149, 1361 },
-    {  766, 1299, 1744, 1161, 1565, 1106, 1045, 1230, 1232,  707 },
-    {  915, 1026, 1404, 1182, 1184,  851, 1428, 2425, 1043,  789 },
-    {  883, 1456,  790, 1082, 1086,  985, 1083, 1484, 1238, 1160 },
-    {  507, 1345, 2261, 1995, 1847, 3636,  653, 1761, 2287,  933 },
-    {  553, 1193, 1470, 2057, 2059, 2059,  833,  779, 2058, 1263 },
-    {  766, 1275, 1515, 1039,  957, 1554, 1286, 1540, 1289,  705 },
-    {  499, 1378, 1496, 1385, 1850, 1850, 1044, 2465, 1515,  720 } },
-  { {  553,  930,  978, 2077, 1968, 1481, 1457,  761, 1957, 2362 },
-    {  694,  864,  905, 1720, 1670, 1621, 1429,  718, 2125, 1477 },
-    {  699,  968,  658, 3190, 2024, 1479, 1865,  750, 2060, 2320 },
-    {  733, 1308, 1296, 1062, 1576, 1322, 1062, 1112, 1172,  816 },
-    {  920,  927, 1052,  939,  947, 1156, 1152, 1073, 3056, 1268 },
-    {  723, 1534,  711, 1547, 1294,  892, 1553,  928, 1815, 1561 },
-    {  663, 1366, 1583, 2111, 1712, 3501,  522, 1155, 2130, 1133 },
-    {  614, 1731, 1188, 2343, 1944, 3733, 1287,  487, 3546, 1758 },
-    {  770, 1585, 1312,  826,  884, 2673, 1185, 1006, 1195, 1195 },
-    {  758, 1333, 1273, 1023, 1621, 1162, 1351,  833, 1479,  862 } },
-  { {  376, 1193, 1446, 1149, 1545, 1577, 1870, 1789, 1175, 1823 },
-    {  803,  633, 1136, 1058, 1350, 1323, 1598, 2247, 1072, 1252 },
-    {  614, 1048,  943,  981, 1152, 1869, 1461, 1020, 1618, 1618 },
-    { 1107, 1085, 1282,  592, 1779, 1933, 1648, 2403,  691, 1246 },
-    {  851, 1309, 1223, 1243,  895, 1593, 1792, 2317,  627, 1076 },
-    {  770, 1216, 1030, 1125,  921,  981, 1629, 1131, 1049, 1646 },
-    {  626, 1469, 1456, 1081, 1489, 3278,  981, 1232, 1498,  733 },
-    {  617, 1201,  812, 1220, 1476, 1476, 1478,  970, 1228, 1488 },
-    { 1179, 1393, 1540,  999, 1243, 1503, 1916, 1925,  414, 1614 },
-    {  943, 1088, 1490,  682, 1112, 1372, 1756, 1505,  966,  966 } },
-  { {  322, 1142, 1589, 1396, 2144, 1859, 1359, 1925, 2084, 1518 },
-    {  617,  625, 1241, 1234, 2121, 1615, 1524, 1858, 1720, 1004 },
-    {  553,  851,  786, 1299, 1452, 1560, 1372, 1561, 1967, 1713 },
-    {  770,  977, 1396,  568, 1893, 1639, 1540, 2108, 1430, 1013 },
-    {  684, 1120, 1375,  982,  930, 2719, 1638, 1643,  933,  993 },
-    {  553, 1103,  996, 1356, 1361, 1005, 1507, 1761, 1184, 1268 },
-    {  419, 1247, 1537, 1554, 1817, 3606, 1026, 1666, 1829,  923 },
-    {  439, 1139, 1101, 1257, 3710, 1922, 1205, 1040, 1931, 1529 },
-    {  979,  935, 1269,  847, 1202, 1286, 1530, 1535,  827, 1036 },
-    {  516, 1378, 1569, 1110, 1798, 1798, 1198, 2199, 1543,  712 } },
+  { {   40, 1151, 1723, 1874, 2103, 2019, 1628, 1777, 2226, 2137 },
+    {  192,  469, 1296, 1308, 1849, 1794, 1781, 1703, 1713, 1522 },
+    {  142,  910,  762, 1684, 1849, 1576, 1460, 1305, 1801, 1657 },
+    {  559,  641, 1370,  421, 1182, 1569, 1612, 1725,  863, 1007 },
+    {  299, 1059, 1256, 1108,  636, 1068, 1581, 1883,  869, 1142 },
+    {  277, 1111,  707, 1362, 1089,  672, 1603, 1541, 1545, 1291 },
+    {  214,  781, 1609, 1303, 1632, 2229,  726, 1560, 1713,  918 },
+    {  152, 1037, 1046, 1759, 1983, 2174, 1358,  742, 1740, 1390 },
+    {  512, 1046, 1420,  753,  752, 1297, 1486, 1613,  460, 1207 },
+    {  424,  827, 1362,  719, 1462, 1202, 1199, 1476, 1199,  538 } },
+  { {  240,  402, 1134, 1491, 1659, 1505, 1517, 1555, 1979, 2099 },
+    {  467,  242,  960, 1232, 1714, 1620, 1834, 1570, 1676, 1391 },
+    {  500,  455,  463, 1507, 1699, 1282, 1564,  982, 2114, 2114 },
+    {  672,  643, 1372,  331, 1589, 1667, 1453, 1938,  996,  876 },
+    {  458,  783, 1037,  911,  738,  968, 1165, 1518,  859, 1033 },
+    {  504,  815,  504, 1139, 1219,  719, 1506, 1085, 1268, 1268 },
+    {  333,  630, 1445, 1239, 1883, 3672,  799, 1548, 1865,  598 },
+    {  399,  644,  746, 1342, 1856, 1350, 1493,  613, 1855, 1015 },
+    {  622,  749, 1205,  608, 1066, 1408, 1290, 1406,  546,  971 },
+    {  500,  753, 1041,  668, 1230, 1617, 1297, 1425, 1383,  523 } },
+  { {  394,  553,  523, 1502, 1536,  981, 1608, 1142, 1666, 2181 },
+    {  655,  430,  375, 1411, 1861, 1220, 1677, 1135, 1978, 1553 },
+    {  690,  640,  245, 1954, 2070, 1194, 1528,  982, 1972, 2232 },
+    {  559,  834,  741,  867, 1131,  980, 1225,  852, 1092,  784 },
+    {  690,  875,  516,  959,  673,  894, 1056, 1190, 1528, 1126 },
+    {  740,  951,  384, 1277, 1177,  492, 1579, 1155, 1846, 1513 },
+    {  323,  775, 1062, 1776, 3062, 1274,  813, 1188, 1372,  655 },
+    {  488,  971,  484, 1767, 1515, 1775, 1115,  503, 1539, 1461 },
+    {  740, 1006,  998,  709,  851, 1230, 1337,  788,  741,  721 },
+    {  522, 1073,  573, 1045, 1346,  887, 1046, 1146, 1203,  697 } },
+  { {  105,  864, 1442, 1009, 1934, 1840, 1519, 1920, 1673, 1579 },
+    {  534,  305, 1193,  683, 1388, 2164, 1802, 1894, 1264, 1170 },
+    {  305,  518,  877, 1108, 1426, 3215, 1425, 1064, 1320, 1242 },
+    {  683,  732, 1927,  257, 1493, 2048, 1858, 1552, 1055,  947 },
+    {  394,  814, 1024,  660,  959, 1556, 1282, 1289,  893, 1047 },
+    {  528,  615,  996,  940, 1201,  635, 1094, 2515,  803, 1358 },
+    {  347,  614, 1609, 1187, 3133, 1345, 1007, 1339, 1017,  667 },
+    {  218,  740,  878, 1605, 3650, 3650, 1345,  758, 1357, 1617 },
+    {  672,  750, 1541,  558, 1257, 1599, 1870, 2135,  402, 1087 },
+    {  592,  684, 1161,  430, 1092, 1497, 1475, 1489, 1095,  822 } },
+  { {  228, 1056, 1059, 1368,  752,  982, 1512, 1518,  987, 1782 },
+    {  494,  514,  818,  942,  965,  892, 1610, 1356, 1048, 1363 },
+    {  512,  648,  591, 1042,  761,  991, 1196, 1454, 1309, 1463 },
+    {  683,  749, 1043,  676,  841, 1396, 1133, 1138,  654,  939 },
+    {  622, 1101, 1126,  994,  361, 1077, 1203, 1318,  877, 1219 },
+    {  631, 1068,  857, 1650,  651,  477, 1650, 1419,  828, 1170 },
+    {  555,  727, 1068, 1335, 3127, 1339,  820, 1331, 1077,  429 },
+    {  504,  879,  624, 1398,  889,  889, 1392,  808,  891, 1406 },
+    {  683, 1602, 1289,  977,  578,  983, 1280, 1708,  406, 1122 },
+    {  399,  865, 1433, 1070, 1072,  764,  968, 1477, 1223,  678 } },
+  { {  333,  760,  935, 1638, 1010,  529, 1646, 1410, 1472, 2219 },
+    {  512,  494,  750, 1160, 1215,  610, 1870, 1868, 1628, 1169 },
+    {  572,  646,  492, 1934, 1208,  603, 1580, 1099, 1398, 1995 },
+    {  786,  789,  942,  581, 1018,  951, 1599, 1207,  731,  768 },
+    {  690, 1015,  672, 1078,  582,  504, 1693, 1438, 1108, 2897 },
+    {  768, 1267,  571, 2005, 1243,  244, 2881, 1380, 1786, 1453 },
+    {  452,  899, 1293,  903, 1311, 3100,  465, 1311, 1319,  813 },
+    {  394,  927,  942, 1103, 1358, 1104,  946,  593, 1363, 1109 },
+    {  559, 1005, 1007, 1016,  658, 1173, 1021, 1164,  623, 1028 },
+    {  564,  796,  632, 1005, 1014,  863, 2316, 1268,  938,  764 } },
+  { {  266,  606, 1098, 1228, 1497, 1243,  948, 1030, 1734, 1461 },
+    {  366,  585,  901, 1060, 1407, 1247,  876, 1134, 1620, 1054 },
+    {  452,  565,  542, 1729, 1479, 1479, 1016,  886, 2938, 1150 },
+    {  555, 1088, 1533,  950, 1354,  895,  834, 1019, 1021,  496 },
+    {  704,  815, 1193,  971,  973,  640, 1217, 2214,  832,  578 },
+    {  672, 1245,  579,  871,  875,  774,  872, 1273, 1027,  949 },
+    {  296, 1134, 2050, 1784, 1636, 3425,  442, 1550, 2076,  722 },
+    {  342,  982, 1259, 1846, 1848, 1848,  622,  568, 1847, 1052 },
+    {  555, 1064, 1304,  828,  746, 1343, 1075, 1329, 1078,  494 },
+    {  288, 1167, 1285, 1174, 1639, 1639,  833, 2254, 1304,  509 } },
+  { {  342,  719,  767, 1866, 1757, 1270, 1246,  550, 1746, 2151 },
+    {  483,  653,  694, 1509, 1459, 1410, 1218,  507, 1914, 1266 },
+    {  488,  757,  447, 2979, 1813, 1268, 1654,  539, 1849, 2109 },
+    {  522, 1097, 1085,  851, 1365, 1111,  851,  901,  961,  605 },
+    {  709,  716,  841,  728,  736,  945,  941,  862, 2845, 1057 },
+    {  512, 1323,  500, 1336, 1083,  681, 1342,  717, 1604, 1350 },
+    {  452, 1155, 1372, 1900, 1501, 3290,  311,  944, 1919,  922 },
+    {  403, 1520,  977, 2132, 1733, 3522, 1076,  276, 3335, 1547 },
+    {  559, 1374, 1101,  615,  673, 2462,  974,  795,  984,  984 },
+    {  547, 1122, 1062,  812, 1410,  951, 1140,  622, 1268,  651 } },
+  { {  165,  982, 1235,  938, 1334, 1366, 1659, 1578,  964, 1612 },
+    {  592,  422,  925,  847, 1139, 1112, 1387, 2036,  861, 1041 },
+    {  403,  837,  732,  770,  941, 1658, 1250,  809, 1407, 1407 },
+    {  896,  874, 1071,  381, 1568, 1722, 1437, 2192,  480, 1035 },
+    {  640, 1098, 1012, 1032,  684, 1382, 1581, 2106,  416,  865 },
+    {  559, 1005,  819,  914,  710,  770, 1418,  920,  838, 1435 },
+    {  415, 1258, 1245,  870, 1278, 3067,  770, 1021, 1287,  522 },
+    {  406,  990,  601, 1009, 1265, 1265, 1267,  759, 1017, 1277 },
+    {  968, 1182, 1329,  788, 1032, 1292, 1705, 1714,  203, 1403 },
+    {  732,  877, 1279,  471,  901, 1161, 1545, 1294,  755,  755 } },
+  { {  111,  931, 1378, 1185, 1933, 1648, 1148, 1714, 1873, 1307 },
+    {  406,  414, 1030, 1023, 1910, 1404, 1313, 1647, 1509,  793 },
+    {  342,  640,  575, 1088, 1241, 1349, 1161, 1350, 1756, 1502 },
+    {  559,  766, 1185,  357, 1682, 1428, 1329, 1897, 1219,  802 },
+    {  473,  909, 1164,  771,  719, 2508, 1427, 1432,  722,  782 },
+    {  342,  892,  785, 1145, 1150,  794, 1296, 1550,  973, 1057 },
+    {  208, 1036, 1326, 1343, 1606, 3395,  815, 1455, 1618,  712 },
+    {  228,  928,  890, 1046, 3499, 1711,  994,  829, 1720, 1318 },
+    {  768,  724, 1058,  636,  991, 1075, 1319, 1324,  616,  825 },
+    {  305, 1167, 1358,  899, 1587, 1587,  987, 1988, 1332,  501 } }
 };
 
 //------------------------------------------------------------------------------
+// helper functions for residuals struct VP8Residual.
+
+void VP8InitResidual(int first, int coeff_type,
+                     VP8Encoder* const enc, VP8Residual* const res) {
+  res->coeff_type = coeff_type;
+  res->prob  = enc->proba_.coeffs_[coeff_type];
+  res->stats = enc->proba_.stats_[coeff_type];
+  res->costs = enc->proba_.remapped_costs_[coeff_type];
+  res->first = first;
+}
+
+//------------------------------------------------------------------------------
+// Mode costs
+
+int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]) {
+  const int x = (it->i4_ & 3), y = (it->i4_ >> 2);
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int R = 0;
+  int ctx;
+
+  VP8InitResidual(0, 3, enc, &res);
+  ctx = it->top_nz_[x] + it->left_nz_[y];
+  VP8SetResidualCoeffs(levels, &res);
+  R += VP8GetResidualCost(ctx, &res);
+  return R;
+}
+
+int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd) {
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int x, y;
+  int R = 0;
+
+  VP8IteratorNzToBytes(it);   // re-import the non-zero context
+
+  // DC
+  VP8InitResidual(0, 1, enc, &res);
+  VP8SetResidualCoeffs(rd->y_dc_levels, &res);
+  R += VP8GetResidualCost(it->top_nz_[8] + it->left_nz_[8], &res);
+
+  // AC
+  VP8InitResidual(1, 0, enc, &res);
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      R += VP8GetResidualCost(ctx, &res);
+      it->top_nz_[x] = it->left_nz_[y] = (res.last >= 0);
+    }
+  }
+  return R;
+}
+
+int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int ch, x, y;
+  int R = 0;
+
+  VP8IteratorNzToBytes(it);  // re-import the non-zero context
+
+  VP8InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        R += VP8GetResidualCost(ctx, &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = (res.last >= 0);
+      }
+    }
+  }
+  return R;
+}
+
+
+//------------------------------------------------------------------------------
+// Recording of token probabilities.
+
+// 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
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
+// Simulate block coding, but only record statistics.
+// Note: no need to record the fixed probas.
+int VP8RecordCoeffs(int ctx, const VP8Residual* const res) {
+  int n = res->first;
+  // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  proba_t* s = res->stats[n][ctx];
+  if (res->last  < 0) {
+    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;
+}
+
+//------------------------------------------------------------------------------
diff --git a/drivers/webp/enc/cost.h b/drivers/webp/enc/cost.h
index 09b75b699d..20960d6d74 100644
--- a/drivers/webp/enc/cost.h
+++ b/drivers/webp/enc/cost.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
@@ -12,14 +14,32 @@
 #ifndef WEBP_ENC_COST_H_
 #define WEBP_ENC_COST_H_
 
+#include <assert.h>
+#include <stdlib.h>
 #include "./vp8enci.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-extern const uint16_t VP8LevelFixedCosts[2048];   // approximate cost per level
-extern const uint16_t VP8EntropyCost[256];        // 8bit fixed-point log(p)
+// On-the-fly info about the current set of residuals. Handy to avoid
+// passing zillions of params.
+typedef struct VP8Residual VP8Residual;
+struct VP8Residual {
+  int first;
+  int last;
+  const int16_t* coeffs;
+
+  int coeff_type;
+  ProbaArray*   prob;
+  StatsArray*   stats;
+  CostArrayPtr  costs;
+};
+
+void VP8InitResidual(int first, int coeff_type,
+                     VP8Encoder* const enc, VP8Residual* const res);
+
+int VP8RecordCoeffs(int ctx, const VP8Residual* const res);
 
 // Cost of coding one event with probability 'proba'.
 static WEBP_INLINE int VP8BitCost(int bit, uint8_t proba) {
@@ -28,7 +48,7 @@ static WEBP_INLINE int VP8BitCost(int bit, uint8_t proba) {
 
 // Level cost calculations
 extern const uint16_t VP8LevelCodes[MAX_VARIABLE_LEVEL][2];
-void VP8CalculateLevelCosts(VP8Proba* const proba);
+void VP8CalculateLevelCosts(VP8EncProba* const proba);
 static WEBP_INLINE int VP8LevelCost(const uint16_t* const table, int level) {
   return VP8LevelFixedCosts[level]
        + table[(level > MAX_VARIABLE_LEVEL) ? MAX_VARIABLE_LEVEL : level];
@@ -41,7 +61,7 @@ extern const uint16_t VP8FixedCostsI4[NUM_BMODES][NUM_BMODES][NUM_BMODES];
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/enc/filter.c b/drivers/webp/enc/filter.c
index 7fb78a3949..1a4dd947fb 100644
--- a/drivers/webp/enc/filter.c
+++ b/drivers/webp/enc/filter.c
@@ -1,194 +1,68 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-// 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;
+#include "../dsp/dsp.h"
+
+// This table gives, for a given sharpness, the filtering strength to be
+// used (at least) in order to filter a given edge step delta.
+// This is constructed by brute force inspection: for all delta, we iterate
+// over all possible filtering strength / thresh until needs_filter() returns
+// true.
+#define MAX_DELTA_SIZE 64
+static const uint8_t kLevelsFromDelta[8][MAX_DELTA_SIZE] = {
+  { 0,   1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+    16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+    32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+    48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63 },
+  { 0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 14, 15, 17, 18,
+    20, 21, 23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42,
+    44, 45, 47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 14, 16, 17, 19,
+    20, 22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43,
+    44, 46, 47, 49, 50, 52, 53, 55, 56, 58, 59, 61, 62, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 15, 16, 18, 19,
+    21, 22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43,
+    45, 46, 48, 49, 51, 52, 54, 55, 57, 58, 60, 61, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 14, 15, 17, 18, 20,
+    21, 23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42, 44,
+    45, 47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 13, 15, 16, 17, 19, 20,
+    22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43, 44,
+    46, 47, 49, 50, 52, 53, 55, 56, 58, 59, 61, 62, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 13, 15, 16, 18, 19, 21,
+    22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43, 45,
+    46, 48, 49, 51, 52, 54, 55, 57, 58, 60, 61, 63, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 14, 15, 17, 18, 20, 21,
+    23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42, 44, 45,
+    47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 }
+};
+
+int VP8FilterStrengthFromDelta(int sharpness, int delta) {
+  const int pos = (delta < MAX_DELTA_SIZE) ? delta : MAX_DELTA_SIZE - 1;
+  assert(sharpness >= 0 && sharpness <= 7);
+  return kLevelsFromDelta[sharpness][pos];
 }
 
 //------------------------------------------------------------------------------
-// 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) {
@@ -211,22 +85,22 @@ static void DoFilter(const VP8EncIterator* const it, int level) {
   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;
+  uint8_t* const y_dst = it->yuv_out2_ + Y_OFF_ENC;
+  uint8_t* const u_dst = it->yuv_out2_ + U_OFF_ENC;
+  uint8_t* const v_dst = it->yuv_out2_ + V_OFF_ENC;
 
   // copy current block to yuv_out2_
-  memcpy(y_dst, it->yuv_out_, YUV_SIZE * sizeof(uint8_t));
+  memcpy(y_dst, it->yuv_out_, YUV_SIZE_ENC * sizeof(uint8_t));
 
   if (enc->filter_hdr_.simple_ == 1) {   // simple
-    VP8EncSimpleHFilter16i(y_dst, BPS, limit);
-    VP8EncSimpleVFilter16i(y_dst, BPS, limit);
+    VP8SimpleHFilter16i(y_dst, BPS, limit);
+    VP8SimpleVFilter16i(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);
+    VP8HFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8HFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+    VP8VFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8VFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
   }
 }
 
@@ -321,13 +195,16 @@ static double GetMBSSIM(const uint8_t* yuv1, const uint8_t* yuv2) {
   // 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);
+      VP8SSIMAccumulate(yuv1 + Y_OFF_ENC, BPS, yuv2 + Y_OFF_ENC, 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);
+      VP8SSIMAccumulate(yuv1 + U_OFF_ENC, BPS, yuv2 + U_OFF_ENC, BPS,
+                        x, y, 8, 8, &s);
+      VP8SSIMAccumulate(yuv1 + V_OFF_ENC, BPS, yuv2 + V_OFF_ENC, BPS,
+                        x, y, 8, 8, &s);
     }
   }
   return VP8SSIMGet(&s);
@@ -338,28 +215,28 @@ static double GetMBSSIM(const uint8_t* yuv1, const uint8_t* yuv2) {
 // loop filter strength
 
 void VP8InitFilter(VP8EncIterator* const it) {
-  int s, i;
-  if (!it->lf_stats_) return;
-
-  InitTables();
-  for (s = 0; s < NUM_MB_SEGMENTS; s++) {
-    for (i = 0; i < MAX_LF_LEVELS; i++) {
-      (*it->lf_stats_)[s][i] = 0;
+  if (it->lf_stats_ != NULL) {
+    int s, i;
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      for (i = 0; i < MAX_LF_LEVELS; i++) {
+        (*it->lf_stats_)[s][i] = 0;
+      }
     }
   }
 }
 
 void VP8StoreFilterStats(VP8EncIterator* const it) {
   int d;
+  VP8Encoder* const enc = it->enc_;
   const int s = it->mb_->segment_;
-  const int level0 = it->enc_->dqm_[s].fstrength_;  // TODO: ref_lf_delta[]
+  const int level0 = enc->dqm_[s].fstrength_;  // TODO: ref_lf_delta[]
 
   // explore +/-quant range of values around level0
-  const int delta_min = -it->enc_->dqm_[s].quant_;
-  const int delta_max = it->enc_->dqm_[s].quant_;
+  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_) return;
+  if (it->lf_stats_ == NULL) return;
 
   // NOTE: Currently we are applying filter only across the sublock edges
   // There are two reasons for that.
@@ -383,27 +260,40 @@ void VP8StoreFilterStats(VP8EncIterator* const it) {
 }
 
 void VP8AdjustFilterStrength(VP8EncIterator* const it) {
-  int s;
   VP8Encoder* const enc = it->enc_;
-
-  if (!it->lf_stats_) {
-    return;
-  }
-  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;
+  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;
     }
-    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;
   }
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
+// -----------------------------------------------------------------------------
diff --git a/drivers/webp/enc/frame.c b/drivers/webp/enc/frame.c
index bdd360069b..5b7a40b9ad 100644
--- a/drivers/webp/enc/frame.c
+++ b/drivers/webp/enc/frame.c
@@ -1,61 +1,98 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "./vp8enci.h"
+#include "../dsp/dsp.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
-};
-
-static const uint8_t kCat3[] = { 173, 148, 140 };
-static const uint8_t kCat4[] = { 176, 155, 140, 135 };
-static const uint8_t kCat5[] = { 180, 157, 141, 134, 130 };
-static const uint8_t kCat6[] =
+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_;
+  VP8EncProba* const proba = &enc->proba_;
   VP8CalculateLevelCosts(proba);
   proba->nb_skip_ = 0;
 }
@@ -71,7 +108,7 @@ static int CalcSkipProba(uint64_t nb, uint64_t total) {
 
 // Returns the bit-cost for coding the skip probability.
 static int FinalizeSkipProba(VP8Encoder* const enc) {
-  VP8Proba* const proba = &enc->proba_;
+  VP8EncProba* const proba = &enc->proba_;
   const int nb_mbs = enc->mb_w_ * enc->mb_h_;
   const int nb_events = proba->nb_skip_;
   int size;
@@ -86,82 +123,6 @@ static int FinalizeSkipProba(VP8Encoder* const enc) {
   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;
-  proba_t* s = res->stats[VP8EncBands[n]][ctx];
-  if (res->last  < 0) {
-    Record(0, s + 0);
-    return 0;
-  }
-  while (n <= res->last) {
-    int v;
-    Record(1, s + 0);
-    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) {
@@ -174,8 +135,12 @@ static int BranchCost(int nb, int total, int proba) {
   return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba);
 }
 
-static int FinalizeTokenProbas(VP8Encoder* const enc) {
-  VP8Proba* const proba = &enc->proba_;
+static void ResetTokenStats(VP8Encoder* const enc) {
+  VP8EncProba* const proba = &enc->proba_;
+  memset(proba->stats_, 0, sizeof(proba->stats_));
+}
+
+static int FinalizeTokenProbas(VP8EncProba* const proba) {
   int has_changed = 0;
   int size = 0;
   int t, b, c, p;
@@ -212,129 +177,44 @@ static int FinalizeTokenProbas(VP8Encoder* const enc) {
 }
 
 //------------------------------------------------------------------------------
-// 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;
-}
+// Finalize Segment probability based on the coding tree
 
-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;
+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
 }
 
-//------------------------------------------------------------------------------
-// Mode costs
-
-static int GetResidualCost(int ctx, const VP8Residual* const res) {
-  int n = res->first;
-  int p0 = res->prob[VP8EncBands[n]][ctx][0];
-  const uint16_t* t = res->cost[VP8EncBands[n]][ctx];
-  int cost;
+static void SetSegmentProbas(VP8Encoder* const enc) {
+  int p[NUM_MB_SEGMENTS] = { 0 };
+  int n;
 
-  if (res->last < 0) {
-    return VP8BitCost(0, p0);
-  }
-  cost = 0;
-  while (n <= res->last) {
-    const int v = res->coeffs[n];
-    const int b = VP8EncBands[n + 1];
-    ++n;
-    if (v == 0) {
-      // short-case for VP8LevelCost(t, 0) (note: VP8LevelFixedCosts[0] == 0):
-      cost += t[0];
-      t = res->cost[b][0];
-      continue;
-    }
-    cost += VP8BitCost(1, p0);
-    if (2u >= (unsigned int)(v + 1)) {   // v = -1 or 1
-      // short-case for "VP8LevelCost(t, 1)" (256 is VP8LevelFixedCosts[1]):
-      cost += 256 + t[1];
-      p0 = res->prob[b][1][0];
-      t = res->cost[b][1];
-    } else {
-      cost += VP8LevelCost(t, abs(v));
-      p0 = res->prob[b][2][0];
-      t = res->cost[b][2];
-    }
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    const VP8MBInfo* const mb = &enc->mb_info_[n];
+    p[mb->segment_]++;
   }
-  if (n < 16) cost += VP8BitCost(0, 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);
+  if (enc->pic_->stats != NULL) {
+    for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
+      enc->pic_->stats->segment_size[n] = p[n];
     }
   }
-  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);
-      }
-    }
+  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;
   }
-  return R;
 }
 
 //------------------------------------------------------------------------------
@@ -342,7 +222,8 @@ int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
 
 static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
   int n = res->first;
-  const uint8_t* p = res->prob[VP8EncBands[n]][ctx];
+  // 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;
   }
@@ -371,30 +252,30 @@ static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
       } else {
         int mask;
         const uint8_t* tab;
-        if (v < 3 + (8 << 1)) {          // kCat3  (3b)
+        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 = kCat3;
-        } else if (v < 3 + (8 << 2)) {   // kCat4  (4b)
+          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 = kCat4;
-        } else if (v < 3 + (8 << 3)) {   // kCat5  (5b)
+          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 = kCat5;
-        } else {                         // kCat6 (11b)
+          tab = VP8Cat5;
+        } else {                         // VP8Cat6 (11b)
           VP8PutBit(bw, 1, p[8]);
           VP8PutBit(bw, 1, p[10]);
           v -= 3 + (8 << 3);
           mask = 1 << 10;
-          tab = kCat6;
+          tab = VP8Cat6;
         }
         while (mask) {
           VP8PutBit(bw, !!(v & mask), *tab++);
@@ -411,8 +292,7 @@ static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
   return 1;
 }
 
-static void CodeResiduals(VP8BitWriter* const bw,
-                          VP8EncIterator* const it,
+static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it,
                           const VP8ModeScore* const rd) {
   int x, y, ch;
   VP8Residual res;
@@ -425,32 +305,32 @@ static void CodeResiduals(VP8BitWriter* const bw,
 
   pos1 = VP8BitWriterPos(bw);
   if (i16) {
-    InitResidual(0, 1, enc, &res);
-    SetResidualCoeffs(rd->y_dc_levels, &res);
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(rd->y_dc_levels, &res);
     it->top_nz_[8] = it->left_nz_[8] =
       PutCoeffs(bw, it->top_nz_[8] + it->left_nz_[8], &res);
-    InitResidual(1, 0, enc, &res);
+    VP8InitResidual(1, 0, enc, &res);
   } else {
-    InitResidual(0, 3, enc, &res);
+    VP8InitResidual(0, 3, enc, &res);
   }
 
   // luma-AC
   for (y = 0; y < 4; ++y) {
     for (x = 0; x < 4; ++x) {
       const int ctx = it->top_nz_[x] + it->left_nz_[y];
-      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
       it->top_nz_[x] = it->left_nz_[y] = PutCoeffs(bw, ctx, &res);
     }
   }
   pos2 = VP8BitWriterPos(bw);
 
   // U/V
-  InitResidual(0, 2, enc, &res);
+  VP8InitResidual(0, 2, enc, &res);
   for (ch = 0; ch <= 2; ch += 2) {
     for (y = 0; y < 2; ++y) {
       for (x = 0; x < 2; ++x) {
         const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
-        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
         it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
             PutCoeffs(bw, ctx, &res);
       }
@@ -475,33 +355,33 @@ static void RecordResiduals(VP8EncIterator* const it,
   VP8IteratorNzToBytes(it);
 
   if (it->mb_->type_ == 1) {   // i16x16
-    InitResidual(0, 1, enc, &res);
-    SetResidualCoeffs(rd->y_dc_levels, &res);
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(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);
+      VP8RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res);
+    VP8InitResidual(1, 0, enc, &res);
   } else {
-    InitResidual(0, 3, enc, &res);
+    VP8InitResidual(0, 3, enc, &res);
   }
 
   // luma-AC
   for (y = 0; y < 4; ++y) {
     for (x = 0; x < 4; ++x) {
       const int ctx = it->top_nz_[x] + it->left_nz_[y];
-      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
-      it->top_nz_[x] = it->left_nz_[y] = RecordCoeffs(ctx, &res);
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] = VP8RecordCoeffs(ctx, &res);
     }
   }
 
   // U/V
-  InitResidual(0, 2, enc, &res);
+  VP8InitResidual(0, 2, enc, &res);
   for (ch = 0; ch <= 2; ch += 2) {
     for (y = 0; y < 2; ++y) {
       for (x = 0; x < 2; ++x) {
         const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
-        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
         it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
-            RecordCoeffs(ctx, &res);
+            VP8RecordCoeffs(ctx, &res);
       }
     }
   }
@@ -512,176 +392,59 @@ static void RecordResiduals(VP8EncIterator* const it,
 //------------------------------------------------------------------------------
 // Token buffer
 
-#ifdef USE_TOKEN_BUFFER
-
-void VP8TBufferInit(VP8TBuffer* const b) {
-  b->rows_ = NULL;
-  b->tokens_ = NULL;
-  b->last_ = &b->rows_;
-  b->left_ = 0;
-  b->error_ = 0;
-}
-
-int VP8TBufferNewPage(VP8TBuffer* const b) {
-  VP8Tokens* const page = b->error_ ? NULL : (VP8Tokens*)malloc(sizeof(*page));
-  if (page == NULL) {
-    b->error_ = 1;
-    return 0;
-  }
-  *b->last_ = page;
-  b->last_ = &page->next_;
-  b->left_ = MAX_NUM_TOKEN;
-  b->tokens_ = page->tokens_;
-  return 1;
-}
-
-void VP8TBufferClear(VP8TBuffer* const b) {
-  if (b != NULL) {
-    const VP8Tokens* p = b->rows_;
-    while (p != NULL) {
-      const VP8Tokens* const next = p->next_;
-      free((void*)p);
-      p = next;
-    }
-    VP8TBufferInit(b);
-  }
-}
-
-int VP8EmitTokens(const VP8TBuffer* const b, VP8BitWriter* const bw,
-                  const uint8_t* const probas) {
-  VP8Tokens* p = b->rows_;
-  if (b->error_) return 0;
-  while (p != NULL) {
-    const int N = (p->next_ == NULL) ? b->left_ : 0;
-    int n = MAX_NUM_TOKEN;
-    while (n-- > N) {
-      VP8PutBit(bw, (p->tokens_[n] >> 15) & 1, probas[p->tokens_[n] & 0x7fff]);
-    }
-    p = p->next_;
-  }
-  return 1;
-}
-
-#define TOKEN_ID(b, ctx, p) ((p) + NUM_PROBAS * ((ctx) + (b) * NUM_CTX))
-
-static int RecordCoeffTokens(int ctx, const VP8Residual* const res,
-                             VP8TBuffer* tokens) {
-  int n = res->first;
-  int b = VP8EncBands[n];
-  if (!VP8AddToken(tokens, res->last >= 0, TOKEN_ID(b, ctx, 0))) {
-    return 0;
-  }
-
-  while (n < 16) {
-    const int c = res->coeffs[n++];
-    const int sign = c < 0;
-    int v = sign ? -c : c;
-    const int base_id = TOKEN_ID(b, ctx, 0);
-    if (!VP8AddToken(tokens, v != 0, base_id + 1)) {
-      b = VP8EncBands[n];
-      ctx = 0;
-      continue;
-    }
-    if (!VP8AddToken(tokens, v > 1, base_id + 2)) {
-      b = VP8EncBands[n];
-      ctx = 1;
-    } else {
-      if (!VP8AddToken(tokens, v > 4, base_id + 3)) {
-        if (VP8AddToken(tokens, v != 2, base_id + 4))
-          VP8AddToken(tokens, v == 4, base_id + 5);
-      } else if (!VP8AddToken(tokens, v > 10, base_id + 6)) {
-        if (!VP8AddToken(tokens, v > 6, base_id + 7)) {
-//          VP8AddToken(tokens, v == 6, 159);
-        } else {
-//          VP8AddToken(tokens, v >= 9, 165);
-//          VP8AddToken(tokens, !(v & 1), 145);
-        }
-      } else {
-        int mask;
-        const uint8_t* tab;
-        if (v < 3 + (8 << 1)) {          // kCat3  (3b)
-          VP8AddToken(tokens, 0, base_id + 8);
-          VP8AddToken(tokens, 0, base_id + 9);
-          v -= 3 + (8 << 0);
-          mask = 1 << 2;
-          tab = kCat3;
-        } else if (v < 3 + (8 << 2)) {   // kCat4  (4b)
-          VP8AddToken(tokens, 0, base_id + 8);
-          VP8AddToken(tokens, 1, base_id + 9);
-          v -= 3 + (8 << 1);
-          mask = 1 << 3;
-          tab = kCat4;
-        } else if (v < 3 + (8 << 3)) {   // kCat5  (5b)
-          VP8AddToken(tokens, 1, base_id + 8);
-          VP8AddToken(tokens, 0, base_id + 10);
-          v -= 3 + (8 << 2);
-          mask = 1 << 4;
-          tab = kCat5;
-        } else {                         // kCat6 (11b)
-          VP8AddToken(tokens, 1, base_id + 8);
-          VP8AddToken(tokens, 1, base_id + 10);
-          v -= 3 + (8 << 3);
-          mask = 1 << 10;
-          tab = kCat6;
-        }
-        while (mask) {
-          // VP8AddToken(tokens, !!(v & mask), *tab++);
-          mask >>= 1;
-        }
-      }
-      ctx = 2;
-    }
-    b = VP8EncBands[n];
-    // VP8PutBitUniform(bw, sign);
-    if (n == 16 || !VP8AddToken(tokens, n <= res->last, TOKEN_ID(b, ctx, 0))) {
-      return 1;   // EOB
-    }
-  }
-  return 1;
-}
+#if !defined(DISABLE_TOKEN_BUFFER)
 
-static void RecordTokens(VP8EncIterator* const it,
-                         const VP8ModeScore* const rd, VP8TBuffer tokens[2]) {
+static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
+                        VP8TBuffer* const tokens) {
   int x, y, ch;
   VP8Residual res;
   VP8Encoder* const enc = it->enc_;
 
   VP8IteratorNzToBytes(it);
   if (it->mb_->type_ == 1) {   // i16x16
-    InitResidual(0, 1, enc, &res);
-    SetResidualCoeffs(rd->y_dc_levels, &res);
-// TODO(skal): FIX ->    it->top_nz_[8] = it->left_nz_[8] =
-      RecordCoeffTokens(it->top_nz_[8] + it->left_nz_[8], &res, &tokens[0]);
-    InitResidual(1, 0, enc, &res);
+    const int ctx = it->top_nz_[8] + it->left_nz_[8];
+    VP8InitResidual(0, 1, enc, &res);
+    VP8SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+        VP8RecordCoeffTokens(ctx, 1,
+                             res.first, res.last, res.coeffs, tokens);
+    VP8RecordCoeffs(ctx, &res);
+    VP8InitResidual(1, 0, enc, &res);
   } else {
-    InitResidual(0, 3, enc, &res);
+    VP8InitResidual(0, 3, enc, &res);
   }
 
   // luma-AC
   for (y = 0; y < 4; ++y) {
     for (x = 0; x < 4; ++x) {
       const int ctx = it->top_nz_[x] + it->left_nz_[y];
-      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
       it->top_nz_[x] = it->left_nz_[y] =
-          RecordCoeffTokens(ctx, &res, &tokens[0]);
+          VP8RecordCoeffTokens(ctx, res.coeff_type,
+                               res.first, res.last, res.coeffs, tokens);
+      VP8RecordCoeffs(ctx, &res);
     }
   }
 
   // U/V
-  InitResidual(0, 2, enc, &res);
+  VP8InitResidual(0, 2, enc, &res);
   for (ch = 0; ch <= 2; ch += 2) {
     for (y = 0; y < 2; ++y) {
       for (x = 0; x < 2; ++x) {
         const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
-        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
         it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
-            RecordCoeffTokens(ctx, &res, &tokens[1]);
+            VP8RecordCoeffTokens(ctx, 2,
+                                 res.first, res.last, res.coeffs, tokens);
+        VP8RecordCoeffs(ctx, &res);
       }
     }
   }
+  VP8IteratorBytesToNz(it);
+  return !tokens->error_;
 }
 
-#endif    // USE_TOKEN_BUFFER
+#endif    // !DISABLE_TOKEN_BUFFER
 
 //------------------------------------------------------------------------------
 // ExtraInfo map / Debug function
@@ -697,7 +460,10 @@ static void SetBlock(uint8_t* p, int value, int size) {
 #endif
 
 static void ResetSSE(VP8Encoder* const enc) {
-  memset(enc->sse_, 0, sizeof(enc->sse_));
+  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;
 }
 
@@ -706,9 +472,9 @@ static void StoreSSE(const VP8EncIterator* const it) {
   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_[0] += VP8SSE16x16(in + Y_OFF_ENC, out + Y_OFF_ENC);
+  enc->sse_[1] += VP8SSE8x8(in + U_OFF_ENC, out + U_OFF_ENC);
+  enc->sse_[2] += VP8SSE8x8(in + V_OFF_ENC, out + V_OFF_ENC);
   enc->sse_count_ += 16 * 16;
 }
 
@@ -736,72 +502,163 @@ static void StoreSideInfo(const VP8EncIterator* const it) {
         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);
+  SetBlock(it->yuv_out_ + Y_OFF_ENC, mb->segment_ * 64, 16);
+  SetBlock(it->yuv_out_ + U_OFF_ENC, it->preds_[0] * 64, 8);
+  SetBlock(it->yuv_out_ + V_OFF_ENC, mb->uv_mode_ * 64, 8);
 #endif
 }
 
-//------------------------------------------------------------------------------
-// Main loops
-//
-//  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
-  }
+static double GetPSNR(uint64_t mse, uint64_t size) {
+  return (mse > 0 && size > 0) ? 10. * log10(255. * 255. * size / mse) : 99;
 }
 
-int VP8EncLoop(VP8Encoder* const enc) {
-  int i, s, p;
-  int ok = 1;
-  VP8EncIterator it;
-  VP8ModeScore info;
-  const int dont_use_skip = !enc->proba_.use_skip_proba_;
-  const int rd_opt = enc->rd_opt_level_;
-  const int kAverageBytesPerMB = 5;     // TODO: have a kTable[quality/10]
-  const int bytes_per_parts =
-    enc->mb_w_ * enc->mb_h_ * kAverageBytesPerMB / enc->num_parts_;
+//------------------------------------------------------------------------------
+//  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.
 
-  // Initialize the bit-writers
-  for (p = 0; p < enc->num_parts_; ++p) {
-    VP8BitWriterInit(enc->parts_ + p, bytes_per_parts);
-  }
+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);
-  VP8InitFilter(&it);
+  SetLoopParams(enc, s->q);
   do {
-    VP8IteratorImport(&it);
-    // 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);
+    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_++;
     }
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-    if (enc->use_layer_) {
-      VP8EncCodeLayerBlock(&it);
+    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
-    StoreSideInfo(&it);
-    VP8StoreFilterStats(&it);
-    VP8IteratorExport(&it);
-    ok = VP8IteratorProgress(&it, 20);
-  } while (ok && VP8IteratorNext(&it, it.yuv_out_));
+    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
+    WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  return ok;
+}
+
+static int PostLoopFinalize(VP8EncIterator* const it, int ok) {
+  VP8Encoder* const enc = it->enc_;
   if (ok) {      // Finalize the partitions, check for extra errors.
+    int p;
     for (p = 0; p < enc->num_parts_; ++p) {
       VP8BitWriterFinish(enc->parts_ + p);
       ok &= !enc->parts_[p].error_;
@@ -809,131 +666,185 @@ int VP8EncLoop(VP8Encoder* const enc) {
   }
 
   if (ok) {      // All good. Finish up.
-    if (enc->pic_->stats) {           // finalize byte counters...
+    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);
+          enc->residual_bytes_[i][s] = (int)((it->bit_count_[s][i] + 7) >> 3);
         }
       }
     }
-    VP8AdjustFilterStrength(&it);     // ...and store filter stats.
+    VP8AdjustFilterStrength(it);     // ...and store filter stats.
   } else {
     // Something bad happened -> need to do some memory cleanup.
     VP8EncFreeBitWriters(enc);
   }
-
   return ok;
 }
 
 //------------------------------------------------------------------------------
-//  VP8StatLoop(): 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, PNSR)
-//                 was specified.
-
-#define kHeaderSizeEstimate (15 + 20 + 10)      // TODO: fix better
-
-static int OneStatPass(VP8Encoder* const enc, float q, int rd_opt, int nb_mbs,
-                       float* const PSNR, int percent_delta) {
-  VP8EncIterator it;
-  uint64_t size = 0;
-  uint64_t distortion = 0;
-  const uint64_t pixel_count = nb_mbs * 384;
+//  VP8EncLoop(): does the final bitstream coding.
 
-  // Make sure the quality parameter is inside valid bounds
-  if (q < 0.) {
-    q = 0;
-  } else if (q > 100.) {
-    q = 100;
+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
   }
+}
 
-  VP8SetSegmentParams(enc, q);      // setup segment quantizations and filters
+int VP8EncLoop(VP8Encoder* const enc) {
+  VP8EncIterator it;
+  int ok = PreLoopInitialize(enc);
+  if (!ok) return 0;
 
-  ResetStats(enc);
-  ResetTokenStats(enc);
+  StatLoop(enc);  // stats-collection loop
 
   VP8IteratorInit(enc, &it);
+  VP8InitFilter(&it);
   do {
     VP8ModeScore info;
-    VP8IteratorImport(&it);
-    if (VP8Decimate(&it, &info, rd_opt)) {
-      // Just record the number of skips and act like skip_proba is not used.
-      enc->proba_.nb_skip_++;
+    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);
     }
-    RecordResiduals(&it, &info);
-    size += info.R;
-    distortion += info.D;
-    if (percent_delta && !VP8IteratorProgress(&it, percent_delta))
-      return 0;
-  } while (VP8IteratorNext(&it, it.yuv_out_) && --nb_mbs > 0);
-  size += FinalizeSkipProba(enc);
-  size += FinalizeTokenProbas(enc);
-  size += enc->segment_hdr_.size_;
-  size = ((size + 1024) >> 11) + kHeaderSizeEstimate;
-
-  if (PSNR) {
-    *PSNR = (float)(10.* log10(255. * 255. * pixel_count / distortion));
-  }
-  return (int)size;
+    StoreSideInfo(&it);
+    VP8StoreFilterStats(&it);
+    VP8IteratorExport(&it);
+    ok = VP8IteratorProgress(&it, 20);
+    VP8IteratorSaveBoundary(&it);
+  } while (ok && VP8IteratorNext(&it));
+
+  return PostLoopFinalize(&it, ok);
 }
 
-// successive refinement increments.
-static const int dqs[] = { 20, 15, 10, 8, 6, 4, 2, 1, 0 };
+//------------------------------------------------------------------------------
+// Single pass using Token Buffer.
 
-int VP8StatLoop(VP8Encoder* const enc) {
-  const int do_search =
-    (enc->config_->target_size > 0 || enc->config_->target_PSNR > 0);
-  const int fast_probe = (enc->method_ < 2 && !do_search);
-  float q = enc->config_->quality;
-  const int max_passes = enc->config_->pass;
-  const int task_percent = 20;
-  const int percent_per_pass = (task_percent + max_passes / 2) / max_passes;
-  const int final_percent = enc->percent_ + task_percent;
-  int pass;
-  int nb_mbs;
+#if !defined(DISABLE_TOKEN_BUFFER)
 
-  // Fast mode: quick analysis pass over few mbs. Better than nothing.
-  nb_mbs = enc->mb_w_ * enc->mb_h_;
-  if (fast_probe && nb_mbs > 100) nb_mbs = 100;
-
-  // No target size: just do several pass without changing 'q'
-  if (!do_search) {
-    for (pass = 0; pass < max_passes; ++pass) {
-      const int rd_opt = (enc->method_ > 2);
-      if (!OneStatPass(enc, q, rd_opt, nb_mbs, NULL, percent_per_pass)) {
-        return 0;
-      }
+#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;
+  VP8EncProba* const proba = &enc->proba_;
+  const VP8RDLevel rd_opt = enc->rd_opt_level_;
+  const uint64_t pixel_count = enc->mb_w_ * enc->mb_h_ * 384;
+  PassStats stats;
+  int ok;
+
+  InitPassStats(enc, &stats);
+  ok = PreLoopInitialize(enc);
+  if (!ok) return 0;
+
+  if (max_count < MIN_COUNT) max_count = MIN_COUNT;
+
+  assert(enc->num_parts_ == 1);
+  assert(enc->use_tokens_);
+  assert(proba->use_skip_proba_ == 0);
+  assert(rd_opt >= RD_OPT_BASIC);   // otherwise, token-buffer won't be useful
+  assert(num_pass_left > 0);
+
+  while (ok && num_pass_left-- > 0) {
+    const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
+                             (num_pass_left == 0) ||
+                             (enc->max_i4_header_bits_ == 0);
+    uint64_t size_p0 = 0;
+    uint64_t distortion = 0;
+    int cnt = max_count;
+    VP8IteratorInit(enc, &it);
+    SetLoopParams(enc, stats.q);
+    if (is_last_pass) {
+      ResetTokenStats(enc);
+      VP8InitFilter(&it);  // don't collect stats until last pass (too costly)
     }
-  } else {
-    // binary search for a size close to target
-    for (pass = 0; pass < max_passes && (dqs[pass] > 0); ++pass) {
-      const int rd_opt = 1;
-      float PSNR;
-      int criterion;
-      const int size = OneStatPass(enc, q, rd_opt, nb_mbs, &PSNR,
-                                   percent_per_pass);
-#if DEBUG_SEARCH
-      printf("#%d size=%d PSNR=%.2f q=%.2f\n", pass, size, PSNR, q);
-#endif
-      if (!size) return 0;
-      if (enc->config_->target_PSNR > 0) {
-        criterion = (PSNR < enc->config_->target_PSNR);
-      } else {
-        criterion = (size < enc->config_->target_size);
+    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;
       }
-      // dichotomize
-      if (criterion) {
-        q += dqs[pass];
-      } else {
-        q -= dqs[pass];
+      VP8Decimate(&it, &info, rd_opt);
+      ok = RecordTokens(&it, &info, &enc->tokens_);
+      if (!ok) {
+        WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+        break;
+      }
+      size_p0 += info.H;
+      distortion += info.D;
+      if (is_last_pass) {
+        StoreSideInfo(&it);
+        VP8StoreFilterStats(&it);
+        VP8IteratorExport(&it);
+        ok = VP8IteratorProgress(&it, 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
     }
   }
-  return WebPReportProgress(enc->pic_, final_percent, &enc->percent_);
+  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
+
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/histogram.c b/drivers/webp/enc/histogram.c
index ca838e064d..62c320d809 100644
--- a/drivers/webp/enc/histogram.c
+++ b/drivers/webp/enc/histogram.c
@@ -1,38 +1,82 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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"
+#include "../webp/config.h"
 #endif
 
 #include <math.h>
-#include <stdio.h>
 
 #include "./backward_references.h"
 #include "./histogram.h"
 #include "../dsp/lossless.h"
 #include "../utils/utils.h"
 
+#define MAX_COST 1.e38
+
+// Number of partitions for the three dominant (literal, red and blue) symbol
+// costs.
+#define NUM_PARTITIONS 4
+// The size of the bin-hash corresponding to the three dominant costs.
+#define BIN_SIZE (NUM_PARTITIONS * NUM_PARTITIONS * NUM_PARTITIONS)
+// Maximum number of histograms allowed in greedy combining algorithm.
+#define MAX_HISTO_GREEDY 100
+
 static void HistogramClear(VP8LHistogram* const p) {
-  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;
+  uint32_t* const literal = p->literal_;
+  const int cache_bits = p->palette_code_bits_;
+  const int histo_size = VP8LGetHistogramSize(cache_bits);
+  memset(p, 0, histo_size);
+  p->palette_code_bits_ = cache_bits;
+  p->literal_ = literal;
+}
+
+// Swap two histogram pointers.
+static void HistogramSwap(VP8LHistogram** const A, VP8LHistogram** const B) {
+  VP8LHistogram* const tmp = *A;
+  *A = *B;
+  *B = tmp;
+}
+
+static void HistogramCopy(const VP8LHistogram* const src,
+                          VP8LHistogram* const dst) {
+  uint32_t* const dst_literal = dst->literal_;
+  const int dst_cache_bits = dst->palette_code_bits_;
+  const int histo_size = VP8LGetHistogramSize(dst_cache_bits);
+  assert(src->palette_code_bits_ == dst_cache_bits);
+  memcpy(dst, src, histo_size);
+  dst->literal_ = dst_literal;
+}
+
+int VP8LGetHistogramSize(int cache_bits) {
+  const int literal_size = VP8LHistogramNumCodes(cache_bits);
+  const size_t total_size = sizeof(VP8LHistogram) + sizeof(int) * literal_size;
+  assert(total_size <= (size_t)0x7fffffff);
+  return (int)total_size;
+}
+
+void VP8LFreeHistogram(VP8LHistogram* const histo) {
+  WebPSafeFree(histo);
+}
+
+void VP8LFreeHistogramSet(VP8LHistogramSet* const histo) {
+  WebPSafeFree(histo);
 }
 
 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
                             VP8LHistogram* const histo) {
-  int i;
-  for (i = 0; i < refs->size; ++i) {
-    VP8LHistogramAddSinglePixOrCopy(histo, &refs->refs[i]);
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  while (VP8LRefsCursorOk(&c)) {
+    VP8LHistogramAddSinglePixOrCopy(histo, c.cur_pos);
+    VP8LRefsCursorNext(&c);
   }
 }
 
@@ -51,13 +95,25 @@ void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) {
   HistogramClear(p);
 }
 
+VP8LHistogram* VP8LAllocateHistogram(int cache_bits) {
+  VP8LHistogram* histo = NULL;
+  const int total_size = VP8LGetHistogramSize(cache_bits);
+  uint8_t* const memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
+  if (memory == NULL) return NULL;
+  histo = (VP8LHistogram*)memory;
+  // literal_ won't necessary be aligned.
+  histo->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
+  VP8LHistogramInit(histo, cache_bits);
+  return histo;
+}
+
 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
   int i;
   VP8LHistogramSet* set;
-  VP8LHistogram* bulk;
-  const uint64_t total_size = (uint64_t)sizeof(*set)
-                            + size * sizeof(*set->histograms)
-                            + size * sizeof(**set->histograms);
+  const int histo_size = VP8LGetHistogramSize(cache_bits);
+  const size_t total_size =
+      sizeof(*set) + size * (sizeof(*set->histograms) +
+      histo_size + WEBP_ALIGN_CST);
   uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
   if (memory == NULL) return NULL;
 
@@ -65,12 +121,15 @@ VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
   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;
+    memory = (uint8_t*)WEBP_ALIGN(memory);
+    set->histograms[i] = (VP8LHistogram*)memory;
+    // literal_ won't necessary be aligned.
+    set->histograms[i]->literal_ = (uint32_t*)(memory + sizeof(VP8LHistogram));
     VP8LHistogramInit(set->histograms[i], cache_bits);
+    memory += histo_size;
   }
   return set;
 }
@@ -85,151 +144,183 @@ void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
     ++histo->literal_[PixOrCopyLiteral(v, 1)];
     ++histo->blue_[PixOrCopyLiteral(v, 0)];
   } else if (PixOrCopyIsCacheIdx(v)) {
-    int literal_ix = 256 + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
+    const int literal_ix =
+        NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
     ++histo->literal_[literal_ix];
   } else {
-    int code, extra_bits_count, extra_bits_value;
-    PrefixEncode(PixOrCopyLength(v),
-                 &code, &extra_bits_count, &extra_bits_value);
-    ++histo->literal_[256 + code];
-    PrefixEncode(PixOrCopyDistance(v),
-                 &code, &extra_bits_count, &extra_bits_value);
+    int code, extra_bits;
+    VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
+    ++histo->literal_[NUM_LITERAL_CODES + code];
+    VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
     ++histo->distance_[code];
   }
 }
 
+// -----------------------------------------------------------------------------
+// Various histogram combine/cost-eval functions
+
+static int GetCombinedHistogramEntropy(const VP8LHistogram* const a,
+                                       const VP8LHistogram* const b,
+                                       double cost_threshold,
+                                       double* cost) {
+  const int palette_code_bits = a->palette_code_bits_;
+  assert(a->palette_code_bits_ == b->palette_code_bits_);
+  *cost += VP8LGetCombinedEntropy(a->literal_, b->literal_,
+                                  VP8LHistogramNumCodes(palette_code_bits));
+  *cost += VP8LExtraCostCombined(a->literal_ + NUM_LITERAL_CODES,
+                                 b->literal_ + NUM_LITERAL_CODES,
+                                 NUM_LENGTH_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  *cost += VP8LGetCombinedEntropy(a->red_, b->red_, NUM_LITERAL_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  *cost += VP8LGetCombinedEntropy(a->blue_, b->blue_, NUM_LITERAL_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  *cost += VP8LGetCombinedEntropy(a->alpha_, b->alpha_, NUM_LITERAL_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  *cost += VP8LGetCombinedEntropy(a->distance_, b->distance_,
+                                  NUM_DISTANCE_CODES);
+  *cost += VP8LExtraCostCombined(a->distance_, b->distance_,
+                                 NUM_DISTANCE_CODES);
+  if (*cost > cost_threshold) return 0;
+
+  return 1;
+}
 
-
-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];
-      }
-    }
+// 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_;
+  cost_threshold += sum_cost;
+
+  if (GetCombinedHistogramEntropy(a, b, cost_threshold, &cost)) {
+    VP8LHistogramAdd(a, b, out);
+    out->bit_cost_ = cost;
+    out->palette_code_bits_ = a->palette_code_bits_;
+    out->trivial_symbol_ = (a->trivial_symbol_ == b->trivial_symbol_) ?
+        a->trivial_symbol_ : VP8L_NON_TRIVIAL_SYM;
   }
-  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;
-  }
+  return cost - sum_cost;
+}
 
-  {
-    double min_limit = 2 * sum - max_val;
-    min_limit = mix * min_limit + (1.0 - mix) * retval;
-    return (retval < min_limit) ? min_limit : retval;
-  }
+// 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) {
+  double cost = -a->bit_cost_;
+  GetCombinedHistogramEntropy(a, b, cost_threshold, &cost);
+  return cost;
 }
 
-double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p) {
-  double retval = BitsEntropy(&p->literal_[0], VP8LHistogramNumCodes(p))
-                + BitsEntropy(&p->red_[0], 256)
-                + BitsEntropy(&p->blue_[0], 256)
-                + BitsEntropy(&p->alpha_[0], 256)
-                + BitsEntropy(&p->distance_[0], NUM_DISTANCE_CODES);
-  // Compute the extra bits cost.
-  int i;
-  for (i = 2; i < NUM_LENGTH_CODES - 2; ++i) {
-    retval +=
-        (i >> 1) * p->literal_[256 + i + 2];
-  }
-  for (i = 2; i < NUM_DISTANCE_CODES - 2; ++i) {
-    retval += (i >> 1) * p->distance_[i + 2];
-  }
-  return 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;
+// -----------------------------------------------------------------------------
+
+// The structure to keep track of cost range for the three dominant entropy
+// symbols.
+// TODO(skal): Evaluate if float can be used here instead of double for
+// representing the entropy costs.
+typedef struct {
+  double literal_max_;
+  double literal_min_;
+  double red_max_;
+  double red_min_;
+  double blue_max_;
+  double blue_min_;
+} DominantCostRange;
+
+static void DominantCostRangeInit(DominantCostRange* const c) {
+  c->literal_max_ = 0.;
+  c->literal_min_ = MAX_COST;
+  c->red_max_ = 0.;
+  c->red_min_ = MAX_COST;
+  c->blue_max_ = 0.;
+  c->blue_min_ = MAX_COST;
+}
+
+static void UpdateDominantCostRange(
+    const VP8LHistogram* const h, DominantCostRange* const c) {
+  if (c->literal_max_ < h->literal_cost_) c->literal_max_ = h->literal_cost_;
+  if (c->literal_min_ > h->literal_cost_) c->literal_min_ = h->literal_cost_;
+  if (c->red_max_ < h->red_cost_) c->red_max_ = h->red_cost_;
+  if (c->red_min_ > h->red_cost_) c->red_min_ = h->red_cost_;
+  if (c->blue_max_ < h->blue_cost_) c->blue_max_ = h->blue_cost_;
+  if (c->blue_min_ > h->blue_cost_) c->blue_min_ = h->blue_cost_;
+}
+
+static void UpdateHistogramCost(VP8LHistogram* const h) {
+  uint32_t alpha_sym, red_sym, blue_sym;
+  const double alpha_cost = VP8LPopulationCost(h->alpha_, NUM_LITERAL_CODES,
+                                               &alpha_sym);
+  const double distance_cost =
+      VP8LPopulationCost(h->distance_, NUM_DISTANCE_CODES, NULL) +
+      VP8LExtraCost(h->distance_, NUM_DISTANCE_CODES);
+  const int num_codes = VP8LHistogramNumCodes(h->palette_code_bits_);
+  h->literal_cost_ = VP8LPopulationCost(h->literal_, num_codes, NULL) +
+                     VP8LExtraCost(h->literal_ + NUM_LITERAL_CODES,
+                                   NUM_LENGTH_CODES);
+  h->red_cost_ = VP8LPopulationCost(h->red_, NUM_LITERAL_CODES, &red_sym);
+  h->blue_cost_ = VP8LPopulationCost(h->blue_, NUM_LITERAL_CODES, &blue_sym);
+  h->bit_cost_ = h->literal_cost_ + h->red_cost_ + h->blue_cost_ +
+                 alpha_cost + distance_cost;
+  if ((alpha_sym | red_sym | blue_sym) == VP8L_NON_TRIVIAL_SYM) {
+    h->trivial_symbol_ = VP8L_NON_TRIVIAL_SYM;
+  } else {
+    h->trivial_symbol_ =
+        ((uint32_t)alpha_sym << 24) | (red_sym << 16) | (blue_sym << 0);
   }
-  return retval;
 }
 
-// Estimates the Huffman dictionary + other block overhead size.
-static double HistogramEstimateBitsHeader(const VP8LHistogram* const p) {
-  return HuffmanCost(&p->alpha_[0], 256) +
-         HuffmanCost(&p->red_[0], 256) +
-         HuffmanCost(&p->literal_[0], VP8LHistogramNumCodes(p)) +
-         HuffmanCost(&p->blue_[0], 256) +
-         HuffmanCost(&p->distance_[0], NUM_DISTANCE_CODES);
+static int GetBinIdForEntropy(double min, double max, double val) {
+  const double range = max - min + 1e-6;
+  const double delta = val - min;
+  return (int)(NUM_PARTITIONS * delta / range);
 }
 
-double VP8LHistogramEstimateBits(const VP8LHistogram* const p) {
-  return HistogramEstimateBitsHeader(p) + VP8LHistogramEstimateBitsBulk(p);
+static int GetHistoBinIndexLowEffort(
+    const VP8LHistogram* const h, const DominantCostRange* const c) {
+  const int bin_id = GetBinIdForEntropy(c->literal_min_, c->literal_max_,
+                                        h->literal_cost_);
+  assert(bin_id < NUM_PARTITIONS);
+  return bin_id;
 }
 
-static void HistogramBuildImage(int xsize, int histo_bits,
-                                const VP8LBackwardRefs* const backward_refs,
-                                VP8LHistogramSet* const image) {
-  int i;
+static int GetHistoBinIndex(
+    const VP8LHistogram* const h, const DominantCostRange* const c) {
+  const int bin_id =
+      GetBinIdForEntropy(c->blue_min_, c->blue_max_, h->blue_cost_) +
+      NUM_PARTITIONS * GetBinIdForEntropy(c->red_min_, c->red_max_,
+                                          h->red_cost_) +
+      NUM_PARTITIONS * NUM_PARTITIONS * GetBinIdForEntropy(c->literal_min_,
+                                                           c->literal_max_,
+                                                           h->literal_cost_);
+  assert(bin_id < BIN_SIZE);
+  return bin_id;
+}
+
+// Construct the histograms from backward references.
+static void HistogramBuild(
+    int xsize, int histo_bits, const VP8LBackwardRefs* const backward_refs,
+    VP8LHistogramSet* const image_histo) {
   int x = 0, y = 0;
   const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
-  VP8LHistogram** const histograms = image->histograms;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  VP8LRefsCursor c = VP8LRefsCursorInit(backward_refs);
   assert(histo_bits > 0);
-  for (i = 0; i < backward_refs->size; ++i) {
-    const PixOrCopy* const v = &backward_refs->refs[i];
+  while (VP8LRefsCursorOk(&c)) {
+    const PixOrCopy* const v = c.cur_pos;
     const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
     VP8LHistogramAddSinglePixOrCopy(histograms[ix], v);
     x += PixOrCopyLength(v);
@@ -237,7 +328,134 @@ static void HistogramBuildImage(int xsize, int histo_bits,
       x -= xsize;
       ++y;
     }
+    VP8LRefsCursorNext(&c);
+  }
+}
+
+// Copies the histograms and computes its bit_cost.
+static void HistogramCopyAndAnalyze(
+    VP8LHistogramSet* const orig_histo, VP8LHistogramSet* const image_histo) {
+  int i;
+  const int histo_size = orig_histo->size;
+  VP8LHistogram** const orig_histograms = orig_histo->histograms;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  for (i = 0; i < histo_size; ++i) {
+    VP8LHistogram* const histo = orig_histograms[i];
+    UpdateHistogramCost(histo);
+    // Copy histograms from orig_histo[] to image_histo[].
+    HistogramCopy(histo, histograms[i]);
+  }
+}
+
+// Partition histograms to different entropy bins for three dominant (literal,
+// red and blue) symbol costs and compute the histogram aggregate bit_cost.
+static void HistogramAnalyzeEntropyBin(VP8LHistogramSet* const image_histo,
+                                       int16_t* const bin_map, int low_effort) {
+  int i;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  const int histo_size = image_histo->size;
+  const int bin_depth = histo_size + 1;
+  DominantCostRange cost_range;
+  DominantCostRangeInit(&cost_range);
+
+  // Analyze the dominant (literal, red and blue) entropy costs.
+  for (i = 0; i < histo_size; ++i) {
+    VP8LHistogram* const histo = histograms[i];
+    UpdateDominantCostRange(histo, &cost_range);
+  }
+
+  // bin-hash histograms on three of the dominant (literal, red and blue)
+  // symbol costs.
+  for (i = 0; i < histo_size; ++i) {
+    int num_histos;
+    VP8LHistogram* const histo = histograms[i];
+    const int16_t bin_id = low_effort ?
+        (int16_t)GetHistoBinIndexLowEffort(histo, &cost_range) :
+        (int16_t)GetHistoBinIndex(histo, &cost_range);
+    const int bin_offset = bin_id * bin_depth;
+    // bin_map[n][0] for every bin 'n' maintains the counter for the number of
+    // histograms in that bin.
+    // Get and increment the num_histos in that bin.
+    num_histos = ++bin_map[bin_offset];
+    assert(bin_offset + num_histos < bin_depth * BIN_SIZE);
+    // Add histogram i'th index at num_histos (last) position in the bin_map.
+    bin_map[bin_offset + num_histos] = i;
+  }
+}
+
+// Compact the histogram set by removing unused entries.
+static void HistogramCompactBins(VP8LHistogramSet* const image_histo) {
+  VP8LHistogram** const histograms = image_histo->histograms;
+  int i, j;
+
+  for (i = 0, j = 0; i < image_histo->size; ++i) {
+    if (histograms[i] != NULL && histograms[i]->bit_cost_ != 0.) {
+      if (j < i) {
+        histograms[j] = histograms[i];
+        histograms[i] = NULL;
+      }
+      ++j;
+    }
   }
+  image_histo->size = j;
+}
+
+static VP8LHistogram* HistogramCombineEntropyBin(
+    VP8LHistogramSet* const image_histo,
+    VP8LHistogram* cur_combo,
+    int16_t* const bin_map, int bin_depth, int num_bins,
+    double combine_cost_factor, int low_effort) {
+  int bin_id;
+  VP8LHistogram** const histograms = image_histo->histograms;
+
+  for (bin_id = 0; bin_id < num_bins; ++bin_id) {
+    const int bin_offset = bin_id * bin_depth;
+    const int num_histos = bin_map[bin_offset];
+    const int idx1 = bin_map[bin_offset + 1];
+    int num_combine_failures = 0;
+    int n;
+    for (n = 2; n <= num_histos; ++n) {
+      const int idx2 = bin_map[bin_offset + n];
+      if (low_effort) {
+        // Merge all histograms with the same bin index, irrespective of cost of
+        // the merged histograms.
+        VP8LHistogramAdd(histograms[idx1], histograms[idx2], histograms[idx1]);
+        histograms[idx2]->bit_cost_ = 0.;
+      } else {
+        const double bit_cost_idx2 = histograms[idx2]->bit_cost_;
+        if (bit_cost_idx2 > 0.) {
+          const double bit_cost_thresh = -bit_cost_idx2 * combine_cost_factor;
+          const double curr_cost_diff =
+              HistogramAddEval(histograms[idx1], histograms[idx2],
+                               cur_combo, bit_cost_thresh);
+          if (curr_cost_diff < bit_cost_thresh) {
+            // Try to merge two histograms only if the combo is a trivial one or
+            // the two candidate histograms are already non-trivial.
+            // For some images, 'try_combine' turns out to be false for a lot of
+            // histogram pairs. In that case, we fallback to combining
+            // histograms as usual to avoid increasing the header size.
+            const int try_combine =
+                (cur_combo->trivial_symbol_ != VP8L_NON_TRIVIAL_SYM) ||
+                ((histograms[idx1]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM) &&
+                 (histograms[idx2]->trivial_symbol_ == VP8L_NON_TRIVIAL_SYM));
+            const int max_combine_failures = 32;
+            if (try_combine || (num_combine_failures >= max_combine_failures)) {
+              HistogramSwap(&cur_combo, &histograms[idx1]);
+              histograms[idx2]->bit_cost_ = 0.;
+            } else {
+              ++num_combine_failures;
+            }
+          }
+        }
+      }
+    }
+    if (low_effort) {
+      // Update the bit_cost for the merged histograms (per bin index).
+      UpdateHistogramCost(histograms[idx1]);
+    }
+  }
+  HistogramCompactBins(image_histo);
+  return cur_combo;
 }
 
 static uint32_t MyRand(uint32_t *seed) {
@@ -248,159 +466,433 @@ static uint32_t MyRand(uint32_t *seed) {
   return *seed;
 }
 
-static int HistogramCombine(const VP8LHistogramSet* const in,
-                            VP8LHistogramSet* const out, int num_pairs) {
+// -----------------------------------------------------------------------------
+// Histogram pairs priority queue
+
+// Pair of histograms. Negative idx1 value means that pair is out-of-date.
+typedef struct {
+  int idx1;
+  int idx2;
+  double cost_diff;
+  double cost_combo;
+} HistogramPair;
+
+typedef struct {
+  HistogramPair* heap;
+  int* positions;
+  int size;
+  int max_index;
+} HistoHeap;
+
+static int HistoHeapInit(HistoHeap* const histo_heap, const int max_index) {
+  histo_heap->size = 0;
+  histo_heap->max_index = max_index;
+  histo_heap->heap = WebPSafeMalloc(max_index * max_index,
+                                    sizeof(*histo_heap->heap));
+  histo_heap->positions = WebPSafeMalloc(max_index * max_index,
+                                         sizeof(*histo_heap->positions));
+  return histo_heap->heap != NULL && histo_heap->positions != NULL;
+}
+
+static void HistoHeapClear(HistoHeap* const histo_heap) {
+  assert(histo_heap != NULL);
+  WebPSafeFree(histo_heap->heap);
+  WebPSafeFree(histo_heap->positions);
+}
+
+static void SwapHistogramPairs(HistogramPair *p1,
+                               HistogramPair *p2) {
+  const HistogramPair tmp = *p1;
+  *p1 = *p2;
+  *p2 = tmp;
+}
+
+// Given a valid min-heap in range [0, heap_size-1) this function places value
+// heap[heap_size-1] into right location within heap and sets its position in
+// positions array.
+static void HeapPush(HistoHeap* const histo_heap) {
+  HistogramPair* const heap = histo_heap->heap - 1;
+  int* const positions = histo_heap->positions;
+  const int max_index = histo_heap->max_index;
+  int v;
+  ++histo_heap->size;
+  v = histo_heap->size;
+  while (v > 1 && heap[v].cost_diff < heap[v >> 1].cost_diff) {
+    SwapHistogramPairs(&heap[v], &heap[v >> 1]);
+    // Change position of moved pair in heap.
+    if (heap[v].idx1 >= 0) {
+      const int pos = heap[v].idx1 * max_index + heap[v].idx2;
+      assert(pos >= 0 && pos < max_index * max_index);
+      positions[pos] = v;
+    }
+    v >>= 1;
+  }
+  positions[heap[v].idx1 * max_index + heap[v].idx2] = v;
+}
+
+// Given a valid min-heap in range [0, heap_size) this function shortens heap
+// range by one and places element with the lowest value to (heap_size-1).
+static void HeapPop(HistoHeap* const histo_heap) {
+  HistogramPair* const heap = histo_heap->heap - 1;
+  int* const positions = histo_heap->positions;
+  const int heap_size = histo_heap->size;
+  const int max_index = histo_heap->max_index;
+  int v = 1;
+  if (heap[v].idx1 >= 0) {
+    positions[heap[v].idx1 * max_index + heap[v].idx2] = -1;
+  }
+  SwapHistogramPairs(&heap[v], &heap[heap_size]);
+  while ((v << 1) < heap_size) {
+    int son = (heap[v << 1].cost_diff < heap[v].cost_diff) ? (v << 1) : v;
+    if (((v << 1) + 1) < heap_size &&
+        heap[(v << 1) + 1].cost_diff < heap[son].cost_diff) {
+      son = (v << 1) + 1;
+    }
+    if (son == v) break;
+    SwapHistogramPairs(&heap[v], &heap[son]);
+    // Change position of moved pair in heap.
+    if (heap[v].idx1 >= 0) {
+      positions[heap[v].idx1 * max_index + heap[v].idx2] = v;
+    }
+    v = son;
+  }
+  if (heap[v].idx1 >= 0) {
+    positions[heap[v].idx1 * max_index + heap[v].idx2] = v;
+  }
+  --histo_heap->size;
+}
+
+// -----------------------------------------------------------------------------
+
+static void PreparePair(VP8LHistogram** histograms, int idx1, int idx2,
+                        HistogramPair* const pair,
+                        VP8LHistogram* const histos) {
+  if (idx1 > idx2) {
+    const int tmp = idx2;
+    idx2 = idx1;
+    idx1 = tmp;
+  }
+  pair->idx1 = idx1;
+  pair->idx2 = idx2;
+  pair->cost_diff =
+      HistogramAddEval(histograms[idx1], histograms[idx2], histos, 0);
+  pair->cost_combo = histos->bit_cost_;
+}
+
+#define POSITION_INVALID (-1)
+
+// Invalidates pairs intersecting (idx1, idx2) in heap.
+static void InvalidatePairs(int idx1, int idx2,
+                            const HistoHeap* const histo_heap) {
+  HistogramPair* const heap = histo_heap->heap - 1;
+  int* const positions = histo_heap->positions;
+  const int max_index = histo_heap->max_index;
+  int i;
+  for (i = 0; i < idx1; ++i) {
+    const int pos = positions[i * max_index + idx1];
+    if (pos >= 0) {
+      heap[pos].idx1 = POSITION_INVALID;
+    }
+  }
+  for (i = idx1 + 1; i < max_index; ++i) {
+    const int pos = positions[idx1 * max_index + i];
+    if (pos >= 0) {
+      heap[pos].idx1 = POSITION_INVALID;
+    }
+  }
+  for (i = 0; i < idx2; ++i) {
+    const int pos = positions[i * max_index + idx2];
+    if (pos >= 0) {
+      heap[pos].idx1 = POSITION_INVALID;
+    }
+  }
+  for (i = idx2 + 1; i < max_index; ++i) {
+    const int pos = positions[idx2 * max_index + i];
+    if (pos >= 0) {
+      heap[pos].idx1 = POSITION_INVALID;
+    }
+  }
+}
+
+// Combines histograms by continuously choosing the one with the highest cost
+// reduction.
+static int HistogramCombineGreedy(VP8LHistogramSet* const image_histo,
+                                  VP8LHistogram* const histos) {
   int ok = 0;
-  int i, iter;
+  int image_histo_size = image_histo->size;
+  int i, j;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  // Indexes of remaining histograms.
+  int* const clusters = WebPSafeMalloc(image_histo_size, sizeof(*clusters));
+  // Heap of histogram pairs.
+  HistoHeap histo_heap;
+
+  if (!HistoHeapInit(&histo_heap, image_histo_size) || clusters == NULL) {
+    goto End;
+  }
+
+  for (i = 0; i < image_histo_size; ++i) {
+    // Initialize clusters indexes.
+    clusters[i] = i;
+    for (j = i + 1; j < image_histo_size; ++j) {
+      // Initialize positions array.
+      histo_heap.positions[i * histo_heap.max_index + j] = POSITION_INVALID;
+      PreparePair(histograms, i, j, &histo_heap.heap[histo_heap.size], histos);
+      if (histo_heap.heap[histo_heap.size].cost_diff < 0) {
+        HeapPush(&histo_heap);
+      }
+    }
+  }
+
+  while (image_histo_size > 1 && histo_heap.size > 0) {
+    const int idx1 = histo_heap.heap[0].idx1;
+    const int idx2 = histo_heap.heap[0].idx2;
+    VP8LHistogramAdd(histograms[idx2], histograms[idx1], histograms[idx1]);
+    histograms[idx1]->bit_cost_ = histo_heap.heap[0].cost_combo;
+    // Remove merged histogram.
+    for (i = 0; i + 1 < image_histo_size; ++i) {
+      if (clusters[i] >= idx2) {
+        clusters[i] = clusters[i + 1];
+      }
+    }
+    --image_histo_size;
+
+    // Invalidate pairs intersecting the just combined best pair.
+    InvalidatePairs(idx1, idx2, &histo_heap);
+
+    // Pop invalid pairs from the top of the heap.
+    while (histo_heap.size > 0 && histo_heap.heap[0].idx1 < 0) {
+      HeapPop(&histo_heap);
+    }
+
+    // Push new pairs formed with combined histogram to the heap.
+    for (i = 0; i < image_histo_size; ++i) {
+      if (clusters[i] != idx1) {
+        PreparePair(histograms, idx1, clusters[i],
+                    &histo_heap.heap[histo_heap.size], histos);
+        if (histo_heap.heap[histo_heap.size].cost_diff < 0) {
+          HeapPush(&histo_heap);
+        }
+      }
+    }
+  }
+  // Move remaining histograms to the beginning of the array.
+  for (i = 0; i < image_histo_size; ++i) {
+    if (i != clusters[i]) {  // swap the two histograms
+      HistogramSwap(&histograms[i], &histograms[clusters[i]]);
+    }
+  }
+
+  image_histo->size = image_histo_size;
+  ok = 1;
+
+ End:
+  WebPSafeFree(clusters);
+  HistoHeapClear(&histo_heap);
+  return ok;
+}
+
+static VP8LHistogram* HistogramCombineStochastic(
+    VP8LHistogramSet* const image_histo,
+    VP8LHistogram* tmp_histo,
+    VP8LHistogram* best_combo,
+    int quality, int min_cluster_size) {
+  int iter;
   uint32_t seed = 0;
   int tries_with_no_success = 0;
-  const int min_cluster_size = 2;
-  int out_size = in->size;
-  const int outer_iters = in->size * 3;
-  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) {
-    // We pick the best pair to be combined out of 'inner_iters' pairs.
+  int image_histo_size = image_histo->size;
+  const int iter_mult = (quality < 25) ? 2 : 2 + (quality - 25) / 8;
+  const int outer_iters = image_histo_size * iter_mult;
+  const int num_pairs = image_histo_size / 2;
+  const int num_tries_no_success = outer_iters / 2;
+  VP8LHistogram** const histograms = image_histo->histograms;
+
+  // Collapse similar histograms in 'image_histo'.
+  ++min_cluster_size;
+  for (iter = 0;
+       iter < outer_iters && image_histo_size >= min_cluster_size;
+       ++iter) {
     double best_cost_diff = 0.;
-    int best_idx1 = 0, best_idx2 = 1;
+    int best_idx1 = -1, best_idx2 = 1;
     int j;
+    const int num_tries =
+        (num_pairs < image_histo_size) ? num_pairs : image_histo_size;
     seed += iter;
-    for (j = 0; j < num_pairs; ++j) {
+    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) % (out_size - 1);
-      const uint32_t diff = (tmp < 3) ? tmp : MyRand(&seed) % (out_size - 1);
-      const uint32_t idx2 = (idx1 + diff + 1) % out_size;
+      const uint32_t idx1 = MyRand(&seed) % image_histo_size;
+      const uint32_t tmp = (j & 7) + 1;
+      const uint32_t diff =
+          (tmp < 3) ? tmp : MyRand(&seed) % (image_histo_size - 1);
+      const uint32_t idx2 = (idx1 + diff + 1) % image_histo_size;
       if (idx1 == idx2) {
         continue;
       }
-      *cur_combo = *out->histograms[idx1];
-      VP8LHistogramAdd(cur_combo, out->histograms[idx2]);
-      cur_combo->bit_cost_ = VP8LHistogramEstimateBits(cur_combo);
+
       // Calculate cost reduction on combining.
-      curr_cost_diff = cur_combo->bit_cost_
-                     - out->histograms[idx1]->bit_cost_
-                     - out->histograms[idx2]->bit_cost_;
-      if (best_cost_diff > curr_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;
-        }
+      curr_cost_diff = HistogramAddEval(histograms[idx1], histograms[idx2],
+                                        tmp_histo, best_cost_diff);
+      if (curr_cost_diff < best_cost_diff) {    // found a better pair?
+        HistogramSwap(&best_combo, &tmp_histo);
         best_cost_diff = curr_cost_diff;
         best_idx1 = idx1;
         best_idx2 = idx2;
       }
     }
 
-    if (best_cost_diff < 0.0) {
-      *out->histograms[best_idx1] = *best_combo;
+    if (best_idx1 >= 0) {
+      HistogramSwap(&best_combo, &histograms[best_idx1]);
       // 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.
+      --image_histo_size;
+      if (best_idx2 != image_histo_size) {
+        HistogramSwap(&histograms[image_histo_size], &histograms[best_idx2]);
+        histograms[image_histo_size] = NULL;
       }
       tries_with_no_success = 0;
     }
-    if (++tries_with_no_success >= 50) {
+    if (++tries_with_no_success >= num_tries_no_success) {
       break;
     }
   }
-  out->size = out_size;
-  ok = 1;
-
- End:
-  free(histos);
-  return ok;
+  image_histo->size = image_histo_size;
+  return best_combo;
 }
 
 // -----------------------------------------------------------------------------
 // Histogram refinement
 
-// What is the bit cost of moving square_histogram from
-// cur_symbol to candidate_symbol.
-// TODO(skal): we don't really need to copy the histogram and Add(). Instead
-// we just need VP8LDualHistogramEstimateBits(A, B) estimation function.
-static double HistogramDistance(const VP8LHistogram* const square_histogram,
-                                const VP8LHistogram* const candidate) {
-  const double previous_bit_cost = candidate->bit_cost_;
-  double new_bit_cost;
-  VP8LHistogram modified_histo;
-  modified_histo = *candidate;
-  VP8LHistogramAdd(&modified_histo, square_histogram);
-  new_bit_cost = VP8LHistogramEstimateBits(&modified_histo);
-
-  return new_bit_cost - previous_bit_cost;
-}
-
 // 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,
+static void HistogramRemap(const VP8LHistogramSet* const orig_histo,
+                           const VP8LHistogramSet* const image_histo,
                            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]);
-    int k;
-    for (k = 1; k < out->size; ++k) {
-      const double cur_bits =
-          HistogramDistance(in->histograms[i], out->histograms[k]);
-      if (cur_bits < best_bits) {
-        best_bits = cur_bits;
-        best_out = k;
+  VP8LHistogram** const orig_histograms = orig_histo->histograms;
+  VP8LHistogram** const histograms = image_histo->histograms;
+  const int orig_histo_size = orig_histo->size;
+  const int image_histo_size = image_histo->size;
+  if (image_histo_size > 1) {
+    for (i = 0; i < orig_histo_size; ++i) {
+      int best_out = 0;
+      double best_bits =
+          HistogramAddThresh(histograms[0], orig_histograms[i], MAX_COST);
+      int k;
+      for (k = 1; k < image_histo_size; ++k) {
+        const double cur_bits =
+            HistogramAddThresh(histograms[k], orig_histograms[i], best_bits);
+        if (cur_bits < best_bits) {
+          best_bits = cur_bits;
+          best_out = k;
+        }
       }
+      symbols[i] = best_out;
+    }
+  } else {
+    assert(image_histo_size == 1);
+    for (i = 0; i < orig_histo_size; ++i) {
+      symbols[i] = 0;
     }
-    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 < image_histo_size; ++i) {
+    HistogramClear(histograms[i]);
   }
-  for (i = 0; i < in->size; ++i) {
-    VP8LHistogramAdd(out->histograms[symbols[i]], in->histograms[i]);
+
+  for (i = 0; i < orig_histo_size; ++i) {
+    const int idx = symbols[i];
+    VP8LHistogramAdd(orig_histograms[i], histograms[idx], histograms[idx]);
   }
 }
 
+static double GetCombineCostFactor(int histo_size, int quality) {
+  double combine_cost_factor = 0.16;
+  if (quality < 90) {
+    if (histo_size > 256) combine_cost_factor /= 2.;
+    if (histo_size > 512) combine_cost_factor /= 2.;
+    if (histo_size > 1024) combine_cost_factor /= 2.;
+    if (quality <= 50) combine_cost_factor /= 2.;
+  }
+  return combine_cost_factor;
+}
+
 int VP8LGetHistoImageSymbols(int xsize, int ysize,
                              const VP8LBackwardRefs* const refs,
-                             int quality, int histo_bits, int cache_bits,
-                             VP8LHistogramSet* const image_in,
+                             int quality, int low_effort,
+                             int histo_bits, int cache_bits,
+                             VP8LHistogramSet* const image_histo,
+                             VP8LHistogramSet* const tmp_histos,
                              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 num_histo_pairs = 10 + quality / 2;  // For HistogramCombine().
-  const int histo_image_raw_size = histo_xsize * histo_ysize;
-  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, num_histo_pairs)) {
-    goto Error;
+  const int image_histo_raw_size = histo_xsize * histo_ysize;
+  const int entropy_combine_num_bins = low_effort ? NUM_PARTITIONS : BIN_SIZE;
+
+  // The bin_map for every bin follows following semantics:
+  // bin_map[n][0] = num_histo; // The number of histograms in that bin.
+  // bin_map[n][1] = index of first histogram in that bin;
+  // bin_map[n][num_histo] = index of last histogram in that bin;
+  // bin_map[n][num_histo + 1] ... bin_map[n][bin_depth - 1] = unused indices.
+  const int bin_depth = image_histo_raw_size + 1;
+  int16_t* bin_map = NULL;
+  VP8LHistogramSet* const orig_histo =
+      VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
+  VP8LHistogram* cur_combo;
+  const int entropy_combine =
+      (orig_histo->size > entropy_combine_num_bins * 2) && (quality < 100);
+
+  if (orig_histo == NULL) goto Error;
+
+  // Don't attempt linear bin-partition heuristic for:
+  // histograms of small sizes, as bin_map will be very sparse and;
+  // Maximum quality (q==100), to preserve the compression gains at that level.
+  if (entropy_combine) {
+    const int bin_map_size = bin_depth * entropy_combine_num_bins;
+    bin_map = (int16_t*)WebPSafeCalloc(bin_map_size, sizeof(*bin_map));
+    if (bin_map == NULL) goto Error;
   }
+
+  // Construct the histograms from backward references.
+  HistogramBuild(xsize, histo_bits, refs, orig_histo);
+  // Copies the histograms and computes its bit_cost.
+  HistogramCopyAndAnalyze(orig_histo, image_histo);
+
+  cur_combo = tmp_histos->histograms[1];  // pick up working slot
+  if (entropy_combine) {
+    const double combine_cost_factor =
+        GetCombineCostFactor(image_histo_raw_size, quality);
+    HistogramAnalyzeEntropyBin(orig_histo, bin_map, low_effort);
+    // Collapse histograms with similar entropy.
+    cur_combo = HistogramCombineEntropyBin(image_histo, cur_combo, bin_map,
+                                           bin_depth, entropy_combine_num_bins,
+                                           combine_cost_factor, low_effort);
+  }
+
+  // Don't combine the histograms using stochastic and greedy heuristics for
+  // low-effort compression mode.
+  if (!low_effort || !entropy_combine) {
+    const float x = quality / 100.f;
+    // cubic ramp between 1 and MAX_HISTO_GREEDY:
+    const int threshold_size = (int)(1 + (x * x * x) * (MAX_HISTO_GREEDY - 1));
+    cur_combo = HistogramCombineStochastic(image_histo,
+                                           tmp_histos->histograms[0],
+                                           cur_combo, quality, threshold_size);
+    if ((image_histo->size <= threshold_size) &&
+        !HistogramCombineGreedy(image_histo, cur_combo)) {
+      goto Error;
+    }
+  }
+
+  // TODO(vikasa): Optimize HistogramRemap for low-effort compression mode also.
   // Find the optimal map from original histograms to the final ones.
-  HistogramRemap(image_out, image_in, histogram_symbols);
+  HistogramRemap(orig_histo, image_histo, histogram_symbols);
+
   ok = 1;
 
-Error:
-  free(image_out);
+ Error:
+  WebPSafeFree(bin_map);
+  VP8LFreeHistogramSet(orig_histo);
   return ok;
 }
diff --git a/drivers/webp/enc/histogram.h b/drivers/webp/enc/histogram.h
index 5b5de25539..adb16c01ca 100644
--- a/drivers/webp/enc/histogram.h
+++ b/drivers/webp/enc/histogram.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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)
@@ -12,17 +14,13 @@
 #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 "../format_constants.h"
-#include "../types.h"
+#include "../webp/format_constants.h"
+#include "../webp/types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -30,18 +28,23 @@ extern "C" {
 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];
+  uint32_t* literal_;         // Pointer to the allocated buffer for literal.
+  uint32_t red_[NUM_LITERAL_CODES];
+  uint32_t blue_[NUM_LITERAL_CODES];
+  uint32_t alpha_[NUM_LITERAL_CODES];
   // Backward reference prefix-code histogram.
-  int distance_[NUM_DISTANCE_CODES];
+  uint32_t distance_[NUM_DISTANCE_CODES];
   int palette_code_bits_;
-  double bit_cost_;   // cached value of VP8LHistogramEstimateBits(this)
+  uint32_t trivial_symbol_;  // True, if histograms for Red, Blue & Alpha
+                             // literal symbols are single valued.
+  double bit_cost_;          // cached value of bit cost.
+  double literal_cost_;      // Cached values of dominant entropy costs:
+  double red_cost_;          // literal, red & blue.
+  double blue_cost_;
 } VP8LHistogram;
 
 // Collection of histograms with fixed capacity, allocated as one
-// big memory chunk. Can be destroyed by simply calling 'free()'.
+// big memory chunk. Can be destroyed by calling WebPSafeFree().
 typedef struct {
   int size;         // number of slots currently in use
   int max_size;     // maximum capacity
@@ -57,6 +60,9 @@ void VP8LHistogramCreate(VP8LHistogram* const p,
                          const VP8LBackwardRefs* const refs,
                          int palette_code_bits);
 
+// Return the size of the histogram for a given palette_code_bits.
+int VP8LGetHistogramSize(int palette_code_bits);
+
 // Set the palette_code_bits and reset the stats.
 void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits);
 
@@ -64,51 +70,40 @@ void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits);
 void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
                             VP8LHistogram* const histo);
 
+// Free the memory allocated for the histogram.
+void VP8LFreeHistogram(VP8LHistogram* const histo);
+
+// Free the memory allocated for the histogram set.
+void VP8LFreeHistogramSet(VP8LHistogramSet* const histo);
+
 // Allocate an array of pointer to histograms, allocated and initialized
 // using 'cache_bits'. Return NULL in case of memory error.
 VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits);
 
+// Allocate and initialize histogram object with specified 'cache_bits'.
+// Returns NULL in case of memory error.
+// Special case of VP8LAllocateHistogramSet, with size equals 1.
+VP8LHistogram* VP8LAllocateHistogram(int cache_bits);
+
 // Accumulate a token 'v' into a histogram.
 void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
                                      const PixOrCopy* const v);
 
-// 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 void VP8LHistogramAdd(VP8LHistogram* const p,
-                                         const VP8LHistogram* const a) {
-  int i;
-  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
-    p->literal_[i] += a->literal_[i];
-  }
-  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
-    p->distance_[i] += a->distance_[i];
-  }
-  for (i = 0; i < 256; ++i) {
-    p->red_[i] += a->red_[i];
-    p->blue_[i] += a->blue_[i];
-    p->alpha_[i] += a->alpha_[i];
-  }
-}
-
-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);
+static WEBP_INLINE int VP8LHistogramNumCodes(int palette_code_bits) {
+  return NUM_LITERAL_CODES + NUM_LENGTH_CODES +
+      ((palette_code_bits > 0) ? (1 << palette_code_bits) : 0);
 }
 
 // Builds the histogram image.
 int VP8LGetHistoImageSymbols(int xsize, int ysize,
                              const VP8LBackwardRefs* const refs,
-                             int quality, int histogram_bits, int cache_bits,
+                             int quality, int low_effort,
+                             int histogram_bits, int cache_bits,
                              VP8LHistogramSet* const image_in,
+                             VP8LHistogramSet* const tmp_histos,
                              uint16_t* const histogram_symbols);
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }
 #endif
 
diff --git a/drivers/webp/enc/iterator.c b/drivers/webp/enc/iterator.c
index 86e473bcf0..99d960a547 100644
--- a/drivers/webp/enc/iterator.c
+++ b/drivers/webp/enc/iterator.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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
@@ -13,21 +15,16 @@
 
 #include "./vp8enci.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
 // VP8Iterator
 //------------------------------------------------------------------------------
 
 static void InitLeft(VP8EncIterator* const it) {
-  const VP8Encoder* const enc = it->enc_;
-  enc->y_left_[-1] = enc->u_left_[-1] = enc->v_left_[-1] =
+  it->y_left_[-1] = it->u_left_[-1] = it->v_left_[-1] =
       (it->y_ > 0) ? 129 : 127;
-  memset(enc->y_left_, 129, 16);
-  memset(enc->u_left_, 129, 8);
-  memset(enc->v_left_, 129, 8);
+  memset(it->y_left_, 129, 16);
+  memset(it->u_left_, 129, 8);
+  memset(it->v_left_, 129, 8);
   it->left_nz_[8] = 0;
 }
 
@@ -38,43 +35,60 @@ static void InitTop(VP8EncIterator* const it) {
   memset(enc->nz_, 0, enc->mb_w_ * sizeof(*enc->nz_));
 }
 
-void VP8IteratorReset(VP8EncIterator* const it) {
+void VP8IteratorSetRow(VP8EncIterator* const it, int y) {
   VP8Encoder* const enc = it->enc_;
   it->x_ = 0;
-  it->y_ = 0;
-  it->y_offset_ = 0;
-  it->uv_offset_ = 0;
-  it->mb_ = enc->mb_info_;
-  it->preds_ = enc->preds_;
+  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->bw_ = &enc->parts_[0];
-  it->done_ = enc->mb_w_* enc->mb_h_;
+  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;
-  // TODO(later): for multithreading, these should be owned by 'it'.
-  it->yuv_in_   = enc->yuv_in_;
-  it->yuv_out_  = enc->yuv_out_;
-  it->yuv_out2_ = enc->yuv_out2_;
-  it->yuv_p_    = enc->yuv_p_;
+  it->yuv_in_   = (uint8_t*)WEBP_ALIGN(it->yuv_mem_);
+  it->yuv_out_  = it->yuv_in_ + YUV_SIZE_ENC;
+  it->yuv_out2_ = it->yuv_out_ + YUV_SIZE_ENC;
+  it->yuv_p_    = it->yuv_out2_ + YUV_SIZE_ENC;
   it->lf_stats_ = enc->lf_stats_;
   it->percent0_ = enc->percent_;
+  it->y_left_ = (uint8_t*)WEBP_ALIGN(it->yuv_left_mem_ + 1);
+  it->u_left_ = it->y_left_ + 16 + 16;
+  it->v_left_ = it->u_left_ + 16;
   VP8IteratorReset(it);
 }
 
 int VP8IteratorProgress(const VP8EncIterator* const it, int delta) {
   VP8Encoder* const enc = it->enc_;
-  if (delta && enc->pic_->progress_hook) {
-    const int percent = (enc->mb_h_ <= 1)
+  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 * it->y_ / (enc->mb_h_ - 1);
+                      : it->percent0_ + delta * done / it->count_down0_;
     return WebPReportProgress(enc->pic_, percent, &enc->percent_);
   }
   return 1;
@@ -84,6 +98,8 @@ int VP8IteratorProgress(const VP8EncIterator* const it, int delta) {
 // 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;
@@ -101,30 +117,55 @@ static void ImportBlock(const uint8_t* src, int src_stride,
   }
 }
 
-void VP8IteratorImport(const VP8EncIterator* const it) {
+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 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;
-  uint8_t* const ydst = it->yuv_in_ + Y_OFF;
-  uint8_t* const udst = it->yuv_in_ + U_OFF;
-  uint8_t* const vdst = it->yuv_in_ + V_OFF;
-  int w = (pic->width - x * 16);
-  int h = (pic->height - y * 16);
-
-  if (w > 16) w = 16;
-  if (h > 16) h = 16;
-
-  // Luma plane
-  ImportBlock(ysrc, pic->y_stride, ydst, w, h, 16);
-
-  {   // U/V planes
-    const int uv_w = (w + 1) >> 1;
-    const int uv_h = (h + 1) >> 1;
-    ImportBlock(usrc, pic->uv_stride, udst, uv_w, uv_h, 8);
-    ImportBlock(vsrc, pic->uv_stride, vdst, uv_w, uv_h, 8);
+  const int w = MinSize(pic->width - x * 16, 16);
+  const int h = MinSize(pic->height - y * 16, 16);
+  const int uv_w = (w + 1) >> 1;
+  const int uv_h = (h + 1) >> 1;
+
+  ImportBlock(ysrc, pic->y_stride,  it->yuv_in_ + Y_OFF_ENC, w, h, 16);
+  ImportBlock(usrc, pic->uv_stride, it->yuv_in_ + U_OFF_ENC, uv_w, uv_h, 8);
+  ImportBlock(vsrc, pic->uv_stride, it->yuv_in_ + V_OFF_ENC, uv_w, uv_h, 8);
+
+  if (tmp_32 == NULL) return;
+
+  // Import source (uncompressed) samples into boundary.
+  if (x == 0) {
+    InitLeft(it);
+  } else {
+    if (y == 0) {
+      it->y_left_[-1] = it->u_left_[-1] = it->v_left_[-1] = 127;
+    } else {
+      it->y_left_[-1] = ysrc[- 1 - pic->y_stride];
+      it->u_left_[-1] = usrc[- 1 - pic->uv_stride];
+      it->v_left_[-1] = vsrc[- 1 - pic->uv_stride];
+    }
+    ImportLine(ysrc - 1, pic->y_stride,  it->y_left_, h,   16);
+    ImportLine(usrc - 1, pic->uv_stride, it->u_left_, uv_h, 8);
+    ImportLine(vsrc - 1, pic->uv_stride, it->v_left_, uv_h, 8);
+  }
+
+  it->y_top_  = tmp_32 + 0;
+  it->uv_top_ = tmp_32 + 16;
+  if (y == 0) {
+    memset(tmp_32, 127, 32 * sizeof(*tmp_32));
+  } else {
+    ImportLine(ysrc - pic->y_stride,  1, tmp_32,          w,   16);
+    ImportLine(usrc - pic->uv_stride, 1, tmp_32 + 16,     uv_w, 8);
+    ImportLine(vsrc - pic->uv_stride, 1, tmp_32 + 16 + 8, uv_w, 8);
   }
 }
 
@@ -144,9 +185,9 @@ 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 uint8_t* const ysrc = it->yuv_out_ + Y_OFF_ENC;
+    const uint8_t* const usrc = it->yuv_out_ + U_OFF_ENC;
+    const uint8_t* const vsrc = it->yuv_out_ + V_OFF_ENC;
     const WebPPicture* const pic = enc->pic_;
     uint8_t* const ydst = pic->y + (y * pic->y_stride + x) * 16;
     uint8_t* const udst = pic->u + (y * pic->uv_stride + x) * 8;
@@ -240,48 +281,44 @@ void VP8IteratorBytesToNz(VP8EncIterator* const it) {
 #undef BIT
 
 //------------------------------------------------------------------------------
-// Advance to the next position, doing the bookeeping.
+// Advance to the next position, doing the bookkeeping.
 
-int VP8IteratorNext(VP8EncIterator* const it,
-                    const uint8_t* const block_to_save) {
+void VP8IteratorSaveBoundary(VP8EncIterator* const it) {
   VP8Encoder* const enc = it->enc_;
-  if (block_to_save) {
-    const int x = it->x_, y = it->y_;
-    const uint8_t* const ysrc = block_to_save + Y_OFF;
-    const uint8_t* const usrc = block_to_save + U_OFF;
-    if (x < enc->mb_w_ - 1) {   // left
-      int i;
-      for (i = 0; i < 16; ++i) {
-        enc->y_left_[i] = ysrc[15 + i * BPS];
-      }
-      for (i = 0; i < 8; ++i) {
-        enc->u_left_[i] = usrc[7 + i * BPS];
-        enc->v_left_[i] = usrc[15 + i * BPS];
-      }
-      // top-left (before 'top'!)
-      enc->y_left_[-1] = enc->y_top_[x * 16 + 15];
-      enc->u_left_[-1] = enc->uv_top_[x * 16 + 0 + 7];
-      enc->v_left_[-1] = enc->uv_top_[x * 16 + 8 + 7];
+  const int x = it->x_, y = it->y_;
+  const uint8_t* const ysrc = it->yuv_out_ + Y_OFF_ENC;
+  const uint8_t* const uvsrc = it->yuv_out_ + U_OFF_ENC;
+  if (x < enc->mb_w_ - 1) {   // left
+    int i;
+    for (i = 0; i < 16; ++i) {
+      it->y_left_[i] = ysrc[15 + i * BPS];
     }
-    if (y < enc->mb_h_ - 1) {  // top
-      memcpy(enc->y_top_ + x * 16, ysrc + 15 * BPS, 16);
-      memcpy(enc->uv_top_ + x * 16, usrc + 7 * BPS, 8 + 8);
+    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);
+  }
+}
 
-  it->mb_++;
+int VP8IteratorNext(VP8EncIterator* const it) {
   it->preds_ += 4;
-  it->nz_++;
-  it->x_++;
-  if (it->x_ == enc->mb_w_) {
-    it->x_ = 0;
-    it->y_++;
-    it->bw_ = &enc->parts_[it->y_ & (enc->num_parts_ - 1)];
-    it->preds_ = enc->preds_ + it->y_ * 4 * enc->preds_w_;
-    it->nz_ = enc->nz_;
-    InitLeft(it);
+  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->done_);
+  return (0 < --it->count_down_);
 }
 
 //------------------------------------------------------------------------------
@@ -368,15 +405,15 @@ void VP8IteratorStartI4(VP8EncIterator* const it) {
 
   // Import the boundary samples
   for (i = 0; i < 17; ++i) {    // left
-    it->i4_boundary_[i] = enc->y_left_[15 - i];
+    it->i4_boundary_[i] = it->y_left_[15 - i];
   }
   for (i = 0; i < 16; ++i) {    // top
-    it->i4_boundary_[17 + i] = enc->y_top_[it->x_ * 16 + i];
+    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] = enc->y_top_[it->x_ * 16 + 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) {
@@ -417,6 +454,3 @@ int VP8IteratorRotateI4(VP8EncIterator* const it,
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/layer.c b/drivers/webp/enc/layer.c
deleted file mode 100644
index 423127df63..0000000000
--- a/drivers/webp/enc/layer.c
+++ /dev/null
@@ -1,49 +0,0 @@
-// Copyright 2011 Google Inc. All Rights Reserved.
-//
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
-// -----------------------------------------------------------------------------
-//
-// Enhancement layer (for YUV444/422)
-//
-// Author: Skal (pascal.massimino@gmail.com)
-
-#include <stdlib.h>
-
-#include "./vp8enci.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-//------------------------------------------------------------------------------
-
-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_);
-}
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/picture.c b/drivers/webp/enc/picture.c
index 44eed06083..26679a72e4 100644
--- a/drivers/webp/enc/picture.c
+++ b/drivers/webp/enc/picture.c
@@ -1,470 +1,179 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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, ...
+// WebPPicture class basis
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include <assert.h>
 #include <stdlib.h>
-#include <math.h>
 
 #include "./vp8enci.h"
-#include "../utils/rescaler.h"
-#include "../utils/utils.h"
 #include "../dsp/dsp.h"
-#include "../dsp/yuv.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-#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)
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 // 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;
-      }
-    } 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;
+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;
+}
 
-      // TODO(skal): align plane to cache line?
-      picture->memory_argb_ = memory;
-      picture->argb = (uint32_t*)memory;
-      picture->argb_stride = width;
-    }
+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;
 }
 
-// Remove reference to the ARGB buffer (doesn't free anything).
-static void PictureResetARGB(WebPPicture* const picture) {
+//------------------------------------------------------------------------------
+
+static void WebPPictureResetBufferARGB(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) {
+static void WebPPictureResetBufferYUVA(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);
+void WebPPictureResetBuffers(WebPPicture* const picture) {
+  WebPPictureResetBufferARGB(picture);
+  WebPPictureResetBufferYUVA(picture);
 }
 
-// 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;
-}
+int WebPPictureAllocARGB(WebPPicture* const picture, int width, int height) {
+  void* memory;
+  const uint64_t argb_size = (uint64_t)width * height;
 
-// 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);
-  }
-}
+  assert(picture != NULL);
 
-//------------------------------------------------------------------------------
-// Picture copying
+  WebPSafeFree(picture->memory_argb_);
+  WebPPictureResetBufferARGB(picture);
 
-// 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;
+  if (width <= 0 || height <= 0) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_BAD_DIMENSION);
   }
-}
-
-// 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;
-    }
+  // allocate a new buffer.
+  memory = WebPSafeMalloc(argb_size, sizeof(*picture->argb));
+  if (memory == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
   }
-}
-
-// 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;
+  // TODO(skal): align plane to cache line?
+  picture->memory_argb_ = memory;
+  picture->argb = (uint32_t*)memory;
+  picture->argb_stride = width;
   return 1;
 }
 
-int WebPPictureCopy(const WebPPicture* src, WebPPicture* dst) {
-  if (src == NULL || dst == NULL) return 0;
-  if (src == dst) return 1;
+int WebPPictureAllocYUVA(WebPPicture* const picture, int width, int height) {
+  const WebPEncCSP uv_csp = picture->colorspace & WEBP_CSP_UV_MASK;
+  const int has_alpha = picture->colorspace & WEBP_CSP_ALPHA_BIT;
+  const int y_stride = width;
+  const int uv_width = (width + 1) >> 1;
+  const int uv_height = (height + 1) >> 1;
+  const int uv_stride = uv_width;
+  int a_width, a_stride;
+  uint64_t y_size, uv_size, a_size, total_size;
+  uint8_t* mem;
 
-  WebPPictureGrabSpecs(src, dst);
-  if (!WebPPictureAlloc(dst)) return 0;
+  assert(picture != NULL);
 
-  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;
-}
+  WebPSafeFree(picture->memory_);
+  WebPPictureResetBufferYUVA(picture);
 
-int WebPPictureIsView(const WebPPicture* picture) {
-  if (picture == NULL) return 0;
-  if (picture->use_argb) {
-    return (picture->memory_argb_ == NULL);
+  if (uv_csp != WEBP_YUV420) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_INVALID_CONFIGURATION);
   }
-  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;
+  // 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;
 
-  // verify rectangle position.
-  if (!AdjustAndCheckRectangle(src, &left, &top, width, height)) return 0;
+  total_size = y_size + a_size + 2 * uv_size;
 
-  if (src != dst) {  // beware of aliasing! We don't want to leak 'memory_'.
-    WebPPictureGrabSpecs(src, dst);
+  // Security and validation checks
+  if (width <= 0 || height <= 0 ||         // luma/alpha param error
+      uv_width < 0 || uv_height < 0) {     // u/v param error
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_BAD_DIMENSION);
   }
-  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);
-    if (src->a != NULL) {
-      dst->a = src->a + top * src->a_stride + left;
-    }
-#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;
-    }
-#endif
-  } else {
-    dst->argb = src->argb + top * src->argb_stride + left;
+  // allocate a new buffer.
+  mem = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*mem));
+  if (mem == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
   }
-  return 1;
-}
 
-//------------------------------------------------------------------------------
-// Picture cropping
+  // 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;
 
-int WebPPictureCrop(WebPPicture* pic,
-                    int left, int top, int width, int height) {
-  WebPPicture tmp;
+  // TODO(skal): we could align the y/u/v planes and adjust stride.
+  picture->y = mem;
+  mem += y_size;
 
-  if (pic == NULL) return 0;
-  if (!AdjustAndCheckRectangle(pic, &left, &top, width, height)) return 0;
+  picture->u = mem;
+  mem += uv_size;
+  picture->v = mem;
+  mem += uv_size;
 
-  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);
+  if (a_size > 0) {
+    picture->a = mem;
+    mem += a_size;
   }
-  WebPPictureFree(pic);
-  *pic = tmp;
+  (void)mem;  // makes the static analyzer happy
   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);
-  }
-}
-
-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;
-    }
+int WebPPictureAlloc(WebPPicture* picture) {
+  if (picture != NULL) {
+    const int width = picture->width;
+    const int height = picture->height;
 
-    RescalePlane(pic->y, prev_width, prev_height, pic->y_stride,
-                 tmp.y, width, height, tmp.y_stride, work, 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);
+    WebPPictureFree(picture);   // erase previous buffer
 
-    if (tmp.a != NULL) {
-      RescalePlane(pic->a, prev_width, prev_height, pic->a_stride,
-                   tmp.a, width, height, tmp.a_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;
+    if (!picture->use_argb) {
+      return WebPPictureAllocYUVA(picture, width, height);
+    } else {
+      return WebPPictureAllocARGB(picture, width, height);
     }
-
-    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);
-
   }
-  WebPPictureFree(pic);
-  free(work);
-  *pic = tmp;
   return 1;
 }
 
+void WebPPictureFree(WebPPicture* picture) {
+  if (picture != NULL) {
+    WebPSafeFree(picture->memory_);
+    WebPSafeFree(picture->memory_argb_);
+    WebPPictureResetBuffers(picture);
+  }
+}
+
 //------------------------------------------------------------------------------
 // WebPMemoryWriter: Write-to-memory
 
@@ -494,7 +203,7 @@ int WebPMemoryWrite(const uint8_t* data, size_t data_size,
     if (w->size > 0) {
       memcpy(new_mem, w->mem, w->size);
     }
-    free(w->mem);
+    WebPSafeFree(w->mem);
     w->mem = new_mem;
     // down-cast is ok, thanks to WebPSafeMalloc
     w->max_size = (size_t)next_max_size;
@@ -506,469 +215,13 @@ int WebPMemoryWrite(const uint8_t* data, size_t 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
-
-// TODO: we can do better than simply 2x2 averaging on U/V samples.
-#define SUM4(ptr) ((ptr)[0] + (ptr)[step] + \
-                   (ptr)[rgb_stride] + (ptr)[rgb_stride + step])
-#define SUM2H(ptr) (2 * (ptr)[0] + 2 * (ptr)[step])
-#define SUM2V(ptr) (2 * (ptr)[0] + 2 * (ptr)[rgb_stride])
-#define SUM1(ptr)  (4 * (ptr)[0])
-#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] = VP8RGBToU(r, g, b);                  \
-  picture->v[dst] = VP8RGBToV(r, g, b);                  \
-}
-
-#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] = VP8RGBToU(r, g, b);                 \
-  picture->v0[dst] = VP8RGBToV(r, g, b);                 \
-}
-
-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
-                              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);
-
-  picture->colorspace = uv_csp;
-  picture->use_argb = 0;
-  if (has_alpha) {
-    picture->colorspace |= WEBP_CSP_ALPHA_BIT;
-  }
-  if (!WebPPictureAlloc(picture)) return 0;
-
-  // 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] =
-          VP8RGBToY(r_ptr[offset], g_ptr[offset], b_ptr[offset]);
-    }
-  }
-
-  // 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);
-    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,
-                              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 =
-            0xff000000u |
-            (r_ptr[offset] << 16) |
-            (g_ptr[offset] <<  8) |
-            (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 = (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->memory_ == NULL || 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(NULL, cur_y, cur_u, cur_v, cur_u, cur_v, NULL, dst, 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 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) {
-          dst[x] = (dst[x] & 0x00ffffffu) | (src[x] << 24);
-        }
-      }
-    }
-  }
-  return 1;
-}
-
-int WebPPictureARGBToYUVA(WebPPicture* picture, WebPEncCSP colorspace) {
-  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, &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;
+void WebPMemoryWriterClear(WebPMemoryWriter* writer) {
+  if (writer != NULL) {
+    WebPSafeFree(writer->mem);
+    writer->mem = NULL;
+    writer->size = 0;
+    writer->max_size = 0;
   }
-  return 1;
-}
-
-//------------------------------------------------------------------------------
-// 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
-
-
-//------------------------------------------------------------------------------
-// Distortion
-
-// Max value returned in case of exact similarity.
-static const double kMinDistortion_dB = 99.;
-
-int WebPPictureDistortion(const WebPPicture* pic1, const WebPPicture* pic2,
-                          int type, float result[5]) {
-  int c;
-  DistoStats stats[5];
-  int has_alpha;
-
-  if (pic1 == NULL || pic2 == NULL ||
-      pic1->width != pic2->width || pic1->height != pic2->height ||
-      pic1->y == NULL || pic2->y == NULL ||
-      pic1->u == NULL || pic2->u == NULL ||
-      pic1->v == NULL || pic2->v == NULL ||
-      result == NULL) {
-    return 0;
-  }
-  // TODO(skal): provide distortion for ARGB too.
-  if (pic1->use_argb == 1 || pic1->use_argb != pic2->use_argb) {
-    return 0;
-  }
-
-  has_alpha = !!(pic1->colorspace & WEBP_CSP_ALPHA_BIT);
-  if (has_alpha != !!(pic2->colorspace & WEBP_CSP_ALPHA_BIT) ||
-      (has_alpha && (pic1->a == NULL || pic2->a == NULL))) {
-    return 0;
-  }
-
-  memset(stats, 0, sizeof(stats));
-  VP8SSIMAccumulatePlane(pic1->y, pic1->y_stride,
-                         pic2->y, pic2->y_stride,
-                         pic1->width, pic1->height, &stats[0]);
-  VP8SSIMAccumulatePlane(pic1->u, pic1->uv_stride,
-                         pic2->u, pic2->uv_stride,
-                         (pic1->width + 1) >> 1, (pic1->height + 1) >> 1,
-                         &stats[1]);
-  VP8SSIMAccumulatePlane(pic1->v, pic1->uv_stride,
-                         pic2->v, pic2->uv_stride,
-                         (pic1->width + 1) >> 1, (pic1->height + 1) >> 1,
-                         &stats[2]);
-  if (has_alpha) {
-    VP8SSIMAccumulatePlane(pic1->a, pic1->a_stride,
-                           pic2->a, pic2->a_stride,
-                           pic1->width, pic1->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] = (float)((v > 0.) ? -4.3429448 * log(v / (255 * 255.))
-                                   : kMinDistortion_dB);
-    }
-    // Accumulate forward
-    if (c < 4) VP8SSIMAddStats(&stats[c], &stats[4]);
-  }
-  return 1;
 }
 
 //------------------------------------------------------------------------------
@@ -1000,7 +253,7 @@ static size_t Encode(const uint8_t* rgba, int width, int height, int stride,
   ok = import(&pic, rgba, stride) && WebPEncode(&config, &pic);
   WebPPictureFree(&pic);
   if (!ok) {
-    free(wrt.mem);
+    WebPMemoryWriterClear(&wrt);
     *output = NULL;
     return 0;
   }
@@ -1014,10 +267,10 @@ size_t NAME(const uint8_t* in, int w, int h, int bps, float q,          \
   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);
+ENCODE_FUNC(WebPEncodeRGB, WebPPictureImportRGB)
+ENCODE_FUNC(WebPEncodeBGR, WebPPictureImportBGR)
+ENCODE_FUNC(WebPEncodeRGBA, WebPPictureImportRGBA)
+ENCODE_FUNC(WebPEncodeBGRA, WebPPictureImportBGRA)
 
 #undef ENCODE_FUNC
 
@@ -1027,15 +280,11 @@ 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);
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGB, WebPPictureImportRGB)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGR, WebPPictureImportBGR)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGBA, WebPPictureImportRGBA)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGRA, WebPPictureImportBGRA)
 
 #undef LOSSLESS_ENCODE_FUNC
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/quant.c b/drivers/webp/enc/quant.c
index ea153849c8..002c326b82 100644
--- a/drivers/webp/enc/quant.c
+++ b/drivers/webp/enc/quant.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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
@@ -11,6 +13,7 @@
 
 #include <assert.h>
 #include <math.h>
+#include <stdlib.h>  // for abs()
 
 #include "./vp8enci.h"
 #include "./cost.h"
@@ -22,16 +25,78 @@
 
 #define MID_ALPHA 64      // neutral value for susceptibility
 #define MIN_ALPHA 30      // lowest usable value for susceptibility
-#define MAX_ALPHA 100     // higher meaninful 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)
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+// #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
 
 //------------------------------------------------------------------------------
 
@@ -100,31 +165,13 @@ static const uint16_t kAcTable2[128] = {
   385, 393, 401, 409, 416, 424, 432, 440
 };
 
-static const uint16_t kCoeffThresh[16] = {
-  0,  10, 20, 30,
-  10, 20, 30, 30,
-  20, 30, 30, 30,
-  30, 30, 30, 30
-};
-
-// TODO(skal): tune more. Coeff thresholding?
-static const uint8_t kBiasMatrices[3][16] = {  // [3] = [luma-ac,luma-dc,chroma]
-  { 96, 96, 96, 96,
-    96, 96, 96, 96,
-    96, 96, 96, 96,
-    96, 96, 96, 96 },
-  { 96, 96, 96, 96,
-    96, 96, 96, 96,
-    96, 96, 96, 96,
-    96, 96, 96, 96 },
-  { 96, 96, 96, 96,
-    96, 96, 96, 96,
-    96, 96, 96, 96,
-    96, 96, 96, 96 }
+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 (only for trellis).
+// 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,
@@ -137,20 +184,30 @@ static const uint8_t kFreqSharpening[16] = {
 
 // Returns the average quantizer
 static int ExpandMatrix(VP8Matrix* const m, int type) {
-  int i;
-  int sum = 0;
+  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 (i = 0; i < 16; ++i) {
-    const int j = kZigzag[i];
-    const int bias = kBiasMatrices[type][j];
-    m->iq_[j] = (1 << QFIX) / m->q_[j];
-    m->bias_[j] = BIAS(bias);
-    // TODO(skal): tune kCoeffThresh[]
-    m->zthresh_[j] = ((256 /*+ kCoeffThresh[j]*/ - bias) * m->q_[j] + 127) >> 8;
-    m->sharpen_[j] = (kFreqSharpening[j] * m->q_[j]) >> 11;
-    sum += m->q_[j];
+  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;
 }
@@ -178,17 +235,17 @@ static void SetupMatrices(VP8Encoder* enc) {
     q16 = ExpandMatrix(&m->y2_, 1);
     quv = ExpandMatrix(&m->uv_, 2);
 
-    // TODO: Switch to kLambda*[] tables?
-    {
-      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->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;
   }
 }
 
@@ -197,16 +254,21 @@ static void SetupMatrices(VP8Encoder* enc) {
 
 // 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 3
+#define FSTRENGTH_CUTOFF 2
 
 static void SetupFilterStrength(VP8Encoder* const enc) {
   int i;
-  const int level0 = enc->config_->filter_strength;
+  // 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) {
-    // Segments with lower quantizer will be less filtered. TODO: tune (wrt SNS)
-    const int level = level0 * 256 * enc->dqm_[i].quant_ / 128;
-    const int f = level / (256 + enc->dqm_[i].beta_);
-    enc->dqm_[i].fstrength_ = (f < FSTRENGTH_CUTOFF) ? 0 : (f > 63) ? 63 : f;
+    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_;
@@ -224,28 +286,90 @@ static void SetupFilterStrength(VP8Encoder* const enc) {
 // 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 q) {
-  const double c = q / 100.;
-  return (c < 0.75) ? c * (2. / 3.) : 2. * c - 1.;
+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->config_->segments;
+  const int num_segments = enc->segment_hdr_.num_segments_;
   const double amp = SNS_TO_DQ * enc->config_->sns_strength / 100. / 128.;
-  const double c_base = QualityToCompression(quality);
+  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) {
-    // 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 modelling for high-quant would make use of kAcTable[]
-    // more explicitely.
-    // Additionally, we modulate the base exponent 1/3 to accommodate for the
-    // quantization susceptibility and allow denser segments to be quantized
-    // more.
-    const double expn = (1. - amp * enc->dqm_[i].alpha_) / 3.;
+    // 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.);
@@ -271,7 +395,7 @@ void VP8SetSegmentParams(VP8Encoder* const enc, float quality) {
   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).
+  // tend to appear, and are unpleasant).
   dq_uv_dc = -4 * enc->config_->sns_strength / 100;
   dq_uv_dc = clip(dq_uv_dc, -15, 15);   // 4bit-signed max allowed
 
@@ -281,9 +405,11 @@ void VP8SetSegmentParams(VP8Encoder* const enc, float quality) {
   enc->dq_uv_dc_ = dq_uv_dc;
   enc->dq_uv_ac_ = dq_uv_ac;
 
-  SetupMatrices(enc);
-
   SetupFilterStrength(enc);   // initialize segments' filtering, eventually
+
+  if (num_segments > 1) SimplifySegments(enc);
+
+  SetupMatrices(enc);         // finalize quantization matrices
 }
 
 //------------------------------------------------------------------------------
@@ -299,16 +425,14 @@ const int VP8I4ModeOffsets[NUM_BMODES] = {
 };
 
 void VP8MakeLuma16Preds(const VP8EncIterator* const it) {
-  const VP8Encoder* const enc = it->enc_;
-  const uint8_t* const left = it->x_ ? enc->y_left_ : NULL;
-  const uint8_t* const top = it->y_ ? enc->y_top_ + it->x_ * 16 : NULL;
+  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 VP8Encoder* const enc = it->enc_;
-  const uint8_t* const left = it->x_ ? enc->u_left_ : NULL;
-  const uint8_t* const top = it->y_ ? enc->uv_top_ + it->x_ * 16 : NULL;
+  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);
 }
 
@@ -320,23 +444,21 @@ void VP8MakeIntra4Preds(const VP8EncIterator* const it) {
 // Quantize
 
 // Layout:
-// +----+
-// |YYYY| 0
-// |YYYY| 4
-// |YYYY| 8
-// |YYYY| 12
-// +----+
-// |UUVV| 16
-// |UUVV| 20
-// +----+
-
-const int VP8Scan[16 + 4 + 4] = {
-  // Luma
+// +----+----+
+// |YYYY|UUVV| 0
+// |YYYY|UUVV| 4
+// |YYYY|....| 8
+// |YYYY|....| 12
+// +----+----+
+
+const int VP8Scan[16] = {  // Luma
   0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
   0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
   0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
   0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS,
+};
 
+static const int VP8ScanUV[4 + 4] = {
   0 + 0 * BPS,   4 + 0 * BPS, 0 + 4 * BPS,  4 + 4 * BPS,    // U
   8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
 };
@@ -364,6 +486,7 @@ 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;
 }
@@ -372,6 +495,7 @@ 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;
 }
@@ -380,6 +504,7 @@ 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;
 }
@@ -387,28 +512,31 @@ static void AddScore(VP8ModeScore* const dst, const VP8ModeScore* const src) {
 //------------------------------------------------------------------------------
 // Performs trellis-optimized quantization.
 
-// Trellis
-
+// Trellis node
 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.)
+  int8_t prev;            // best previous node
+  int8_t sign;            // sign of coeff_i
+  int16_t level;          // level
 } Node;
 
+// Score state
+typedef struct {
+  score_t score;          // partial RD score
+  const uint16_t* costs;  // shortcut to cost tables
+} ScoreState;
+
 // If a coefficient was quantized to a value Q (using a neutral bias),
 // we test all alternate possibilities between [Q-MIN_DELTA, Q+MAX_DELTA]
 // We don't test negative values though.
 #define MIN_DELTA 0   // how much lower level to try
 #define MAX_DELTA 1   // how much higher
 #define NUM_NODES (MIN_DELTA + 1 + MAX_DELTA)
-#define NODE(n, l) (nodes[(n) + 1][(l) + MIN_DELTA])
+#define NODE(n, l) (nodes[(n)][(l) + MIN_DELTA])
+#define SCORE_STATE(n, l) (score_states[n][(l) + MIN_DELTA])
 
 static WEBP_INLINE void SetRDScore(int lambda, VP8ModeScore* const rd) {
   // TODO: incorporate the "* 256" in the tables?
-  rd->score = rd->R * lambda + 256 * (rd->D + rd->SD);
+  rd->score = (rd->R + rd->H) * lambda + 256 * (rd->D + rd->SD);
 }
 
 static WEBP_INLINE score_t RDScoreTrellis(int lambda, score_t rate,
@@ -416,34 +544,37 @@ static WEBP_INLINE score_t RDScoreTrellis(int lambda, score_t rate,
   return rate * lambda + 256 * distortion;
 }
 
-static int TrellisQuantizeBlock(const VP8EncIterator* const it,
+static int TrellisQuantizeBlock(const VP8Encoder* const enc,
                                 int16_t in[16], int16_t out[16],
                                 int ctx0, int coeff_type,
                                 const VP8Matrix* const mtx,
                                 int lambda) {
-  ProbaArray* const last_costs = it->enc_->proba_.coeffs_[coeff_type];
-  CostArray* const costs = it->enc_->proba_.level_cost_[coeff_type];
+  const ProbaArray* const probas = enc->proba_.coeffs_[coeff_type];
+  CostArrayPtr const costs =
+      (CostArrayPtr)enc->proba_.remapped_costs_[coeff_type];
   const int first = (coeff_type == 0) ? 1 : 0;
-  Node nodes[17][NUM_NODES];
+  Node nodes[16][NUM_NODES];
+  ScoreState score_states[2][NUM_NODES];
+  ScoreState* ss_cur = &SCORE_STATE(0, MIN_DELTA);
+  ScoreState* ss_prev = &SCORE_STATE(1, MIN_DELTA);
   int best_path[3] = {-1, -1, -1};   // store best-last/best-level/best-previous
   score_t best_score;
-  int best_node;
-  int last = first - 1;
-  int n, m, p, nz;
+  int n, m, p, last;
 
   {
     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];
+    const int last_proba = probas[VP8EncBands[first]][ctx0][0];
 
-    // compute maximal distortion.
-    max_error = 0;
-    for (n = first; n < 16; ++n) {
-      const int j  = kZigzag[n];
+    // compute the position of the last interesting coefficient
+    last = first - 1;
+    for (n = 15; n >= first; --n) {
+      const int j = kZigzag[n];
       const int err = in[j] * in[j];
-      max_error += kWeightTrellis[j] * err;
-      if (err > thresh) last = n;
+      if (err > thresh) {
+        last = n;
+        break;
+      }
     }
     // we don't need to go inspect up to n = 16 coeffs. We can just go up
     // to last + 1 (inclusive) without losing much.
@@ -451,93 +582,95 @@ static int TrellisQuantizeBlock(const VP8EncIterator* const it,
 
     // compute 'skip' score. This is the max score one can do.
     cost = VP8BitCost(0, last_proba);
-    best_score = RDScoreTrellis(lambda, cost, max_error);
+    best_score = RDScoreTrellis(lambda, cost, 0);
 
     // 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;
+      const score_t rate = (ctx0 == 0) ? VP8BitCost(1, last_proba) : 0;
+      ss_cur[m].score = RDScoreTrellis(lambda, rate, 0);
+      ss_cur[m].costs = costs[first][ctx0];
     }
   }
 
   // traverse trellis.
   for (n = first; n <= last; ++n) {
-    const int j  = kZigzag[n];
-    const int Q  = mtx->q_[j];
-    const int iQ = mtx->iq_[j];
-    const int B = BIAS(0x00);     // neutral bias
+    const int j = kZigzag[n];
+    const uint32_t Q  = mtx->q_[j];
+    const uint32_t iQ = mtx->iq_[j];
+    const uint32_t B = BIAS(0x00);     // neutral bias
     // note: it's important to take sign of the _original_ coeff,
     // so we don't have to consider level < 0 afterward.
     const int sign = (in[j] < 0);
-    int coeff0 = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
-    int level0;
-    if (coeff0 > 2047) coeff0 = 2047;
+    const uint32_t coeff0 = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
+    int level0 = QUANTDIV(coeff0, iQ, B);
+    if (level0 > MAX_LEVEL) level0 = MAX_LEVEL;
+
+    {   // Swap current and previous score states
+      ScoreState* const tmp = ss_cur;
+      ss_cur = ss_prev;
+      ss_prev = tmp;
+    }
 
-    level0 = QUANTDIV(coeff0, iQ, B);
     // 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 >= 2048 || level < 0) {   // node is dead?
-        cur->cost = MAX_COST;
+      const int ctx = (level > 2) ? 2 : level;
+      const int band = VP8EncBands[n + 1];
+      score_t base_score, last_pos_score;
+      score_t best_cur_score = MAX_COST;
+      int best_prev = 0;   // default, in case
+
+      ss_cur[m].score = MAX_COST;
+      ss_cur[m].costs = costs[n + 1][ctx];
+      if (level > MAX_LEVEL || level < 0) {   // node is dead?
         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);
+      // Compute extra rate cost if last coeff's position is < 15
+      {
+        const score_t last_pos_cost =
+            (n < 15) ? VP8BitCost(0, probas[band][ctx][0]) : 0;
+        last_pos_score = RDScoreTrellis(lambda, last_pos_cost, 0);
+      }
+
+      {
+        // Compute delta_error = how much coding this level will
+        // subtract to max_error as distortion.
+        // Here, distortion = sum of (|coeff_i| - level_i * Q_i)^2
+        const int new_error = coeff0 - level * Q;
+        const int delta_error =
+            kWeightTrellis[j] * (new_error * new_error - coeff0 * coeff0);
+        base_score = RDScoreTrellis(lambda, 0, delta_error);
+      }
 
       // Inspect all possible non-dead predecessors. Retain only the best one.
       for (p = -MIN_DELTA; p <= MAX_DELTA; ++p) {
-        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);
-
+        // Dead nodes (with ss_prev[p].score >= MAX_COST) are automatically
+        // eliminated since their score can't be better than the current best.
+        const score_t cost = VP8LevelCost(ss_prev[p].costs, level);
         // Examine node assuming it's a non-terminal one.
-        cost = base_cost;
-        if (level && n < 15) {
-          cost += VP8BitCost(1, last_proba);
+        const score_t score =
+            base_score + ss_prev[p].score + RDScoreTrellis(lambda, cost, 0);
+        if (score < best_cur_score) {
+          best_cur_score = score;
+          best_prev = p;
         }
-        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
-          }
+      }
+      // Store best finding in current node.
+      cur->sign = sign;
+      cur->level = level;
+      cur->prev = best_prev;
+      ss_cur[m].score = best_cur_score;
+
+      // Now, record best terminal node (and thus best entry in the graph).
+      if (level != 0) {
+        const score_t score = best_cur_score + last_pos_score;
+        if (score < best_score) {
+          best_score = score;
+          best_path[0] = n;                     // best eob position
+          best_path[1] = m;                     // best node index
+          best_path[2] = best_prev;             // best predecessor
         }
       }
     }
@@ -550,23 +683,25 @@ static int TrellisQuantizeBlock(const VP8EncIterator* const it,
     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;
+  {
+    // Unwind the best path.
+    // Note: best-prev on terminal node is not necessarily equal to the
+    // best_prev for non-terminal. So we patch best_path[2] in.
+    int nz = 0;
+    int best_node = best_path[1];
+    n = best_path[0];
+    NODE(n, best_node).prev = best_path[2];   // force best-prev for terminal
+
+    for (; n >= first; --n) {
+      const Node* const node = &NODE(n, best_node);
+      const int j = kZigzag[n];
+      out[n] = node->sign ? -node->level : node->level;
+      nz |= node->level;
+      in[j] = out[n] * mtx->q_[j];
+      best_node = node->prev;
+    }
+    return (nz != 0);
   }
-  return nz;
 }
 
 #undef NODE
@@ -582,17 +717,17 @@ static int ReconstructIntra16(VP8EncIterator* const it,
                               int mode) {
   const VP8Encoder* const enc = it->enc_;
   const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
-  const uint8_t* const src = it->yuv_in_ + Y_OFF;
+  const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC;
   const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
   int nz = 0;
   int n;
   int16_t tmp[16][16], dc_tmp[16];
 
-  for (n = 0; n < 16; ++n) {
-    VP8FTransform(src + VP8Scan[n], ref + VP8Scan[n], tmp[n]);
+  for (n = 0; n < 16; n += 2) {
+    VP8FTransform2(src + VP8Scan[n], ref + VP8Scan[n], tmp[n]);
   }
   VP8FTransformWHT(tmp[0], dc_tmp);
-  nz |= VP8EncQuantizeBlock(dc_tmp, rd->y_dc_levels, 0, &dqm->y2_) << 24;
+  nz |= VP8EncQuantizeBlockWHT(dc_tmp, rd->y_dc_levels, &dqm->y2_) << 24;
 
   if (DO_TRELLIS_I16 && it->do_trellis_) {
     int x, y;
@@ -601,20 +736,26 @@ static int ReconstructIntra16(VP8EncIterator* const it,
       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_);
+            TrellisQuantizeBlock(enc, tmp[n], rd->y_ac_levels[n], ctx, 0,
+                                 &dqm->y1_, dqm->lambda_trellis_i16_);
         it->top_nz_[x] = it->left_nz_[y] = non_zero;
+        rd->y_ac_levels[n][0] = 0;
         nz |= non_zero << n;
       }
     }
   } else {
-    for (n = 0; n < 16; ++n) {
-      nz |= VP8EncQuantizeBlock(tmp[n], rd->y_ac_levels[n], 1, &dqm->y1_) << n;
+    for (n = 0; n < 16; n += 2) {
+      // Zero-out the first coeff, so that: a) nz is correct below, and
+      // b) finding 'last' non-zero coeffs in SetResidualCoeffs() is simplified.
+      tmp[n][0] = tmp[n + 1][0] = 0;
+      nz |= VP8EncQuantize2Blocks(tmp[n], rd->y_ac_levels[n], &dqm->y1_) << n;
+      assert(rd->y_ac_levels[n + 0][0] == 0);
+      assert(rd->y_ac_levels[n + 1][0] == 0);
     }
   }
 
   // Transform back
-  VP8ITransformWHT(dc_tmp, tmp[0]);
+  VP8TransformWHT(dc_tmp, tmp[0]);
   for (n = 0; n < 16; n += 2) {
     VP8ITransform(ref + VP8Scan[n], tmp[n], yuv_out + VP8Scan[n], 1);
   }
@@ -637,10 +778,10 @@ static int ReconstructIntra4(VP8EncIterator* const it,
   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_,
+    nz = TrellisQuantizeBlock(enc, tmp, levels, ctx, 3, &dqm->y1_,
                               dqm->lambda_trellis_i4_);
   } else {
-    nz = VP8EncQuantizeBlock(tmp, levels, 0, &dqm->y1_);
+    nz = VP8EncQuantizeBlock(tmp, levels, &dqm->y1_);
   }
   VP8ITransform(ref, tmp, yuv_out, 0);
   return nz;
@@ -650,14 +791,14 @@ 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 uint8_t* const src = it->yuv_in_ + U_OFF_ENC;
   const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
   int nz = 0;
   int n;
   int16_t tmp[8][16];
 
-  for (n = 0; n < 8; ++n) {
-    VP8FTransform(src + VP8Scan[16 + n], ref + VP8Scan[16 + n], tmp[n]);
+  for (n = 0; n < 8; n += 2) {
+    VP8FTransform2(src + VP8ScanUV[n], ref + VP8ScanUV[n], tmp[n]);
   }
   if (DO_TRELLIS_UV && it->do_trellis_) {
     int ch, x, y;
@@ -666,28 +807,45 @@ static int ReconstructUV(VP8EncIterator* const it, VP8ModeScore* const rd,
         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_);
+              TrellisQuantizeBlock(enc, tmp[n], rd->uv_levels[n], ctx, 2,
+                                   &dqm->uv_, dqm->lambda_trellis_uv_);
           it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = non_zero;
           nz |= non_zero << n;
         }
       }
     }
   } else {
-    for (n = 0; n < 8; ++n) {
-      nz |= VP8EncQuantizeBlock(tmp[n], rd->uv_levels[n], 0, &dqm->uv_) << n;
+    for (n = 0; n < 8; n += 2) {
+      nz |= VP8EncQuantize2Blocks(tmp[n], rd->uv_levels[n], &dqm->uv_) << n;
     }
   }
 
   for (n = 0; n < 8; n += 2) {
-    VP8ITransform(ref + VP8Scan[16 + n], tmp[n], yuv_out + VP8Scan[16 + n], 1);
+    VP8ITransform(ref + VP8ScanUV[n], tmp[n], yuv_out + VP8ScanUV[n], 1);
   }
   return (nz << 16);
 }
 
 //------------------------------------------------------------------------------
 // RD-opt decision. Reconstruct each modes, evalue distortion and bit-cost.
-// Pick the mode is lower RD-cost = Rate + lamba * Distortion.
+// 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 SwapModeScore(VP8ModeScore** a, VP8ModeScore** b) {
+  VP8ModeScore* const tmp = *a;
+  *a = *b;
+  *b = tmp;
+}
 
 static void SwapPtr(uint8_t** a, uint8_t** b) {
   uint8_t* const tmp = *a;
@@ -699,43 +857,69 @@ static void SwapOut(VP8EncIterator* const it) {
   SwapPtr(&it->yuv_out_, &it->yuv_out2_);
 }
 
-static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* const rd) {
-  const VP8Encoder* const enc = it->enc_;
-  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+static score_t IsFlat(const int16_t* levels, int num_blocks, score_t thresh) {
+  score_t score = 0;
+  while (num_blocks-- > 0) {      // TODO(skal): refine positional scoring?
+    int i;
+    for (i = 1; i < 16; ++i) {    // omit DC, we're only interested in AC
+      score += (levels[i] != 0);
+      if (score > thresh) return 0;
+    }
+    levels += 16;
+  }
+  return 1;
+}
+
+static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* rd) {
+  const int kNumBlocks = 16;
+  VP8SegmentInfo* const dqm = &it->enc_->dqm_[it->mb_->segment_];
   const int lambda = dqm->lambda_i16_;
   const int tlambda = dqm->tlambda_;
-  const uint8_t* const src = it->yuv_in_ + Y_OFF;
-  VP8ModeScore rd16;
+  const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC;
+  VP8ModeScore rd_tmp;
+  VP8ModeScore* rd_cur = &rd_tmp;
+  VP8ModeScore* rd_best = rd;
   int mode;
 
   rd->mode_i16 = -1;
-  for (mode = 0; mode < 4; ++mode) {
-    uint8_t* const tmp_dst = it->yuv_out2_ + Y_OFF;  // scratch buffer
-    int nz;
+  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
+    uint8_t* const tmp_dst = it->yuv_out2_ + Y_OFF_ENC;  // scratch buffer
+    rd_cur->mode_i16 = mode;
 
     // Reconstruct
-    nz = ReconstructIntra16(it, &rd16, tmp_dst, mode);
+    rd_cur->nz = ReconstructIntra16(it, rd_cur, 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.R = VP8GetCostLuma16(it, &rd16);
-    rd16.R += VP8FixedCostsI16[mode];
+    rd_cur->D = VP8SSE16x16(src, tmp_dst);
+    rd_cur->SD =
+        tlambda ? MULT_8B(tlambda, VP8TDisto16x16(src, tmp_dst, kWeightY)) : 0;
+    rd_cur->H = VP8FixedCostsI16[mode];
+    rd_cur->R = VP8GetCostLuma16(it, rd_cur);
+    if (mode > 0 &&
+        IsFlat(rd_cur->y_ac_levels[0], kNumBlocks, FLATNESS_LIMIT_I16)) {
+      // penalty to avoid flat area to be mispredicted by complex mode
+      rd_cur->R += FLATNESS_PENALTY * kNumBlocks;
+    }
 
     // Since we always examine Intra16 first, we can overwrite *rd directly.
-    SetRDScore(lambda, &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));
+    SetRDScore(lambda, rd_cur);
+    if (mode == 0 || rd_cur->score < rd_best->score) {
+      SwapModeScore(&rd_cur, &rd_best);
       SwapOut(it);
     }
   }
+  if (rd_best != rd) {
+    memcpy(rd, rd_best, sizeof(*rd));
+  }
   SetRDScore(dqm->lambda_mode_, rd);   // finalize score for mode decision.
   VP8SetIntra16Mode(it, rd->mode_i16);
+
+  // we have a blocky macroblock (only DCs are non-zero) with fairly high
+  // distortion, record max delta so we can later adjust the minimal filtering
+  // strength needed to smooth these blocks out.
+  if ((rd->nz & 0xffff) == 0 && rd->D > dqm->min_disto_) {
+    StoreMaxDelta(dqm, rd->y_dc_levels);
+  }
 }
 
 //------------------------------------------------------------------------------
@@ -755,8 +939,8 @@ static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
   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;
+  const uint8_t* const src0 = it->yuv_in_ + Y_OFF_ENC;
+  uint8_t* const best_blocks = it->yuv_out2_ + Y_OFF_ENC;
   int total_header_bits = 0;
   VP8ModeScore rd_best;
 
@@ -765,9 +949,11 @@ static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
   }
 
   InitScore(&rd_best);
-  rd_best.score = 211;  // '211' is the value of VP8BitCost(0, 145)
+  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;
@@ -791,27 +977,44 @@ static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
       rd_tmp.SD =
           tlambda ? MULT_8B(tlambda, VP8TDisto4x4(src, tmp_dst, kWeightY))
                   : 0;
-      rd_tmp.R = VP8GetCostLuma4(it, tmp_levels);
-      rd_tmp.R += mode_costs[mode];
+      rd_tmp.H = mode_costs[mode];
+
+      // Add flatness penalty
+      if (mode > 0 && IsFlat(tmp_levels, kNumBlocks, FLATNESS_LIMIT_I4)) {
+        rd_tmp.R = FLATNESS_PENALTY * kNumBlocks;
+      } else {
+        rd_tmp.R = 0;
+      }
 
+      // early-out check
       SetRDScore(lambda, &rd_tmp);
+      if (best_mode >= 0 && rd_tmp.score >= rd_i4.score) continue;
+
+      // finish computing score
+      rd_tmp.R += VP8GetCostLuma4(it, tmp_levels);
+      SetRDScore(lambda, &rd_tmp);
+
       if (best_mode < 0 || rd_tmp.score < rd_i4.score) {
         CopyScore(&rd_i4, &rd_tmp);
         best_mode = mode;
         SwapPtr(&tmp_dst, &best_block);
-        memcpy(rd_best.y_ac_levels[it->i4_], tmp_levels, sizeof(tmp_levels));
+        memcpy(rd_best.y_ac_levels[it->i4_], tmp_levels,
+               sizeof(rd_best.y_ac_levels[it->i4_]));
       }
     }
     SetRDScore(dqm->lambda_mode_, &rd_i4);
     AddScore(&rd_best, &rd_i4);
-    total_header_bits += mode_costs[best_mode];
-    if (rd_best.score >= rd->score ||
-        total_header_bits > enc->max_i4_header_bits_) {
+    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_])
+    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));
@@ -827,18 +1030,19 @@ static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
 //------------------------------------------------------------------------------
 
 static void PickBestUV(VP8EncIterator* const it, VP8ModeScore* const rd) {
-  const VP8Encoder* const enc = it->enc_;
-  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  const int kNumBlocks = 8;
+  const VP8SegmentInfo* const dqm = &it->enc_->dqm_[it->mb_->segment_];
   const int lambda = dqm->lambda_uv_;
-  const uint8_t* const src = it->yuv_in_ + U_OFF;
-  uint8_t* const tmp_dst = it->yuv_out2_ + U_OFF;  // scratch buffer
-  uint8_t* const dst0 = it->yuv_out_ + U_OFF;
+  const uint8_t* const src = it->yuv_in_ + U_OFF_ENC;
+  uint8_t* tmp_dst = it->yuv_out2_ + U_OFF_ENC;  // scratch buffer
+  uint8_t* dst0 = it->yuv_out_ + U_OFF_ENC;
+  uint8_t* dst = dst0;
   VP8ModeScore rd_best;
   int mode;
 
   rd->mode_uv = -1;
   InitScore(&rd_best);
-  for (mode = 0; mode < 4; ++mode) {
+  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
     VP8ModeScore rd_uv;
 
     // Reconstruct
@@ -847,19 +1051,25 @@ static void PickBestUV(VP8EncIterator* const it, VP8ModeScore* const rd) {
     // 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);
-    rd_uv.R += VP8FixedCostsUV[mode];
+    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() ?
+      SwapPtr(&dst, &tmp_dst);
     }
   }
   VP8SetIntraUVMode(it, rd->mode_uv);
   AddScore(rd, &rd_best);
+  if (dst != dst0) {   // copy 16x8 block if needed
+    VP8Copy16x8(dst, dst0);
+  }
 }
 
 //------------------------------------------------------------------------------
@@ -867,33 +1077,88 @@ static void PickBestUV(VP8EncIterator* const it, VP8ModeScore* const rd) {
 
 static void SimpleQuantize(VP8EncIterator* const it, VP8ModeScore* const rd) {
   const VP8Encoder* const enc = it->enc_;
-  const int i16 = (it->mb_->type_ == 1);
+  const int is_i16 = (it->mb_->type_ == 1);
   int nz = 0;
 
-  if (i16) {
-    nz = ReconstructIntra16(it, rd, it->yuv_out_ + Y_OFF, it->preds_[0]);
+  if (is_i16) {
+    nz = ReconstructIntra16(it, rd, it->yuv_out_ + Y_OFF_ENC, it->preds_[0]);
   } else {
     VP8IteratorStartI4(it);
     do {
       const int mode =
           it->preds_[(it->i4_ & 3) + (it->i4_ >> 2) * enc->preds_w_];
-      const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
-      uint8_t* const dst = it->yuv_out_ + Y_OFF + VP8Scan[it->i4_];
+      const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC + VP8Scan[it->i4_];
+      uint8_t* const dst = it->yuv_out_ + Y_OFF_ENC + VP8Scan[it->i4_];
       VP8MakeIntra4Preds(it);
       nz |= ReconstructIntra4(it, rd->y_ac_levels[it->i4_],
                               src, dst, mode) << it->i4_;
-    } while (VP8IteratorRotateI4(it, it->yuv_out_ + Y_OFF));
+    } while (VP8IteratorRotateI4(it, it->yuv_out_ + Y_OFF_ENC));
   }
 
-  nz |= ReconstructUV(it, rd, it->yuv_out_ + U_OFF, it->mb_->uv_mode_);
+  nz |= ReconstructUV(it, rd, it->yuv_out_ + U_OFF_ENC, it->mb_->uv_mode_);
   rd->nz = nz;
 }
 
+// Refine intra16/intra4 sub-modes based on distortion only (not rate).
+static void 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_ENC;
+      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_ENC + 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_ENC));
+    if (score_i4 < best_score) {
+      VP8SetIntra4Mode(it, modes_i4);
+    }
+  }
+}
+
 //------------------------------------------------------------------------------
 // Entry point
 
-int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd, int rd_opt) {
+int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd,
+                VP8RDLevel rd_opt) {
   int is_skipped;
+  const int method = it->enc_->method_;
 
   InitScore(rd);
 
@@ -902,22 +1167,21 @@ int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd, int rd_opt) {
   VP8MakeLuma16Preds(it);
   VP8MakeChroma8Preds(it);
 
-  // for rd_opt = 2, we perform trellis-quant on the final decision only.
-  // for rd_opt > 2, we use it for every scoring (=much slower).
-  if (rd_opt > 0) {
-    it->do_trellis_ = (rd_opt > 2);
+  if (rd_opt > RD_OPT_NONE) {
+    it->do_trellis_ = (rd_opt >= RD_OPT_TRELLIS_ALL);
     PickBestIntra16(it, rd);
-    if (it->enc_->method_ >= 2) {
+    if (method >= 2) {
       PickBestIntra4(it, rd);
     }
     PickBestUV(it, rd);
-    if (rd_opt == 2) {
+    if (rd_opt == RD_OPT_TRELLIS) {   // finish off with trellis-optim now
       it->do_trellis_ = 1;
       SimpleQuantize(it, rd);
     }
   } else {
-    // TODO: for method_ == 2, pick the best intra4/intra16 based on SSE
-    it->do_trellis_ = (it->enc_->method_ == 2);
+    // 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);
@@ -925,6 +1189,3 @@ int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd, int rd_opt) {
   return is_skipped;
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/syntax.c b/drivers/webp/enc/syntax.c
index 4221436ff9..a0e79ef404 100644
--- a/drivers/webp/enc/syntax.c
+++ b/drivers/webp/enc/syntax.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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
@@ -11,35 +13,20 @@
 
 #include <assert.h>
 
-#include "../format_constants.h"
+#include "../utils/utils.h"
+#include "../webp/format_constants.h"  // RIFF constants
+#include "../webp/mux_types.h"         // ALPHA_FLAG
 #include "./vp8enci.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
 // Helper functions
 
-// TODO(later): Move to webp/format_constants.h?
-static void PutLE24(uint8_t* const data, uint32_t val) {
-  data[0] = (val >>  0) & 0xff;
-  data[1] = (val >>  8) & 0xff;
-  data[2] = (val >> 16) & 0xff;
-}
-
-static void PutLE32(uint8_t* const data, uint32_t val) {
-  PutLE24(data, val);
-  data[3] = (val >> 24) & 0xff;
-}
-
 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);
 }
@@ -73,14 +60,14 @@ static WebPEncodingError PutVP8XHeader(const VP8Encoder* const enc) {
   assert(pic->width <= MAX_CANVAS_SIZE && pic->height <= MAX_CANVAS_SIZE);
 
   if (enc->has_alpha_) {
-    flags |= ALPHA_FLAG_BIT;
+    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)) {
+  if (!pic->writer(vp8x, sizeof(vp8x), pic)) {
     return VP8_ENC_ERROR_BAD_WRITE;
   }
   return VP8_ENC_OK;
@@ -199,8 +186,8 @@ static int PutWebPHeaders(const VP8Encoder* const enc, size_t size0,
 // 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_;
+  const VP8EncSegmentHeader* const hdr = &enc->segment_hdr_;
+  const VP8EncProba* const proba = &enc->proba_;
   if (VP8PutBitUniform(bw, (hdr->num_segments_ > 1))) {
     // We always 'update' the quant and filter strength values
     const int update_data = 1;
@@ -210,16 +197,16 @@ static void PutSegmentHeader(VP8BitWriter* const bw,
       // 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);
+        VP8PutSignedBits(bw, enc->dqm_[s].quant_, 7);
       }
       for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
-        VP8PutSignedValue(bw, enc->dqm_[s].fstrength_, 6);
+        VP8PutSignedBits(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);
+          VP8PutBits(bw, proba->segments_[s], 8);
         }
       }
     }
@@ -228,20 +215,20 @@ static void PutSegmentHeader(VP8BitWriter* const bw,
 
 // Filtering parameters header
 static void PutFilterHeader(VP8BitWriter* const bw,
-                            const VP8FilterHeader* const hdr) {
+                            const VP8EncFilterHeader* 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);
+  VP8PutBits(bw, hdr->level_, 6);
+  VP8PutBits(bw, hdr->sharpness_, 3);
   if (VP8PutBitUniform(bw, use_lf_delta)) {
     // '0' is the default value for i4x4_lf_delta_ at frame #0.
     const int need_update = (hdr->i4x4_lf_delta_ != 0);
     if (VP8PutBitUniform(bw, need_update)) {
       // we don't use ref_lf_delta => emit four 0 bits
-      VP8PutValue(bw, 0, 4);
+      VP8PutBits(bw, 0, 4);
       // we use mode_lf_delta for i4x4
-      VP8PutSignedValue(bw, hdr->i4x4_lf_delta_, 6);
-      VP8PutValue(bw, 0, 3);    // all others unused
+      VP8PutSignedBits(bw, hdr->i4x4_lf_delta_, 6);
+      VP8PutBits(bw, 0, 3);    // all others unused
     }
   }
 }
@@ -249,12 +236,12 @@ static void PutFilterHeader(VP8BitWriter* const bw,
 // 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);
+  VP8PutBits(bw, enc->base_quant_, 7);
+  VP8PutSignedBits(bw, enc->dq_y1_dc_, 4);
+  VP8PutSignedBits(bw, enc->dq_y2_dc_, 4);
+  VP8PutSignedBits(bw, enc->dq_y2_ac_, 4);
+  VP8PutSignedBits(bw, enc->dq_uv_dc_, 4);
+  VP8PutSignedBits(bw, enc->dq_uv_ac_, 4);
 }
 
 // Partition sizes
@@ -276,58 +263,23 @@ static int EmitPartitionsSize(const VP8Encoder* const enc,
 
 //------------------------------------------------------------------------------
 
-#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) {
+static int GeneratePartition0(VP8Encoder* const enc) {
   VP8BitWriter* const bw = &enc->bw_;
   const int mb_size = enc->mb_w_ * enc->mb_h_;
   uint64_t pos1, pos2, pos3;
-#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
+  if (!VP8BitWriterInit(bw, mb_size * 7 / 8)) {        // ~7 bits per macroblock
+    return WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
   VP8PutBitUniform(bw, 0);   // colorspace
-#endif
   VP8PutBitUniform(bw, 0);   // clamp type
 
   PutSegmentHeader(bw, enc);
   PutFilterHeader(bw, &enc->filter_hdr_);
-  VP8PutValue(bw, enc->config_->partitions, 2);
+  VP8PutBits(bw, enc->num_parts_ == 8 ? 3 :
+                 enc->num_parts_ == 4 ? 2 :
+                 enc->num_parts_ == 2 ? 1 : 0, 2);
   PutQuant(bw, enc);
   VP8PutBitUniform(bw, 0);   // no proba update
   VP8WriteProbas(bw, &enc->proba_);
@@ -335,21 +287,17 @@ static size_t GeneratePartition0(VP8Encoder* const enc) {
   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_;
+  if (bw->error_) {
+    return WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  return 1;
 }
 
 void VP8EncFreeBitWriters(VP8Encoder* const enc) {
@@ -371,7 +319,8 @@ int VP8EncWrite(VP8Encoder* const enc) {
   int p;
 
   // Partition #0 with header and partition sizes
-  ok = !!GeneratePartition0(enc);
+  ok = GeneratePartition0(enc);
+  if (!ok) return 0;
 
   // Compute VP8 size
   vp8_size = VP8_FRAME_HEADER_SIZE +
@@ -432,6 +381,3 @@ int VP8EncWrite(VP8Encoder* const enc) {
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/tree.c b/drivers/webp/enc/tree.c
index 8b25e5e488..f141006d19 100644
--- a/drivers/webp/enc/tree.c
+++ b/drivers/webp/enc/tree.c
@@ -1,27 +1,24 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
-// Token probabilities
+// Coding of token probabilities, intra modes and segments.
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include "./vp8enci.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
 // Default probabilities
 
 // Paragraph 13.5
 const uint8_t
   VP8CoeffsProba0[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS] = {
-  // genereated using vp8_default_coef_probs() in entropy.c:129
   { { { 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 }
@@ -157,7 +154,7 @@ const uint8_t
 };
 
 void VP8DefaultProbas(VP8Encoder* const enc) {
-  VP8Proba* const probas = &enc->proba_;
+  VP8EncProba* const probas = &enc->proba_;
   probas->use_skip_proba_ = 0;
   memset(probas->segments_, 255u, sizeof(probas->segments_));
   memcpy(probas->coeffs_, VP8CoeffsProba0, sizeof(VP8CoeffsProba0));
@@ -318,7 +315,7 @@ void VP8CodeIntraModes(VP8Encoder* const enc) {
   VP8EncIterator it;
   VP8IteratorInit(enc, &it);
   do {
-    const VP8MBInfo* mb = it.mb_;
+    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_);
@@ -343,7 +340,7 @@ void VP8CodeIntraModes(VP8Encoder* const enc) {
       }
     }
     PutUVMode(bw, mb->uv_mode_);
-  } while (VP8IteratorNext(&it, 0));
+  } while (VP8IteratorNext(&it));
 }
 
 //------------------------------------------------------------------------------
@@ -485,7 +482,7 @@ const uint8_t
   }
 };
 
-void VP8WriteProbas(VP8BitWriter* const bw, const VP8Proba* const probas) {
+void VP8WriteProbas(VP8BitWriter* const bw, const VP8EncProba* const probas) {
   int t, b, c, p;
   for (t = 0; t < NUM_TYPES; ++t) {
     for (b = 0; b < NUM_BANDS; ++b) {
@@ -494,17 +491,14 @@ void VP8WriteProbas(VP8BitWriter* const bw, const VP8Proba* const probas) {
           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);
+            VP8PutBits(bw, p0, 8);
           }
         }
       }
     }
   }
   if (VP8PutBitUniform(bw, probas->use_skip_proba_)) {
-    VP8PutValue(bw, probas->skip_proba_, 8);
+    VP8PutBits(bw, probas->skip_proba_, 8);
   }
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/vp8enci.h b/drivers/webp/enc/vp8enci.h
index 936e1c18ce..1a7ebe5703 100644
--- a/drivers/webp/enc/vp8enci.h
+++ b/drivers/webp/enc/vp8enci.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
@@ -13,11 +15,18 @@
 #define WEBP_ENC_VP8ENCI_H_
 
 #include <string.h>     // for memcpy()
-#include "../encode.h"
+#include "../dec/common.h"
 #include "../dsp/dsp.h"
 #include "../utils/bit_writer.h"
+#include "../utils/thread.h"
+#include "../utils/utils.h"
+#include "../webp/encode.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+#include "./vp8li.h"
+#endif  // WEBP_EXPERIMENTAL_FEATURES
+
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -26,141 +35,94 @@ extern "C" {
 
 // version numbers
 #define ENC_MAJ_VERSION 0
-#define ENC_MIN_VERSION 2
-#define ENC_REV_VERSION 0
-
-// size of histogram used by CollectHistogram.
-#define MAX_COEFF_THRESH   64
-
-// 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
-     };
+#define ENC_MIN_VERSION 4
+#define ENC_REV_VERSION 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
+enum { MAX_LF_LEVELS = 64,       // Maximum loop filter level
+       MAX_VARIABLE_LEVEL = 67,  // last (inclusive) level with variable cost
+       MAX_LEVEL = 2047          // max level (note: max codable is 2047 + 67)
      };
 
-// YUV-cache parameters. Cache is 16-pixels wide.
-// The original or reconstructed samples can be accessed using VP8Scan[]
+typedef enum {   // Rate-distortion optimization levels
+  RD_OPT_NONE        = 0,  // no rd-opt
+  RD_OPT_BASIC       = 1,  // basic scoring (no trellis)
+  RD_OPT_TRELLIS     = 2,  // perform trellis-quant on the final decision only
+  RD_OPT_TRELLIS_ALL = 3   // trellis-quant for every scoring (much slower)
+} VP8RDLevel;
+
+// YUV-cache parameters. Cache is 32-bytes wide (= one cacheline).
+// The original or reconstructed samples can be accessed using VP8Scan[].
 // The predicted blocks can be accessed using offsets to yuv_p_ and
-// the arrays VP8*ModeOffsets[];
-//         +----+      YUV Samples area. See VP8Scan[] for accessing the blocks.
-//  Y_OFF  |YYYY| <- original samples  (enc->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
+// the arrays VP8*ModeOffsets[].
+// * YUV Samples area (yuv_in_/yuv_out_/yuv_out2_)
+//   (see VP8Scan[] for accessing the blocks, along with
+//   Y_OFF_ENC/U_OFF_ENC/V_OFF_ENC):
+//             +----+----+
+//  Y_OFF_ENC  |YYYY|UUVV|
+//  U_OFF_ENC  |YYYY|UUVV|
+//  V_OFF_ENC  |YYYY|....| <- 25% wasted U/V area
+//             |YYYY|....|
+//             +----+----+
+// * Prediction area ('yuv_p_', size = PRED_SIZE_ENC)
+//   Intra16 predictions (16x16 block each, two per row):
+//         |I16DC16|I16TM16|
+//         |I16VE16|I16HE16|
+//   Chroma U/V predictions (16x8 block each, two per row):
+//         |C8DC8|C8TM8|
+//         |C8VE8|C8HE8|
+//   Intra 4x4 predictions (4x4 block each)
+//         |I4DC4 I4TM4 I4VE4 I4HE4|I4RD4 I4VR4 I4LD4 I4VL4|
+//         |I4HD4 I4HU4 I4TMP .....|.......................| <- ~31% wasted
+#define YUV_SIZE_ENC (BPS * 16)
+#define PRED_SIZE_ENC (32 * BPS + 16 * BPS + 8 * BPS)   // I16+Chroma+I4 preds
+#define Y_OFF_ENC    (0)
+#define U_OFF_ENC    (16)
+#define V_OFF_ENC    (16 + 8)
+
+extern const int VP8Scan[16];           // 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)
+#define I16TM16 (I16DC16 + 16)
+#define I16VE16 (1 * 16 * BPS)
+#define I16HE16 (I16VE16 + 16)
 // chroma 8x8, two U/V blocks side by side (hence: 16x8 each)
-#define C8DC8 (4 * 16 * BPS)
-#define C8TM8 (4 * 16 * BPS + 8 * BPS)
-#define C8VE8 (5 * 16 * BPS)
-#define C8HE8 (5 * 16 * BPS + 8 * BPS)
+#define C8DC8 (2 * 16 * BPS)
+#define C8TM8 (C8DC8 + 1 * 16)
+#define C8VE8 (2 * 16 * BPS + 8 * BPS)
+#define C8HE8 (C8VE8 + 1 * 16)
 // intra 4x4
-#define I4DC4 (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)
+#define I4DC4 (3 * 16 * BPS +  0)
+#define I4TM4 (I4DC4 +  4)
+#define I4VE4 (I4DC4 +  8)
+#define I4HE4 (I4DC4 + 12)
+#define I4RD4 (I4DC4 + 16)
+#define I4VR4 (I4DC4 + 20)
+#define I4LD4 (I4DC4 + 24)
+#define I4VL4 (I4DC4 + 28)
+#define I4HD4 (3 * 16 * BPS + 4 * BPS)
+#define I4HU4 (I4HD4 + 4)
+#define I4TMP (I4HD4 + 8)
 
 typedef int64_t score_t;     // type used for scores, rate, distortion
+// Note that MAX_COST is not the maximum allowed by sizeof(score_t),
+// in order to allow overflowing computations.
 #define MAX_COST ((score_t)0x7fffffffffffffLL)
 
 #define QFIX 17
 #define BIAS(b)  ((b) << (QFIX - 8))
 // Fun fact: this is the _only_ line where we're actually being lossy and
 // discarding bits.
-static WEBP_INLINE int QUANTDIV(int n, int iQ, int B) {
-  return (n * iQ + B) >> QFIX;
+static WEBP_INLINE int QUANTDIV(uint32_t n, uint32_t iQ, uint32_t B) {
+  return (int)((n * iQ + B) >> QFIX);
 }
-extern const uint8_t VP8Zigzag[16];
+
+// Uncomment the following to remove token-buffer code:
+// #define DISABLE_TOKEN_BUFFER
 
 //------------------------------------------------------------------------------
 // Headers
@@ -169,6 +131,8 @@ typedef uint32_t proba_t;   // 16b + 16b
 typedef uint8_t ProbaArray[NUM_CTX][NUM_PROBAS];
 typedef proba_t StatsArray[NUM_CTX][NUM_PROBAS];
 typedef uint16_t CostArray[NUM_CTX][MAX_VARIABLE_LEVEL + 1];
+typedef const uint16_t* (*CostArrayPtr)[NUM_CTX];   // for easy casting
+typedef const uint16_t* CostArrayMap[16][NUM_CTX];
 typedef double LFStats[NUM_MB_SEGMENTS][MAX_LF_LEVELS];  // filter stats
 
 typedef struct VP8Encoder VP8Encoder;
@@ -179,19 +143,20 @@ typedef struct {
   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;
+} VP8EncSegmentHeader;
 
 // Struct collecting all frame-persistent probabilities.
 typedef struct {
   uint8_t segments_[3];     // probabilities for segment tree
   uint8_t skip_proba_;      // final probability of being skipped.
-  ProbaArray coeffs_[NUM_TYPES][NUM_BANDS];      // 924 bytes
+  ProbaArray coeffs_[NUM_TYPES][NUM_BANDS];      // 1056 bytes
   StatsArray stats_[NUM_TYPES][NUM_BANDS];       // 4224 bytes
-  CostArray level_cost_[NUM_TYPES][NUM_BANDS];   // 11.4k
+  CostArray level_cost_[NUM_TYPES][NUM_BANDS];   // 13056 bytes
+  CostArrayMap remapped_costs_[NUM_TYPES];       // 1536 bytes
   int dirty_;               // if true, need to call VP8CalculateLevelCosts()
   int use_skip_proba_;      // Note: we always use skip_proba for now.
   int nb_skip_;             // number of skipped blocks
-} VP8Proba;
+} VP8EncProba;
 
 // Filter parameters. Not actually used in the code (we don't perform
 // the in-loop filtering), but filled from user's config
@@ -200,7 +165,7 @@ typedef struct {
   int level_;              // base filter level [0..63]
   int sharpness_;          // [0..7]
   int i4x4_lf_delta_;      // delta filter level for i4x4 relative to i16x16
-} VP8FilterHeader;
+} VP8EncFilterHeader;
 
 //------------------------------------------------------------------------------
 // Informations about the macroblocks.
@@ -217,8 +182,8 @@ typedef struct {
 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
+  uint32_t bias_[16];     // rounding bias
+  uint32_t zthresh_[16];  // value below which a coefficient is zeroed
   uint16_t sharpen_[16];  // frequency boosters for slight sharpening
 } VP8Matrix;
 
@@ -229,16 +194,19 @@ typedef struct {
   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 transcient struct to accumulate score and info during RD-optimization
+// Handy transient struct to accumulate score and info during RD-optimization
 // and mode evaluation.
 typedef struct {
-  score_t D, SD, R, score;    // Distortion, spectral distortion, rate, score.
+  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];
@@ -252,12 +220,11 @@ typedef struct {
 // right neighbouring data (samples, predictions, contexts, ...)
 typedef struct {
   int x_, y_;                      // current macroblock
-  int y_offset_, uv_offset_;       // offset to the luma / chroma planes
   int y_stride_, uv_stride_;       // respective strides
-  uint8_t*      yuv_in_;           // borrowed from enc_ (for now)
-  uint8_t*      yuv_out_;          // ''
-  uint8_t*      yuv_out2_;         // ''
-  uint8_t*      yuv_p_;            // ''
+  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
@@ -273,24 +240,44 @@ typedef struct {
   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           done_;             // true when scan is finished
+  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_
+  uint8_t yuv_left_mem_[17 + 16 + 16 + 8 + WEBP_ALIGN_CST];
+  // memory for yuv_*
+  uint8_t yuv_mem_[3 * YUV_SIZE_ENC + PRED_SIZE_ENC + WEBP_ALIGN_CST];
 } VP8EncIterator;
 
   // in iterator.c
-// must be called first.
+// must be called first
 void VP8IteratorInit(VP8Encoder* const enc, VP8EncIterator* const it);
-// restart a scan.
+// restart a scan
 void VP8IteratorReset(VP8EncIterator* const it);
-// import samples from source
-void VP8IteratorImport(const 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 !done_. If *block_to_save is non-null, will
-// save the boundary values to top_/left_ arrays. block_to_save can be
-// it->yuv_out_ or it->yuv_in_.
-int VP8IteratorNext(VP8EncIterator* const it,
-                    const uint8_t* const block_to_save);
+// 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);
@@ -314,44 +301,43 @@ void VP8SetSegment(const VP8EncIterator* const it, int segment);
 //------------------------------------------------------------------------------
 // Paginated token buffer
 
-// WIP: #define USE_TOKEN_BUFFER
+typedef struct VP8Tokens VP8Tokens;  // struct details in token.c
 
-#ifdef USE_TOKEN_BUFFER
+typedef struct {
+#if !defined(DISABLE_TOKEN_BUFFER)
+  VP8Tokens* pages_;        // first page
+  VP8Tokens** last_page_;   // last page
+  uint16_t* tokens_;        // set to (*last_page_)->tokens_
+  int left_;                // how many free tokens left before the page is full
+  int page_size_;           // number of tokens per page
+#endif
+  int error_;         // true in case of malloc error
+} VP8TBuffer;
 
-#define MAX_NUM_TOKEN 2048
+// initialize an empty buffer
+void VP8TBufferInit(VP8TBuffer* const b, int page_size);
+void VP8TBufferClear(VP8TBuffer* const b);   // de-allocate pages memory
 
-typedef struct VP8Tokens VP8Tokens;
-struct VP8Tokens {
-  uint16_t tokens_[MAX_NUM_TOKEN];  // bit#15: bit, bits 0..14: slot
-  int left_;
-  VP8Tokens* next_;
-};
+#if !defined(DISABLE_TOKEN_BUFFER)
 
-typedef struct {
-  VP8Tokens* rows_;
-  uint16_t* tokens_;    // set to (*last_)->tokens_
-  VP8Tokens** last_;
-  int left_;
-  int error_;  // true in case of malloc error
-} VP8TBuffer;
+// 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);
 
-void VP8TBufferInit(VP8TBuffer* const b);    // initialize an empty buffer
-int VP8TBufferNewPage(VP8TBuffer* const b);  // allocate a new page
-void VP8TBufferClear(VP8TBuffer* const b);   // de-allocate memory
+// record the coding of coefficients without knowing the probabilities yet
+int VP8RecordCoeffTokens(const int ctx, const int coeff_type,
+                         int first, int last,
+                         const int16_t* const coeffs,
+                         VP8TBuffer* const tokens);
 
-int VP8EmitTokens(const VP8TBuffer* const b, VP8BitWriter* const bw,
-                  const uint8_t* const probas);
+// Estimate the final coded size given a set of 'probas'.
+size_t VP8EstimateTokenSize(VP8TBuffer* const b, const uint8_t* const probas);
 
-static WEBP_INLINE int VP8AddToken(VP8TBuffer* const b,
-                                   int bit, int proba_idx) {
-  if (b->left_ > 0 || VP8TBufferNewPage(b)) {
-    const int slot = --b->left_;
-    b->tokens_[slot] = (bit << 15) | proba_idx;
-  }
-  return bit;
-}
+// unused for now
+void VP8TokenToStats(const VP8TBuffer* const b, proba_t* const stats);
 
-#endif  // USE_TOKEN_BUFFER
+#endif  // !DISABLE_TOKEN_BUFFER
 
 //------------------------------------------------------------------------------
 // VP8Encoder
@@ -361,8 +347,8 @@ struct VP8Encoder {
   WebPPicture* pic_;            // input / output picture
 
   // headers
-  VP8FilterHeader   filter_hdr_;     // filtering information
-  VP8SegmentHeader  segment_hdr_;    // segment information
+  VP8EncFilterHeader   filter_hdr_;     // filtering information
+  VP8EncSegmentHeader  segment_hdr_;    // segment information
 
   int profile_;                      // VP8's profile, deduced from Config.
 
@@ -376,6 +362,7 @@ struct VP8Encoder {
   // per-partition boolean decoders.
   VP8BitWriter bw_;                         // part0
   VP8BitWriter parts_[MAX_NUM_PARTITIONS];  // token partitions
+  VP8TBuffer tokens_;                       // token buffer
 
   int percent_;                             // for progress
 
@@ -383,17 +370,13 @@ struct VP8Encoder {
   int has_alpha_;
   uint8_t* alpha_data_;       // non-NULL if transparency is present
   uint32_t alpha_data_size_;
-
-  // enhancement layer
-  int use_layer_;
-  VP8BitWriter layer_bw_;
-  uint8_t* layer_data_;
-  size_t layer_data_size_;
+  WebPWorker alpha_worker_;
 
   // quantization info (one set of DC/AC dequant factor per segment)
   VP8SegmentInfo dqm_[NUM_MB_SEGMENTS];
   int base_quant_;                 // nominal quantizer value. Only used
                                    // for relative coding of segments' quant.
+  int alpha_;                      // global susceptibility (<=> complexity)
   int uv_alpha_;                   // U/V quantization susceptibility
   // global offset of quantizers, shared by all segments
   int dq_y1_dc_;
@@ -401,34 +384,29 @@ struct VP8Encoder {
   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];
+  VP8EncProba proba_;
+  uint64_t    sse_[4];      // sum of Y/U/V/A squared errors for all macroblocks
+  uint64_t    sse_count_;   // pixel count for the sse_[] stats
+  int         coded_size_;
+  int         residual_bytes_[3][4];
+  int         block_count_[3];
 
   // quality/speed settings
-  int method_;              // 0=fastest, 6=best/slowest.
-  int rd_opt_level_;        // Deduced from method_.
-  int max_i4_header_bits_;  // partition #0 safeness factor
+  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*   yuv_in_;    // input samples
-  uint8_t*   yuv_out_;   // output samples
-  uint8_t*   yuv_out2_;  // secondary scratch out-buffer. swapped with yuv_out_.
-  uint8_t*   yuv_p_;     // scratch buffer for prediction
-  uint8_t   *y_top_;     // top luma samples.
-  uint8_t   *uv_top_;    // top u/v samples.
-                         // U and V are packed into 16 pixels (8 U + 8 V)
-  uint8_t   *y_left_;    // left luma samples (adressable from index -1 to 15).
-  uint8_t   *u_left_;    // left u samples (adressable from index -1 to 7)
-  uint8_t   *v_left_;    // left v samples (adressable from index -1 to 7)
-
-  LFStats   *lf_stats_;  // autofilter stats (if NULL, autofilter is off)
+  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)
 };
 
 //------------------------------------------------------------------------------
@@ -441,7 +419,7 @@ extern const uint8_t
 // 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);
+void VP8WriteProbas(VP8BitWriter* const bw, const VP8EncProba* const probas);
 // Writes the partition #0 modes (that is: all intra modes)
 void VP8CodeIntraModes(VP8Encoder* const enc);
 
@@ -454,7 +432,11 @@ int VP8EncWrite(VP8Encoder* const enc);
 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
@@ -466,9 +448,9 @@ void VP8MakeIntra4Preds(const VP8EncIterator* const it);
 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 stat / coding passes
+// Main coding calls
 int VP8EncLoop(VP8Encoder* const enc);
-int VP8StatLoop(VP8Encoder* const enc);
+int VP8EncTokenLoop(VP8Encoder* const enc);
 
   // in webpenc.c
 // Assign an error code to a picture. Return false for convenience.
@@ -485,18 +467,14 @@ int VP8EncAnalyze(VP8Encoder* const enc);
 // 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, int rd_opt);
+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
-void 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
+int VP8EncDeleteAlpha(VP8Encoder* const enc);   // delete compressed data
 
   // in filter.c
 
@@ -516,9 +494,38 @@ void VP8InitFilter(VP8EncIterator* const it);
 void VP8StoreFilterStats(VP8EncIterator* const it);
 void VP8AdjustFilterStrength(VP8EncIterator* const it);
 
+// returns the approximate filtering strength needed to smooth a edge
+// step of 'delta', given a sharpness parameter 'sharpness'.
+int VP8FilterStrengthFromDelta(int sharpness, int delta);
+
+  // misc utils for picture_*.c:
+
+// Remove reference to the ARGB/YUVA buffer (doesn't free anything).
+void WebPPictureResetBuffers(WebPPicture* const picture);
+
+// Allocates ARGB buffer of given dimension (previous one is always free'd).
+// Preserves the YUV(A) buffer. Returns false in case of error (invalid param,
+// out-of-memory).
+int WebPPictureAllocARGB(WebPPicture* const picture, int width, int height);
+
+// Allocates YUVA buffer of given dimension (previous one is always free'd).
+// Uses picture->csp to determine whether an alpha buffer is needed.
+// Preserves the ARGB buffer.
+// Returns false in case of error (invalid param, out-of-memory).
+int WebPPictureAllocYUVA(WebPPicture* const picture, int width, int height);
+
+  // in near_lossless.c
+// Near lossless preprocessing in RGB color-space.
+int VP8ApplyNearLossless(int xsize, int ysize, uint32_t* argb, int quality);
+// Near lossless adjustment for predictors.
+void VP8ApplyNearLosslessPredict(int xsize, int ysize, int pred_bits,
+                                 const uint32_t* argb_orig,
+                                 uint32_t* argb, uint32_t* argb_scratch,
+                                 const uint32_t* const transform_data,
+                                 int quality, int subtract_green);
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/enc/vp8l.c b/drivers/webp/enc/vp8l.c
index f4eb6e783f..047c9032ac 100644
--- a/drivers/webp/enc/vp8l.c
+++ b/drivers/webp/enc/vp8l.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
@@ -11,7 +13,6 @@
 //
 
 #include <assert.h>
-#include <stdio.h>
 #include <stdlib.h>
 
 #include "./backward_references.h"
@@ -21,28 +22,107 @@
 #include "../utils/bit_writer.h"
 #include "../utils/huffman_encode.h"
 #include "../utils/utils.h"
-#include "../format_constants.h"
+#include "../webp/format_constants.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "./delta_palettization.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
+// Maximum number of histogram images (sub-blocks).
+#define MAX_HUFF_IMAGE_SIZE       2600
 
-// -----------------------------------------------------------------------------
-// Palette
+// Palette reordering for smaller sum of deltas (and for smaller storage).
 
-static int CompareColors(const void* p1, const void* p2) {
+static int PaletteCompareColorsForQsort(const void* p1, const void* p2) {
   const uint32_t a = *(const uint32_t*)p1;
   const uint32_t b = *(const uint32_t*)p2;
-  return (a < b) ? -1 : (a > b) ? 1 : 0;
+  assert(a != b);
+  return (a < b) ? -1 : 1;
+}
+
+static WEBP_INLINE uint32_t PaletteComponentDistance(uint32_t v) {
+  return (v <= 128) ? v : (256 - v);
+}
+
+// Computes a value that is related to the entropy created by the
+// palette entry diff.
+//
+// Note that the last & 0xff is a no-operation in the next statement, but
+// removed by most compilers and is here only for regularity of the code.
+static WEBP_INLINE uint32_t PaletteColorDistance(uint32_t col1, uint32_t col2) {
+  const uint32_t diff = VP8LSubPixels(col1, col2);
+  const int kMoreWeightForRGBThanForAlpha = 9;
+  uint32_t score;
+  score =  PaletteComponentDistance((diff >>  0) & 0xff);
+  score += PaletteComponentDistance((diff >>  8) & 0xff);
+  score += PaletteComponentDistance((diff >> 16) & 0xff);
+  score *= kMoreWeightForRGBThanForAlpha;
+  score += PaletteComponentDistance((diff >> 24) & 0xff);
+  return score;
+}
+
+static WEBP_INLINE void SwapColor(uint32_t* const col1, uint32_t* const col2) {
+  const uint32_t tmp = *col1;
+  *col1 = *col2;
+  *col2 = tmp;
+}
+
+static void GreedyMinimizeDeltas(uint32_t palette[], int num_colors) {
+  // Find greedily always the closest color of the predicted color to minimize
+  // deltas in the palette. This reduces storage needs since the
+  // palette is stored with delta encoding.
+  uint32_t predict = 0x00000000;
+  int i, k;
+  for (i = 0; i < num_colors; ++i) {
+    int best_ix = i;
+    uint32_t best_score = ~0U;
+    for (k = i; k < num_colors; ++k) {
+      const uint32_t cur_score = PaletteColorDistance(palette[k], predict);
+      if (best_score > cur_score) {
+        best_score = cur_score;
+        best_ix = k;
+      }
+    }
+    SwapColor(&palette[best_ix], &palette[i]);
+    predict = palette[i];
+  }
+}
+
+// The palette has been sorted by alpha. This function checks if the other
+// components of the palette have a monotonic development with regards to
+// position in the palette. If all have monotonic development, there is
+// no benefit to re-organize them greedily. A monotonic development
+// would be spotted in green-only situations (like lossy alpha) or gray-scale
+// images.
+static int PaletteHasNonMonotonousDeltas(uint32_t palette[], int num_colors) {
+  uint32_t predict = 0x000000;
+  int i;
+  uint8_t sign_found = 0x00;
+  for (i = 0; i < num_colors; ++i) {
+    const uint32_t diff = VP8LSubPixels(palette[i], predict);
+    const uint8_t rd = (diff >> 16) & 0xff;
+    const uint8_t gd = (diff >>  8) & 0xff;
+    const uint8_t bd = (diff >>  0) & 0xff;
+    if (rd != 0x00) {
+      sign_found |= (rd < 0x80) ? 1 : 2;
+    }
+    if (gd != 0x00) {
+      sign_found |= (gd < 0x80) ? 8 : 16;
+    }
+    if (bd != 0x00) {
+      sign_found |= (bd < 0x80) ? 64 : 128;
+    }
+    predict = palette[i];
+  }
+  return (sign_found & (sign_found << 1)) != 0;  // two consequent signs.
 }
 
+// -----------------------------------------------------------------------------
+// Palette
+
 // 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,
+                                   int low_effort,
                                    uint32_t palette[MAX_PALETTE_SIZE],
                                    int* const palette_size) {
   int i, x, y, key;
@@ -85,7 +165,7 @@ static int AnalyzeAndCreatePalette(const WebPPicture* const pic,
     argb += pic->argb_stride;
   }
 
-  // TODO(skal): could we reuse in_use[] to speed up ApplyPalette()?
+  // 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]) {
@@ -93,106 +173,272 @@ static int AnalyzeAndCreatePalette(const WebPPicture* const pic,
       ++num_colors;
     }
   }
-
-  qsort(palette, num_colors, sizeof(*palette), CompareColors);
   *palette_size = num_colors;
+  qsort(palette, num_colors, sizeof(*palette), PaletteCompareColorsForQsort);
+  if (!low_effort && PaletteHasNonMonotonousDeltas(palette, num_colors)) {
+    GreedyMinimizeDeltas(palette, num_colors);
+  }
   return 1;
 }
 
+// These five modes are evaluated and their respective entropy is computed.
+typedef enum {
+  kDirect = 0,
+  kSpatial = 1,
+  kSubGreen = 2,
+  kSpatialSubGreen = 3,
+  kPalette = 4,
+  kNumEntropyIx = 5
+} EntropyIx;
+
+typedef enum {
+  kHistoAlpha = 0,
+  kHistoAlphaPred,
+  kHistoGreen,
+  kHistoGreenPred,
+  kHistoRed,
+  kHistoRedPred,
+  kHistoBlue,
+  kHistoBluePred,
+  kHistoRedSubGreen,
+  kHistoRedPredSubGreen,
+  kHistoBlueSubGreen,
+  kHistoBluePredSubGreen,
+  kHistoPalette,
+  kHistoTotal  // Must be last.
+} HistoIx;
+
+static void AddSingleSubGreen(uint32_t p, uint32_t* r, uint32_t* b) {
+  const uint32_t green = p >> 8;  // The upper bits are masked away later.
+  ++r[((p >> 16) - green) & 0xff];
+  ++b[(p - green) & 0xff];
+}
+
+static void AddSingle(uint32_t p,
+                      uint32_t* a, uint32_t* r, uint32_t* g, uint32_t* b) {
+  ++a[p >> 24];
+  ++r[(p >> 16) & 0xff];
+  ++g[(p >> 8) & 0xff];
+  ++b[(p & 0xff)];
+}
+
 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;
+                          int use_palette,
+                          EntropyIx* const min_entropy_ix,
+                          int* const red_and_blue_always_zero) {
+  // Allocate histogram set with cache_bits = 0.
+  uint32_t* const histo =
+      (uint32_t*)WebPSafeCalloc(kHistoTotal, sizeof(*histo) * 256);
+  if (histo != NULL) {
+    int i, x, y;
+    const uint32_t* prev_row = argb;
+    const uint32_t* curr_row = argb + argb_stride;
+    for (y = 1; y < height; ++y) {
+      uint32_t prev_pix = curr_row[0];
+      for (x = 1; x < width; ++x) {
+        const uint32_t pix = curr_row[x];
+        const uint32_t pix_diff = VP8LSubPixels(pix, prev_pix);
+        if ((pix_diff == 0) || (pix == prev_row[x])) continue;
+        prev_pix = pix;
+        AddSingle(pix,
+                  &histo[kHistoAlpha * 256],
+                  &histo[kHistoRed * 256],
+                  &histo[kHistoGreen * 256],
+                  &histo[kHistoBlue * 256]);
+        AddSingle(pix_diff,
+                  &histo[kHistoAlphaPred * 256],
+                  &histo[kHistoRedPred * 256],
+                  &histo[kHistoGreenPred * 256],
+                  &histo[kHistoBluePred * 256]);
+        AddSingleSubGreen(pix,
+                          &histo[kHistoRedSubGreen * 256],
+                          &histo[kHistoBlueSubGreen * 256]);
+        AddSingleSubGreen(pix_diff,
+                          &histo[kHistoRedPredSubGreen * 256],
+                          &histo[kHistoBluePredSubGreen * 256]);
+        {
+          // Approximate the palette by the entropy of the multiplicative hash.
+          const int hash = ((pix + (pix >> 19)) * 0x39c5fba7) >> 24;
+          ++histo[kHistoPalette * 256 + (hash & 0xff)];
+        }
+      }
+      prev_row = curr_row;
+      curr_row += argb_stride;
+    }
+    {
+      double entropy_comp[kHistoTotal];
+      double entropy[kNumEntropyIx];
+      EntropyIx k;
+      EntropyIx last_mode_to_analyze =
+          use_palette ? kPalette : kSpatialSubGreen;
+      int j;
+      // Let's add one zero to the predicted histograms. The zeros are removed
+      // too efficiently by the pix_diff == 0 comparison, at least one of the
+      // zeros is likely to exist.
+      ++histo[kHistoRedPredSubGreen * 256];
+      ++histo[kHistoBluePredSubGreen * 256];
+      ++histo[kHistoRedPred * 256];
+      ++histo[kHistoGreenPred * 256];
+      ++histo[kHistoBluePred * 256];
+      ++histo[kHistoAlphaPred * 256];
+
+      for (j = 0; j < kHistoTotal; ++j) {
+        entropy_comp[j] = VP8LBitsEntropy(&histo[j * 256], 256, NULL);
+      }
+      entropy[kDirect] = entropy_comp[kHistoAlpha] +
+          entropy_comp[kHistoRed] +
+          entropy_comp[kHistoGreen] +
+          entropy_comp[kHistoBlue];
+      entropy[kSpatial] = entropy_comp[kHistoAlphaPred] +
+          entropy_comp[kHistoRedPred] +
+          entropy_comp[kHistoGreenPred] +
+          entropy_comp[kHistoBluePred];
+      entropy[kSubGreen] = entropy_comp[kHistoAlpha] +
+          entropy_comp[kHistoRedSubGreen] +
+          entropy_comp[kHistoGreen] +
+          entropy_comp[kHistoBlueSubGreen];
+      entropy[kSpatialSubGreen] = entropy_comp[kHistoAlphaPred] +
+          entropy_comp[kHistoRedPredSubGreen] +
+          entropy_comp[kHistoGreenPred] +
+          entropy_comp[kHistoBluePredSubGreen];
+      // Palette mode seems more efficient in a breakeven case. Bias with 1.0.
+      entropy[kPalette] = entropy_comp[kHistoPalette] - 1.0;
+
+      *min_entropy_ix = kDirect;
+      for (k = kDirect + 1; k <= last_mode_to_analyze; ++k) {
+        if (entropy[*min_entropy_ix] > entropy[k]) {
+          *min_entropy_ix = k;
+        }
       }
-      last_pix = pix;
+      *red_and_blue_always_zero = 1;
+      // Let's check if the histogram of the chosen entropy mode has
+      // non-zero red and blue values. If all are zero, we can later skip
+      // the cross color optimization.
       {
-        const PixOrCopy pix_token = PixOrCopyCreateLiteral(pix);
-        const PixOrCopy pix_diff_token = PixOrCopyCreateLiteral(pix_diff);
-        VP8LHistogramAddSinglePixOrCopy(nonpredicted, &pix_token);
-        VP8LHistogramAddSinglePixOrCopy(predicted, &pix_diff_token);
+        static const uint8_t kHistoPairs[5][2] = {
+          { kHistoRed, kHistoBlue },
+          { kHistoRedPred, kHistoBluePred },
+          { kHistoRedSubGreen, kHistoBlueSubGreen },
+          { kHistoRedPredSubGreen, kHistoBluePredSubGreen },
+          { kHistoRed, kHistoBlue }
+        };
+        const uint32_t* const red_histo =
+            &histo[256 * kHistoPairs[*min_entropy_ix][0]];
+        const uint32_t* const blue_histo =
+            &histo[256 * kHistoPairs[*min_entropy_ix][1]];
+        for (i = 1; i < 256; ++i) {
+          if ((red_histo[i] | blue_histo[i]) != 0) {
+            *red_and_blue_always_zero = 0;
+            break;
+          }
+        }
       }
     }
-    last_line = argb;
-    argb += argb_stride;
+    free(histo);
+    return 1;
+  } else {
+    return 0;
   }
-  *nonpredicted_bits = VP8LHistogramEstimateBitsBulk(nonpredicted);
-  *predicted_bits = VP8LHistogramEstimateBitsBulk(predicted);
-  free(predicted);
-  return 1;
 }
 
-static int VP8LEncAnalyze(VP8LEncoder* const enc, WebPImageHint image_hint) {
+static int GetHistoBits(int method, int use_palette, int width, int height) {
+  // Make tile size a function of encoding method (Range: 0 to 6).
+  int histo_bits = (use_palette ? 9 : 7) - method;
+  while (1) {
+    const int huff_image_size = VP8LSubSampleSize(width, histo_bits) *
+                                VP8LSubSampleSize(height, histo_bits);
+    if (huff_image_size <= MAX_HUFF_IMAGE_SIZE) break;
+    ++histo_bits;
+  }
+  return (histo_bits < MIN_HUFFMAN_BITS) ? MIN_HUFFMAN_BITS :
+         (histo_bits > MAX_HUFFMAN_BITS) ? MAX_HUFFMAN_BITS : histo_bits;
+}
+
+static int GetTransformBits(int method, int histo_bits) {
+  const int max_transform_bits = (method < 4) ? 6 : (method > 4) ? 4 : 5;
+  return (histo_bits > max_transform_bits) ? max_transform_bits : histo_bits;
+}
+
+static int AnalyzeAndInit(VP8LEncoder* const enc) {
   const WebPPicture* const pic = enc->pic_;
+  const int width = pic->width;
+  const int height = pic->height;
+  const int pix_cnt = width * height;
+  const WebPConfig* const config = enc->config_;
+  const int method = config->method;
+  const int low_effort = (config->method == 0);
+  // we round the block size up, so we're guaranteed to have
+  // at max MAX_REFS_BLOCK_PER_IMAGE blocks used:
+  int refs_block_size = (pix_cnt - 1) / MAX_REFS_BLOCK_PER_IMAGE + 1;
   assert(pic != NULL && pic->argb != NULL);
 
+  enc->use_cross_color_ = 0;
+  enc->use_predict_ = 0;
+  enc->use_subtract_green_ = 0;
   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;
+      AnalyzeAndCreatePalette(pic, low_effort,
+                              enc->palette_, &enc->palette_size_);
+
+  // TODO(jyrki): replace the decision to be based on an actual estimate
+  // of entropy, or even spatial variance of entropy.
+  enc->histo_bits_ = GetHistoBits(method, enc->use_palette_,
+                                  pic->width, pic->height);
+  enc->transform_bits_ = GetTransformBits(method, enc->histo_bits_);
+
+  if (low_effort) {
+    // AnalyzeEntropy is somewhat slow.
+    enc->use_predict_ = !enc->use_palette_;
+    enc->use_subtract_green_ = !enc->use_palette_;
+    enc->use_cross_color_ = 0;
+  } else {
+    int red_and_blue_always_zero;
+    EntropyIx min_entropy_ix;
+    if (!AnalyzeEntropy(pic->argb, width, height, pic->argb_stride,
+                        enc->use_palette_, &min_entropy_ix,
+                        &red_and_blue_always_zero)) {
+      return 0;
     }
+    enc->use_palette_ = (min_entropy_ix == kPalette);
+    enc->use_subtract_green_ =
+        (min_entropy_ix == kSubGreen) || (min_entropy_ix == kSpatialSubGreen);
+    enc->use_predict_ =
+        (min_entropy_ix == kSpatial) || (min_entropy_ix == kSpatialSubGreen);
+    enc->use_cross_color_ = red_and_blue_always_zero ? 0 : enc->use_predict_;
   }
 
-  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;
-        // TODO(vikasa): Observed some correlation of cross_color transform with
-        // predict. Need to investigate this further and add separate heuristic
-        // for setting use_cross_color flag.
-        enc->use_cross_color_ = 1;
-      }
-    }
-  }
+  if (!VP8LHashChainInit(&enc->hash_chain_, pix_cnt)) return 0;
+
+  // palette-friendly input typically uses less literals
+  //  -> reduce block size a bit
+  if (enc->use_palette_) refs_block_size /= 2;
+  VP8LBackwardRefsInit(&enc->refs_[0], refs_block_size);
+  VP8LBackwardRefsInit(&enc->refs_[1], refs_block_size);
 
   return 1;
 }
 
+// Returns false in case of memory error.
 static int GetHuffBitLengthsAndCodes(
     const VP8LHistogramSet* const histogram_image,
     HuffmanTreeCode* const huffman_codes) {
   int i, k;
-  int ok = 1;
+  int ok = 0;
   uint64_t total_length_size = 0;
   uint8_t* mem_buf = NULL;
   const int histogram_image_size = histogram_image->size;
+  int max_num_symbols = 0;
+  uint8_t* buf_rle = NULL;
+  HuffmanTree* huff_tree = NULL;
 
   // Iterate over all histograms and get the aggregate number of codes used.
   for (i = 0; i < histogram_image_size; ++i) {
     const VP8LHistogram* const histo = histogram_image->histograms[i];
     HuffmanTreeCode* const codes = &huffman_codes[5 * i];
     for (k = 0; k < 5; ++k) {
-      const int num_symbols = (k == 0) ? VP8LHistogramNumCodes(histo)
-                            : (k == 4) ? NUM_DISTANCE_CODES
-                            : 256;
+      const int num_symbols =
+          (k == 0) ? VP8LHistogramNumCodes(histo->palette_code_bits_) :
+          (k == 4) ? NUM_DISTANCE_CODES : 256;
       codes[k].num_symbols = num_symbols;
       total_length_size += num_symbols;
     }
@@ -204,10 +450,8 @@ static int GetHuffBitLengthsAndCodes(
     uint8_t* lengths;
     mem_buf = (uint8_t*)WebPSafeCalloc(total_length_size,
                                        sizeof(*lengths) + sizeof(*codes));
-    if (mem_buf == NULL) {
-      ok = 0;
-      goto End;
-    }
+    if (mem_buf == NULL) goto End;
+
     codes = (uint16_t*)mem_buf;
     lengths = (uint8_t*)&codes[total_length_size];
     for (i = 0; i < 5 * histogram_image_size; ++i) {
@@ -216,22 +460,35 @@ static int GetHuffBitLengthsAndCodes(
       huffman_codes[i].code_lengths = lengths;
       codes += bit_length;
       lengths += bit_length;
+      if (max_num_symbols < bit_length) {
+        max_num_symbols = bit_length;
+      }
     }
   }
 
+  buf_rle = (uint8_t*)WebPSafeMalloc(1ULL, max_num_symbols);
+  huff_tree = (HuffmanTree*)WebPSafeMalloc(3ULL * max_num_symbols,
+                                           sizeof(*huff_tree));
+  if (buf_rle == NULL || huff_tree == NULL) goto End;
+
   // Create Huffman trees.
   for (i = 0; i < histogram_image_size; ++i) {
     HuffmanTreeCode* const codes = &huffman_codes[5 * i];
     VP8LHistogram* const histo = histogram_image->histograms[i];
-    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);
+    VP8LCreateHuffmanTree(histo->literal_, 15, buf_rle, huff_tree, codes + 0);
+    VP8LCreateHuffmanTree(histo->red_, 15, buf_rle, huff_tree, codes + 1);
+    VP8LCreateHuffmanTree(histo->blue_, 15, buf_rle, huff_tree, codes + 2);
+    VP8LCreateHuffmanTree(histo->alpha_, 15, buf_rle, huff_tree, codes + 3);
+    VP8LCreateHuffmanTree(histo->distance_, 15, buf_rle, huff_tree, codes + 4);
   }
-
+  ok = 1;
  End:
-  if (!ok) free(mem_buf);
+  WebPSafeFree(huff_tree);
+  WebPSafeFree(buf_rle);
+  if (!ok) {
+    WebPSafeFree(mem_buf);
+    memset(huffman_codes, 0, 5 * histogram_image_size * sizeof(*huffman_codes));
+  }
   return ok;
 }
 
@@ -251,9 +508,9 @@ static void StoreHuffmanTreeOfHuffmanTreeToBitMask(
       break;
     }
   }
-  VP8LWriteBits(bw, 4, codes_to_store - 4);
+  VP8LPutBits(bw, codes_to_store - 4, 4);
   for (i = 0; i < codes_to_store; ++i) {
-    VP8LWriteBits(bw, 3, code_length_bitdepth[kStorageOrder[i]]);
+    VP8LPutBits(bw, code_length_bitdepth[kStorageOrder[i]], 3);
   }
 }
 
@@ -281,49 +538,46 @@ static void StoreHuffmanTreeToBitMask(
   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]);
+    VP8LPutBits(bw, huffman_code->codes[ix], huffman_code->code_lengths[ix]);
     switch (ix) {
       case 16:
-        VP8LWriteBits(bw, 2, extra_bits);
+        VP8LPutBits(bw, extra_bits, 2);
         break;
       case 17:
-        VP8LWriteBits(bw, 3, extra_bits);
+        VP8LPutBits(bw, extra_bits, 3);
         break;
       case 18:
-        VP8LWriteBits(bw, 7, extra_bits);
+        VP8LPutBits(bw, extra_bits, 7);
         break;
     }
   }
 }
 
-static int StoreFullHuffmanCode(VP8LBitWriter* const bw,
-                                const HuffmanTreeCode* const tree) {
-  int ok = 0;
+// 'huff_tree' and 'tokens' are pre-alloacted buffers.
+static void StoreFullHuffmanCode(VP8LBitWriter* const bw,
+                                 HuffmanTree* const huff_tree,
+                                 HuffmanTreeToken* const tokens,
+                                 const HuffmanTreeCode* const tree) {
   uint8_t code_length_bitdepth[CODE_LENGTH_CODES] = { 0 };
   uint16_t code_length_bitdepth_symbols[CODE_LENGTH_CODES] = { 0 };
   const int max_tokens = tree->num_symbols;
   int num_tokens;
   HuffmanTreeCode huffman_code;
-  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);
+  VP8LPutBits(bw, 0, 1);
   num_tokens = VP8LCreateCompressedHuffmanTree(tree, tokens, max_tokens);
   {
-    int histogram[CODE_LENGTH_CODES] = { 0 };
+    uint32_t histogram[CODE_LENGTH_CODES] = { 0 };
+    uint8_t buf_rle[CODE_LENGTH_CODES] = { 0 };
     int i;
     for (i = 0; i < num_tokens; ++i) {
       ++histogram[tokens[i].code];
     }
 
-    if (!VP8LCreateHuffmanTree(histogram, 7, &huffman_code)) {
-      goto End;
-    }
+    VP8LCreateHuffmanTree(histogram, 7, buf_rle, huff_tree, &huffman_code);
   }
 
   StoreHuffmanTreeOfHuffmanTreeToBitMask(bw, code_length_bitdepth);
@@ -350,24 +604,23 @@ static int StoreFullHuffmanCode(VP8LBitWriter* const bw,
     }
     write_trimmed_length = (trimmed_length > 1 && trailing_zero_bits > 12);
     length = write_trimmed_length ? trimmed_length : num_tokens;
-    VP8LWriteBits(bw, 1, write_trimmed_length);
+    VP8LPutBits(bw, write_trimmed_length, 1);
     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);
+      VP8LPutBits(bw, nbitpairs - 1, 3);
       assert(trimmed_length >= 2);
-      VP8LWriteBits(bw, nbitpairs * 2, trimmed_length - 2);
+      VP8LPutBits(bw, trimmed_length - 2, nbitpairs * 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) {
+// 'huff_tree' and 'tokens' are pre-alloacted buffers.
+static void StoreHuffmanCode(VP8LBitWriter* const bw,
+                             HuffmanTree* const huff_tree,
+                             HuffmanTreeToken* const tokens,
+                             const HuffmanTreeCode* const huffman_code) {
   int i;
   int count = 0;
   int symbols[2] = { 0, 0 };
@@ -384,163 +637,248 @@ static int StoreHuffmanCode(VP8LBitWriter* const bw,
 
   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;
+    VP8LPutBits(bw, 0x01, 4);
   } 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);
+    VP8LPutBits(bw, 1, 1);  // Small tree marker to encode 1 or 2 symbols.
+    VP8LPutBits(bw, count - 1, 1);
     if (symbols[0] <= 1) {
-      VP8LWriteBits(bw, 1, 0);  // Code bit for small (1 bit) symbol value.
-      VP8LWriteBits(bw, 1, symbols[0]);
+      VP8LPutBits(bw, 0, 1);  // Code bit for small (1 bit) symbol value.
+      VP8LPutBits(bw, symbols[0], 1);
     } else {
-      VP8LWriteBits(bw, 1, 1);
-      VP8LWriteBits(bw, 8, symbols[0]);
+      VP8LPutBits(bw, 1, 1);
+      VP8LPutBits(bw, symbols[0], 8);
     }
     if (count == 2) {
-      VP8LWriteBits(bw, 8, symbols[1]);
+      VP8LPutBits(bw, symbols[1], 8);
     }
-    return 1;
   } else {
-    return StoreFullHuffmanCode(bw, huffman_code);
+    StoreFullHuffmanCode(bw, huff_tree, tokens, huffman_code);
   }
 }
 
-static void WriteHuffmanCode(VP8LBitWriter* const bw,
-                             const HuffmanTreeCode* const code, int index) {
-  const int depth = code->code_lengths[index];
-  const int symbol = code->codes[index];
-  VP8LWriteBits(bw, depth, symbol);
+static WEBP_INLINE void WriteHuffmanCode(VP8LBitWriter* const bw,
+                             const HuffmanTreeCode* const code,
+                             int code_index) {
+  const int depth = code->code_lengths[code_index];
+  const int symbol = code->codes[code_index];
+  VP8LPutBits(bw, symbol, depth);
 }
 
-static void StoreImageToBitMask(
+static WEBP_INLINE void WriteHuffmanCodeWithExtraBits(
+    VP8LBitWriter* const bw,
+    const HuffmanTreeCode* const code,
+    int code_index,
+    int bits,
+    int n_bits) {
+  const int depth = code->code_lengths[code_index];
+  const int symbol = code->codes[code_index];
+  VP8LPutBits(bw, (bits << depth) | symbol, depth + n_bits);
+}
+
+static WebPEncodingError StoreImageToBitMask(
     VP8LBitWriter* const bw, int width, int histo_bits,
-    const VP8LBackwardRefs* const refs,
+    VP8LBackwardRefs* const refs,
     const uint16_t* histogram_symbols,
     const HuffmanTreeCode* const huffman_codes) {
+  const int histo_xsize = histo_bits ? VP8LSubSampleSize(width, histo_bits) : 1;
+  const int tile_mask = (histo_bits == 0) ? 0 : -(1 << histo_bits);
   // x and y trace the position in the image.
   int x = 0;
   int y = 0;
-  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)) {
+  int tile_x = x & tile_mask;
+  int tile_y = y & tile_mask;
+  int histogram_ix = histogram_symbols[0];
+  const HuffmanTreeCode* codes = huffman_codes + 5 * histogram_ix;
+  VP8LRefsCursor c = VP8LRefsCursorInit(refs);
+  while (VP8LRefsCursorOk(&c)) {
+    const PixOrCopy* const v = c.cur_pos;
+    if ((tile_x != (x & tile_mask)) || (tile_y != (y & tile_mask))) {
+      tile_x = x & tile_mask;
+      tile_y = y & tile_mask;
+      histogram_ix = histogram_symbols[(y >> histo_bits) * histo_xsize +
+                                       (x >> histo_bits)];
+      codes = huffman_codes + 5 * histogram_ix;
+    }
+    if (PixOrCopyIsLiteral(v)) {
       static const 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 if (PixOrCopyIsCacheIdx(v)) {
+      const int code = PixOrCopyCacheIdx(v);
+      const int literal_ix = 256 + NUM_LENGTH_CODES + code;
+      WriteHuffmanCode(bw, codes, literal_ix);
     } else {
       int bits, n_bits;
-      int code, distance;
+      int code;
 
-      PrefixEncode(v->len, &code, &n_bits, &bits);
-      WriteHuffmanCode(bw, codes, 256 + code);
-      VP8LWriteBits(bw, n_bits, bits);
+      const int distance = PixOrCopyDistance(v);
+      VP8LPrefixEncode(v->len, &code, &n_bits, &bits);
+      WriteHuffmanCodeWithExtraBits(bw, codes, 256 + code, bits, n_bits);
 
-      distance = PixOrCopyDistance(v);
-      PrefixEncode(distance, &code, &n_bits, &bits);
+      // Don't write the distance with the extra bits code since
+      // the distance can be up to 18 bits of extra bits, and the prefix
+      // 15 bits, totaling to 33, and our PutBits only supports up to 32 bits.
+      // TODO(jyrki): optimize this further.
+      VP8LPrefixEncode(distance, &code, &n_bits, &bits);
       WriteHuffmanCode(bw, codes + 4, code);
-      VP8LWriteBits(bw, n_bits, bits);
+      VP8LPutBits(bw, bits, n_bits);
     }
     x += PixOrCopyLength(v);
     while (x >= width) {
       x -= width;
       ++y;
     }
+    VP8LRefsCursorNext(&c);
   }
+  return bw->error_ ? VP8_ENC_ERROR_OUT_OF_MEMORY : VP8_ENC_OK;
 }
 
 // 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) {
+static WebPEncodingError EncodeImageNoHuffman(VP8LBitWriter* const bw,
+                                              const uint32_t* const argb,
+                                              VP8LHashChain* const hash_chain,
+                                              VP8LBackwardRefs refs_array[2],
+                                              int width, int height,
+                                              int quality) {
   int i;
-  int ok = 0;
-  VP8LBackwardRefs refs;
+  int max_tokens = 0;
+  WebPEncodingError err = VP8_ENC_OK;
+  VP8LBackwardRefs* refs;
+  HuffmanTreeToken* tokens = NULL;
   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;
+  int cache_bits = 0;
+  VP8LHistogramSet* histogram_image = NULL;
+  HuffmanTree* const huff_tree = (HuffmanTree*)WebPSafeMalloc(
+        3ULL * CODE_LENGTH_CODES, sizeof(*huff_tree));
+  if (huff_tree == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
 
   // Calculate backward references from ARGB image.
-  if (!VP8LGetBackwardReferences(width, height, argb, quality, 0, 1, &refs)) {
+  refs = VP8LGetBackwardReferences(width, height, argb, quality, 0, &cache_bits,
+                                   hash_chain, refs_array);
+  if (refs == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
+  histogram_image = VP8LAllocateHistogramSet(1, cache_bits);
+  if (histogram_image == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
   // Build histogram image and symbols from backward references.
-  VP8LHistogramStoreRefs(&refs, histogram_image->histograms[0]);
+  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)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
 
   // No color cache, no Huffman image.
-  VP8LWriteBits(bw, 1, 0);
+  VP8LPutBits(bw, 0, 1);
 
-  // Store Huffman codes.
+  // Find maximum number of symbols for the huffman tree-set.
   for (i = 0; i < 5; ++i) {
     HuffmanTreeCode* const codes = &huffman_codes[i];
-    if (!StoreHuffmanCode(bw, codes)) {
-      goto Error;
+    if (max_tokens < codes->num_symbols) {
+      max_tokens = codes->num_symbols;
     }
+  }
+
+  tokens = (HuffmanTreeToken*)WebPSafeMalloc(max_tokens, sizeof(*tokens));
+  if (tokens == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  // Store Huffman codes.
+  for (i = 0; i < 5; ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[i];
+    StoreHuffmanCode(bw, huff_tree, tokens, codes);
     ClearHuffmanTreeIfOnlyOneSymbol(codes);
   }
 
   // Store actual literals.
-  StoreImageToBitMask(bw, width, 0, &refs, histogram_symbols, huffman_codes);
-  ok = 1;
+  err = StoreImageToBitMask(bw, width, 0, refs, histogram_symbols,
+                            huffman_codes);
 
  Error:
-  free(histogram_image);
-  VP8LClearBackwardRefs(&refs);
-  free(huffman_codes[0].codes);
-  return ok;
+  WebPSafeFree(tokens);
+  WebPSafeFree(huff_tree);
+  VP8LFreeHistogramSet(histogram_image);
+  WebPSafeFree(huffman_codes[0].codes);
+  return err;
 }
 
-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);
+static WebPEncodingError EncodeImageInternal(VP8LBitWriter* const bw,
+                                             const uint32_t* const argb,
+                                             VP8LHashChain* const hash_chain,
+                                             VP8LBackwardRefs refs_array[2],
+                                             int width, int height, int quality,
+                                             int low_effort, int* cache_bits,
+                                             int histogram_bits,
+                                             size_t init_byte_position,
+                                             int* const hdr_size,
+                                             int* const data_size) {
+  WebPEncodingError err = VP8_ENC_OK;
   const uint32_t histogram_image_xysize =
       VP8LSubSampleSize(width, histogram_bits) *
       VP8LSubSampleSize(height, histogram_bits);
-  VP8LHistogramSet* histogram_image =
-      VP8LAllocateHistogramSet(histogram_image_xysize, 0);
+  VP8LHistogramSet* histogram_image = NULL;
+  VP8LHistogramSet* tmp_histos = NULL;
   int histogram_image_size = 0;
   size_t bit_array_size = 0;
+  HuffmanTree* huff_tree = NULL;
+  HuffmanTreeToken* tokens = NULL;
   HuffmanTreeCode* huffman_codes = NULL;
   VP8LBackwardRefs refs;
+  VP8LBackwardRefs* best_refs;
   uint16_t* const histogram_symbols =
-      (uint16_t*)WebPSafeMalloc((uint64_t)histogram_image_xysize,
+      (uint16_t*)WebPSafeMalloc(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) goto Error;
+  assert(hdr_size != NULL);
+  assert(data_size != NULL);
 
+  VP8LBackwardRefsInit(&refs, refs_array[0].block_size_);
+  if (histogram_symbols == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  *cache_bits = MAX_COLOR_CACHE_BITS;
+  // 'best_refs' is the reference to the best backward refs and points to one
+  // of refs_array[0] or refs_array[1].
   // Calculate backward references from ARGB image.
-  if (!VP8LGetBackwardReferences(width, height, argb, quality, cache_bits,
-                                 use_2d_locality, &refs)) {
+  best_refs = VP8LGetBackwardReferences(width, height, argb, quality,
+                                        low_effort, cache_bits, hash_chain,
+                                        refs_array);
+  if (best_refs == NULL || !VP8LBackwardRefsCopy(best_refs, &refs)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
+  histogram_image =
+      VP8LAllocateHistogramSet(histogram_image_xysize, *cache_bits);
+  tmp_histos = VP8LAllocateHistogramSet(2, *cache_bits);
+  if (histogram_image == NULL || tmp_histos == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
   // Build histogram image and symbols from backward references.
-  if (!VP8LGetHistoImageSymbols(width, height, &refs,
-                                quality, histogram_bits, cache_bits,
-                                histogram_image,
-                                histogram_symbols)) {
+  if (!VP8LGetHistoImageSymbols(width, height, &refs, quality, low_effort,
+                                histogram_bits, *cache_bits, histogram_image,
+                                tmp_histos, histogram_symbols)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
   // Create Huffman bit lengths and codes for each histogram image.
@@ -548,171 +886,167 @@ static int EncodeImageInternal(VP8LBitWriter* const bw,
   bit_array_size = 5 * histogram_image_size;
   huffman_codes = (HuffmanTreeCode*)WebPSafeCalloc(bit_array_size,
                                                    sizeof(*huffman_codes));
+  // Note: some histogram_image entries may point to tmp_histos[], so the latter
+  // need to outlive the following call to GetHuffBitLengthsAndCodes().
   if (huffman_codes == NULL ||
       !GetHuffBitLengthsAndCodes(histogram_image, huffman_codes)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
+  // Free combined histograms.
+  VP8LFreeHistogramSet(histogram_image);
+  histogram_image = NULL;
+
+  // Free scratch histograms.
+  VP8LFreeHistogramSet(tmp_histos);
+  tmp_histos = NULL;
 
   // Color Cache parameters.
-  VP8LWriteBits(bw, 1, use_color_cache);
-  if (use_color_cache) {
-    VP8LWriteBits(bw, 4, cache_bits);
+  if (*cache_bits > 0) {
+    VP8LPutBits(bw, 1, 1);
+    VP8LPutBits(bw, *cache_bits, 4);
+  } else {
+    VP8LPutBits(bw, 0, 1);
   }
 
   // Huffman image + meta huffman.
   {
     const int write_histogram_image = (histogram_image_size > 1);
-    VP8LWriteBits(bw, 1, write_histogram_image);
+    VP8LPutBits(bw, write_histogram_image, 1);
     if (write_histogram_image) {
       uint32_t* const histogram_argb =
-          (uint32_t*)WebPSafeMalloc((uint64_t)histogram_image_xysize,
+          (uint32_t*)WebPSafeMalloc(histogram_image_xysize,
                                     sizeof(*histogram_argb));
       int max_index = 0;
       uint32_t i;
-      if (histogram_argb == NULL) goto Error;
+      if (histogram_argb == NULL) {
+        err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+        goto Error;
+      }
       for (i = 0; i < histogram_image_xysize; ++i) {
-        const int index = histogram_symbols[i] & 0xffff;
-        histogram_argb[i] = 0xff000000 | (index << 8);
-        if (index >= max_index) {
-          max_index = index + 1;
+        const int symbol_index = histogram_symbols[i] & 0xffff;
+        histogram_argb[i] = (symbol_index << 8);
+        if (symbol_index >= max_index) {
+          max_index = symbol_index + 1;
         }
       }
       histogram_image_size = max_index;
 
-      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;
+      VP8LPutBits(bw, histogram_bits - 2, 3);
+      err = EncodeImageNoHuffman(bw, histogram_argb, hash_chain, refs_array,
+                                 VP8LSubSampleSize(width, histogram_bits),
+                                 VP8LSubSampleSize(height, histogram_bits),
+                                 quality);
+      WebPSafeFree(histogram_argb);
+      if (err != VP8_ENC_OK) goto Error;
     }
   }
 
   // Store Huffman codes.
   {
     int i;
+    int max_tokens = 0;
+    huff_tree = (HuffmanTree*)WebPSafeMalloc(3ULL * CODE_LENGTH_CODES,
+                                             sizeof(*huff_tree));
+    if (huff_tree == NULL) {
+      err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+      goto Error;
+    }
+    // Find maximum number of symbols for the huffman tree-set.
     for (i = 0; i < 5 * histogram_image_size; ++i) {
       HuffmanTreeCode* const codes = &huffman_codes[i];
-      if (!StoreHuffmanCode(bw, codes)) goto Error;
+      if (max_tokens < codes->num_symbols) {
+        max_tokens = codes->num_symbols;
+      }
+    }
+    tokens = (HuffmanTreeToken*)WebPSafeMalloc(max_tokens,
+                                               sizeof(*tokens));
+    if (tokens == NULL) {
+      err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+      goto Error;
+    }
+    for (i = 0; i < 5 * histogram_image_size; ++i) {
+      HuffmanTreeCode* const codes = &huffman_codes[i];
+      StoreHuffmanCode(bw, huff_tree, tokens, codes);
       ClearHuffmanTreeIfOnlyOneSymbol(codes);
     }
   }
-  // Free combined histograms.
-  free(histogram_image);
-  histogram_image = NULL;
 
+  *hdr_size = (int)(VP8LBitWriterNumBytes(bw) - init_byte_position);
   // Store actual literals.
-  StoreImageToBitMask(bw, width, histogram_bits, &refs,
-                      histogram_symbols, huffman_codes);
-  ok = 1;
+  err = StoreImageToBitMask(bw, width, histogram_bits, &refs,
+                            histogram_symbols, huffman_codes);
+  *data_size =
+        (int)(VP8LBitWriterNumBytes(bw) - init_byte_position - *hdr_size);
 
  Error:
-  if (!ok) free(histogram_image);
-
-  VP8LClearBackwardRefs(&refs);
+  WebPSafeFree(tokens);
+  WebPSafeFree(huff_tree);
+  VP8LFreeHistogramSet(histogram_image);
+  VP8LFreeHistogramSet(tmp_histos);
+  VP8LBackwardRefsClear(&refs);
   if (huffman_codes != NULL) {
-    free(huffman_codes->codes);
-    free(huffman_codes);
+    WebPSafeFree(huffman_codes->codes);
+    WebPSafeFree(huffman_codes);
   }
-  free(histogram_symbols);
-  return ok;
+  WebPSafeFree(histogram_symbols);
+  return err;
 }
 
 // -----------------------------------------------------------------------------
 // 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 void ApplySubtractGreen(VP8LEncoder* const enc, int width, int height,
+                               VP8LBitWriter* const bw) {
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, SUBTRACT_GREEN, 2);
+  VP8LSubtractGreenFromBlueAndRed(enc->argb_, width * height);
 }
 
-static int ApplyPredictFilter(const VP8LEncoder* const enc,
-                              int width, int height, int quality,
-                              VP8LBitWriter* const bw) {
+static WebPEncodingError ApplyPredictFilter(const VP8LEncoder* const enc,
+                                            int width, int height,
+                                            int quality, int low_effort,
+                                            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);
+  VP8LResidualImage(width, height, pred_bits, low_effort, enc->argb_,
+                    enc->argb_scratch_, enc->transform_data_);
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, PREDICTOR_TRANSFORM, 2);
   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;
+  VP8LPutBits(bw, pred_bits - 2, 3);
+  return EncodeImageNoHuffman(bw, enc->transform_data_,
+                              (VP8LHashChain*)&enc->hash_chain_,
+                              (VP8LBackwardRefs*)enc->refs_,  // cast const away
+                              transform_width, transform_height,
+                              quality);
 }
 
-static int ApplyCrossColorFilter(const VP8LEncoder* const enc,
-                                 int width, int height, int quality,
-                                 VP8LBitWriter* const bw) {
+static WebPEncodingError 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 == 0) ? 32 : 8;
 
-  VP8LColorSpaceTransform(width, height, ccolor_transform_bits, step,
+  VP8LColorSpaceTransform(width, height, ccolor_transform_bits, quality,
                           enc->argb_, enc->transform_data_);
-  VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
-  VP8LWriteBits(bw, 2, CROSS_COLOR_TRANSFORM);
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, CROSS_COLOR_TRANSFORM, 2);
   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;
+  VP8LPutBits(bw, ccolor_transform_bits - 2, 3);
+  return EncodeImageNoHuffman(bw, enc->transform_data_,
+                              (VP8LHashChain*)&enc->hash_chain_,
+                              (VP8LBackwardRefs*)enc->refs_,  // cast const away
+                              transform_width, transform_height,
+                              quality);
 }
 
 // -----------------------------------------------------------------------------
 
-static void PutLE32(uint8_t* const data, uint32_t val) {
-  data[0] = (val >>  0) & 0xff;
-  data[1] = (val >>  8) & 0xff;
-  data[2] = (val >> 16) & 0xff;
-  data[3] = (val >> 24) & 0xff;
-}
-
 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] = {
@@ -733,14 +1067,14 @@ static int WriteImageSize(const WebPPicture* const pic,
   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);
+  VP8LPutBits(bw, width, VP8L_IMAGE_SIZE_BITS);
+  VP8LPutBits(bw, height, VP8L_IMAGE_SIZE_BITS);
   return !bw->error_;
 }
 
 static int WriteRealAlphaAndVersion(VP8LBitWriter* const bw, int has_alpha) {
-  VP8LWriteBits(bw, 1, has_alpha);
-  VP8LWriteBits(bw, VP8L_VERSION_BITS, VP8L_VERSION);
+  VP8LPutBits(bw, has_alpha, 1);
+  VP8LPutBits(bw, VP8L_VERSION, VP8L_VERSION_BITS);
   return !bw->error_;
 }
 
@@ -780,166 +1114,261 @@ static WebPEncodingError WriteImage(const WebPPicture* const pic,
 
 // Allocates the memory for argb (W x H) buffer, 2 rows of context for
 // prediction and transform data.
+// Flags influencing the memory allocated:
+//  enc->transform_bits_
+//  enc->use_predict_, enc->use_cross_color_
 static 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;
+  if (enc->argb_ == NULL) {
+    const int tile_size = 1 << enc->transform_bits_;
+    const uint64_t image_size = width * height;
+    // Ensure enough size for tiles, as well as for two scanlines and two
+    // extra pixels for CopyImageWithPrediction.
+    const uint64_t argb_scratch_size =
+        enc->use_predict_ ? tile_size * width + width + 2 : 0;
+    const int transform_data_size =
+        (enc->use_predict_ || enc->use_cross_color_)
+            ? VP8LSubSampleSize(width, enc->transform_bits_) *
+              VP8LSubSampleSize(height, enc->transform_bits_)
+            : 0;
+    const uint64_t total_size =
+        image_size + WEBP_ALIGN_CST +
+        argb_scratch_size + WEBP_ALIGN_CST +
+        (uint64_t)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 = (uint32_t*)WEBP_ALIGN(mem + image_size);
+    enc->argb_scratch_ = mem;
+    mem = (uint32_t*)WEBP_ALIGN(mem + argb_scratch_size);
+    enc->transform_data_ = mem;
+    enc->current_width_ = width;
   }
-  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;
 }
 
-// Bundles multiple (2, 4 or 8) pixels into a single pixel.
-// Returns the new xsize.
-static void BundleColorMap(const WebPPicture* const pic,
-                           int xbits, uint32_t* bundled_argb, int xs) {
-  int y;
-  const int bit_depth = 1 << (3 - xbits);
-  uint32_t code = 0;
-  const uint32_t* argb = pic->argb;
-  const int width = pic->width;
-  const int height = pic->height;
+static void ClearTransformBuffer(VP8LEncoder* const enc) {
+  WebPSafeFree(enc->argb_);
+  enc->argb_ = NULL;
+}
 
+static WebPEncodingError MakeInputImageCopy(VP8LEncoder* const enc) {
+  WebPEncodingError err = VP8_ENC_OK;
+  const WebPPicture* const picture = enc->pic_;
+  const int width = picture->width;
+  const int height = picture->height;
+  int y;
+  err = AllocateTransformBuffer(enc, width, height);
+  if (err != VP8_ENC_OK) return err;
   for (y = 0; y < height; ++y) {
-    int x;
-    for (x = 0; x < width; ++x) {
-      const int mask = (1 << xbits) - 1;
-      const int xsub = x & mask;
-      if (xsub == 0) {
-        code = 0;
-      }
-      // TODO(vikasa): simplify the bundling logic.
-      code |= (argb[x] & 0xff00) << (bit_depth * xsub);
-      bundled_argb[y * xs + (x >> xbits)] = 0xff000000 | code;
-    }
-    argb += pic->argb_stride;
+    memcpy(enc->argb_ + y * width,
+           picture->argb + y * picture->argb_stride,
+           width * sizeof(*enc->argb_));
   }
+  assert(enc->current_width_ == width);
+  return VP8_ENC_OK;
 }
 
-// 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 ApplyPalette(VP8LBitWriter* const bw,
-                                      VP8LEncoder* const enc, int quality) {
-  WebPEncodingError err = VP8_ENC_OK;
-  int i, x, y;
-  const WebPPicture* const pic = enc->pic_;
-  uint32_t* argb = pic->argb;
-  const int width = pic->width;
-  const int height = pic->height;
-  uint32_t* const palette = enc->palette_;
-  const int palette_size = enc->palette_size_;
+// -----------------------------------------------------------------------------
 
-  // Replace each input pixel by corresponding palette index.
-  for (y = 0; y < height; ++y) {
-    for (x = 0; x < width; ++x) {
-      const uint32_t pix = argb[x];
-      for (i = 0; i < palette_size; ++i) {
+static void MapToPalette(const uint32_t palette[], int num_colors,
+                         uint32_t* const last_pix, int* const last_idx,
+                         const uint32_t* src, uint8_t* dst, int width) {
+  int x;
+  int prev_idx = *last_idx;
+  uint32_t prev_pix = *last_pix;
+  for (x = 0; x < width; ++x) {
+    const uint32_t pix = src[x];
+    if (pix != prev_pix) {
+      int i;
+      for (i = 0; i < num_colors; ++i) {
         if (pix == palette[i]) {
-          argb[x] = 0xff000000u | (i << 8);
+          prev_idx = i;
+          prev_pix = pix;
           break;
         }
       }
     }
-    argb += pic->argb_stride;
+    dst[x] = prev_idx;
   }
+  *last_idx = prev_idx;
+  *last_pix = prev_pix;
+}
 
-  // 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;
+// Remap argb values in src[] to packed palettes entries in dst[]
+// using 'row' as a temporary buffer of size 'width'.
+// We assume that all src[] values have a corresponding entry in the palette.
+// Note: src[] can be the same as dst[]
+static WebPEncodingError ApplyPalette(const uint32_t* src, uint32_t src_stride,
+                                      uint32_t* dst, uint32_t dst_stride,
+                                      const uint32_t* palette, int palette_size,
+                                      int width, int height, int xbits) {
+  // TODO(skal): this tmp buffer is not needed if VP8LBundleColorMap() can be
+  // made to work in-place.
+  uint8_t* const tmp_row = (uint8_t*)WebPSafeMalloc(width, sizeof(*tmp_row));
+  int i, x, y;
+  int use_LUT = 1;
+
+  if (tmp_row == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
+  for (i = 0; i < palette_size; ++i) {
+    if ((palette[i] & 0xffff00ffu) != 0) {
+      use_LUT = 0;
+      break;
+    }
   }
 
-  if (palette_size <= 16) {
-    // Image can be packed (multiple pixels per uint32_t).
-    int xbits = 1;
-    if (palette_size <= 2) {
-      xbits = 3;
-    } else if (palette_size <= 4) {
-      xbits = 2;
+  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;
+        tmp_row[x] = inv_palette[color];
+      }
+      VP8LBundleColorMap(tmp_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) {
+      MapToPalette(palette, palette_size, &last_pix, &last_idx,
+                   src, tmp_row, width);
+      VP8LBundleColorMap(tmp_row, width, xbits, dst);
+      src += src_stride;
+      dst += dst_stride;
     }
-    err = AllocateTransformBuffer(enc, VP8LSubSampleSize(width, xbits), height);
-    if (err != VP8_ENC_OK) goto Error;
-    BundleColorMap(pic, xbits, enc->argb_, enc->current_width_);
   }
+  WebPSafeFree(tmp_row);
+  return VP8_ENC_OK;
+}
 
- Error:
+// Note: Expects "enc->palette_" to be set properly.
+static WebPEncodingError MapImageFromPalette(VP8LEncoder* const enc,
+                                             int in_place) {
+  WebPEncodingError err = VP8_ENC_OK;
+  const WebPPicture* const pic = enc->pic_;
+  const int width = pic->width;
+  const int height = pic->height;
+  const uint32_t* const palette = enc->palette_;
+  const uint32_t* src = in_place ? enc->argb_ : pic->argb;
+  const int src_stride = in_place ? enc->current_width_ : pic->argb_stride;
+  const int palette_size = enc->palette_size_;
+  int xbits;
+
+  // Replace each input pixel by corresponding palette index.
+  // This is done line by line.
+  if (palette_size <= 4) {
+    xbits = (palette_size <= 2) ? 3 : 2;
+  } else {
+    xbits = (palette_size <= 16) ? 1 : 0;
+  }
+
+  err = AllocateTransformBuffer(enc, VP8LSubSampleSize(width, xbits), height);
+  if (err != VP8_ENC_OK) return err;
+
+  err = ApplyPalette(src, src_stride,
+                     enc->argb_, enc->current_width_,
+                     palette, palette_size, width, height, xbits);
   return err;
 }
 
-// -----------------------------------------------------------------------------
+// Save palette_[] to bitstream.
+static WebPEncodingError EncodePalette(VP8LBitWriter* const bw,
+                                       VP8LEncoder* const enc) {
+  int i;
+  uint32_t tmp_palette[MAX_PALETTE_SIZE];
+  const int palette_size = enc->palette_size_;
+  const uint32_t* const palette = enc->palette_;
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, COLOR_INDEXING_TRANSFORM, 2);
+  assert(palette_size >= 1 && palette_size <= MAX_PALETTE_SIZE);
+  VP8LPutBits(bw, palette_size - 1, 8);
+  for (i = palette_size - 1; i >= 1; --i) {
+    tmp_palette[i] = VP8LSubPixels(palette[i], palette[i - 1]);
+  }
+  tmp_palette[0] = palette[0];
+  return EncodeImageNoHuffman(bw, tmp_palette, &enc->hash_chain_, enc->refs_,
+                              palette_size, 1, 20 /* quality */);
+}
 
-static int GetHistoBits(const WebPConfig* const config,
-                        const WebPPicture* const pic) {
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+
+static WebPEncodingError EncodeDeltaPalettePredictorImage(
+    VP8LBitWriter* const bw, VP8LEncoder* const enc, int quality) {
+  const WebPPicture* const pic = enc->pic_;
   const int width = pic->width;
   const int height = pic->height;
-  const size_t hist_size = sizeof(VP8LHistogram);
-  // Make tile size a function of encoding method (Range: 0 to 6).
-  int histo_bits = 7 - config->method;
-  while (1) {
-    const size_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;
+
+  const int pred_bits = 5;
+  const int transform_width = VP8LSubSampleSize(width, pred_bits);
+  const int transform_height = VP8LSubSampleSize(height, pred_bits);
+  const int pred = 7;   // default is Predictor7 (Top/Left Average)
+  const int tiles_per_row = VP8LSubSampleSize(width, pred_bits);
+  const int tiles_per_col = VP8LSubSampleSize(height, pred_bits);
+  uint32_t* predictors;
+  int tile_x, tile_y;
+  WebPEncodingError err = VP8_ENC_OK;
+
+  predictors = (uint32_t*)WebPSafeMalloc(tiles_per_col * tiles_per_row,
+                                         sizeof(*predictors));
+  if (predictors == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
+
+  for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) {
+    for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) {
+      predictors[tile_y * tiles_per_row + tile_x] = 0xff000000u | (pred << 8);
+    }
   }
-  return (histo_bits < MIN_HUFFMAN_BITS) ? MIN_HUFFMAN_BITS :
-         (histo_bits > MAX_HUFFMAN_BITS) ? MAX_HUFFMAN_BITS : histo_bits;
-}
 
-static void InitEncParams(VP8LEncoder* const enc) {
-  const WebPConfig* const config = enc->config_;
-  const WebPPicture* const picture = enc->pic_;
-  const int method = config->method;
-  const float quality = config->quality;
-  enc->transform_bits_ = (method < 4) ? 5 : (method > 4) ? 3 : 4;
-  enc->histo_bits_ = GetHistoBits(config, picture);
-  enc->cache_bits_ = (quality <= 25.f) ? 0 : 7;
+  VP8LPutBits(bw, TRANSFORM_PRESENT, 1);
+  VP8LPutBits(bw, PREDICTOR_TRANSFORM, 2);
+  VP8LPutBits(bw, pred_bits - 2, 3);
+  err = EncodeImageNoHuffman(bw, predictors, &enc->hash_chain_,
+                             (VP8LBackwardRefs*)enc->refs_,  // cast const away
+                             transform_width, transform_height,
+                             quality);
+  WebPSafeFree(predictors);
+  return err;
 }
 
+#endif // WEBP_EXPERIMENTAL_FEATURES
+
 // -----------------------------------------------------------------------------
 // VP8LEncoder
 
 static VP8LEncoder* VP8LEncoderNew(const WebPConfig* const config,
                                    const WebPPicture* const picture) {
-  VP8LEncoder* const enc = (VP8LEncoder*)calloc(1, sizeof(*enc));
+  VP8LEncoder* const enc = (VP8LEncoder*)WebPSafeCalloc(1ULL, sizeof(*enc));
   if (enc == NULL) {
     WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
     return NULL;
   }
   enc->config_ = config;
   enc->pic_ = picture;
+
+  VP8LEncDspInit();
+
   return enc;
 }
 
 static void VP8LEncoderDelete(VP8LEncoder* enc) {
-  free(enc->argb_);
-  free(enc);
+  if (enc != NULL) {
+    VP8LHashChainClear(&enc->hash_chain_);
+    VP8LBackwardRefsClear(&enc->refs_[0]);
+    VP8LBackwardRefsClear(&enc->refs_[1]);
+    ClearTransformBuffer(enc);
+    WebPSafeFree(enc);
+  }
 }
 
 // -----------------------------------------------------------------------------
@@ -950,89 +1379,102 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
                                    VP8LBitWriter* const bw) {
   WebPEncodingError err = VP8_ENC_OK;
   const int quality = (int)config->quality;
+  const int low_effort = (config->method == 0);
   const int width = picture->width;
   const int height = picture->height;
   VP8LEncoder* const enc = VP8LEncoderNew(config, picture);
   const size_t byte_position = VP8LBitWriterNumBytes(bw);
+  int use_near_lossless = 0;
+  int hdr_size = 0;
+  int data_size = 0;
+  int use_delta_palettization = 0;
 
   if (enc == NULL) {
     err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
 
-  InitEncParams(enc);
-
   // ---------------------------------------------------------------------------
   // Analyze image (entropy, num_palettes etc)
 
-  if (!VP8LEncAnalyze(enc, config->image_hint)) {
+  if (!AnalyzeAndInit(enc)) {
     err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
 
-  if (enc->use_palette_) {
-    err = ApplyPalette(bw, enc, quality);
-    if (err != VP8_ENC_OK) goto Error;
-    // Color cache is disabled for palette.
-    enc->cache_bits_ = 0;
+  // Apply near-lossless preprocessing.
+  use_near_lossless = !enc->use_palette_ && (config->near_lossless < 100);
+  if (use_near_lossless) {
+    if (!VP8ApplyNearLossless(width, height, picture->argb,
+                              config->near_lossless)) {
+      err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+      goto Error;
+    }
   }
 
-  // In case image is not packed.
-  if (enc->argb_ == NULL) {
-    int y;
-    err = AllocateTransformBuffer(enc, width, height);
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  if (config->delta_palettization) {
+    enc->use_predict_ = 1;
+    enc->use_cross_color_ = 0;
+    enc->use_subtract_green_ = 0;
+    enc->use_palette_ = 1;
+    err = MakeInputImageCopy(enc);
     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_));
+    err = WebPSearchOptimalDeltaPalette(enc);
+    if (err != VP8_ENC_OK) goto Error;
+    if (enc->use_palette_) {
+      err = AllocateTransformBuffer(enc, width, height);
+      if (err != VP8_ENC_OK) goto Error;
+      err = EncodeDeltaPalettePredictorImage(bw, enc, quality);
+      if (err != VP8_ENC_OK) goto Error;
+      use_delta_palettization = 1;
     }
-    enc->current_width_ = width;
   }
+#endif  // WEBP_EXPERIMENTAL_FEATURES
 
-  // ---------------------------------------------------------------------------
-  // Apply transforms and write transform data.
-
-  if (!EvalAndApplySubtractGreen(enc, enc->current_width_, height, bw)) {
-    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-    goto Error;
+  // Encode palette
+  if (enc->use_palette_) {
+    err = EncodePalette(bw, enc);
+    if (err != VP8_ENC_OK) goto Error;
+    err = MapImageFromPalette(enc, use_delta_palettization);
+    if (err != VP8_ENC_OK) goto Error;
   }
-
-  if (enc->use_predict_) {
-    if (!ApplyPredictFilter(enc, enc->current_width_, height, quality, bw)) {
-      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
-      goto Error;
+  if (!use_delta_palettization) {
+    // In case image is not packed.
+    if (enc->argb_ == NULL) {
+      err = MakeInputImageCopy(enc);
+      if (err != VP8_ENC_OK) goto Error;
     }
-  }
 
-  if (enc->use_cross_color_) {
-    if (!ApplyCrossColorFilter(enc, enc->current_width_, height, quality, bw)) {
-      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
-      goto Error;
-    }
-  }
+    // -------------------------------------------------------------------------
+    // Apply transforms and write transform data.
 
-  VP8LWriteBits(bw, 1, !TRANSFORM_PRESENT);  // No more transforms.
+    if (enc->use_subtract_green_) {
+      ApplySubtractGreen(enc, enc->current_width_, height, bw);
+    }
 
-  // ---------------------------------------------------------------------------
-  // Estimate the color cache size.
+    if (enc->use_predict_) {
+      err = ApplyPredictFilter(enc, enc->current_width_, height, quality,
+                               low_effort, bw);
+      if (err != VP8_ENC_OK) goto Error;
+    }
 
-  if (enc->cache_bits_ > 0) {
-    if (!VP8LCalculateEstimateForCacheSize(enc->argb_, enc->current_width_,
-                                           height, &enc->cache_bits_)) {
-      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
-      goto Error;
+    if (enc->use_cross_color_) {
+      err = ApplyCrossColorFilter(enc, enc->current_width_,
+                                  height, quality, bw);
+      if (err != VP8_ENC_OK) goto Error;
     }
   }
 
+  VP8LPutBits(bw, !TRANSFORM_PRESENT, 1);  // No more transforms.
+
   // ---------------------------------------------------------------------------
   // 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;
-  }
+  err = EncodeImageInternal(bw, enc->argb_, &enc->hash_chain_, enc->refs_,
+                            enc->current_width_, height, quality, low_effort,
+                            &enc->cache_bits_, enc->histo_bits_, byte_position,
+                            &hdr_size, &data_size);
+  if (err != VP8_ENC_OK) goto Error;
 
   if (picture->stats != NULL) {
     WebPAuxStats* const stats = picture->stats;
@@ -1046,6 +1488,8 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
     stats->cache_bits = enc->cache_bits_;
     stats->palette_size = enc->palette_size_;
     stats->lossless_size = (int)(VP8LBitWriterNumBytes(bw) - byte_position);
+    stats->lossless_hdr_size = hdr_size;
+    stats->lossless_data_size = data_size;
   }
 
  Error:
@@ -1059,6 +1503,7 @@ int VP8LEncodeImage(const WebPConfig* const config,
   int has_alpha;
   size_t coded_size;
   int percent = 0;
+  int initial_size;
   WebPEncodingError err = VP8_ENC_OK;
   VP8LBitWriter bw;
 
@@ -1072,7 +1517,11 @@ int VP8LEncodeImage(const WebPConfig* const config,
 
   width = picture->width;
   height = picture->height;
-  if (!VP8LBitWriterInit(&bw, (width * height) >> 1)) {
+  // Initialize BitWriter with size corresponding to 16 bpp to photo images and
+  // 8 bpp for graphical images.
+  initial_size = (config->image_hint == WEBP_HINT_GRAPH) ?
+      width * height : width * height * 2;
+  if (!VP8LBitWriterInit(&bw, initial_size)) {
     err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
@@ -1135,7 +1584,7 @@ int VP8LEncodeImage(const WebPConfig* const config,
 
  Error:
   if (bw.error_) err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-  VP8LBitWriterDestroy(&bw);
+  VP8LBitWriterWipeOut(&bw);
   if (err != VP8_ENC_OK) {
     WebPEncodingSetError(picture, err);
     return 0;
@@ -1144,7 +1593,3 @@ int VP8LEncodeImage(const WebPConfig* const config,
 }
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/enc/vp8li.h b/drivers/webp/enc/vp8li.h
index bb111aec33..6b6db127db 100644
--- a/drivers/webp/enc/vp8li.h
+++ b/drivers/webp/enc/vp8li.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
@@ -12,12 +14,13 @@
 #ifndef WEBP_ENC_VP8LI_H_
 #define WEBP_ENC_VP8LI_H_
 
+#include "./backward_references.h"
 #include "./histogram.h"
 #include "../utils/bit_writer.h"
-#include "../encode.h"
-#include "../format_constants.h"
+#include "../webp/encode.h"
+#include "../webp/format_constants.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -43,6 +46,12 @@ typedef struct {
   int use_palette_;
   int palette_size_;
   uint32_t palette_[MAX_PALETTE_SIZE];
+
+  // Some 'scratch' (potentially large) objects.
+  struct VP8LBackwardRefs refs_[2];  // Backward Refs array corresponding to
+                                     // LZ77 & RLE coding.
+  VP8LHashChain hash_chain_;         // HashChain data for constructing
+                                     // backward references.
 } VP8LEncoder;
 
 //------------------------------------------------------------------------------
@@ -61,7 +70,7 @@ WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/enc/webpenc.c b/drivers/webp/enc/webpenc.c
index 3c275589fc..8ced07a2a3 100644
--- a/drivers/webp/enc/webpenc.c
+++ b/drivers/webp/enc/webpenc.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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
@@ -14,16 +16,13 @@
 #include <string.h>
 #include <math.h>
 
+#include "./cost.h"
 #include "./vp8enci.h"
 #include "./vp8li.h"
 #include "../utils/utils.h"
 
 // #define PRINT_MEMORY_INFO
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 #ifdef PRINT_MEMORY_INFO
 #include <stdio.h>
 #endif
@@ -35,43 +34,18 @@ int WebPGetEncoderVersion(void) {
 }
 
 //------------------------------------------------------------------------------
-// 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_;
+  VP8EncSegmentHeader* const hdr = &enc->segment_hdr_;
   hdr->num_segments_ = enc->config_->segments;
   hdr->update_map_  = (hdr->num_segments_ > 1);
   hdr->size_ = 0;
 }
 
 static void ResetFilterHeader(VP8Encoder* const enc) {
-  VP8FilterHeader* const hdr = &enc->filter_hdr_;
+  VP8EncFilterHeader* const hdr = &enc->filter_hdr_;
   hdr->simple_ = 1;
   hdr->level_ = 0;
   hdr->sharpness_ = 0;
@@ -93,56 +67,73 @@ static void ResetBoundaryPredictions(VP8Encoder* const enc) {
   enc->nz_[-1] = 0;   // constant
 }
 
-// Map configured quality level to coding tools used.
-//-------------+---+---+---+---+---+---+
-//   Quality   | 0 | 1 | 2 | 3 | 4 | 5 +
-//-------------+---+---+---+---+---+---+
-// dynamic prob| ~ | x | x | x | x | x |
-//-------------+---+---+---+---+---+---+
-// rd-opt modes|   |   | x | x | x | x |
-//-------------+---+---+---+---+---+---+
-// fast i4/i16 | x | x |   |   |   |   |
-//-------------+---+---+---+---+---+---+
-// rd-opt i4/16|   |   | x | x | x | x |
-//-------------+---+---+---+---+---+---+
-// Trellis     |   | x |   |   | x | x |
-//-------------+---+---+---+---+---+---+
-// full-SNS    |   |   |   |   |   | x |
-//-------------+---+---+---+---+---+---+
+// 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 int method = enc->config_->method;
-  const int limit = 100 - enc->config_->partition_limit;
+  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) ? 3
-                     : (method >= 5) ? 2
-                     : (method >= 3) ? 1
-                     : 0;
+  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
-// Transcient object sizes:
-//       VP8EncIterator: 352
-//         VP8ModeScore: 912
-//       VP8SegmentInfo: 532
-//             VP8Proba: 31032
+// Typical memory footprint (614x440 picture)
+//              encoder: 22111
+//                 info: 4368
+//                preds: 17741
+//          top samples: 1263
+//             non-zero: 175
+//             lf-stats: 0
+//                total: 45658
+// Transient object sizes:
+//       VP8EncIterator: 3360
+//         VP8ModeScore: 872
+//       VP8SegmentInfo: 732
+//          VP8EncProba: 18352
 //              LFStats: 2048
-// Picture size (yuv): 589824
+// Picture size (yuv): 419328
 
 static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
                                   WebPPicture* const picture) {
@@ -154,20 +145,16 @@ static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
   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);
-  const size_t cache_size = (3 * YUV_SIZE + PRED_SIZE) * sizeof(uint8_t);
+  const size_t nz_size = (mb_w + 1) * sizeof(uint32_t) + WEBP_ALIGN_CST;
   const size_t info_size = mb_w * mb_h * sizeof(VP8MBInfo);
-  const size_t samples_size = (2 * top_stride +         // top-luma/u/v
-                               16 + 16 + 16 + 8 + 1 +   // left y/u/v
-                               2 * ALIGN_CST)           // align all
-                               * sizeof(uint8_t);
+  const size_t samples_size = 2 * top_stride * sizeof(uint8_t)  // top-luma/u/v
+                            + WEBP_ALIGN_CST;                   // align all
   const size_t lf_stats_size =
-      config->autofilter ? sizeof(LFStats) + ALIGN_CST : 0;
+      config->autofilter ? sizeof(LFStats) + WEBP_ALIGN_CST : 0;
   VP8Encoder* enc;
   uint8_t* mem;
   const uint64_t size = (uint64_t)sizeof(VP8Encoder)   // main struct
-                      + ALIGN_CST                      // cache alignment
-                      + cache_size                     // working caches
+                      + WEBP_ALIGN_CST                 // cache alignment
                       + info_size                      // modes info
                       + preds_size                     // prediction modes
                       + samples_size                   // top/left samples
@@ -178,23 +165,22 @@ static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
   printf("===================================\n");
   printf("Memory used:\n"
          "             encoder: %ld\n"
-         "         block cache: %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, cache_size, info_size,
+         sizeof(VP8Encoder) + WEBP_ALIGN_CST, info_size,
          preds_size, samples_size, nz_size, lf_stats_size, size);
-  printf("Transcient object sizes:\n"
+  printf("Transient object sizes:\n"
          "      VP8EncIterator: %ld\n"
          "        VP8ModeScore: %ld\n"
          "      VP8SegmentInfo: %ld\n"
-         "            VP8Proba: %ld\n"
+         "         VP8EncProba: %ld\n"
          "             LFStats: %ld\n",
          sizeof(VP8EncIterator), sizeof(VP8ModeScore),
-         sizeof(VP8SegmentInfo), sizeof(VP8Proba),
+         sizeof(VP8SegmentInfo), sizeof(VP8EncProba),
          sizeof(LFStats));
   printf("Picture size (yuv): %ld\n",
          mb_w * mb_h * 384 * sizeof(uint8_t));
@@ -206,41 +192,27 @@ static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
     return NULL;
   }
   enc = (VP8Encoder*)mem;
-  mem = (uint8_t*)DO_ALIGN(mem + sizeof(*enc));
+  mem = (uint8_t*)WEBP_ALIGN(mem + sizeof(*enc));
   memset(enc, 0, sizeof(*enc));
   enc->num_parts_ = 1 << config->partitions;
   enc->mb_w_ = mb_w;
   enc->mb_h_ = mb_h;
   enc->preds_w_ = preds_w;
-  enc->yuv_in_ = (uint8_t*)mem;
-  mem += YUV_SIZE;
-  enc->yuv_out_ = (uint8_t*)mem;
-  mem += YUV_SIZE;
-  enc->yuv_out2_ = (uint8_t*)mem;
-  mem += YUV_SIZE;
-  enc->yuv_p_ = (uint8_t*)mem;
-  mem += PRED_SIZE;
   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*)mem;
+  enc->nz_ = 1 + (uint32_t*)WEBP_ALIGN(mem);
   mem += nz_size;
-  enc->lf_stats_ = lf_stats_size ? (LFStats*)DO_ALIGN(mem) : NULL;
+  enc->lf_stats_ = lf_stats_size ? (LFStats*)WEBP_ALIGN(mem) : NULL;
   mem += lf_stats_size;
 
   // top samples (all 16-aligned)
-  mem = (uint8_t*)DO_ALIGN(mem);
+  mem = (uint8_t*)WEBP_ALIGN(mem);
   enc->y_top_ = (uint8_t*)mem;
   enc->uv_top_ = enc->y_top_ + top_stride;
   mem += 2 * top_stride;
-  mem = (uint8_t*)DO_ALIGN(mem + 1);
-  enc->y_left_ = (uint8_t*)mem;
-  mem += 16 + 16;
-  enc->u_left_ = (uint8_t*)mem;
-  mem += 16;
-  enc->v_left_ = (uint8_t*)mem;
-  mem += 8;
+  assert(mem <= (uint8_t*)enc + size);
 
   enc->config_ = config;
   enc->profile_ = use_filter ? ((config->filter_type == 1) ? 0 : 1) : 2;
@@ -253,29 +225,32 @@ static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
   ResetSegmentHeader(enc);
   ResetFilterHeader(enc);
   ResetBoundaryPredictions(enc);
-
+  VP8EncDspCostInit();
   VP8EncInitAlpha(enc);
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-  VP8EncInitLayer(enc);
-#endif
 
+  // lower quality means smaller output -> we modulate a little the page
+  // size based on quality. This is just a crude 1rst-order prediction.
+  {
+    const float scale = 1.f + config->quality * 5.f / 100.f;  // in [1,6]
+    VP8TBufferInit(&enc->tokens_, (int)(mb_w * mb_h * 4 * scale));
+  }
   return enc;
 }
 
-static void DeleteVP8Encoder(VP8Encoder* enc) {
+static int DeleteVP8Encoder(VP8Encoder* enc) {
+  int ok = 1;
   if (enc != NULL) {
-    VP8EncDeleteAlpha(enc);
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-    VP8EncDeleteLayer(enc);
-#endif
-    free(enc);
+    ok = VP8EncDeleteAlpha(enc);
+    VP8TBufferClear(&enc->tokens_);
+    WebPSafeFree(enc);
   }
+  return ok;
 }
 
 //------------------------------------------------------------------------------
 
 static double GetPSNR(uint64_t err, uint64_t size) {
-  return err ? 10. * log10(255. * 255. * size / err) : 99.;
+  return (err > 0 && size > 0) ? 10. * log10(255. * 255. * size / err) : 99.;
 }
 
 static void FinalizePSNR(const VP8Encoder* const enc) {
@@ -332,7 +307,7 @@ int WebPReportProgress(const WebPPicture* const pic,
 //------------------------------------------------------------------------------
 
 int WebPEncode(const WebPConfig* config, WebPPicture* pic) {
-  int ok;
+  int ok = 0;
 
   if (pic == NULL)
     return 0;
@@ -346,44 +321,63 @@ int WebPEncode(const WebPConfig* config, WebPPicture* pic) {
   if (pic->width > WEBP_MAX_DIMENSION || pic->height > WEBP_MAX_DIMENSION)
     return WebPEncodingSetError(pic, VP8_ENC_ERROR_BAD_DIMENSION);
 
+  if (!config->exact) {
+    WebPCleanupTransparentArea(pic);
+  }
+
   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) {
-      if (pic->argb != NULL) {
-        if (!WebPPictureARGBToYUVA(pic, WEBP_YUV420)) return 0;
+    if (pic->use_argb || pic->y == NULL || pic->u == NULL || pic->v == NULL) {
+      // Make sure we have YUVA samples.
+      if (config->preprocessing & 4) {
+        if (!WebPPictureSmartARGBToYUVA(pic)) {
+          return 0;
+        }
       } else {
-        return WebPEncodingSetError(pic, VP8_ENC_ERROR_NULL_PARAMETER);
+        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)
-      && VP8StatLoop(enc)
-      && VP8EncLoop(enc)
-      && VP8EncFinishAlpha(enc)
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-      && VP8EncFinishLayer(enc)
-#endif
-      && VP8EncWrite(enc);
+    ok = VP8EncAnalyze(enc);
+
+    // Analysis is done, proceed to actual coding.
+    ok = ok && VP8EncStartAlpha(enc);   // possibly done in parallel
+    if (!enc->use_tokens_) {
+      ok = ok && VP8EncLoop(enc);
+    } else {
+      ok = ok && VP8EncTokenLoop(enc);
+    }
+    ok = ok && VP8EncFinishAlpha(enc);
+
+    ok = ok && VP8EncWrite(enc);
     StoreStats(enc);
     if (!ok) {
       VP8EncFreeBitWriters(enc);
     }
-    DeleteVP8Encoder(enc);
+    ok &= DeleteVP8Encoder(enc);  // must always be called, even if !ok
   } else {
-    if (pic->argb == NULL)
-      return WebPEncodingSetError(pic, VP8_ENC_ERROR_NULL_PARAMETER);
+    // 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;
 }
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/encode.h b/drivers/webp/encode.h
index 2e37cfabe7..c382ea7608 100644
--- a/drivers/webp/encode.h
+++ b/drivers/webp/encode.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 interface
@@ -14,11 +16,22 @@
 
 #include "./types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-#define WEBP_ENCODER_ABI_VERSION 0x0200    // MAJOR(8b) + MINOR(8b)
+#define WEBP_ENCODER_ABI_VERSION 0x0209    // MAJOR(8b) + MINOR(8b)
+
+// Note: forward declaring enumerations is not allowed in (strict) C and C++,
+// the types are left here for reference.
+// typedef enum WebPImageHint WebPImageHint;
+// typedef enum WebPEncCSP WebPEncCSP;
+// typedef enum WebPPreset WebPPreset;
+// typedef enum WebPEncodingError WebPEncodingError;
+typedef struct WebPConfig WebPConfig;
+typedef struct WebPPicture WebPPicture;   // main structure for I/O
+typedef struct WebPAuxStats WebPAuxStats;
+typedef struct WebPMemoryWriter WebPMemoryWriter;
 
 // Return the encoder's version number, packed in hexadecimal using 8bits for
 // each of major/minor/revision. E.g: v2.5.7 is 0x020507.
@@ -29,7 +42,7 @@ WEBP_EXTERN(int) WebPGetEncoderVersion(void);
 
 // Returns the size of the compressed data (pointed to by *output), or 0 if
 // an error occurred. The compressed data must be released by the caller
-// using the call 'free(*output)'.
+// using the call 'WebPFree(*output)'.
 // These functions compress using the lossy format, and the quality_factor
 // can go from 0 (smaller output, lower quality) to 100 (best quality,
 // larger output).
@@ -62,11 +75,14 @@ WEBP_EXTERN(size_t) WebPEncodeLosslessBGRA(const uint8_t* bgra,
                                            int width, int height, int stride,
                                            uint8_t** output);
 
+// Releases memory returned by the WebPEncode*() functions above.
+WEBP_EXTERN(void) WebPFree(void* ptr);
+
 //------------------------------------------------------------------------------
 // Coding parameters
 
 // Image characteristics hint for the underlying encoder.
-typedef enum {
+typedef enum WebPImageHint {
   WEBP_HINT_DEFAULT = 0,  // default preset.
   WEBP_HINT_PICTURE,      // digital picture, like portrait, inner shot
   WEBP_HINT_PHOTO,        // outdoor photograph, with natural lighting
@@ -74,7 +90,8 @@ typedef enum {
   WEBP_HINT_LAST
 } WebPImageHint;
 
-typedef struct {
+// Compression parameters.
+struct WebPConfig {
   int lossless;           // Lossless encoding (0=lossy(default), 1=lossless).
   float quality;          // between 0 (smallest file) and 100 (biggest)
   int method;             // quality/speed trade-off (0=fast, 6=slower-better)
@@ -103,19 +120,38 @@ typedef struct {
 
   int show_compressed;    // if true, export the compressed picture back.
                           // In-loop filtering is not applied.
-  int preprocessing;      // preprocessing filter (0=none, 1=segment-smooth)
+  int preprocessing;      // preprocessing filter:
+                          // 0=none, 1=segment-smooth, 2=pseudo-random dithering
   int partitions;         // log2(number of token partitions) in [0..3]. Default
                           // is set to 0 for easier progressive decoding.
   int partition_limit;    // quality degradation allowed to fit the 512k limit
                           // on prediction modes coding (0: no degradation,
                           // 100: maximum possible degradation).
-
-  uint32_t pad[8];        // padding for later use
-} WebPConfig;
+  int emulate_jpeg_size;  // If true, compression parameters will be remapped
+                          // to better match the expected output size from
+                          // JPEG compression. Generally, the output size will
+                          // be similar but the degradation will be lower.
+  int thread_level;       // If non-zero, try and use multi-threaded encoding.
+  int low_memory;         // If set, reduce memory usage (but increase CPU use).
+
+  int near_lossless;      // Near lossless encoding [0 = off(default) .. 100].
+                          // This feature is experimental.
+  int exact;              // if non-zero, preserve the exact RGB values under
+                          // transparent area. Otherwise, discard this invisible
+                          // RGB information for better compression. The default
+                          // value is 0.
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  int delta_palettization;
+  uint32_t pad[2];        // padding for later use
+#else
+  uint32_t pad[3];        // padding for later use
+#endif  // WEBP_EXPERIMENTAL_FEATURES
+};
 
 // Enumerate some predefined settings for WebPConfig, depending on the type
 // of source picture. These presets are used when calling WebPConfigPreset().
-typedef enum {
+typedef enum WebPPreset {
   WEBP_PRESET_DEFAULT = 0,  // default preset.
   WEBP_PRESET_PICTURE,      // digital picture, like portrait, inner shot
   WEBP_PRESET_PHOTO,        // outdoor photograph, with natural lighting
@@ -146,17 +182,23 @@ static WEBP_INLINE int WebPConfigPreset(WebPConfig* config,
                                 WEBP_ENCODER_ABI_VERSION);
 }
 
+// Activate the lossless compression mode with the desired efficiency level
+// between 0 (fastest, lowest compression) and 9 (slower, best compression).
+// A good default level is '6', providing a fair tradeoff between compression
+// speed and final compressed size.
+// This function will overwrite several fields from config: 'method', 'quality'
+// and 'lossless'. Returns false in case of parameter error.
+WEBP_EXTERN(int) WebPConfigLosslessPreset(WebPConfig* config, int level);
+
 // Returns true if 'config' is non-NULL and all configuration parameters are
 // within their valid ranges.
 WEBP_EXTERN(int) WebPValidateConfig(const WebPConfig* config);
 
 //------------------------------------------------------------------------------
 // Input / Output
-
-typedef struct WebPPicture WebPPicture;   // main structure for I/O
-
 // Structure for storing auxiliary statistics (mostly for lossy encoding).
-typedef struct {
+
+struct WebPAuxStats {
   int coded_size;         // final size
 
   float PSNR[5];          // peak-signal-to-noise ratio for Y/U/V/All/Alpha
@@ -180,9 +222,11 @@ typedef struct {
   int cache_bits;              // number of bits for color cache lookup
   int palette_size;            // number of color in palette, if used
   int lossless_size;           // final lossless size
+  int lossless_hdr_size;       // lossless header (transform, huffman etc) size
+  int lossless_data_size;      // lossless image data size
 
-  uint32_t pad[4];        // padding for later use
-} WebPAuxStats;
+  uint32_t pad[2];        // padding for later use
+};
 
 // Signature for output function. Should return true if writing was successful.
 // data/data_size is the segment of data to write, and 'picture' is for
@@ -192,18 +236,22 @@ typedef int (*WebPWriterFunction)(const uint8_t* data, size_t data_size,
 
 // WebPMemoryWrite: a special WebPWriterFunction that writes to memory using
 // the following WebPMemoryWriter object (to be set as a custom_ptr).
-typedef struct {
+struct WebPMemoryWriter {
   uint8_t* mem;       // final buffer (of size 'max_size', larger than 'size').
   size_t   size;      // final size
   size_t   max_size;  // total capacity
   uint32_t pad[1];    // padding for later use
-} WebPMemoryWriter;
+};
 
 // The following must be called first before any use.
 WEBP_EXTERN(void) WebPMemoryWriterInit(WebPMemoryWriter* writer);
 
+// The following must be called to deallocate writer->mem memory. The 'writer'
+// object itself is not deallocated.
+WEBP_EXTERN(void) WebPMemoryWriterClear(WebPMemoryWriter* writer);
 // The custom writer to be used with WebPMemoryWriter as custom_ptr. Upon
 // completion, writer.mem and writer.size will hold the coded data.
+// writer.mem must be freed by calling WebPMemoryWriterClear.
 WEBP_EXTERN(int) WebPMemoryWrite(const uint8_t* data, size_t data_size,
                                  const WebPPicture* picture);
 
@@ -212,23 +260,17 @@ WEBP_EXTERN(int) WebPMemoryWrite(const uint8_t* data, size_t data_size,
 // everything is OK.
 typedef int (*WebPProgressHook)(int percent, const WebPPicture* picture);
 
-typedef enum {
+// Color spaces.
+typedef enum WebPEncCSP {
   // chroma sampling
-  WEBP_YUV420 = 0,   // 4:2:0
-  WEBP_YUV422 = 1,   // 4:2:2
-  WEBP_YUV444 = 2,   // 4:4:4
-  WEBP_YUV400 = 3,   // grayscale
-  WEBP_CSP_UV_MASK = 3,   // bit-mask to get the UV sampling factors
-  // alpha channel variants
-  WEBP_YUV420A = 4,
-  WEBP_YUV422A = 5,
-  WEBP_YUV444A = 6,
-  WEBP_YUV400A = 7,   // grayscale + alpha
+  WEBP_YUV420  = 0,        // 4:2:0
+  WEBP_YUV420A = 4,        // alpha channel variant
+  WEBP_CSP_UV_MASK = 3,    // bit-mask to get the UV sampling factors
   WEBP_CSP_ALPHA_BIT = 4   // bit that is set if alpha is present
 } WebPEncCSP;
 
 // Encoding error conditions.
-typedef enum {
+typedef enum WebPEncodingError {
   VP8_ENC_OK = 0,
   VP8_ENC_ERROR_OUT_OF_MEMORY,            // memory error allocating objects
   VP8_ENC_ERROR_BITSTREAM_OUT_OF_MEMORY,  // memory error while flushing bits
@@ -248,7 +290,6 @@ typedef enum {
 
 // Main exchange structure (input samples, output bytes, statistics)
 struct WebPPicture {
-
   //   INPUT
   //////////////
   // Main flag for encoder selecting between ARGB or YUV input.
@@ -303,17 +344,15 @@ struct WebPPicture {
 
   uint32_t pad3[3];       // padding for later use
 
-  // Unused for now: original samples (for non-YUV420 modes)
-  uint8_t *u0, *v0;
-  int uv0_stride;
-
-  uint32_t pad4[7];       // padding for later use
+  // Unused for now
+  uint8_t *pad4, *pad5;
+  uint32_t pad6[8];       // padding for later use
 
   // PRIVATE FIELDS
   ////////////////////
   void* memory_;          // row chunk of memory for yuva planes
   void* memory_argb_;     // and for argb too.
-  void* pad5[2];          // padding for later use
+  void* pad7[2];          // padding for later use
 };
 
 // Internal, version-checked, entry point
@@ -343,18 +382,19 @@ WEBP_EXTERN(int) WebPPictureAlloc(WebPPicture* picture);
 // preserved.
 WEBP_EXTERN(void) WebPPictureFree(WebPPicture* picture);
 
-// Copy the pixels of *src into *dst, using WebPPictureAlloc. Upon return,
-// *dst will fully own the copied pixels (this is not a view).
+// Copy the pixels of *src into *dst, using WebPPictureAlloc. Upon return, *dst
+// will fully own the copied pixels (this is not a view). The 'dst' picture need
+// not be initialized as its content is overwritten.
 // Returns false in case of memory allocation error.
 WEBP_EXTERN(int) WebPPictureCopy(const WebPPicture* src, WebPPicture* dst);
 
-// Compute PSNR or SSIM distortion between two pictures.
-// Result is in dB, stores in result[] in the Y/U/V/Alpha/All order.
-// Returns false in case of error (pic1 and pic2 don't have same dimension, ...)
+// Compute PSNR, SSIM or LSIM distortion metric between two pictures. Results
+// are in dB, stored in result[] in the Y/U/V/Alpha/All or B/G/R/A/All order.
+// Returns false in case of error (src and ref don't have same dimension, ...)
 // Warning: this function is rather CPU-intensive.
 WEBP_EXTERN(int) WebPPictureDistortion(
-    const WebPPicture* pic1, const WebPPicture* pic2,
-    int metric_type,           // 0 = PSNR, 1 = SSIM
+    const WebPPicture* src, const WebPPicture* ref,
+    int metric_type,           // 0 = PSNR, 1 = SSIM, 2 = LSIM
     float result[5]);
 
 // self-crops a picture to the rectangle defined by top/left/width/height.
@@ -375,7 +415,9 @@ WEBP_EXTERN(int) WebPPictureCrop(WebPPicture* picture,
 // the top and left coordinates will be snapped to even values.
 // Picture 'src' must out-live 'dst' picture. Self-extraction of view is allowed
 // ('src' equal to 'dst') as a mean of fast-cropping (but note that doing so,
-// the original dimension will be lost).
+// the original dimension will be lost). Picture 'dst' need not be initialized
+// with WebPPictureInit() if it is different from 'src', since its content will
+// be overwritten.
 // Returns false in case of memory allocation error or invalid parameters.
 WEBP_EXTERN(int) WebPPictureView(const WebPPicture* src,
                                  int left, int top, int width, int height,
@@ -386,7 +428,9 @@ WEBP_EXTERN(int) WebPPictureView(const WebPPicture* src,
 WEBP_EXTERN(int) WebPPictureIsView(const WebPPicture* picture);
 
 // Rescale a picture to new dimension width x height.
-// Now gamma correction is applied.
+// If either 'width' or 'height' (but not both) is 0 the corresponding
+// dimension will be calculated preserving the aspect ratio.
+// No gamma correction is applied.
 // Returns false in case of error (invalid parameter or insufficient memory).
 WEBP_EXTERN(int) WebPPictureRescale(WebPPicture* pic, int width, int height);
 
@@ -413,13 +457,28 @@ WEBP_EXTERN(int) WebPPictureImportBGRA(
 WEBP_EXTERN(int) WebPPictureImportBGRX(
     WebPPicture* picture, const uint8_t* bgrx, int bgrx_stride);
 
-// Converts picture->argb data to the YUVA format specified by 'colorspace'.
+// Converts picture->argb data to the YUV420A format. The 'colorspace'
+// parameter is deprecated and should be equal to WEBP_YUV420.
 // Upon return, picture->use_argb is set to false. The presence of real
 // non-opaque transparent values is detected, and 'colorspace' will be
 // adjusted accordingly. Note that this method is lossy.
 // Returns false in case of error.
 WEBP_EXTERN(int) WebPPictureARGBToYUVA(WebPPicture* picture,
-                                       WebPEncCSP colorspace);
+                                       WebPEncCSP /*colorspace = WEBP_YUV420*/);
+
+// Same as WebPPictureARGBToYUVA(), but the conversion is done using
+// pseudo-random dithering with a strength 'dithering' between
+// 0.0 (no dithering) and 1.0 (maximum dithering). This is useful
+// for photographic picture.
+WEBP_EXTERN(int) WebPPictureARGBToYUVADithered(
+    WebPPicture* picture, WebPEncCSP colorspace, float dithering);
+
+// Performs 'smart' RGBA->YUVA420 downsampling and colorspace conversion.
+// Downsampling is handled with extra care in case of color clipping. This
+// method is roughly 2x slower than WebPPictureARGBToYUVA() but produces better
+// YUV representation.
+// Returns false in case of error.
+WEBP_EXTERN(int) WebPPictureSmartARGBToYUVA(WebPPicture* picture);
 
 // Converts picture->yuv to picture->argb and sets picture->use_argb to true.
 // The input format must be YUV_420 or YUV_420A.
@@ -429,9 +488,9 @@ WEBP_EXTERN(int) WebPPictureARGBToYUVA(WebPPicture* picture,
 // Returns false in case of error.
 WEBP_EXTERN(int) WebPPictureYUVAToARGB(WebPPicture* picture);
 
-// Helper function: given a width x height plane of YUV(A) samples
-// (with stride 'stride'), clean-up the YUV samples under fully transparent
-// area, to help compressibility (no guarantee, though).
+// Helper function: given a width x height plane of RGBA or YUV(A) samples
+// clean-up the YUV or RGB samples under fully transparent area, to help
+// compressibility (no guarantee, though).
 WEBP_EXTERN(void) WebPCleanupTransparentArea(WebPPicture* picture);
 
 // Scan the picture 'picture' for the presence of non fully opaque alpha values.
@@ -439,6 +498,11 @@ WEBP_EXTERN(void) WebPCleanupTransparentArea(WebPPicture* picture);
 // alpha plane can be ignored altogether e.g.).
 WEBP_EXTERN(int) WebPPictureHasTransparency(const WebPPicture* picture);
 
+// Remove the transparency information (if present) by blending the color with
+// the background color 'background_rgb' (specified as 24bit RGB triplet).
+// After this call, all alpha values are reset to 0xff.
+WEBP_EXTERN(void) WebPBlendAlpha(WebPPicture* pic, uint32_t background_rgb);
+
 //------------------------------------------------------------------------------
 // Main call
 
@@ -456,7 +520,7 @@ WEBP_EXTERN(int) WebPEncode(const WebPConfig* config, WebPPicture* picture);
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/format_constants.h b/drivers/webp/format_constants.h
index 7ce498f672..b6e78a643e 100644
--- a/drivers/webp/format_constants.h
+++ b/drivers/webp/format_constants.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 //  Internal header for constants related to WebP file format.
@@ -12,6 +14,9 @@
 #ifndef WEBP_WEBP_FORMAT_CONSTANTS_H_
 #define WEBP_WEBP_FORMAT_CONSTANTS_H_
 
+// Create fourcc of the chunk from the chunk tag characters.
+#define MKFOURCC(a, b, c, d) ((a) | (b) << 8 | (c) << 16 | (uint32_t)(d) << 24)
+
 // VP8 related constants.
 #define VP8_SIGNATURE 0x9d012a              // Signature in VP8 data.
 #define VP8_MAX_PARTITION0_SIZE (1 << 19)   // max size of mode partition
@@ -65,23 +70,16 @@ typedef enum {
 #define CHUNK_SIZE_BYTES   4     // Size needed to store chunk's size.
 #define CHUNK_HEADER_SIZE  8     // Size of a chunk header.
 #define RIFF_HEADER_SIZE   12    // Size of the RIFF header ("RIFFnnnnWEBP").
-#define FRAME_CHUNK_SIZE   15    // Size of a FRM chunk.
-#define LOOP_CHUNK_SIZE    2     // Size of a LOOP chunk.
-#define TILE_CHUNK_SIZE    6     // Size of a TILE chunk.
+#define ANMF_CHUNK_SIZE    16    // Size of an ANMF chunk.
+#define ANIM_CHUNK_SIZE    6     // Size of an ANIM chunk.
+#define FRGM_CHUNK_SIZE    6     // Size of a FRGM chunk.
 #define VP8X_CHUNK_SIZE    10    // Size of a VP8X chunk.
 
-#define TILING_FLAG_BIT    0x01  // Set if tiles are possibly used.
-#define ANIMATION_FLAG_BIT 0x02  // Set if some animation is expected
-#define ICC_FLAG_BIT       0x04  // Whether ICC is present or not.
-#define METADATA_FLAG_BIT  0x08  // Set if some META chunk is possibly present.
-#define ALPHA_FLAG_BIT     0x10  // Should be same as the ALPHA_FLAG in mux.h
-#define ROTATION_FLAG_BITS 0xe0  // all 3 bits for rotation + symmetry
-
-#define MAX_CANVAS_SIZE     (1 << 24)    // 24-bit max for VP8X width/height.
-#define MAX_IMAGE_AREA      (1ULL << 32) // 32-bit max for width x height.
-#define MAX_LOOP_COUNT      (1 << 16)    // maximum value for loop-count
-#define MAX_DURATION        (1 << 24)    // maximum duration
-#define MAX_POSITION_OFFSET (1 << 24)    // maximum frame/tile x/y offset
+#define MAX_CANVAS_SIZE     (1 << 24)     // 24-bit max for VP8X width/height.
+#define MAX_IMAGE_AREA      (1ULL << 32)  // 32-bit max for width x height.
+#define MAX_LOOP_COUNT      (1 << 16)     // maximum value for loop-count
+#define MAX_DURATION        (1 << 24)     // maximum duration
+#define MAX_POSITION_OFFSET (1 << 24)     // maximum frame/fragment x/y offset
 
 // Maximum chunk payload is such that adding the header and padding won't
 // overflow a uint32_t.
diff --git a/drivers/webp/mux.h b/drivers/webp/mux.h
index 5139af80fa..1fddfb76d4 100644
--- a/drivers/webp/mux.h
+++ b/drivers/webp/mux.h
@@ -1,60 +1,76 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
-//  RIFF container manipulation for WEBP images.
+//  RIFF container manipulation and encoding for WebP images.
 //
 // Authors: Urvang (urvang@google.com)
 //          Vikas (vikasa@google.com)
 
-// This API allows manipulation of WebP container images containing features
-// like Color profile, XMP metadata, Animation and Tiling.
-//
-// Code Example#1: Creating a MUX with image data, color profile and XMP
-// metadata.
-//
-//   int copy_data = 0;
-//   WebPMux* mux = WebPMuxNew();
-//   // ... (Prepare image data).
-//   WebPMuxSetImage(mux, &image, copy_data);
-//   // ... (Prepare ICCP color profile data).
-//   WebPMuxSetColorProfile(mux, &icc_profile, copy_data);
-//   // ... (Prepare XMP metadata).
-//   WebPMuxSetMetadata(mux, &xmp, copy_data);
-//   // Get data from mux in WebP RIFF format.
-//   WebPMuxAssemble(mux, &output_data);
-//   WebPMuxDelete(mux);
-//   // ... (Consume output_data; e.g. write output_data.bytes_ to file).
-//   WebPDataClear(&output_data);
-//
-// Code Example#2: Get image and color profile data from a WebP file.
-//
-//   int copy_data = 0;
-//   // ... (Read data from file).
-//   WebPMux* mux = WebPMuxCreate(&data, copy_data);
-//   WebPMuxGetImage(mux, &image);
-//   // ... (Consume image; e.g. call WebPDecode() to decode the data).
-//   WebPMuxGetColorProfile(mux, &icc_profile);
-//   // ... (Consume icc_data).
-//   WebPMuxDelete(mux);
-//   free(data);
-
 #ifndef WEBP_WEBP_MUX_H_
 #define WEBP_WEBP_MUX_H_
 
-#include "./types.h"
+#include "./mux_types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-#define WEBP_MUX_ABI_VERSION 0x0100        // MAJOR(8b) + MINOR(8b)
+#define WEBP_MUX_ABI_VERSION 0x0106        // MAJOR(8b) + MINOR(8b)
+
+//------------------------------------------------------------------------------
+// Mux API
+//
+// This API allows manipulation of WebP container images containing features
+// like color profile, metadata, animation and fragmented images.
+//
+// Code Example#1: Create a WebPMux object with image data, color profile and
+// XMP metadata.
+/*
+  int copy_data = 0;
+  WebPMux* mux = WebPMuxNew();
+  // ... (Prepare image data).
+  WebPMuxSetImage(mux, &image, copy_data);
+  // ... (Prepare ICCP color profile data).
+  WebPMuxSetChunk(mux, "ICCP", &icc_profile, copy_data);
+  // ... (Prepare XMP metadata).
+  WebPMuxSetChunk(mux, "XMP ", &xmp, copy_data);
+  // Get data from mux in WebP RIFF format.
+  WebPMuxAssemble(mux, &output_data);
+  WebPMuxDelete(mux);
+  // ... (Consume output_data; e.g. write output_data.bytes to file).
+  WebPDataClear(&output_data);
+*/
+
+// Code Example#2: Get image and color profile data from a WebP file.
+/*
+  int copy_data = 0;
+  // ... (Read data from file).
+  WebPMux* mux = WebPMuxCreate(&data, copy_data);
+  WebPMuxGetFrame(mux, 1, &image);
+  // ... (Consume image; e.g. call WebPDecode() to decode the data).
+  WebPMuxGetChunk(mux, "ICCP", &icc_profile);
+  // ... (Consume icc_data).
+  WebPMuxDelete(mux);
+  free(data);
+*/
+
+// Note: forward declaring enumerations is not allowed in (strict) C and C++,
+// the types are left here for reference.
+// typedef enum WebPMuxError WebPMuxError;
+// typedef enum WebPChunkId WebPChunkId;
+typedef struct WebPMux WebPMux;   // main opaque object.
+typedef struct WebPMuxFrameInfo WebPMuxFrameInfo;
+typedef struct WebPMuxAnimParams WebPMuxAnimParams;
+typedef struct WebPAnimEncoderOptions WebPAnimEncoderOptions;
 
 // Error codes
-typedef enum {
+typedef enum WebPMuxError {
   WEBP_MUX_OK                 =  1,
   WEBP_MUX_NOT_FOUND          =  0,
   WEBP_MUX_INVALID_ARGUMENT   = -1,
@@ -63,51 +79,26 @@ typedef enum {
   WEBP_MUX_NOT_ENOUGH_DATA    = -4
 } WebPMuxError;
 
-// Flag values for different features used in VP8X chunk.
-typedef enum {
-  TILE_FLAG       = 0x00000001,
-  ANIMATION_FLAG  = 0x00000002,
-  ICCP_FLAG       = 0x00000004,
-  META_FLAG       = 0x00000008,
-  ALPHA_FLAG      = 0x00000010
-} WebPFeatureFlags;
-
 // IDs for different types of chunks.
-typedef enum {
+typedef enum WebPChunkId {
   WEBP_CHUNK_VP8X,     // VP8X
   WEBP_CHUNK_ICCP,     // ICCP
-  WEBP_CHUNK_LOOP,     // LOOP
-  WEBP_CHUNK_FRAME,    // FRM
-  WEBP_CHUNK_TILE,     // TILE
+  WEBP_CHUNK_ANIM,     // ANIM
+  WEBP_CHUNK_ANMF,     // ANMF
+  WEBP_CHUNK_FRGM,     // FRGM
   WEBP_CHUNK_ALPHA,    // ALPH
   WEBP_CHUNK_IMAGE,    // VP8/VP8L
-  WEBP_CHUNK_META,     // META
+  WEBP_CHUNK_EXIF,     // EXIF
+  WEBP_CHUNK_XMP,      // XMP
   WEBP_CHUNK_UNKNOWN,  // Other chunks.
   WEBP_CHUNK_NIL
 } WebPChunkId;
 
-typedef struct WebPMux WebPMux;   // main opaque object.
-
-// Data type used to describe 'raw' data, e.g., chunk data
-// (ICC profile, metadata) and WebP compressed image data.
-typedef struct {
-  const uint8_t* bytes_;
-  size_t size_;
-} WebPData;
-
 //------------------------------------------------------------------------------
-// Manipulation of a WebPData object.
-
-// Initializes the contents of the 'webp_data' object with default values.
-WEBP_EXTERN(void) WebPDataInit(WebPData* webp_data);
 
-// Clears the contents of the 'webp_data' object by calling free(). Does not
-// deallocate the object itself.
-WEBP_EXTERN(void) WebPDataClear(WebPData* webp_data);
-
-// Allocates necessary storage for 'dst' and copies the contents of 'src'.
-// Returns true on success.
-WEBP_EXTERN(int) WebPDataCopy(const WebPData* src, WebPData* dst);
+// Returns the version number of the mux library, packed in hexadecimal using
+// 8bits for each of major/minor/revision. E.g: v2.5.7 is 0x020507.
+WEBP_EXTERN(int) WebPGetMuxVersion(void);
 
 //------------------------------------------------------------------------------
 // Life of a Mux object
@@ -118,6 +109,7 @@ WEBP_EXTERN(WebPMux*) WebPNewInternal(int);
 // Creates an empty mux object.
 // Returns:
 //   A pointer to the newly created empty mux object.
+//   Or NULL in case of memory error.
 static WEBP_INLINE WebPMux* WebPMuxNew(void) {
   return WebPNewInternal(WEBP_MUX_ABI_VERSION);
 }
@@ -136,8 +128,8 @@ WEBP_EXTERN(WebPMux*) WebPMuxCreateInternal(const WebPData*, int, int);
 // Creates a mux object from raw data given in WebP RIFF format.
 // Parameters:
 //   bitstream - (in) the bitstream data in WebP RIFF format
-//   copy_data - (in) value 1 indicates given data WILL copied to the mux, and
-//               value 0 indicates data will NOT be copied.
+//   copy_data - (in) value 1 indicates given data WILL be copied to the mux
+//               object and value 0 indicates data will NOT be copied.
 // Returns:
 //   A pointer to the mux object created from given data - on success.
 //   NULL - In case of invalid data or memory error.
@@ -147,295 +139,237 @@ static WEBP_INLINE WebPMux* WebPMuxCreate(const WebPData* bitstream,
 }
 
 //------------------------------------------------------------------------------
-// Single Image.
-
-// Sets the image in the mux object. Any existing images (including frame/tile)
-// will be removed.
-// Parameters:
-//   mux - (in/out) object in which the image is to be set
-//   bitstream - (in) can either be a raw VP8/VP8L bitstream or a single-image
-//               WebP file (non-animated and non-tiled)
-//   copy_data - (in) value 1 indicates given data WILL copied to the mux, and
-//               value 0 indicates data will NOT be copied.
-// Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL or bitstream is NULL.
-//   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
-//   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxSetImage(WebPMux* mux,
-                                          const WebPData* bitstream,
-                                          int copy_data);
-
-// Gets image data from the mux object.
-// The content of 'bitstream' is allocated using malloc(), and NOT
-// owned by the 'mux' object. It MUST be deallocated by the caller by calling
-// WebPDataClear().
-// Parameters:
-//   mux - (in) object from which the image is to be fetched
-//   bitstream - (out) the image data
-// Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if either mux or bitstream is NULL
-//                               OR mux contains animation/tiling.
-//   WEBP_MUX_NOT_FOUND - if image is not present in mux object.
-//   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxGetImage(const WebPMux* mux,
-                                          WebPData* bitstream);
-
-// Deletes the image in the mux object.
-// Parameters:
-//   mux - (in/out) object from which the image is to be deleted
-// Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL
-//                               OR if mux contains animation/tiling.
-//   WEBP_MUX_NOT_FOUND - if image is not present in mux object.
-//   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxDeleteImage(WebPMux* mux);
+// Non-image chunks.
 
-//------------------------------------------------------------------------------
-// XMP Metadata.
+// Note: Only non-image related chunks should be managed through chunk APIs.
+// (Image related chunks are: "ANMF", "FRGM", "VP8 ", "VP8L" and "ALPH").
+// To add, get and delete images, use WebPMuxSetImage(), WebPMuxPushFrame(),
+// WebPMuxGetFrame() and WebPMuxDeleteFrame().
 
-// Sets the XMP metadata in the mux object. Any existing metadata chunk(s) will
-// be removed.
+// Adds a chunk with id 'fourcc' and data 'chunk_data' in the mux object.
+// Any existing chunk(s) with the same id will be removed.
 // Parameters:
-//   mux - (in/out) object to which the XMP metadata is to be added
-//   metadata - (in) the XMP metadata data to be added
-//   copy_data - (in) value 1 indicates given data WILL copied to the mux, and
-//               value 0 indicates data will NOT be copied.
+//   mux - (in/out) object to which the chunk is to be added
+//   fourcc - (in) a character array containing the fourcc of the given chunk;
+//                 e.g., "ICCP", "XMP ", "EXIF" etc.
+//   chunk_data - (in) the chunk data to be added
+//   copy_data - (in) value 1 indicates given data WILL be copied to the mux
+//               object and value 0 indicates data will NOT be copied.
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux or metadata is NULL.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux, fourcc or chunk_data is NULL
+//                               or if fourcc corresponds to an image chunk.
 //   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
 //   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxSetMetadata(WebPMux* mux,
-                                             const WebPData* metadata,
-                                             int copy_data);
+WEBP_EXTERN(WebPMuxError) WebPMuxSetChunk(
+    WebPMux* mux, const char fourcc[4], const WebPData* chunk_data,
+    int copy_data);
 
-// Gets a reference to the XMP metadata in the mux object.
+// Gets a reference to the data of the chunk with id 'fourcc' in the mux object.
 // The caller should NOT free the returned data.
 // Parameters:
-//   mux - (in) object from which the XMP metadata is to be fetched
-//   metadata - (out) XMP metadata
+//   mux - (in) object from which the chunk data is to be fetched
+//   fourcc - (in) a character array containing the fourcc of the chunk;
+//                 e.g., "ICCP", "XMP ", "EXIF" etc.
+//   chunk_data - (out) returned chunk data
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if either mux or metadata is NULL.
-//   WEBP_MUX_NOT_FOUND - if metadata is not present in mux object.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux, fourcc or chunk_data is NULL
+//                               or if fourcc corresponds to an image chunk.
+//   WEBP_MUX_NOT_FOUND - If mux does not contain a chunk with the given id.
 //   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxGetMetadata(const WebPMux* mux,
-                                             WebPData* metadata);
+WEBP_EXTERN(WebPMuxError) WebPMuxGetChunk(
+    const WebPMux* mux, const char fourcc[4], WebPData* chunk_data);
 
-// Deletes the XMP metadata in the mux object.
+// Deletes the chunk with the given 'fourcc' from the mux object.
 // Parameters:
-//   mux - (in/out) object from which XMP metadata is to be deleted
+//   mux - (in/out) object from which the chunk is to be deleted
+//   fourcc - (in) a character array containing the fourcc of the chunk;
+//                 e.g., "ICCP", "XMP ", "EXIF" etc.
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL
-//   WEBP_MUX_NOT_FOUND - If mux does not contain metadata.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux or fourcc is NULL
+//                               or if fourcc corresponds to an image chunk.
+//   WEBP_MUX_NOT_FOUND - If mux does not contain a chunk with the given fourcc.
 //   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxDeleteMetadata(WebPMux* mux);
+WEBP_EXTERN(WebPMuxError) WebPMuxDeleteChunk(
+    WebPMux* mux, const char fourcc[4]);
 
 //------------------------------------------------------------------------------
-// ICC Color Profile.
-
-// Sets the color profile in the mux object. Any existing color profile chunk(s)
-// will be removed.
+// Images.
+
+// Encapsulates data about a single frame/fragment.
+struct WebPMuxFrameInfo {
+  WebPData    bitstream;  // image data: can be a raw VP8/VP8L bitstream
+                          // or a single-image WebP file.
+  int         x_offset;   // x-offset of the frame.
+  int         y_offset;   // y-offset of the frame.
+  int         duration;   // duration of the frame (in milliseconds).
+
+  WebPChunkId id;         // frame type: should be one of WEBP_CHUNK_ANMF,
+                          // WEBP_CHUNK_FRGM or WEBP_CHUNK_IMAGE
+  WebPMuxAnimDispose dispose_method;  // Disposal method for the frame.
+  WebPMuxAnimBlend   blend_method;    // Blend operation for the frame.
+  uint32_t    pad[1];     // padding for later use
+};
+
+// Sets the (non-animated and non-fragmented) image in the mux object.
+// Note: Any existing images (including frames/fragments) will be removed.
 // Parameters:
-//   mux - (in/out) object to which the color profile is to be added
-//   color_profile - (in) the color profile data to be added
-//   copy_data - (in) value 1 indicates given data WILL copied to the mux, and
-//               value 0 indicates data will NOT be copied.
-// Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux or color_profile is NULL
-//   WEBP_MUX_MEMORY_ERROR - on memory allocation error
-//   WEBP_MUX_OK - on success
-WEBP_EXTERN(WebPMuxError) WebPMuxSetColorProfile(WebPMux* mux,
-                                                 const WebPData* color_profile,
-                                                 int copy_data);
-
-// Gets a reference to the color profile in the mux object.
-// The caller should NOT free the returned data.
-// Parameters:
-//   mux - (in) object from which the color profile data is to be fetched
-//   color_profile - (out) color profile data
+//   mux - (in/out) object in which the image is to be set
+//   bitstream - (in) can be a raw VP8/VP8L bitstream or a single-image
+//               WebP file (non-animated and non-fragmented)
+//   copy_data - (in) value 1 indicates given data WILL be copied to the mux
+//               object and value 0 indicates data will NOT be copied.
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if either mux or color_profile is NULL.
-//   WEBP_MUX_NOT_FOUND - if color profile is not present in mux object.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL or bitstream is NULL.
+//   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
 //   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxGetColorProfile(const WebPMux* mux,
-                                                 WebPData* color_profile);
-
-// Deletes the color profile in the mux object.
+WEBP_EXTERN(WebPMuxError) WebPMuxSetImage(
+    WebPMux* mux, const WebPData* bitstream, int copy_data);
+
+// Adds a frame at the end of the mux object.
+// Notes: (1) frame.id should be one of WEBP_CHUNK_ANMF or WEBP_CHUNK_FRGM
+//        (2) For setting a non-animated non-fragmented image, use
+//            WebPMuxSetImage() instead.
+//        (3) Type of frame being pushed must be same as the frames in mux.
+//        (4) As WebP only supports even offsets, any odd offset will be snapped
+//            to an even location using: offset &= ~1
 // Parameters:
-//   mux - (in/out) object from which color profile is to be deleted
+//   mux - (in/out) object to which the frame is to be added
+//   frame - (in) frame data.
+//   copy_data - (in) value 1 indicates given data WILL be copied to the mux
+//               object and value 0 indicates data will NOT be copied.
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL
-//   WEBP_MUX_NOT_FOUND - If mux does not contain color profile.
-//   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxDeleteColorProfile(WebPMux* mux);
-
-//------------------------------------------------------------------------------
-// Animation.
-
-// Adds an animation frame at the end of the mux object.
-// Note: as WebP only supports even offsets, any odd offset will be snapped to
-// an even location using: offset &= ~1
-// Parameters:
-//   mux - (in/out) object to which an animation frame is to be added
-//   bitstream - (in) the image data corresponding to the frame. It can either
-//               be a raw VP8/VP8L bitstream or a single-image WebP file
-//               (non-animated and non-tiled)
-//   x_offset - (in) x-offset of the frame to be added
-//   y_offset - (in) y-offset of the frame to be added
-//   duration - (in) duration of the frame to be added (in milliseconds)
-//   copy_data - (in) value 1 indicates given data WILL copied to the mux, and
-//               value 0 indicates data will NOT be copied.
-// Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL or bitstream is NULL
+//   WEBP_MUX_INVALID_ARGUMENT - if mux or frame is NULL
+//                               or if content of 'frame' is invalid.
 //   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
 //   WEBP_MUX_OK - on success.
 WEBP_EXTERN(WebPMuxError) WebPMuxPushFrame(
-    WebPMux* mux, const WebPData* bitstream,
-    int x_offset, int y_offset, int duration, int copy_data);
-
-// TODO(urvang): Create a struct as follows to reduce argument list size:
-// typedef struct {
-//  WebPData bitstream;
-//  int x_offset, y_offset;
-//  int duration;
-// } FrameInfo;
-
-// Gets the nth animation frame from the mux object.
-// The content of 'bitstream' is allocated using malloc(), and NOT
+    WebPMux* mux, const WebPMuxFrameInfo* frame, int copy_data);
+
+// Gets the nth frame from the mux object.
+// The content of 'frame->bitstream' is allocated using malloc(), and NOT
 // owned by the 'mux' object. It MUST be deallocated by the caller by calling
 // WebPDataClear().
 // nth=0 has a special meaning - last position.
 // Parameters:
 //   mux - (in) object from which the info is to be fetched
 //   nth - (in) index of the frame in the mux object
-//   bitstream - (out) the image data
-//   x_offset - (out) x-offset of the returned frame
-//   y_offset - (out) y-offset of the returned frame
-//   duration - (out) duration of the returned frame (in milliseconds)
+//   frame - (out) data of the returned frame
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if either mux, bitstream, x_offset,
-//                               y_offset, or duration is NULL
+//   WEBP_MUX_INVALID_ARGUMENT - if mux or frame is NULL.
 //   WEBP_MUX_NOT_FOUND - if there are less than nth frames in the mux object.
 //   WEBP_MUX_BAD_DATA - if nth frame chunk in mux is invalid.
+//   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
 //   WEBP_MUX_OK - on success.
 WEBP_EXTERN(WebPMuxError) WebPMuxGetFrame(
-    const WebPMux* mux, uint32_t nth, WebPData* bitstream,
-    int* x_offset, int* y_offset, int* duration);
+    const WebPMux* mux, uint32_t nth, WebPMuxFrameInfo* frame);
 
-// Deletes an animation frame from the mux object.
+// Deletes a frame from the mux object.
 // nth=0 has a special meaning - last position.
 // Parameters:
 //   mux - (in/out) object from which a frame is to be deleted
 //   nth - (in) The position from which the frame is to be deleted
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL
+//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL.
 //   WEBP_MUX_NOT_FOUND - If there are less than nth frames in the mux object
 //                        before deletion.
 //   WEBP_MUX_OK - on success.
 WEBP_EXTERN(WebPMuxError) WebPMuxDeleteFrame(WebPMux* mux, uint32_t nth);
 
-// Sets the animation loop count in the mux object. Any existing loop count
-// value(s) will be removed.
+//------------------------------------------------------------------------------
+// Animation.
+
+// Animation parameters.
+struct WebPMuxAnimParams {
+  uint32_t bgcolor;  // Background color of the canvas stored (in MSB order) as:
+                     // Bits 00 to 07: Alpha.
+                     // Bits 08 to 15: Red.
+                     // Bits 16 to 23: Green.
+                     // Bits 24 to 31: Blue.
+  int loop_count;    // Number of times to repeat the animation [0 = infinite].
+};
+
+// Sets the animation parameters in the mux object. Any existing ANIM chunks
+// will be removed.
 // Parameters:
-//   mux - (in/out) object in which loop chunk is to be set/added
-//   loop_count - (in) animation loop count value.
-//                Note that loop_count of zero denotes infinite loop.
+//   mux - (in/out) object in which ANIM chunk is to be set/added
+//   params - (in) animation parameters.
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL
+//   WEBP_MUX_INVALID_ARGUMENT - if mux or params is NULL.
 //   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
 //   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxSetLoopCount(WebPMux* mux, int loop_count);
+WEBP_EXTERN(WebPMuxError) WebPMuxSetAnimationParams(
+    WebPMux* mux, const WebPMuxAnimParams* params);
 
-// Gets the animation loop count from the mux object.
+// Gets the animation parameters from the mux object.
 // Parameters:
-//   mux - (in) object from which the loop count is to be fetched
-//   loop_count - (out) the loop_count value present in the LOOP chunk
+//   mux - (in) object from which the animation parameters to be fetched
+//   params - (out) animation parameters extracted from the ANIM chunk
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if either of mux or loop_count is NULL
-//   WEBP_MUX_NOT_FOUND - if loop chunk is not present in mux object.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux or params is NULL.
+//   WEBP_MUX_NOT_FOUND - if ANIM chunk is not present in mux object.
 //   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxGetLoopCount(const WebPMux* mux,
-                                              int* loop_count);
+WEBP_EXTERN(WebPMuxError) WebPMuxGetAnimationParams(
+    const WebPMux* mux, WebPMuxAnimParams* params);
 
 //------------------------------------------------------------------------------
-// Tiling.
-
-// Adds a tile at the end of the mux object.
-// Note: as WebP only supports even offsets, any odd offset will be snapped to
-// an even location using: offset &= ~1
-// Parameters:
-//   mux - (in/out) object to which a tile is to be added.
-//   bitstream - (in) the image data corresponding to the frame. It can either
-//               be a raw VP8/VP8L bitstream or a single-image WebP file
-//               (non-animated and non-tiled)
-//   x_offset - (in) x-offset of the tile to be added
-//   y_offset - (in) y-offset of the tile to be added
-//   copy_data - (in) value 1 indicates given data WILL copied to the mux, and
-//               value 0 indicates data will NOT be copied.
-// Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL or bitstream is NULL
-//   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
-//   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxPushTile(
-    WebPMux* mux, const WebPData* bitstream,
-    int x_offset, int y_offset, int copy_data);
+// Misc Utilities.
 
-// Gets the nth tile from the mux object.
-// The content of 'bitstream' is allocated using malloc(), and NOT
-// owned by the 'mux' object. It MUST be deallocated by the caller by calling
-// WebPDataClear().
-// nth=0 has a special meaning - last position.
+// Sets the canvas size for the mux object. The width and height can be
+// specified explicitly or left as zero (0, 0).
+// * When width and height are specified explicitly, then this frame bound is
+//   enforced during subsequent calls to WebPMuxAssemble() and an error is
+//   reported if any animated frame does not completely fit within the canvas.
+// * When unspecified (0, 0), the constructed canvas will get the frame bounds
+//   from the bounding-box over all frames after calling WebPMuxAssemble().
 // Parameters:
-//   mux - (in) object from which the info is to be fetched
-//   nth - (in) index of the tile in the mux object
-//   bitstream - (out) the image data
-//   x_offset - (out) x-offset of the returned tile
-//   y_offset - (out) y-offset of the returned tile
+//   mux - (in) object to which the canvas size is to be set
+//   width - (in) canvas width
+//   height - (in) canvas height
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if either mux, bitstream, x_offset or
-//                               y_offset is NULL
-//   WEBP_MUX_NOT_FOUND - if there are less than nth tiles in the mux object.
-//   WEBP_MUX_BAD_DATA - if nth tile chunk in mux is invalid.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL; or
+//                               width or height are invalid or out of bounds
 //   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxGetTile(
-    const WebPMux* mux, uint32_t nth, WebPData* bitstream,
-    int* x_offset, int* y_offset);
+WEBP_EXTERN(WebPMuxError) WebPMuxSetCanvasSize(WebPMux* mux,
+                                               int width, int height);
 
-// Deletes a tile from the mux object.
-// nth=0 has a special meaning - last position
+// Gets the canvas size from the mux object.
+// Note: This method assumes that the VP8X chunk, if present, is up-to-date.
+// That is, the mux object hasn't been modified since the last call to
+// WebPMuxAssemble() or WebPMuxCreate().
 // Parameters:
-//   mux - (in/out) object from which a tile is to be deleted
-//   nth - (in) The position from which the tile is to be deleted
+//   mux - (in) object from which the canvas size is to be fetched
+//   width - (out) canvas width
+//   height - (out) canvas height
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux is NULL
-//   WEBP_MUX_NOT_FOUND - If there are less than nth tiles in the mux object
-//                        before deletion.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux, width or height is NULL.
+//   WEBP_MUX_BAD_DATA - if VP8X/VP8/VP8L chunk or canvas size is invalid.
 //   WEBP_MUX_OK - on success.
-WEBP_EXTERN(WebPMuxError) WebPMuxDeleteTile(WebPMux* mux, uint32_t nth);
-
-//------------------------------------------------------------------------------
-// Misc Utilities.
+WEBP_EXTERN(WebPMuxError) WebPMuxGetCanvasSize(const WebPMux* mux,
+                                               int* width, int* height);
 
 // Gets the feature flags from the mux object.
+// Note: This method assumes that the VP8X chunk, if present, is up-to-date.
+// That is, the mux object hasn't been modified since the last call to
+// WebPMuxAssemble() or WebPMuxCreate().
 // Parameters:
 //   mux - (in) object from which the features are to be fetched
 //   flags - (out) the flags specifying which features are present in the
 //           mux object. This will be an OR of various flag values.
 //           Enum 'WebPFeatureFlags' can be used to test individual flag values.
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if mux or flags is NULL
-//   WEBP_MUX_NOT_FOUND - if VP8X chunk is not present in mux object.
-//   WEBP_MUX_BAD_DATA - if VP8X chunk in mux is invalid.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux or flags is NULL.
+//   WEBP_MUX_BAD_DATA - if VP8X/VP8/VP8L chunk or canvas size is invalid.
 //   WEBP_MUX_OK - on success.
 WEBP_EXTERN(WebPMuxError) WebPMuxGetFeatures(const WebPMux* mux,
                                              uint32_t* flags);
 
-// Gets number of chunks having tag value tag in the mux object.
+// Gets number of chunks with the given 'id' in the mux object.
 // Parameters:
 //   mux - (in) object from which the info is to be fetched
 //   id - (in) chunk id specifying the type of chunk
 //   num_elements - (out) number of chunks with the given chunk id
 // Returns:
-//   WEBP_MUX_INVALID_ARGUMENT - if either mux, or num_elements is NULL
+//   WEBP_MUX_INVALID_ARGUMENT - if mux, or num_elements is NULL.
 //   WEBP_MUX_OK - on success.
 WEBP_EXTERN(WebPMuxError) WebPMuxNumChunks(const WebPMux* mux,
                                            WebPChunkId id, int* num_elements);
@@ -445,159 +379,151 @@ WEBP_EXTERN(WebPMuxError) WebPMuxNumChunks(const WebPMux* mux,
 // Note: The content of 'assembled_data' will be ignored and overwritten.
 // Also, the content of 'assembled_data' is allocated using malloc(), and NOT
 // owned by the 'mux' object. It MUST be deallocated by the caller by calling
-// WebPDataClear().
+// WebPDataClear(). It's always safe to call WebPDataClear() upon return,
+// even in case of error.
 // Parameters:
 //   mux - (in/out) object whose chunks are to be assembled
 //   assembled_data - (out) assembled WebP data
 // Returns:
 //   WEBP_MUX_BAD_DATA - if mux object is invalid.
-//   WEBP_MUX_INVALID_ARGUMENT - if either mux, output_data or output_size is
-//                               NULL.
+//   WEBP_MUX_INVALID_ARGUMENT - if mux or assembled_data is NULL.
 //   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
-//   WEBP_MUX_OK - on success
+//   WEBP_MUX_OK - on success.
 WEBP_EXTERN(WebPMuxError) WebPMuxAssemble(WebPMux* mux,
                                           WebPData* assembled_data);
 
 //------------------------------------------------------------------------------
-// Demux API.
-// Enables extraction of image and extended format data from WebP files.
-
-#define WEBP_DEMUX_ABI_VERSION 0x0100    // MAJOR(8b) + MINOR(8b)
-
-typedef struct WebPDemuxer WebPDemuxer;
-
-typedef enum {
-  WEBP_DEMUX_PARSING_HEADER,  // Not enough data to parse full header.
-  WEBP_DEMUX_PARSED_HEADER,   // Header parsing complete, data may be available.
-  WEBP_DEMUX_DONE             // Entire file has been parsed.
-} WebPDemuxState;
-
-//------------------------------------------------------------------------------
-// Life of a Demux object
-
-// Internal, version-checked, entry point
-WEBP_EXTERN(WebPDemuxer*) WebPDemuxInternal(
-    const WebPData*, int, WebPDemuxState*, int);
-
-// Parses the WebP file given by 'data'.
-// A complete WebP file must be present in 'data' for the function to succeed.
-// Returns a WebPDemuxer object on successful parse, NULL otherwise.
-static WEBP_INLINE WebPDemuxer* WebPDemux(const WebPData* data) {
-  return WebPDemuxInternal(data, 0, NULL, WEBP_DEMUX_ABI_VERSION);
-}
-
-// Parses the WebP file given by 'data'.
-// If 'state' is non-NULL it will be set to indicate the status of the demuxer.
-// Returns a WebPDemuxer object on successful parse, NULL otherwise.
-static WEBP_INLINE WebPDemuxer* WebPDemuxPartial(
-    const WebPData* data, WebPDemuxState* state) {
-  return WebPDemuxInternal(data, 1, state, WEBP_DEMUX_ABI_VERSION);
+// WebPAnimEncoder API
+//
+// This API allows encoding (possibly) animated WebP images.
+//
+// Code Example:
+/*
+  WebPAnimEncoderOptions enc_options;
+  WebPAnimEncoderOptionsInit(&enc_options);
+  // Tune 'enc_options' as needed.
+  WebPAnimEncoder* enc = WebPAnimEncoderNew(width, height, &enc_options);
+  while(<there are more frames>) {
+    WebPConfig config;
+    WebPConfigInit(&config);
+    // Tune 'config' as needed.
+    WebPAnimEncoderAdd(enc, frame, timestamp_ms, &config);
+  }
+  WebPAnimEncoderAdd(enc, NULL, timestamp_ms, NULL);
+  WebPAnimEncoderAssemble(enc, webp_data);
+  WebPAnimEncoderDelete(enc);
+  // Write the 'webp_data' to a file, or re-mux it further.
+*/
+
+typedef struct WebPAnimEncoder WebPAnimEncoder;  // Main opaque object.
+
+// Forward declarations. Defined in encode.h.
+struct WebPPicture;
+struct WebPConfig;
+
+// Global options.
+struct WebPAnimEncoderOptions {
+  WebPMuxAnimParams anim_params;  // Animation parameters.
+  int minimize_size;    // If true, minimize the output size (slow). Implicitly
+                        // disables key-frame insertion.
+  int kmin;
+  int kmax;             // Minimum and maximum distance between consecutive key
+                        // frames in the output. The library may insert some key
+                        // frames as needed to satisfy this criteria.
+                        // Note that these conditions should hold: kmax > kmin
+                        // and kmin >= kmax / 2 + 1. Also, if kmin == 0, then
+                        // key-frame insertion is disabled; and if kmax == 0,
+                        // then all frames will be key-frames.
+  int allow_mixed;      // If true, use mixed compression mode; may choose
+                        // either lossy and lossless for each frame.
+  int verbose;          // If true, print info and warning messages to stderr.
+
+  uint32_t padding[4];  // Padding for later use.
+};
+
+// Internal, version-checked, entry point.
+WEBP_EXTERN(int) WebPAnimEncoderOptionsInitInternal(
+    WebPAnimEncoderOptions*, int);
+
+// Should always be called, to initialize a fresh WebPAnimEncoderOptions
+// structure before modification. Returns false in case of version mismatch.
+// WebPAnimEncoderOptionsInit() must have succeeded before using the
+// 'enc_options' object.
+static WEBP_INLINE int WebPAnimEncoderOptionsInit(
+    WebPAnimEncoderOptions* enc_options) {
+  return WebPAnimEncoderOptionsInitInternal(enc_options, WEBP_MUX_ABI_VERSION);
 }
 
-// Frees memory associated with 'dmux'.
-WEBP_EXTERN(void) WebPDemuxDelete(WebPDemuxer* dmux);
-
-//------------------------------------------------------------------------------
-// Data/information extraction.
+// Internal, version-checked, entry point.
+WEBP_EXTERN(WebPAnimEncoder*) WebPAnimEncoderNewInternal(
+    int, int, const WebPAnimEncoderOptions*, int);
 
-typedef enum {
-  WEBP_FF_FORMAT_FLAGS,  // Extended format flags present in the 'VP8X' chunk.
-  WEBP_FF_CANVAS_WIDTH,
-  WEBP_FF_CANVAS_HEIGHT,
-  WEBP_FF_LOOP_COUNT
-} WebPFormatFeature;
+// Creates and initializes a WebPAnimEncoder object.
+// Parameters:
+//   width/height - (in) canvas width and height of the animation.
+//   enc_options - (in) encoding options; can be passed NULL to pick
+//                     reasonable defaults.
+// Returns:
+//   A pointer to the newly created WebPAnimEncoder object.
+//   Or NULL in case of memory error.
+static WEBP_INLINE WebPAnimEncoder* WebPAnimEncoderNew(
+    int width, int height, const WebPAnimEncoderOptions* enc_options) {
+  return WebPAnimEncoderNewInternal(width, height, enc_options,
+                                    WEBP_MUX_ABI_VERSION);
+}
 
-// Get the 'feature' value from the 'dmux'.
-// NOTE: values are only valid if WebPDemux() was used or WebPDemuxPartial()
-// returned a state > WEBP_DEMUX_PARSING_HEADER.
-WEBP_EXTERN(uint32_t) WebPDemuxGetI(
-    const WebPDemuxer* dmux, WebPFormatFeature feature);
+// Optimize the given frame for WebP, encode it and add it to the
+// WebPAnimEncoder object.
+// The last call to 'WebPAnimEncoderAdd' should be with frame = NULL, which
+// indicates that no more frames are to be added. This call is also used to
+// determine the duration of the last frame.
+// Parameters:
+//   enc - (in/out) object to which the frame is to be added.
+//   frame - (in/out) frame data in ARGB or YUV(A) format. If it is in YUV(A)
+//           format, it will be converted to ARGB, which incurs a small loss.
+//   timestamp_ms - (in) timestamp of this frame in milliseconds.
+//                       Duration of a frame would be calculated as
+//                       "timestamp of next frame - timestamp of this frame".
+//                       Hence, timestamps should be in non-decreasing order.
+//   config - (in) encoding options; can be passed NULL to pick
+//            reasonable defaults.
+// Returns:
+//   On error, returns false and frame->error_code is set appropriately.
+//   Otherwise, returns true.
+WEBP_EXTERN(int) WebPAnimEncoderAdd(
+    WebPAnimEncoder* enc, struct WebPPicture* frame, int timestamp_ms,
+    const struct WebPConfig* config);
+
+// Assemble all frames added so far into a WebP bitstream.
+// This call should be preceded by  a call to 'WebPAnimEncoderAdd' with
+// frame = NULL; if not, the duration of the last frame will be internally
+// estimated.
+// Parameters:
+//   enc - (in/out) object from which the frames are to be assembled.
+//   webp_data - (out) generated WebP bitstream.
+// Returns:
+//   True on success.
+WEBP_EXTERN(int) WebPAnimEncoderAssemble(WebPAnimEncoder* enc,
+                                         WebPData* webp_data);
 
-//------------------------------------------------------------------------------
-// Frame iteration.
-
-typedef struct {
-  int frame_num_;
-  int num_frames_;
-  int tile_num_;
-  int num_tiles_;
-  int x_offset_, y_offset_;  // offset relative to the canvas.
-  int width_, height_;       // dimensions of this frame or tile.
-  int duration_;   // display duration in milliseconds.
-  int complete_;   // true if 'tile_' contains a full frame. partial images may
-                   // still be decoded with the WebP incremental decoder.
-  WebPData tile_;  // The frame or tile given by 'frame_num_' and 'tile_num_'.
-
-  uint32_t pad[4];           // padding for later use
-  void* private_;
-} WebPIterator;
-
-// Retrieves frame 'frame_number' from 'dmux'.
-// 'iter->tile_' points to the first tile on return from this function.
-// Individual tiles may be extracted using WebPDemuxSetTile().
-// Setting 'frame_number' equal to 0 will return the last frame of the image.
-// Returns false if 'dmux' is NULL or frame 'frame_number' is not present.
-// Call WebPDemuxReleaseIterator() when use of the iterator is complete.
-// NOTE: 'dmux' must persist for the lifetime of 'iter'.
-WEBP_EXTERN(int) WebPDemuxGetFrame(
-    const WebPDemuxer* dmux, int frame_number, WebPIterator* iter);
-
-// Sets 'iter->tile_' to point to the next ('iter->frame_num_' + 1) or previous
-// ('iter->frame_num_' - 1) frame. These functions do not loop.
-// Returns true on success, false otherwise.
-WEBP_EXTERN(int) WebPDemuxNextFrame(WebPIterator* iter);
-WEBP_EXTERN(int) WebPDemuxPrevFrame(WebPIterator* iter);
-
-// Sets 'iter->tile_' to reflect tile number 'tile_number'.
-// Returns true if tile 'tile_number' is present, false otherwise.
-WEBP_EXTERN(int) WebPDemuxSelectTile(WebPIterator* iter, int tile_number);
-
-// Releases any memory associated with 'iter'.
-// Must be called before destroying the associated WebPDemuxer with
-// WebPDemuxDelete().
-WEBP_EXTERN(void) WebPDemuxReleaseIterator(WebPIterator* iter);
+// Get error string corresponding to the most recent call using 'enc'. The
+// returned string is owned by 'enc' and is valid only until the next call to
+// WebPAnimEncoderAdd() or WebPAnimEncoderAssemble() or WebPAnimEncoderDelete().
+// Parameters:
+//   enc - (in/out) object from which the error string is to be fetched.
+// Returns:
+//   NULL if 'enc' is NULL. Otherwise, returns the error string if the last call
+//   to 'enc' had an error, or an empty string if the last call was a success.
+WEBP_EXTERN(const char*) WebPAnimEncoderGetError(WebPAnimEncoder* enc);
 
-//------------------------------------------------------------------------------
-// Chunk iteration.
-
-typedef struct {
-  // The current and total number of chunks with the fourcc given to
-  // WebPDemuxGetChunk().
-  int chunk_num_;
-  int num_chunks_;
-  WebPData chunk_;    // The payload of the chunk.
-
-  uint32_t pad[6];    // padding for later use
-  void* private_;
-} WebPChunkIterator;
-
-// Retrieves the 'chunk_number' instance of the chunk with id 'fourcc' from
-// 'dmux'.
-// 'fourcc' is a character array containing the fourcc of the chunk to return,
-// e.g., "ICCP", "META", "EXIF", etc.
-// Setting 'chunk_number' equal to 0 will return the last chunk in a set.
-// Returns true if the chunk is found, false otherwise. Image related chunk
-// payloads are accessed through WebPDemuxGetFrame() and related functions.
-// Call WebPDemuxReleaseChunkIterator() when use of the iterator is complete.
-// NOTE: 'dmux' must persist for the lifetime of the iterator.
-WEBP_EXTERN(int) WebPDemuxGetChunk(const WebPDemuxer* dmux,
-                                   const char fourcc[4], int chunk_number,
-                                   WebPChunkIterator* iter);
-
-// Sets 'iter->chunk_' to point to the next ('iter->chunk_num_' + 1) or previous
-// ('iter->chunk_num_' - 1) chunk. These functions do not loop.
-// Returns true on success, false otherwise.
-WEBP_EXTERN(int) WebPDemuxNextChunk(WebPChunkIterator* iter);
-WEBP_EXTERN(int) WebPDemuxPrevChunk(WebPChunkIterator* iter);
-
-// Releases any memory associated with 'iter'.
-// Must be called before destroying the associated WebPDemuxer with
-// WebPDemuxDelete().
-WEBP_EXTERN(void) WebPDemuxReleaseChunkIterator(WebPChunkIterator* iter);
+// Deletes the WebPAnimEncoder object.
+// Parameters:
+//   enc - (in/out) object to be deleted
+WEBP_EXTERN(void) WebPAnimEncoderDelete(WebPAnimEncoder* enc);
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/mux/demux.c b/drivers/webp/mux/demux.c
deleted file mode 100644
index 501d08f41d..0000000000
--- a/drivers/webp/mux/demux.c
+++ /dev/null
@@ -1,902 +0,0 @@
-// Copyright 2012 Google Inc. All Rights Reserved.
-//
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
-// -----------------------------------------------------------------------------
-//
-//  WebP container demux.
-//
-
-#include "../mux.h"
-
-#include <stdlib.h>
-#include <string.h>
-
-#include "../decode.h"  // WebPGetInfo
-#include "../format_constants.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-#define MKFOURCC(a, b, c, d) ((uint32_t)(a) | (b) << 8 | (c) << 16 | (d) << 24)
-
-typedef struct {
-  size_t start_;        // start location of the data
-  size_t end_;          // end location
-  size_t riff_end_;     // riff chunk end location, can be > end_.
-  size_t buf_size_;     // size of the buffer
-  const uint8_t* buf_;
-} MemBuffer;
-
-typedef struct {
-  size_t offset_;
-  size_t size_;
-} ChunkData;
-
-typedef struct Frame {
-  int x_offset_, y_offset_;
-  int width_, height_;
-  int duration_;
-  int is_tile_;    // this is an image fragment from a 'TILE'.
-  int frame_num_;  // the referent frame number for use in assembling tiles.
-  int complete_;   // img_components_ contains a full image.
-  ChunkData img_components_[2];  // 0=VP8{,L} 1=ALPH
-  struct Frame* next_;
-} Frame;
-
-typedef struct Chunk {
-  ChunkData data_;
-  struct Chunk* next_;
-} Chunk;
-
-struct WebPDemuxer {
-  MemBuffer mem_;
-  WebPDemuxState state_;
-  int is_ext_format_;
-  uint32_t feature_flags_;
-  int canvas_width_, canvas_height_;
-  int loop_count_;
-  int num_frames_;
-  Frame* frames_;
-  Chunk* chunks_;  // non-image chunks
-};
-
-typedef enum {
-  PARSE_OK,
-  PARSE_NEED_MORE_DATA,
-  PARSE_ERROR
-} ParseStatus;
-
-typedef struct ChunkParser {
-  uint8_t id[4];
-  ParseStatus (*parse)(WebPDemuxer* const dmux);
-  int (*valid)(const WebPDemuxer* const dmux);
-} ChunkParser;
-
-static ParseStatus ParseSingleImage(WebPDemuxer* const dmux);
-static ParseStatus ParseVP8X(WebPDemuxer* const dmux);
-static int IsValidSimpleFormat(const WebPDemuxer* const dmux);
-static int IsValidExtendedFormat(const WebPDemuxer* const dmux);
-
-static const ChunkParser kMasterChunks[] = {
-  { { 'V', 'P', '8', ' ' }, ParseSingleImage, IsValidSimpleFormat },
-  { { 'V', 'P', '8', 'L' }, ParseSingleImage, IsValidSimpleFormat },
-  { { 'V', 'P', '8', 'X' }, ParseVP8X,        IsValidExtendedFormat },
-  { { '0', '0', '0', '0' }, NULL,             NULL },
-};
-
-// -----------------------------------------------------------------------------
-// MemBuffer
-
-static int RemapMemBuffer(MemBuffer* const mem,
-                          const uint8_t* data, size_t size) {
-  if (size < mem->buf_size_) return 0;  // can't remap to a shorter buffer!
-
-  mem->buf_ = data;
-  mem->end_ = mem->buf_size_ = size;
-  return 1;
-}
-
-static int InitMemBuffer(MemBuffer* const mem,
-                         const uint8_t* data, size_t size) {
-  memset(mem, 0, sizeof(*mem));
-  return RemapMemBuffer(mem, data, size);
-}
-
-// Return the remaining data size available in 'mem'.
-static WEBP_INLINE size_t MemDataSize(const MemBuffer* const mem) {
-  return (mem->end_ - mem->start_);
-}
-
-// Return true if 'size' exceeds the end of the RIFF chunk.
-static WEBP_INLINE int SizeIsInvalid(const MemBuffer* const mem, size_t size) {
-  return (size > mem->riff_end_ - mem->start_);
-}
-
-static WEBP_INLINE void Skip(MemBuffer* const mem, size_t size) {
-  mem->start_ += size;
-}
-
-static WEBP_INLINE void Rewind(MemBuffer* const mem, size_t size) {
-  mem->start_ -= size;
-}
-
-static WEBP_INLINE const uint8_t* GetBuffer(MemBuffer* const mem) {
-  return mem->buf_ + mem->start_;
-}
-
-static WEBP_INLINE uint8_t GetByte(MemBuffer* const mem) {
-  const uint8_t byte = mem->buf_[mem->start_];
-  Skip(mem, 1);
-  return byte;
-}
-
-// Read 16, 24 or 32 bits stored in little-endian order.
-static WEBP_INLINE int ReadLE16s(const uint8_t* const data) {
-  return (int)(data[0] << 0) | (data[1] << 8);
-}
-
-static WEBP_INLINE int ReadLE24s(const uint8_t* const data) {
-  return ReadLE16s(data) | (data[2] << 16);
-}
-
-static WEBP_INLINE uint32_t ReadLE32(const uint8_t* const data) {
-  return (uint32_t)ReadLE24s(data) | (data[3] << 24);
-}
-
-// In addition to reading, skip the read bytes.
-static WEBP_INLINE int GetLE16s(MemBuffer* const mem) {
-  const uint8_t* const data = mem->buf_ + mem->start_;
-  const int val = ReadLE16s(data);
-  Skip(mem, 2);
-  return val;
-}
-
-static WEBP_INLINE int GetLE24s(MemBuffer* const mem) {
-  const uint8_t* const data = mem->buf_ + mem->start_;
-  const int val = ReadLE24s(data);
-  Skip(mem, 3);
-  return val;
-}
-
-static WEBP_INLINE uint32_t GetLE32(MemBuffer* const mem) {
-  const uint8_t* const data = mem->buf_ + mem->start_;
-  const uint32_t val = ReadLE32(data);
-  Skip(mem, 4);
-  return val;
-}
-
-// -----------------------------------------------------------------------------
-// Secondary chunk parsing
-
-static void AddChunk(WebPDemuxer* const dmux, Chunk* const chunk) {
-  Chunk** c = &dmux->chunks_;
-  while (*c != NULL) c = &(*c)->next_;
-  *c = chunk;
-  chunk->next_ = NULL;
-}
-
-// Add a frame to the end of the list, ensuring the last frame is complete.
-// Returns true on success, false otherwise.
-static int AddFrame(WebPDemuxer* const dmux, Frame* const frame) {
-  const Frame* last_frame = NULL;
-  Frame** f = &dmux->frames_;
-  while (*f != NULL) {
-    last_frame = *f;
-    f = &(*f)->next_;
-  }
-  if (last_frame != NULL && !last_frame->complete_) return 0;
-  *f = frame;
-  frame->next_ = NULL;
-  return 1;
-}
-
-// Store image bearing chunks to 'frame'.
-static ParseStatus StoreFrame(int frame_num, MemBuffer* const mem,
-                              Frame* const frame) {
-  int alpha_chunks = 0;
-  int image_chunks = 0;
-  int done = (MemDataSize(mem) < CHUNK_HEADER_SIZE);
-  ParseStatus status = PARSE_OK;
-
-  if (done) return PARSE_NEED_MORE_DATA;
-
-  do {
-    const size_t chunk_start_offset = mem->start_;
-    const uint32_t fourcc = GetLE32(mem);
-    const uint32_t payload_size = GetLE32(mem);
-    const uint32_t payload_size_padded = payload_size + (payload_size & 1);
-    const size_t payload_available = (payload_size_padded > MemDataSize(mem))
-                                   ? MemDataSize(mem) : payload_size_padded;
-    const size_t chunk_size = CHUNK_HEADER_SIZE + payload_available;
-
-    if (payload_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
-    if (SizeIsInvalid(mem, payload_size_padded)) return PARSE_ERROR;
-    if (payload_size_padded > MemDataSize(mem)) status = PARSE_NEED_MORE_DATA;
-
-    switch (fourcc) {
-      case MKFOURCC('A', 'L', 'P', 'H'):
-        if (alpha_chunks == 0) {
-          ++alpha_chunks;
-          frame->img_components_[1].offset_ = chunk_start_offset;
-          frame->img_components_[1].size_ = chunk_size;
-          frame->frame_num_ = frame_num;
-          Skip(mem, payload_available);
-        } else {
-          goto Done;
-        }
-        break;
-      case MKFOURCC('V', 'P', '8', ' '):
-      case MKFOURCC('V', 'P', '8', 'L'):
-        if (image_chunks == 0) {
-          int width = 0, height = 0;
-          ++image_chunks;
-          frame->img_components_[0].offset_ = chunk_start_offset;
-          frame->img_components_[0].size_ = chunk_size;
-          // Extract the width and height from the bitstream, tolerating
-          // failures when the data is incomplete.
-          if (!WebPGetInfo(mem->buf_ + frame->img_components_[0].offset_,
-                           frame->img_components_[0].size_, &width, &height) &&
-              status != PARSE_NEED_MORE_DATA) {
-            return PARSE_ERROR;
-          }
-
-          frame->width_ = width;
-          frame->height_ = height;
-          frame->frame_num_ = frame_num;
-          frame->complete_ = (status == PARSE_OK);
-          Skip(mem, payload_available);
-        } else {
-          goto Done;
-        }
-        break;
- Done:
-      default:
-        // Restore fourcc/size when moving up one level in parsing.
-        Rewind(mem, CHUNK_HEADER_SIZE);
-        done = 1;
-        break;
-    }
-
-    if (mem->start_ == mem->riff_end_) {
-      done = 1;
-    } else if (MemDataSize(mem) < CHUNK_HEADER_SIZE) {
-      status = PARSE_NEED_MORE_DATA;
-    }
-  } while (!done && status == PARSE_OK);
-
-  return status;
-}
-
-// Creates a new Frame if 'actual_size' is within bounds and 'mem' contains
-// enough data ('min_size') to parse the payload.
-// Returns PARSE_OK on success with *frame pointing to the new Frame.
-// Returns PARSE_NEED_MORE_DATA with insufficient data, PARSE_ERROR otherwise.
-static ParseStatus NewFrame(const MemBuffer* const mem,
-                            uint32_t min_size, uint32_t expected_size,
-                            uint32_t actual_size, Frame** frame) {
-  if (SizeIsInvalid(mem, min_size)) return PARSE_ERROR;
-  if (actual_size < expected_size) return PARSE_ERROR;
-  if (MemDataSize(mem) < min_size)  return PARSE_NEED_MORE_DATA;
-
-  *frame = (Frame*)calloc(1, sizeof(**frame));
-  return (*frame == NULL) ? PARSE_ERROR : PARSE_OK;
-}
-
-// Parse a 'FRM ' chunk and any image bearing chunks that immediately follow.
-// 'frame_chunk_size' is the previously validated, padded chunk size.
-static ParseStatus ParseFrame(
-    WebPDemuxer* const dmux, uint32_t frame_chunk_size) {
-  const int has_frames = !!(dmux->feature_flags_ & ANIMATION_FLAG);
-  const uint32_t min_size = frame_chunk_size + CHUNK_HEADER_SIZE;
-  int added_frame = 0;
-  MemBuffer* const mem = &dmux->mem_;
-  Frame* frame;
-  ParseStatus status =
-      NewFrame(mem, min_size, FRAME_CHUNK_SIZE, frame_chunk_size, &frame);
-  if (status != PARSE_OK) return status;
-
-  frame->x_offset_ = 2 * GetLE24s(mem);
-  frame->y_offset_ = 2 * GetLE24s(mem);
-  frame->width_    = 1 + GetLE24s(mem);
-  frame->height_   = 1 + GetLE24s(mem);
-  frame->duration_ = 1 + GetLE24s(mem);
-  Skip(mem, frame_chunk_size - FRAME_CHUNK_SIZE);  // skip any trailing data.
-  if (frame->width_ * (uint64_t)frame->height_ >= MAX_IMAGE_AREA) {
-    return PARSE_ERROR;
-  }
-
-  // Store a (potentially partial) frame only if the animation flag is set
-  // and there is some data in 'frame'.
-  status = StoreFrame(dmux->num_frames_ + 1, mem, frame);
-  if (status != PARSE_ERROR && has_frames && frame->frame_num_ > 0) {
-    added_frame = AddFrame(dmux, frame);
-    if (added_frame) {
-      ++dmux->num_frames_;
-    } else {
-      status = PARSE_ERROR;
-    }
-  }
-
-  if (!added_frame) free(frame);
-  return status;
-}
-
-// Parse a 'TILE' chunk and any image bearing chunks that immediately follow.
-// 'tile_chunk_size' is the previously validated, padded chunk size.
-static ParseStatus ParseTile(WebPDemuxer* const dmux,
-                             uint32_t tile_chunk_size) {
-  const int has_tiles = !!(dmux->feature_flags_ & TILE_FLAG);
-  const uint32_t min_size = tile_chunk_size + CHUNK_HEADER_SIZE;
-  int added_tile = 0;
-  MemBuffer* const mem = &dmux->mem_;
-  Frame* frame;
-  ParseStatus status =
-      NewFrame(mem, min_size, TILE_CHUNK_SIZE, tile_chunk_size, &frame);
-  if (status != PARSE_OK) return status;
-
-  frame->is_tile_  = 1;
-  frame->x_offset_ = 2 * GetLE24s(mem);
-  frame->y_offset_ = 2 * GetLE24s(mem);
-  Skip(mem, tile_chunk_size - TILE_CHUNK_SIZE);  // skip any trailing data.
-
-  // Store a (potentially partial) tile only if the tile flag is set
-  // and the tile contains some data.
-  status = StoreFrame(dmux->num_frames_, mem, frame);
-  if (status != PARSE_ERROR && has_tiles && frame->frame_num_ > 0) {
-    // Note num_frames_ is incremented only when all tiles have been consumed.
-    added_tile = AddFrame(dmux, frame);
-    if (!added_tile) status = PARSE_ERROR;
-  }
-
-  if (!added_tile) free(frame);
-  return status;
-}
-
-// General chunk storage starting with the header at 'start_offset' allowing
-// the user to request the payload via a fourcc string. 'size' includes the
-// header and the unpadded payload size.
-// Returns true on success, false otherwise.
-static int StoreChunk(WebPDemuxer* const dmux,
-                      size_t start_offset, uint32_t size) {
-  Chunk* const chunk = (Chunk*)calloc(1, sizeof(*chunk));
-  if (chunk == NULL) return 0;
-
-  chunk->data_.offset_ = start_offset;
-  chunk->data_.size_ = size;
-  AddChunk(dmux, chunk);
-  return 1;
-}
-
-// -----------------------------------------------------------------------------
-// Primary chunk parsing
-
-static int ReadHeader(MemBuffer* const mem) {
-  const size_t min_size = RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE;
-  uint32_t riff_size;
-
-  // Basic file level validation.
-  if (MemDataSize(mem) < min_size) return 0;
-  if (memcmp(GetBuffer(mem), "RIFF", CHUNK_SIZE_BYTES) ||
-      memcmp(GetBuffer(mem) + CHUNK_HEADER_SIZE, "WEBP", CHUNK_SIZE_BYTES)) {
-    return 0;
-  }
-
-  riff_size = ReadLE32(GetBuffer(mem) + TAG_SIZE);
-  if (riff_size < CHUNK_HEADER_SIZE) return 0;
-  if (riff_size > MAX_CHUNK_PAYLOAD) return 0;
-
-  // There's no point in reading past the end of the RIFF chunk
-  mem->riff_end_ = riff_size + CHUNK_HEADER_SIZE;
-  if (mem->buf_size_ > mem->riff_end_) {
-    mem->buf_size_ = mem->end_ = mem->riff_end_;
-  }
-
-  Skip(mem, RIFF_HEADER_SIZE);
-  return 1;
-}
-
-static ParseStatus ParseSingleImage(WebPDemuxer* const dmux) {
-  const size_t min_size = CHUNK_HEADER_SIZE;
-  MemBuffer* const mem = &dmux->mem_;
-  Frame* frame;
-  ParseStatus status;
-
-  if (dmux->frames_ != NULL) return PARSE_ERROR;
-  if (SizeIsInvalid(mem, min_size)) return PARSE_ERROR;
-  if (MemDataSize(mem) < min_size) return PARSE_NEED_MORE_DATA;
-
-  frame = (Frame*)calloc(1, sizeof(*frame));
-  if (frame == NULL) return PARSE_ERROR;
-
-  status = StoreFrame(1, &dmux->mem_, frame);
-  if (status != PARSE_ERROR) {
-    const int has_alpha = !!(dmux->feature_flags_ & ALPHA_FLAG);
-    // Clear any alpha when the alpha flag is missing.
-    if (!has_alpha && frame->img_components_[1].size_ > 0) {
-      frame->img_components_[1].offset_ = 0;
-      frame->img_components_[1].size_ = 0;
-    }
-
-    // Use the frame width/height as the canvas values for non-vp8x files.
-    if (!dmux->is_ext_format_ && frame->width_ > 0 && frame->height_ > 0) {
-      dmux->state_ = WEBP_DEMUX_PARSED_HEADER;
-      dmux->canvas_width_ = frame->width_;
-      dmux->canvas_height_ = frame->height_;
-    }
-    AddFrame(dmux, frame);
-    dmux->num_frames_ = 1;
-  } else {
-    free(frame);
-  }
-
-  return status;
-}
-
-static ParseStatus ParseVP8X(WebPDemuxer* const dmux) {
-  MemBuffer* const mem = &dmux->mem_;
-  int loop_chunks = 0;
-  uint32_t vp8x_size;
-  ParseStatus status = PARSE_OK;
-
-  if (MemDataSize(mem) < CHUNK_HEADER_SIZE) return PARSE_NEED_MORE_DATA;
-
-  dmux->is_ext_format_ = 1;
-  Skip(mem, TAG_SIZE);  // VP8X
-  vp8x_size = GetLE32(mem);
-  if (vp8x_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
-  if (vp8x_size < VP8X_CHUNK_SIZE) return PARSE_ERROR;
-  vp8x_size += vp8x_size & 1;
-  if (SizeIsInvalid(mem, vp8x_size)) return PARSE_ERROR;
-  if (MemDataSize(mem) < vp8x_size) return PARSE_NEED_MORE_DATA;
-
-  dmux->feature_flags_ = GetByte(mem);
-  Skip(mem, 3);  // Reserved.
-  dmux->canvas_width_  = 1 + GetLE24s(mem);
-  dmux->canvas_height_ = 1 + GetLE24s(mem);
-  if (dmux->canvas_width_ * (uint64_t)dmux->canvas_height_ >= MAX_IMAGE_AREA) {
-    return PARSE_ERROR;  // image final dimension is too large
-  }
-  Skip(mem, vp8x_size - VP8X_CHUNK_SIZE);  // skip any trailing data.
-  dmux->state_ = WEBP_DEMUX_PARSED_HEADER;
-
-  if (SizeIsInvalid(mem, CHUNK_HEADER_SIZE)) return PARSE_ERROR;
-  if (MemDataSize(mem) < CHUNK_HEADER_SIZE) return PARSE_NEED_MORE_DATA;
-
-  do {
-    int store_chunk = 1;
-    const size_t chunk_start_offset = mem->start_;
-    const uint32_t fourcc = GetLE32(mem);
-    const uint32_t chunk_size = GetLE32(mem);
-    const uint32_t chunk_size_padded = chunk_size + (chunk_size & 1);
-
-    if (chunk_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
-    if (SizeIsInvalid(mem, chunk_size_padded)) return PARSE_ERROR;
-
-    switch (fourcc) {
-      case MKFOURCC('V', 'P', '8', 'X'): {
-        return PARSE_ERROR;
-      }
-      case MKFOURCC('A', 'L', 'P', 'H'):
-      case MKFOURCC('V', 'P', '8', ' '):
-      case MKFOURCC('V', 'P', '8', 'L'): {
-        Rewind(mem, CHUNK_HEADER_SIZE);
-        status = ParseSingleImage(dmux);
-        break;
-      }
-      case MKFOURCC('L', 'O', 'O', 'P'): {
-        if (chunk_size_padded < LOOP_CHUNK_SIZE) return PARSE_ERROR;
-
-        if (MemDataSize(mem) < chunk_size_padded) {
-          status = PARSE_NEED_MORE_DATA;
-        } else if (loop_chunks == 0) {
-          ++loop_chunks;
-          dmux->loop_count_ = GetLE16s(mem);
-          Skip(mem, chunk_size_padded - LOOP_CHUNK_SIZE);
-        } else {
-          store_chunk = 0;
-          goto Skip;
-        }
-        break;
-      }
-      case MKFOURCC('F', 'R', 'M', ' '): {
-        status = ParseFrame(dmux, chunk_size_padded);
-        break;
-      }
-      case MKFOURCC('T', 'I', 'L', 'E'): {
-        if (dmux->num_frames_ == 0) dmux->num_frames_ = 1;
-        status = ParseTile(dmux, chunk_size_padded);
-        break;
-      }
-      case MKFOURCC('I', 'C', 'C', 'P'): {
-        store_chunk = !!(dmux->feature_flags_ & ICCP_FLAG);
-        goto Skip;
-      }
-      case MKFOURCC('M', 'E', 'T', 'A'): {
-        store_chunk = !!(dmux->feature_flags_ & META_FLAG);
-        goto Skip;
-      }
- Skip:
-      default: {
-        if (chunk_size_padded <= MemDataSize(mem)) {
-          if (store_chunk) {
-            // Store only the chunk header and unpadded size as only the payload
-            // will be returned to the user.
-            if (!StoreChunk(dmux, chunk_start_offset,
-                            CHUNK_HEADER_SIZE + chunk_size)) {
-              return PARSE_ERROR;
-            }
-          }
-          Skip(mem, chunk_size_padded);
-        } else {
-          status = PARSE_NEED_MORE_DATA;
-        }
-      }
-    }
-
-    if (mem->start_ == mem->riff_end_) {
-      break;
-    } else if (MemDataSize(mem) < CHUNK_HEADER_SIZE) {
-      status = PARSE_NEED_MORE_DATA;
-    }
-  } while (status == PARSE_OK);
-
-  return status;
-}
-
-// -----------------------------------------------------------------------------
-// Format validation
-
-static int IsValidSimpleFormat(const WebPDemuxer* const dmux) {
-  const Frame* const frame = dmux->frames_;
-  if (dmux->state_ == WEBP_DEMUX_PARSING_HEADER) return 1;
-
-  if (dmux->canvas_width_ <= 0 || dmux->canvas_height_ <= 0) return 0;
-  if (dmux->state_ == WEBP_DEMUX_DONE && frame == NULL) return 0;
-
-  if (frame->width_ <= 0 || frame->height_ <= 0) return 0;
-  return 1;
-}
-
-static int IsValidExtendedFormat(const WebPDemuxer* const dmux) {
-  const int has_tiles = !!(dmux->feature_flags_ & TILE_FLAG);
-  const int has_frames = !!(dmux->feature_flags_ & ANIMATION_FLAG);
-  const Frame* f;
-
-  if (dmux->state_ == WEBP_DEMUX_PARSING_HEADER) return 1;
-
-  if (dmux->canvas_width_ <= 0 || dmux->canvas_height_ <= 0) return 0;
-  if (dmux->loop_count_ < 0) return 0;
-  if (dmux->state_ == WEBP_DEMUX_DONE && dmux->frames_ == NULL) return 0;
-
-  for (f = dmux->frames_; f != NULL; f = f->next_) {
-    const int cur_frame_set = f->frame_num_;
-    int frame_count = 0, tile_count = 0;
-
-    // Check frame properties and if the image is composed of tiles that each
-    // fragment came from a 'TILE'.
-    for (; f != NULL && f->frame_num_ == cur_frame_set; f = f->next_) {
-      const ChunkData* const image = f->img_components_;
-      const ChunkData* const alpha = f->img_components_ + 1;
-
-      if (!has_tiles && f->is_tile_) return 0;
-      if (!has_frames && f->frame_num_ > 1) return 0;
-      if (f->x_offset_ < 0 || f->y_offset_ < 0) return 0;
-      if (f->complete_) {
-        if (alpha->size_ == 0 && image->size_ == 0) return 0;
-        // Ensure alpha precedes image bitstream.
-        if (alpha->size_ > 0 && alpha->offset_ > image->offset_) {
-          return 0;
-        }
-
-        if (f->width_ <= 0 || f->height_ <= 0) return 0;
-      } else {
-        // Ensure alpha precedes image bitstream.
-        if (alpha->size_ > 0 && image->size_ > 0 &&
-            alpha->offset_ > image->offset_) {
-          return 0;
-        }
-        // There shouldn't be any frames after an incomplete one.
-        if (f->next_ != NULL) return 0;
-      }
-
-      tile_count += f->is_tile_;
-      ++frame_count;
-    }
-    if (!has_tiles && frame_count > 1) return 0;
-    if (tile_count > 0 && frame_count != tile_count) return 0;
-    if (f == NULL) break;
-  }
-  return 1;
-}
-
-// -----------------------------------------------------------------------------
-// WebPDemuxer object
-
-static void InitDemux(WebPDemuxer* const dmux, const MemBuffer* const mem) {
-  dmux->state_ = WEBP_DEMUX_PARSING_HEADER;
-  dmux->loop_count_ = 1;
-  dmux->canvas_width_ = -1;
-  dmux->canvas_height_ = -1;
-  dmux->mem_ = *mem;
-}
-
-WebPDemuxer* WebPDemuxInternal(const WebPData* data, int allow_partial,
-                               WebPDemuxState* state, int version) {
-  const ChunkParser* parser;
-  int partial;
-  ParseStatus status = PARSE_ERROR;
-  MemBuffer mem;
-  WebPDemuxer* dmux;
-
-  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_DEMUX_ABI_VERSION)) return NULL;
-  if (data == NULL || data->bytes_ == NULL || data->size_ == 0) return NULL;
-
-  if (!InitMemBuffer(&mem, data->bytes_, data->size_)) return NULL;
-  if (!ReadHeader(&mem)) return NULL;
-
-  partial = (mem.buf_size_ < mem.riff_end_);
-  if (!allow_partial && partial) return NULL;
-
-  dmux = (WebPDemuxer*)calloc(1, sizeof(*dmux));
-  if (dmux == NULL) return NULL;
-  InitDemux(dmux, &mem);
-
-  for (parser = kMasterChunks; parser->parse != NULL; ++parser) {
-    if (!memcmp(parser->id, GetBuffer(&dmux->mem_), TAG_SIZE)) {
-      status = parser->parse(dmux);
-      if (status == PARSE_OK) dmux->state_ = WEBP_DEMUX_DONE;
-      if (status != PARSE_ERROR && !parser->valid(dmux)) status = PARSE_ERROR;
-      break;
-    }
-  }
-  if (state) *state = dmux->state_;
-
-  if (status == PARSE_ERROR) {
-    WebPDemuxDelete(dmux);
-    return NULL;
-  }
-  return dmux;
-}
-
-void WebPDemuxDelete(WebPDemuxer* dmux) {
-  Chunk* c;
-  Frame* f;
-  if (dmux == NULL) return;
-
-  for (f = dmux->frames_; f != NULL;) {
-    Frame* const cur_frame = f;
-    f = f->next_;
-    free(cur_frame);
-  }
-  for (c = dmux->chunks_; c != NULL;) {
-    Chunk* const cur_chunk = c;
-    c = c->next_;
-    free(cur_chunk);
-  }
-  free(dmux);
-}
-
-// -----------------------------------------------------------------------------
-
-uint32_t WebPDemuxGetI(const WebPDemuxer* dmux, WebPFormatFeature feature) {
-  if (dmux == NULL) return 0;
-
-  switch (feature) {
-    case WEBP_FF_FORMAT_FLAGS:  return dmux->feature_flags_;
-    case WEBP_FF_CANVAS_WIDTH:  return (uint32_t)dmux->canvas_width_;
-    case WEBP_FF_CANVAS_HEIGHT: return (uint32_t)dmux->canvas_height_;
-    case WEBP_FF_LOOP_COUNT:    return (uint32_t)dmux->loop_count_;
-  }
-  return 0;
-}
-
-// -----------------------------------------------------------------------------
-// Frame iteration
-
-// Find the first 'frame_num' frame. There may be multiple in a tiled frame.
-static const Frame* GetFrame(const WebPDemuxer* const dmux, int frame_num) {
-  const Frame* f;
-  for (f = dmux->frames_; f != NULL; f = f->next_) {
-    if (frame_num == f->frame_num_) break;
-  }
-  return f;
-}
-
-// Returns tile 'tile_num' and the total count.
-static const Frame* GetTile(
-    const Frame* const frame_set, int tile_num, int* const count) {
-  const int this_frame = frame_set->frame_num_;
-  const Frame* f = frame_set;
-  const Frame* tile = NULL;
-  int total;
-
-  for (total = 0; f != NULL && f->frame_num_ == this_frame; f = f->next_) {
-    if (++total == tile_num) tile = f;
-  }
-  *count = total;
-  return tile;
-}
-
-static const uint8_t* GetFramePayload(const uint8_t* const mem_buf,
-                                      const Frame* const frame,
-                                      size_t* const data_size) {
-  *data_size = 0;
-  if (frame != NULL) {
-    const ChunkData* const image = frame->img_components_;
-    const ChunkData* const alpha = frame->img_components_ + 1;
-    size_t start_offset = image->offset_;
-    *data_size = image->size_;
-
-    // if alpha exists it precedes image, update the size allowing for
-    // intervening chunks.
-    if (alpha->size_ > 0) {
-      const size_t inter_size = (image->offset_ > 0)
-                              ? image->offset_ - (alpha->offset_ + alpha->size_)
-                              : 0;
-      start_offset = alpha->offset_;
-      *data_size  += alpha->size_ + inter_size;
-    }
-    return mem_buf + start_offset;
-  }
-  return NULL;
-}
-
-// Create a whole 'frame' from VP8 (+ alpha) or lossless.
-static int SynthesizeFrame(const WebPDemuxer* const dmux,
-                           const Frame* const first_frame,
-                           int tile_num, WebPIterator* const iter) {
-  const uint8_t* const mem_buf = dmux->mem_.buf_;
-  int num_tiles;
-  size_t payload_size = 0;
-  const Frame* const tile = GetTile(first_frame, tile_num, &num_tiles);
-  const uint8_t* const payload = GetFramePayload(mem_buf, tile, &payload_size);
-  if (payload == NULL) return 0;
-
-  iter->frame_num_   = first_frame->frame_num_;
-  iter->num_frames_  = dmux->num_frames_;
-  iter->tile_num_    = tile_num;
-  iter->num_tiles_   = num_tiles;
-  iter->x_offset_    = tile->x_offset_;
-  iter->y_offset_    = tile->y_offset_;
-  iter->width_       = tile->width_;
-  iter->height_      = tile->height_;
-  iter->duration_    = tile->duration_;
-  iter->complete_    = tile->complete_;
-  iter->tile_.bytes_ = payload;
-  iter->tile_.size_  = payload_size;
-  // TODO(jzern): adjust offsets for 'TILE's embedded in 'FRM 's
-  return 1;
-}
-
-static int SetFrame(int frame_num, WebPIterator* const iter) {
-  const Frame* frame;
-  const WebPDemuxer* const dmux = (WebPDemuxer*)iter->private_;
-  if (dmux == NULL || frame_num < 0) return 0;
-  if (frame_num > dmux->num_frames_) return 0;
-  if (frame_num == 0) frame_num = dmux->num_frames_;
-
-  frame = GetFrame(dmux, frame_num);
-  return SynthesizeFrame(dmux, frame, 1, iter);
-}
-
-int WebPDemuxGetFrame(const WebPDemuxer* dmux, int frame, WebPIterator* iter) {
-  if (iter == NULL) return 0;
-
-  memset(iter, 0, sizeof(*iter));
-  iter->private_ = (void*)dmux;
-  return SetFrame(frame, iter);
-}
-
-int WebPDemuxNextFrame(WebPIterator* iter) {
-  if (iter == NULL) return 0;
-  return SetFrame(iter->frame_num_ + 1, iter);
-}
-
-int WebPDemuxPrevFrame(WebPIterator* iter) {
-  if (iter == NULL) return 0;
-  if (iter->frame_num_ <= 1) return 0;
-  return SetFrame(iter->frame_num_ - 1, iter);
-}
-
-int WebPDemuxSelectTile(WebPIterator* iter, int tile) {
-  if (iter != NULL && iter->private_ != NULL && tile > 0) {
-    const WebPDemuxer* const dmux = (WebPDemuxer*)iter->private_;
-    const Frame* const frame = GetFrame(dmux, iter->frame_num_);
-    if (frame == NULL) return 0;
-
-    return SynthesizeFrame(dmux, frame, tile, iter);
-  }
-  return 0;
-}
-
-void WebPDemuxReleaseIterator(WebPIterator* iter) {
-  (void)iter;
-}
-
-// -----------------------------------------------------------------------------
-// Chunk iteration
-
-static int ChunkCount(const WebPDemuxer* const dmux, const char fourcc[4]) {
-  const uint8_t* const mem_buf = dmux->mem_.buf_;
-  const Chunk* c;
-  int count = 0;
-  for (c = dmux->chunks_; c != NULL; c = c->next_) {
-    const uint8_t* const header = mem_buf + c->data_.offset_;
-    if (!memcmp(header, fourcc, TAG_SIZE)) ++count;
-  }
-  return count;
-}
-
-static const Chunk* GetChunk(const WebPDemuxer* const dmux,
-                             const char fourcc[4], int chunk_num) {
-  const uint8_t* const mem_buf = dmux->mem_.buf_;
-  const Chunk* c;
-  int count = 0;
-  for (c = dmux->chunks_; c != NULL; c = c->next_) {
-    const uint8_t* const header = mem_buf + c->data_.offset_;
-    if (!memcmp(header, fourcc, TAG_SIZE)) ++count;
-    if (count == chunk_num) break;
-  }
-  return c;
-}
-
-static int SetChunk(const char fourcc[4], int chunk_num,
-                    WebPChunkIterator* const iter) {
-  const WebPDemuxer* const dmux = (WebPDemuxer*)iter->private_;
-  int count;
-
-  if (dmux == NULL || fourcc == NULL || chunk_num < 0) return 0;
-  count = ChunkCount(dmux, fourcc);
-  if (count == 0) return 0;
-  if (chunk_num == 0) chunk_num = count;
-
-  if (chunk_num <= count) {
-    const uint8_t* const mem_buf = dmux->mem_.buf_;
-    const Chunk* const chunk = GetChunk(dmux, fourcc, chunk_num);
-    iter->chunk_.bytes_ = mem_buf + chunk->data_.offset_ + CHUNK_HEADER_SIZE;
-    iter->chunk_.size_  = chunk->data_.size_ - CHUNK_HEADER_SIZE;
-    iter->num_chunks_   = count;
-    iter->chunk_num_    = chunk_num;
-    return 1;
-  }
-  return 0;
-}
-
-int WebPDemuxGetChunk(const WebPDemuxer* dmux,
-                      const char fourcc[4], int chunk_num,
-                      WebPChunkIterator* iter) {
-  if (iter == NULL) return 0;
-
-  memset(iter, 0, sizeof(*iter));
-  iter->private_ = (void*)dmux;
-  return SetChunk(fourcc, chunk_num, iter);
-}
-
-int WebPDemuxNextChunk(WebPChunkIterator* iter) {
-  if (iter != NULL) {
-    const char* const fourcc =
-        (const char*)iter->chunk_.bytes_ - CHUNK_HEADER_SIZE;
-    return SetChunk(fourcc, iter->chunk_num_ + 1, iter);
-  }
-  return 0;
-}
-
-int WebPDemuxPrevChunk(WebPChunkIterator* iter) {
-  if (iter != NULL && iter->chunk_num_ > 1) {
-    const char* const fourcc =
-        (const char*)iter->chunk_.bytes_ - CHUNK_HEADER_SIZE;
-    return SetChunk(fourcc, iter->chunk_num_ - 1, iter);
-  }
-  return 0;
-}
-
-void WebPDemuxReleaseChunkIterator(WebPChunkIterator* iter) {
-  (void)iter;
-}
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}  // extern "C"
-#endif
diff --git a/drivers/webp/mux/muxedit.c b/drivers/webp/mux/muxedit.c
index 08629d4ae2..b27663f87a 100644
--- a/drivers/webp/mux/muxedit.c
+++ b/drivers/webp/mux/muxedit.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Set and delete APIs for mux.
@@ -12,50 +14,51 @@
 
 #include <assert.h>
 #include "./muxi.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 // Life of a mux object.
 
 static void MuxInit(WebPMux* const mux) {
-  if (mux == NULL) return;
+  assert(mux != NULL);
   memset(mux, 0, sizeof(*mux));
+  mux->canvas_width_ = 0;     // just to be explicit
+  mux->canvas_height_ = 0;
 }
 
 WebPMux* WebPNewInternal(int version) {
   if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_MUX_ABI_VERSION)) {
     return NULL;
   } else {
-    WebPMux* const mux = (WebPMux*)malloc(sizeof(WebPMux));
-    // If mux is NULL MuxInit is a noop.
-    MuxInit(mux);
+    WebPMux* const mux = (WebPMux*)WebPSafeMalloc(1ULL, sizeof(WebPMux));
+    if (mux != NULL) MuxInit(mux);
     return mux;
   }
 }
 
-static void DeleteAllChunks(WebPChunk** const chunk_list) {
-  while (*chunk_list) {
-    *chunk_list = ChunkDelete(*chunk_list);
+// Delete all images in 'wpi_list'.
+static void DeleteAllImages(WebPMuxImage** const wpi_list) {
+  while (*wpi_list != NULL) {
+    *wpi_list = MuxImageDelete(*wpi_list);
   }
 }
 
 static void MuxRelease(WebPMux* const mux) {
-  if (mux == NULL) return;
-  MuxImageDeleteAll(&mux->images_);
-  DeleteAllChunks(&mux->vp8x_);
-  DeleteAllChunks(&mux->iccp_);
-  DeleteAllChunks(&mux->loop_);
-  DeleteAllChunks(&mux->meta_);
-  DeleteAllChunks(&mux->unknown_);
+  assert(mux != NULL);
+  DeleteAllImages(&mux->images_);
+  ChunkListDelete(&mux->vp8x_);
+  ChunkListDelete(&mux->iccp_);
+  ChunkListDelete(&mux->anim_);
+  ChunkListDelete(&mux->exif_);
+  ChunkListDelete(&mux->xmp_);
+  ChunkListDelete(&mux->unknown_);
 }
 
 void WebPMuxDelete(WebPMux* mux) {
-  // If mux is NULL MuxRelease is a noop.
-  MuxRelease(mux);
-  free(mux);
+  if (mux != NULL) {
+    MuxRelease(mux);
+    WebPSafeFree(mux);
+  }
 }
 
 //------------------------------------------------------------------------------
@@ -64,81 +67,60 @@ void WebPMuxDelete(WebPMux* mux) {
 // Handy MACRO, makes MuxSet() very symmetric to MuxGet().
 #define SWITCH_ID_LIST(INDEX, LIST)                                            \
   if (idx == (INDEX)) {                                                        \
-    err = ChunkAssignData(&chunk, data, copy_data, kChunks[(INDEX)].tag);      \
+    err = ChunkAssignData(&chunk, data, copy_data, tag);                       \
     if (err == WEBP_MUX_OK) {                                                  \
       err = ChunkSetNth(&chunk, (LIST), nth);                                  \
     }                                                                          \
     return err;                                                                \
   }
 
-static WebPMuxError MuxSet(WebPMux* const mux, CHUNK_INDEX idx, uint32_t nth,
+static WebPMuxError MuxSet(WebPMux* const mux, uint32_t tag, uint32_t nth,
                            const WebPData* const data, int copy_data) {
   WebPChunk chunk;
   WebPMuxError err = WEBP_MUX_NOT_FOUND;
+  const CHUNK_INDEX idx = ChunkGetIndexFromTag(tag);
   assert(mux != NULL);
   assert(!IsWPI(kChunks[idx].id));
 
   ChunkInit(&chunk);
-  SWITCH_ID_LIST(IDX_VP8X, &mux->vp8x_);
-  SWITCH_ID_LIST(IDX_ICCP, &mux->iccp_);
-  SWITCH_ID_LIST(IDX_LOOP, &mux->loop_);
-  SWITCH_ID_LIST(IDX_META, &mux->meta_);
-  if (idx == IDX_UNKNOWN && data->size_ > TAG_SIZE) {
-    // For raw-data unknown chunk, the first four bytes should be the tag to be
-    // used for the chunk.
-    const WebPData tmp = { data->bytes_ + TAG_SIZE, data->size_ - TAG_SIZE };
-    err = ChunkAssignData(&chunk, &tmp, copy_data, GetLE32(data->bytes_ + 0));
-    if (err == WEBP_MUX_OK)
-      err = ChunkSetNth(&chunk, &mux->unknown_, nth);
-  }
+  SWITCH_ID_LIST(IDX_VP8X,    &mux->vp8x_);
+  SWITCH_ID_LIST(IDX_ICCP,    &mux->iccp_);
+  SWITCH_ID_LIST(IDX_ANIM,    &mux->anim_);
+  SWITCH_ID_LIST(IDX_EXIF,    &mux->exif_);
+  SWITCH_ID_LIST(IDX_XMP,     &mux->xmp_);
+  SWITCH_ID_LIST(IDX_UNKNOWN, &mux->unknown_);
   return err;
 }
 #undef SWITCH_ID_LIST
 
-static WebPMuxError MuxAddChunk(WebPMux* const mux, uint32_t nth, uint32_t tag,
-                                const uint8_t* data, size_t size,
-                                int copy_data) {
-  const CHUNK_INDEX idx = ChunkGetIndexFromTag(tag);
-  const WebPData chunk_data = { data, size };
-  assert(mux != NULL);
-  assert(size <= MAX_CHUNK_PAYLOAD);
-  assert(idx != IDX_NIL);
-  return MuxSet(mux, idx, nth, &chunk_data, copy_data);
-}
+// Create data for frame/fragment given image data, offsets and duration.
+static WebPMuxError CreateFrameFragmentData(
+    int width, int height, const WebPMuxFrameInfo* const info, int is_frame,
+    WebPData* const frame_frgm) {
+  uint8_t* frame_frgm_bytes;
+  const size_t frame_frgm_size = kChunks[is_frame ? IDX_ANMF : IDX_FRGM].size;
 
-// Create data for frame/tile given image data, offsets and duration.
-static WebPMuxError CreateFrameTileData(const WebPData* const image,
-                                        int x_offset, int y_offset,
-                                        int duration, int is_lossless,
-                                        int is_frame,
-                                        WebPData* const frame_tile) {
-  int width;
-  int height;
-  uint8_t* frame_tile_bytes;
-  const size_t frame_tile_size = kChunks[is_frame ? IDX_FRAME : IDX_TILE].size;
-
-  const int ok = is_lossless ?
-      VP8LGetInfo(image->bytes_, image->size_, &width, &height, NULL) :
-      VP8GetInfo(image->bytes_, image->size_, image->size_, &width, &height);
-  if (!ok) return WEBP_MUX_INVALID_ARGUMENT;
-
-  assert(width > 0 && height > 0 && duration > 0);
+  assert(width > 0 && height > 0 && info->duration >= 0);
+  assert(info->dispose_method == (info->dispose_method & 1));
   // Note: assertion on upper bounds is done in PutLE24().
 
-  frame_tile_bytes = (uint8_t*)malloc(frame_tile_size);
-  if (frame_tile_bytes == NULL) return WEBP_MUX_MEMORY_ERROR;
+  frame_frgm_bytes = (uint8_t*)WebPSafeMalloc(1ULL, frame_frgm_size);
+  if (frame_frgm_bytes == NULL) return WEBP_MUX_MEMORY_ERROR;
 
-  PutLE24(frame_tile_bytes + 0, x_offset / 2);
-  PutLE24(frame_tile_bytes + 3, y_offset / 2);
+  PutLE24(frame_frgm_bytes + 0, info->x_offset / 2);
+  PutLE24(frame_frgm_bytes + 3, info->y_offset / 2);
 
   if (is_frame) {
-    PutLE24(frame_tile_bytes + 6, width - 1);
-    PutLE24(frame_tile_bytes + 9, height - 1);
-    PutLE24(frame_tile_bytes + 12, duration - 1);
+    PutLE24(frame_frgm_bytes + 6, width - 1);
+    PutLE24(frame_frgm_bytes + 9, height - 1);
+    PutLE24(frame_frgm_bytes + 12, info->duration);
+    frame_frgm_bytes[15] =
+        (info->blend_method == WEBP_MUX_NO_BLEND ? 2 : 0) |
+        (info->dispose_method == WEBP_MUX_DISPOSE_BACKGROUND ? 1 : 0);
   }
 
-  frame_tile->bytes_ = frame_tile_bytes;
-  frame_tile->size_ = frame_tile_size;
+  frame_frgm->bytes = frame_frgm_bytes;
+  frame_frgm->size = frame_frgm_size;
   return WEBP_MUX_OK;
 }
 
@@ -149,8 +131,8 @@ static WebPMuxError GetImageData(const WebPData* const bitstream,
                                  WebPData* const image, WebPData* const alpha,
                                  int* const is_lossless) {
   WebPDataInit(alpha);  // Default: no alpha.
-  if (bitstream->size_ < TAG_SIZE ||
-      memcmp(bitstream->bytes_, "RIFF", TAG_SIZE)) {
+  if (bitstream->size < TAG_SIZE ||
+      memcmp(bitstream->bytes, "RIFF", TAG_SIZE)) {
     // It is NOT webp file data. Return input data as is.
     *image = *bitstream;
   } else {
@@ -166,7 +148,7 @@ static WebPMuxError GetImageData(const WebPData* const bitstream,
     }
     WebPMuxDelete(mux);
   }
-  *is_lossless = VP8LCheckSignature(image->bytes_, image->size_);
+  *is_lossless = VP8LCheckSignature(image->bytes, image->size);
   return WEBP_MUX_OK;
 }
 
@@ -185,204 +167,166 @@ static WebPMuxError DeleteChunks(WebPChunk** chunk_list, uint32_t tag) {
   return err;
 }
 
-static WebPMuxError MuxDeleteAllNamedData(WebPMux* const mux, CHUNK_INDEX idx) {
-  const WebPChunkId id = kChunks[idx].id;
-  WebPChunk** chunk_list;
-
-  if (mux == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+static WebPMuxError MuxDeleteAllNamedData(WebPMux* const mux, uint32_t tag) {
+  const WebPChunkId id = ChunkGetIdFromTag(tag);
+  assert(mux != NULL);
   if (IsWPI(id)) return WEBP_MUX_INVALID_ARGUMENT;
-
-  chunk_list = MuxGetChunkListFromId(mux, id);
-  if (chunk_list == NULL) return WEBP_MUX_INVALID_ARGUMENT;
-
-  return DeleteChunks(chunk_list, kChunks[idx].tag);
-}
-
-static WebPMuxError DeleteLoopCount(WebPMux* const mux) {
-  return MuxDeleteAllNamedData(mux, IDX_LOOP);
+  return DeleteChunks(MuxGetChunkListFromId(mux, id), tag);
 }
 
 //------------------------------------------------------------------------------
 // Set API(s).
 
-WebPMuxError WebPMuxSetImage(WebPMux* mux,
-                             const WebPData* bitstream, int copy_data) {
+WebPMuxError WebPMuxSetChunk(WebPMux* mux, const char fourcc[4],
+                             const WebPData* chunk_data, int copy_data) {
+  uint32_t tag;
   WebPMuxError err;
-  WebPChunk chunk;
-  WebPMuxImage wpi;
-  WebPData image;
-  WebPData alpha;
-  int is_lossless;
-  int image_tag;
-
-  if (mux == NULL || bitstream == NULL || bitstream->bytes_ == NULL ||
-      bitstream->size_ > MAX_CHUNK_PAYLOAD) {
+  if (mux == NULL || fourcc == NULL || chunk_data == NULL ||
+      chunk_data->bytes == NULL || chunk_data->size > MAX_CHUNK_PAYLOAD) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
+  tag = ChunkGetTagFromFourCC(fourcc);
 
-  // If given data is for a whole webp file,
-  // extract only the VP8/VP8L data from it.
-  err = GetImageData(bitstream, &image, &alpha, &is_lossless);
-  if (err != WEBP_MUX_OK) return err;
-  image_tag = is_lossless ? kChunks[IDX_VP8L].tag : kChunks[IDX_VP8].tag;
-
-  // Delete the existing images.
-  MuxImageDeleteAll(&mux->images_);
-
-  MuxImageInit(&wpi);
+  // Delete existing chunk(s) with the same 'fourcc'.
+  err = MuxDeleteAllNamedData(mux, tag);
+  if (err != WEBP_MUX_OK && err != WEBP_MUX_NOT_FOUND) return err;
 
-  if (alpha.bytes_ != NULL) {  // Add alpha chunk.
-    ChunkInit(&chunk);
-    err = ChunkAssignData(&chunk, &alpha, copy_data, kChunks[IDX_ALPHA].tag);
-    if (err != WEBP_MUX_OK) goto Err;
-    err = ChunkSetNth(&chunk, &wpi.alpha_, 1);
-    if (err != WEBP_MUX_OK) goto Err;
-  }
+  // Add the given chunk.
+  return MuxSet(mux, tag, 1, chunk_data, copy_data);
+}
 
-  // Add image chunk.
+// Creates a chunk from given 'data' and sets it as 1st chunk in 'chunk_list'.
+static WebPMuxError AddDataToChunkList(
+    const WebPData* const data, int copy_data, uint32_t tag,
+    WebPChunk** chunk_list) {
+  WebPChunk chunk;
+  WebPMuxError err;
   ChunkInit(&chunk);
-  err = ChunkAssignData(&chunk, &image, copy_data, image_tag);
+  err = ChunkAssignData(&chunk, data, copy_data, tag);
   if (err != WEBP_MUX_OK) goto Err;
-  err = ChunkSetNth(&chunk, &wpi.img_, 1);
-  if (err != WEBP_MUX_OK) goto Err;
-
-  // Add this image to mux.
-  err = MuxImagePush(&wpi, &mux->images_);
+  err = ChunkSetNth(&chunk, chunk_list, 1);
   if (err != WEBP_MUX_OK) goto Err;
-
-  // All OK.
   return WEBP_MUX_OK;
-
  Err:
-  // Something bad happened.
   ChunkRelease(&chunk);
-  MuxImageRelease(&wpi);
   return err;
 }
 
-WebPMuxError WebPMuxSetMetadata(WebPMux* mux, const WebPData* metadata,
-                                int copy_data) {
-  WebPMuxError err;
-
-  if (mux == NULL || metadata == NULL || metadata->bytes_ == NULL ||
-      metadata->size_ > MAX_CHUNK_PAYLOAD) {
-    return WEBP_MUX_INVALID_ARGUMENT;
+// Extracts image & alpha data from the given bitstream and then sets wpi.alpha_
+// and wpi.img_ appropriately.
+static WebPMuxError SetAlphaAndImageChunks(
+    const WebPData* const bitstream, int copy_data, WebPMuxImage* const wpi) {
+  int is_lossless = 0;
+  WebPData image, alpha;
+  WebPMuxError err = GetImageData(bitstream, &image, &alpha, &is_lossless);
+  const int image_tag =
+      is_lossless ? kChunks[IDX_VP8L].tag : kChunks[IDX_VP8].tag;
+  if (err != WEBP_MUX_OK) return err;
+  if (alpha.bytes != NULL) {
+    err = AddDataToChunkList(&alpha, copy_data, kChunks[IDX_ALPHA].tag,
+                             &wpi->alpha_);
+    if (err != WEBP_MUX_OK) return err;
   }
-
-  // Delete the existing metadata chunk(s).
-  err = WebPMuxDeleteMetadata(mux);
-  if (err != WEBP_MUX_OK && err != WEBP_MUX_NOT_FOUND) return err;
-
-  // Add the given metadata chunk.
-  return MuxSet(mux, IDX_META, 1, metadata, copy_data);
+  err = AddDataToChunkList(&image, copy_data, image_tag, &wpi->img_);
+  if (err != WEBP_MUX_OK) return err;
+  return MuxImageFinalize(wpi) ? WEBP_MUX_OK : WEBP_MUX_INVALID_ARGUMENT;
 }
 
-WebPMuxError WebPMuxSetColorProfile(WebPMux* mux, const WebPData* color_profile,
-                                    int copy_data) {
+WebPMuxError WebPMuxSetImage(WebPMux* mux, const WebPData* bitstream,
+                             int copy_data) {
+  WebPMuxImage wpi;
   WebPMuxError err;
 
-  if (mux == NULL || color_profile == NULL || color_profile->bytes_ == NULL ||
-      color_profile->size_ > MAX_CHUNK_PAYLOAD) {
+  // Sanity checks.
+  if (mux == NULL || bitstream == NULL || bitstream->bytes == NULL ||
+      bitstream->size > MAX_CHUNK_PAYLOAD) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
 
-  // Delete the existing ICCP chunk(s).
-  err = WebPMuxDeleteColorProfile(mux);
-  if (err != WEBP_MUX_OK && err != WEBP_MUX_NOT_FOUND) return err;
-
-  // Add the given ICCP chunk.
-  return MuxSet(mux, IDX_ICCP, 1, color_profile, copy_data);
-}
-
-WebPMuxError WebPMuxSetLoopCount(WebPMux* mux, int loop_count) {
-  WebPMuxError err;
-  uint8_t* data = NULL;
+  if (mux->images_ != NULL) {
+    // Only one 'simple image' can be added in mux. So, remove present images.
+    DeleteAllImages(&mux->images_);
+  }
 
-  if (mux == NULL) return WEBP_MUX_INVALID_ARGUMENT;
-  if (loop_count >= MAX_LOOP_COUNT) return WEBP_MUX_INVALID_ARGUMENT;
+  MuxImageInit(&wpi);
+  err = SetAlphaAndImageChunks(bitstream, copy_data, &wpi);
+  if (err != WEBP_MUX_OK) goto Err;
 
-  // Delete the existing LOOP chunk(s).
-  err = DeleteLoopCount(mux);
-  if (err != WEBP_MUX_OK && err != WEBP_MUX_NOT_FOUND) return err;
+  // Add this WebPMuxImage to mux.
+  err = MuxImagePush(&wpi, &mux->images_);
+  if (err != WEBP_MUX_OK) goto Err;
 
-  // Add the given loop count.
-  data = (uint8_t*)malloc(kChunks[IDX_LOOP].size);
-  if (data == NULL) return WEBP_MUX_MEMORY_ERROR;
+  // All is well.
+  return WEBP_MUX_OK;
 
-  PutLE16(data, loop_count);
-  err = MuxAddChunk(mux, 1, kChunks[IDX_LOOP].tag, data,
-                    kChunks[IDX_LOOP].size, 1);
-  free(data);
+ Err:  // Something bad happened.
+  MuxImageRelease(&wpi);
   return err;
 }
 
-static WebPMuxError MuxPushFrameTileInternal(
-    WebPMux* const mux, const WebPData* const bitstream, int x_offset,
-    int y_offset, int duration, int copy_data, uint32_t tag) {
-  WebPChunk chunk;
-  WebPData image;
-  WebPData alpha;
+WebPMuxError WebPMuxPushFrame(WebPMux* mux, const WebPMuxFrameInfo* frame,
+                              int copy_data) {
   WebPMuxImage wpi;
   WebPMuxError err;
-  WebPData frame_tile;
-  const int is_frame = (tag == kChunks[IDX_FRAME].tag) ? 1 : 0;
-  int is_lossless;
-  int image_tag;
+  int is_frame;
+  const WebPData* const bitstream = &frame->bitstream;
 
   // Sanity checks.
-  if (mux == NULL || bitstream == NULL || bitstream->bytes_ == NULL ||
-      bitstream->size_ > MAX_CHUNK_PAYLOAD) {
+  if (mux == NULL || frame == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+
+  is_frame = (frame->id == WEBP_CHUNK_ANMF);
+  if (!(is_frame || (frame->id == WEBP_CHUNK_FRGM))) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
-  if (x_offset < 0 || x_offset >= MAX_POSITION_OFFSET ||
-      y_offset < 0 || y_offset >= MAX_POSITION_OFFSET ||
-      duration <= 0 || duration > MAX_DURATION) {
+  if (frame->id == WEBP_CHUNK_FRGM) {     // Dead experiment.
     return WEBP_MUX_INVALID_ARGUMENT;
   }
 
-  // Snap offsets to even positions.
-  x_offset &= ~1;
-  y_offset &= ~1;
+  if (bitstream->bytes == NULL || bitstream->size > MAX_CHUNK_PAYLOAD) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
 
-  // If given data is for a whole webp file,
-  // extract only the VP8/VP8L data from it.
-  err = GetImageData(bitstream, &image, &alpha, &is_lossless);
-  if (err != WEBP_MUX_OK) return err;
-  image_tag = is_lossless ? kChunks[IDX_VP8L].tag : kChunks[IDX_VP8].tag;
+  if (mux->images_ != NULL) {
+    const WebPMuxImage* const image = mux->images_;
+    const uint32_t image_id = (image->header_ != NULL) ?
+        ChunkGetIdFromTag(image->header_->tag_) : WEBP_CHUNK_IMAGE;
+    if (image_id != frame->id) {
+      return WEBP_MUX_INVALID_ARGUMENT;  // Conflicting frame types.
+    }
+  }
 
-  WebPDataInit(&frame_tile);
-  ChunkInit(&chunk);
   MuxImageInit(&wpi);
-
-  if (alpha.bytes_ != NULL) {
-    // Add alpha chunk.
-    err = ChunkAssignData(&chunk, &alpha, copy_data, kChunks[IDX_ALPHA].tag);
+  err = SetAlphaAndImageChunks(bitstream, copy_data, &wpi);
+  if (err != WEBP_MUX_OK) goto Err;
+  assert(wpi.img_ != NULL);  // As SetAlphaAndImageChunks() was successful.
+
+  {
+    WebPData frame_frgm;
+    const uint32_t tag = kChunks[is_frame ? IDX_ANMF : IDX_FRGM].tag;
+    WebPMuxFrameInfo tmp = *frame;
+    tmp.x_offset &= ~1;  // Snap offsets to even.
+    tmp.y_offset &= ~1;
+    if (!is_frame) {  // Reset unused values.
+      tmp.duration = 1;
+      tmp.dispose_method = WEBP_MUX_DISPOSE_NONE;
+      tmp.blend_method = WEBP_MUX_BLEND;
+    }
+    if (tmp.x_offset < 0 || tmp.x_offset >= MAX_POSITION_OFFSET ||
+        tmp.y_offset < 0 || tmp.y_offset >= MAX_POSITION_OFFSET ||
+        (tmp.duration < 0 || tmp.duration >= MAX_DURATION) ||
+        tmp.dispose_method != (tmp.dispose_method & 1)) {
+      err = WEBP_MUX_INVALID_ARGUMENT;
+      goto Err;
+    }
+    err = CreateFrameFragmentData(wpi.width_, wpi.height_, &tmp, is_frame,
+                                  &frame_frgm);
     if (err != WEBP_MUX_OK) goto Err;
-    err = ChunkSetNth(&chunk, &wpi.alpha_, 1);
+    // Add frame/fragment chunk (with copy_data = 1).
+    err = AddDataToChunkList(&frame_frgm, 1, tag, &wpi.header_);
+    WebPDataClear(&frame_frgm);  // frame_frgm owned by wpi.header_ now.
     if (err != WEBP_MUX_OK) goto Err;
-    ChunkInit(&chunk);  // chunk owned by wpi.alpha_ now.
   }
 
-  // Add image chunk.
-  err = ChunkAssignData(&chunk, &image, copy_data, image_tag);
-  if (err != WEBP_MUX_OK) goto Err;
-  err = ChunkSetNth(&chunk, &wpi.img_, 1);
-  if (err != WEBP_MUX_OK) goto Err;
-  ChunkInit(&chunk);  // chunk owned by wpi.img_ now.
-
-  // Create frame/tile data.
-  err = CreateFrameTileData(&image, x_offset, y_offset, duration, is_lossless,
-                            is_frame, &frame_tile);
-  if (err != WEBP_MUX_OK) goto Err;
-
-  // Add frame/tile chunk (with copy_data = 1).
-  err = ChunkAssignData(&chunk, &frame_tile, 1, tag);
-  if (err != WEBP_MUX_OK) goto Err;
-  WebPDataClear(&frame_tile);
-  err = ChunkSetNth(&chunk, &wpi.header_, 1);
-  if (err != WEBP_MUX_OK) goto Err;
-  ChunkInit(&chunk);  // chunk owned by wpi.header_ now.
-
   // Add this WebPMuxImage to mux.
   err = MuxImagePush(&wpi, &mux->images_);
   if (err != WEBP_MUX_OK) goto Err;
@@ -391,128 +335,114 @@ static WebPMuxError MuxPushFrameTileInternal(
   return WEBP_MUX_OK;
 
  Err:  // Something bad happened.
-  WebPDataClear(&frame_tile);
-  ChunkRelease(&chunk);
   MuxImageRelease(&wpi);
   return err;
 }
 
-WebPMuxError WebPMuxPushFrame(WebPMux* mux, const WebPData* bitstream,
-                              int x_offset, int y_offset,
-                              int duration, int copy_data) {
-  return MuxPushFrameTileInternal(mux, bitstream, x_offset, y_offset,
-                                  duration, copy_data, kChunks[IDX_FRAME].tag);
-}
-
-WebPMuxError WebPMuxPushTile(WebPMux* mux, const WebPData* bitstream,
-                             int x_offset, int y_offset,
-                             int copy_data) {
-  return MuxPushFrameTileInternal(mux, bitstream, x_offset, y_offset,
-                                  1 /* unused duration */, copy_data,
-                                  kChunks[IDX_TILE].tag);
-}
-
-//------------------------------------------------------------------------------
-// Delete API(s).
-
-WebPMuxError WebPMuxDeleteImage(WebPMux* mux) {
+WebPMuxError WebPMuxSetAnimationParams(WebPMux* mux,
+                                       const WebPMuxAnimParams* params) {
   WebPMuxError err;
+  uint8_t data[ANIM_CHUNK_SIZE];
+  const WebPData anim = { data, ANIM_CHUNK_SIZE };
 
-  if (mux == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+  if (mux == NULL || params == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+  if (params->loop_count < 0 || params->loop_count >= MAX_LOOP_COUNT) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
 
-  err = MuxValidateForImage(mux);
-  if (err != WEBP_MUX_OK) return err;
+  // Delete any existing ANIM chunk(s).
+  err = MuxDeleteAllNamedData(mux, kChunks[IDX_ANIM].tag);
+  if (err != WEBP_MUX_OK && err != WEBP_MUX_NOT_FOUND) return err;
 
-  // All well, delete image.
-  MuxImageDeleteAll(&mux->images_);
-  return WEBP_MUX_OK;
+  // Set the animation parameters.
+  PutLE32(data, params->bgcolor);
+  PutLE16(data + 4, params->loop_count);
+  return MuxSet(mux, kChunks[IDX_ANIM].tag, 1, &anim, 1);
 }
 
-WebPMuxError WebPMuxDeleteMetadata(WebPMux* mux) {
-  return MuxDeleteAllNamedData(mux, IDX_META);
-}
+WebPMuxError WebPMuxSetCanvasSize(WebPMux* mux,
+                                  int width, int height) {
+  WebPMuxError err;
+  if (mux == NULL) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+  if (width < 0 || height < 0 ||
+      width > MAX_CANVAS_SIZE || height > MAX_CANVAS_SIZE) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+  if (width * (uint64_t)height >= MAX_IMAGE_AREA) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+  if ((width * height) == 0 && (width | height) != 0) {
+    // one of width / height is zero, but not both -> invalid!
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+  // If we already assembled a VP8X chunk, invalidate it.
+  err = MuxDeleteAllNamedData(mux, kChunks[IDX_VP8X].tag);
+  if (err != WEBP_MUX_OK && err != WEBP_MUX_NOT_FOUND) return err;
 
-WebPMuxError WebPMuxDeleteColorProfile(WebPMux* mux) {
-  return MuxDeleteAllNamedData(mux, IDX_ICCP);
+  mux->canvas_width_ = width;
+  mux->canvas_height_ = height;
+  return WEBP_MUX_OK;
 }
 
-static WebPMuxError DeleteFrameTileInternal(WebPMux* const mux, uint32_t nth,
-                                            CHUNK_INDEX idx) {
-  const WebPChunkId id = kChunks[idx].id;
-  if (mux == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+//------------------------------------------------------------------------------
+// Delete API(s).
 
-  assert(idx == IDX_FRAME || idx == IDX_TILE);
-  return MuxImageDeleteNth(&mux->images_, nth, id);
+WebPMuxError WebPMuxDeleteChunk(WebPMux* mux, const char fourcc[4]) {
+  if (mux == NULL || fourcc == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+  return MuxDeleteAllNamedData(mux, ChunkGetTagFromFourCC(fourcc));
 }
 
 WebPMuxError WebPMuxDeleteFrame(WebPMux* mux, uint32_t nth) {
-  return DeleteFrameTileInternal(mux, nth, IDX_FRAME);
-}
-
-WebPMuxError WebPMuxDeleteTile(WebPMux* mux, uint32_t nth) {
-  return DeleteFrameTileInternal(mux, nth, IDX_TILE);
+  if (mux == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+  return MuxImageDeleteNth(&mux->images_, nth);
 }
 
 //------------------------------------------------------------------------------
 // Assembly of the WebP RIFF file.
 
-static WebPMuxError GetFrameTileInfo(const WebPChunk* const frame_tile_chunk,
-                                     int* const x_offset, int* const y_offset,
-                                     int* const duration) {
-  const uint32_t tag = frame_tile_chunk->tag_;
-  const int is_frame = (tag == kChunks[IDX_FRAME].tag);
-  const WebPData* const data = &frame_tile_chunk->data_;
+static WebPMuxError GetFrameFragmentInfo(
+    const WebPChunk* const frame_frgm_chunk,
+    int* const x_offset, int* const y_offset, int* const duration) {
+  const uint32_t tag = frame_frgm_chunk->tag_;
+  const int is_frame = (tag == kChunks[IDX_ANMF].tag);
+  const WebPData* const data = &frame_frgm_chunk->data_;
   const size_t expected_data_size =
-      is_frame ? FRAME_CHUNK_SIZE : TILE_CHUNK_SIZE;
-  assert(frame_tile_chunk != NULL);
-  assert(tag == kChunks[IDX_FRAME].tag || tag ==  kChunks[IDX_TILE].tag);
-  if (data->size_ != expected_data_size) return WEBP_MUX_INVALID_ARGUMENT;
-
-  *x_offset = 2 * GetLE24(data->bytes_ + 0);
-  *y_offset = 2 * GetLE24(data->bytes_ + 3);
-  if (is_frame) *duration = 1 + GetLE24(data->bytes_ + 12);
+      is_frame ? ANMF_CHUNK_SIZE : FRGM_CHUNK_SIZE;
+  assert(frame_frgm_chunk != NULL);
+  assert(tag == kChunks[IDX_ANMF].tag || tag ==  kChunks[IDX_FRGM].tag);
+  if (data->size != expected_data_size) return WEBP_MUX_INVALID_ARGUMENT;
+
+  *x_offset = 2 * GetLE24(data->bytes + 0);
+  *y_offset = 2 * GetLE24(data->bytes + 3);
+  if (is_frame) *duration = GetLE24(data->bytes + 12);
   return WEBP_MUX_OK;
 }
 
-WebPMuxError MuxGetImageWidthHeight(const WebPChunk* const image_chunk,
-                                    int* const width, int* const height) {
-  const uint32_t tag = image_chunk->tag_;
-  const WebPData* const data = &image_chunk->data_;
-  int w, h;
-  int ok;
-  assert(image_chunk != NULL);
-  assert(tag == kChunks[IDX_VP8].tag || tag ==  kChunks[IDX_VP8L].tag);
-  ok = (tag == kChunks[IDX_VP8].tag) ?
-      VP8GetInfo(data->bytes_, data->size_, data->size_, &w, &h) :
-      VP8LGetInfo(data->bytes_, data->size_, &w, &h, NULL);
-  if (ok) {
-    *width = w;
-    *height = h;
-    return WEBP_MUX_OK;
-  } else {
-    return WEBP_MUX_BAD_DATA;
-  }
-}
-
 static WebPMuxError GetImageInfo(const WebPMuxImage* const wpi,
                                  int* const x_offset, int* const y_offset,
                                  int* const duration,
                                  int* const width, int* const height) {
-  const WebPChunk* const image_chunk = wpi->img_;
-  const WebPChunk* const frame_tile_chunk = wpi->header_;
+  const WebPChunk* const frame_frgm_chunk = wpi->header_;
+  WebPMuxError err;
+  assert(wpi != NULL);
+  assert(frame_frgm_chunk != NULL);
 
-  // Get offsets and duration from FRM/TILE chunk.
-  const WebPMuxError err =
-      GetFrameTileInfo(frame_tile_chunk, x_offset, y_offset, duration);
+  // Get offsets and duration from ANMF/FRGM chunk.
+  err = GetFrameFragmentInfo(frame_frgm_chunk, x_offset, y_offset, duration);
   if (err != WEBP_MUX_OK) return err;
 
   // Get width and height from VP8/VP8L chunk.
-  return MuxGetImageWidthHeight(image_chunk, width, height);
+  if (width != NULL) *width = wpi->width_;
+  if (height != NULL) *height = wpi->height_;
+  return WEBP_MUX_OK;
 }
 
-static WebPMuxError GetImageCanvasWidthHeight(
-    const WebPMux* const mux, uint32_t flags,
-    int* const width, int* const height) {
+// Returns the tightest dimension for the canvas considering the image list.
+static WebPMuxError GetAdjustedCanvasSize(const WebPMux* const mux,
+                                          uint32_t flags,
+                                          int* const width, int* const height) {
   WebPMuxImage* wpi = NULL;
   assert(mux != NULL);
   assert(width != NULL && height != NULL);
@@ -521,13 +451,15 @@ static WebPMuxError GetImageCanvasWidthHeight(
   assert(wpi != NULL);
   assert(wpi->img_ != NULL);
 
-  if (wpi->next_) {
+  if (wpi->next_ != NULL) {
     int max_x = 0;
     int max_y = 0;
     int64_t image_area = 0;
-    // Aggregate the bounding box for animation frames & tiled images.
+    // if we have a chain of wpi's, header_ is necessarily set
+    assert(wpi->header_ != NULL);
+    // Aggregate the bounding box for animation frames & fragmented images.
     for (; wpi != NULL; wpi = wpi->next_) {
-      int x_offset, y_offset, duration, w, h;
+      int x_offset = 0, y_offset = 0, duration = 0, w = 0, h = 0;
       const WebPMuxError err = GetImageInfo(wpi, &x_offset, &y_offset,
                                             &duration, &w, &h);
       const int max_x_pos = x_offset + w;
@@ -542,23 +474,19 @@ static WebPMuxError GetImageCanvasWidthHeight(
     }
     *width = max_x;
     *height = max_y;
-    // Crude check to validate that there are no image overlaps/holes for tile
-    // images. Check that the aggregated image area for individual tiles exactly
-    // matches the image area of the constructed canvas. However, the area-match
-    // is necessary but not sufficient condition.
-    if ((flags & TILE_FLAG) && (image_area != (max_x * max_y))) {
+    // Crude check to validate that there are no image overlaps/holes for
+    // fragmented images. Check that the aggregated image area for individual
+    // fragments exactly matches the image area of the constructed canvas.
+    // However, the area-match is necessary but not sufficient condition.
+    if ((flags & FRAGMENTS_FLAG) && (image_area != (max_x * max_y))) {
       *width = 0;
       *height = 0;
       return WEBP_MUX_INVALID_ARGUMENT;
     }
   } else {
-    // For a single image, extract the width & height from VP8/VP8L image-data.
-    int w, h;
-    const WebPChunk* const image_chunk = wpi->img_;
-    const WebPMuxError err = MuxGetImageWidthHeight(image_chunk, &w, &h);
-    if (err != WEBP_MUX_OK) return err;
-    *width = w;
-    *height = h;
+    // For a single image, canvas dimensions are same as image dimensions.
+    *width = wpi->width_;
+    *height = wpi->height_;
   }
   return WEBP_MUX_OK;
 }
@@ -574,50 +502,45 @@ static WebPMuxError CreateVP8XChunk(WebPMux* const mux) {
   int width = 0;
   int height = 0;
   uint8_t data[VP8X_CHUNK_SIZE];
-  const size_t data_size = VP8X_CHUNK_SIZE;
+  const WebPData vp8x = { data, VP8X_CHUNK_SIZE };
   const WebPMuxImage* images = NULL;
 
   assert(mux != NULL);
   images = mux->images_;  // First image.
   if (images == NULL || images->img_ == NULL ||
-      images->img_->data_.bytes_ == NULL) {
+      images->img_->data_.bytes == NULL) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
 
   // If VP8X chunk(s) is(are) already present, remove them (and later add new
   // VP8X chunk with updated flags).
-  err = MuxDeleteAllNamedData(mux, IDX_VP8X);
+  err = MuxDeleteAllNamedData(mux, kChunks[IDX_VP8X].tag);
   if (err != WEBP_MUX_OK && err != WEBP_MUX_NOT_FOUND) return err;
 
   // Set flags.
-  if (mux->iccp_ != NULL && mux->iccp_->data_.bytes_ != NULL) {
+  if (mux->iccp_ != NULL && mux->iccp_->data_.bytes != NULL) {
     flags |= ICCP_FLAG;
   }
-
-  if (mux->meta_ != NULL && mux->meta_->data_.bytes_ != NULL) {
-    flags |= META_FLAG;
+  if (mux->exif_ != NULL && mux->exif_->data_.bytes != NULL) {
+    flags |= EXIF_FLAG;
+  }
+  if (mux->xmp_ != NULL && mux->xmp_->data_.bytes != NULL) {
+    flags |= XMP_FLAG;
   }
-
   if (images->header_ != NULL) {
-    if (images->header_->tag_ == kChunks[IDX_TILE].tag) {
-      // This is a tiled image.
-      flags |= TILE_FLAG;
-    } else if (images->header_->tag_ == kChunks[IDX_FRAME].tag) {
+    if (images->header_->tag_ == kChunks[IDX_FRGM].tag) {
+      // This is a fragmented image.
+      flags |= FRAGMENTS_FLAG;
+    } else if (images->header_->tag_ == kChunks[IDX_ANMF].tag) {
       // This is an image with animation.
       flags |= ANIMATION_FLAG;
     }
   }
-
   if (MuxImageCount(images, WEBP_CHUNK_ALPHA) > 0) {
     flags |= ALPHA_FLAG;  // Some images have an alpha channel.
   }
 
-  if (flags == 0) {
-    // For Simple Image, VP8X chunk should not be added.
-    return WEBP_MUX_OK;
-  }
-
-  err = GetImageCanvasWidthHeight(mux, flags, &width, &height);
+  err = GetAdjustedCanvasSize(mux, flags, &width, &height);
   if (err != WEBP_MUX_OK) return err;
 
   if (width <= 0 || height <= 0) {
@@ -627,9 +550,21 @@ static WebPMuxError CreateVP8XChunk(WebPMux* const mux) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
 
-  if (MuxHasLosslessImages(images)) {
-    // We have a file with a VP8X chunk having some lossless images.
-    // As lossless images implicitly contain alpha, force ALPHA_FLAG to be true.
+  if (mux->canvas_width_ != 0 || mux->canvas_height_ != 0) {
+    if (width > mux->canvas_width_ || height > mux->canvas_height_) {
+      return WEBP_MUX_INVALID_ARGUMENT;
+    }
+    width = mux->canvas_width_;
+    height = mux->canvas_height_;
+  }
+
+  if (flags == 0) {
+    // For Simple Image, VP8X chunk should not be added.
+    return WEBP_MUX_OK;
+  }
+
+  if (MuxHasAlpha(images)) {
+    // This means some frames explicitly/implicitly contain alpha.
     // Note: This 'flags' update must NOT be done for a lossless image
     // without a VP8X chunk!
     flags |= ALPHA_FLAG;
@@ -639,74 +574,123 @@ static WebPMuxError CreateVP8XChunk(WebPMux* const mux) {
   PutLE24(data + 4, width - 1);   // canvas width.
   PutLE24(data + 7, height - 1);  // canvas height.
 
-  err = MuxAddChunk(mux, 1, kChunks[IDX_VP8X].tag, data, data_size, 1);
-  return err;
+  return MuxSet(mux, kChunks[IDX_VP8X].tag, 1, &vp8x, 1);
+}
+
+// Cleans up 'mux' by removing any unnecessary chunks.
+static WebPMuxError MuxCleanup(WebPMux* const mux) {
+  int num_frames;
+  int num_fragments;
+  int num_anim_chunks;
+
+  // If we have an image with a single fragment or frame, and its rectangle
+  // covers the whole canvas, convert it to a non-animated non-fragmented image
+  // (to avoid writing FRGM/ANMF chunk unnecessarily).
+  WebPMuxError err = WebPMuxNumChunks(mux, kChunks[IDX_ANMF].id, &num_frames);
+  if (err != WEBP_MUX_OK) return err;
+  err = WebPMuxNumChunks(mux, kChunks[IDX_FRGM].id, &num_fragments);
+  if (err != WEBP_MUX_OK) return err;
+  if (num_frames == 1 || num_fragments == 1) {
+    WebPMuxImage* frame_frag;
+    err = MuxImageGetNth((const WebPMuxImage**)&mux->images_, 1, &frame_frag);
+    assert(err == WEBP_MUX_OK);  // We know that one frame/fragment does exist.
+    assert(frame_frag != NULL);
+    if (frame_frag->header_ != NULL &&
+        ((mux->canvas_width_ == 0 && mux->canvas_height_ == 0) ||
+         (frame_frag->width_ == mux->canvas_width_ &&
+          frame_frag->height_ == mux->canvas_height_))) {
+      assert(frame_frag->header_->tag_ == kChunks[IDX_ANMF].tag ||
+             frame_frag->header_->tag_ == kChunks[IDX_FRGM].tag);
+      ChunkDelete(frame_frag->header_);  // Removes ANMF/FRGM chunk.
+      frame_frag->header_ = NULL;
+      num_frames = 0;
+      num_fragments = 0;
+    }
+  }
+  // Remove ANIM chunk if this is a non-animated image.
+  err = WebPMuxNumChunks(mux, kChunks[IDX_ANIM].id, &num_anim_chunks);
+  if (err != WEBP_MUX_OK) return err;
+  if (num_anim_chunks >= 1 && num_frames == 0) {
+    err = MuxDeleteAllNamedData(mux, kChunks[IDX_ANIM].tag);
+    if (err != WEBP_MUX_OK) return err;
+  }
+  return WEBP_MUX_OK;
+}
+
+// Total size of a list of images.
+static size_t ImageListDiskSize(const WebPMuxImage* wpi_list) {
+  size_t size = 0;
+  while (wpi_list != NULL) {
+    size += MuxImageDiskSize(wpi_list);
+    wpi_list = wpi_list->next_;
+  }
+  return size;
+}
+
+// Write out the given list of images into 'dst'.
+static uint8_t* ImageListEmit(const WebPMuxImage* wpi_list, uint8_t* dst) {
+  while (wpi_list != NULL) {
+    dst = MuxImageEmit(wpi_list, dst);
+    wpi_list = wpi_list->next_;
+  }
+  return dst;
 }
 
 WebPMuxError WebPMuxAssemble(WebPMux* mux, WebPData* assembled_data) {
   size_t size = 0;
   uint8_t* data = NULL;
   uint8_t* dst = NULL;
-  int num_frames;
-  int num_loop_chunks;
   WebPMuxError err;
 
-  if (mux == NULL || assembled_data == NULL) {
+  if (assembled_data == NULL) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
+  // Clean up returned data, in case something goes wrong.
+  memset(assembled_data, 0, sizeof(*assembled_data));
 
-  // Remove LOOP chunk if unnecessary.
-  err = WebPMuxNumChunks(mux, kChunks[IDX_LOOP].id, &num_loop_chunks);
-  if (err != WEBP_MUX_OK) return err;
-  if (num_loop_chunks >= 1) {
-    err = WebPMuxNumChunks(mux, kChunks[IDX_FRAME].id, &num_frames);
-    if (err != WEBP_MUX_OK) return err;
-    if (num_frames == 0) {
-      err = DeleteLoopCount(mux);
-      if (err != WEBP_MUX_OK) return err;
-    }
+  if (mux == NULL) {
+    return WEBP_MUX_INVALID_ARGUMENT;
   }
 
-  // Create VP8X chunk.
+  // Finalize mux.
+  err = MuxCleanup(mux);
+  if (err != WEBP_MUX_OK) return err;
   err = CreateVP8XChunk(mux);
   if (err != WEBP_MUX_OK) return err;
 
   // Allocate data.
-  size = ChunksListDiskSize(mux->vp8x_) + ChunksListDiskSize(mux->iccp_)
-       + ChunksListDiskSize(mux->loop_) + MuxImageListDiskSize(mux->images_)
-       + ChunksListDiskSize(mux->meta_) + ChunksListDiskSize(mux->unknown_)
-       + RIFF_HEADER_SIZE;
+  size = ChunkListDiskSize(mux->vp8x_) + ChunkListDiskSize(mux->iccp_)
+       + ChunkListDiskSize(mux->anim_) + ImageListDiskSize(mux->images_)
+       + ChunkListDiskSize(mux->exif_) + ChunkListDiskSize(mux->xmp_)
+       + ChunkListDiskSize(mux->unknown_) + RIFF_HEADER_SIZE;
 
-  data = (uint8_t*)malloc(size);
+  data = (uint8_t*)WebPSafeMalloc(1ULL, size);
   if (data == NULL) return WEBP_MUX_MEMORY_ERROR;
 
   // Emit header & chunks.
   dst = MuxEmitRiffHeader(data, size);
   dst = ChunkListEmit(mux->vp8x_, dst);
   dst = ChunkListEmit(mux->iccp_, dst);
-  dst = ChunkListEmit(mux->loop_, dst);
-  dst = MuxImageListEmit(mux->images_, dst);
-  dst = ChunkListEmit(mux->meta_, dst);
+  dst = ChunkListEmit(mux->anim_, dst);
+  dst = ImageListEmit(mux->images_, dst);
+  dst = ChunkListEmit(mux->exif_, dst);
+  dst = ChunkListEmit(mux->xmp_, dst);
   dst = ChunkListEmit(mux->unknown_, dst);
   assert(dst == data + size);
 
   // Validate mux.
   err = MuxValidate(mux);
   if (err != WEBP_MUX_OK) {
-    free(data);
+    WebPSafeFree(data);
     data = NULL;
     size = 0;
   }
 
-  // Finalize.
-  assembled_data->bytes_ = data;
-  assembled_data->size_ = size;
+  // Finalize data.
+  assembled_data->bytes = data;
+  assembled_data->size = size;
 
   return err;
 }
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/mux/muxi.h b/drivers/webp/mux/muxi.h
index 2f06f3ed03..718b2f5d58 100644
--- a/drivers/webp/mux/muxi.h
+++ b/drivers/webp/mux/muxi.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Internal header for mux library.
@@ -15,34 +17,41 @@
 #include <stdlib.h>
 #include "../dec/vp8i.h"
 #include "../dec/vp8li.h"
-#include "../format_constants.h"
-#include "../mux.h"
+#include "../webp/mux.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
 //------------------------------------------------------------------------------
 // Defines and constants.
 
+#define MUX_MAJ_VERSION 0
+#define MUX_MIN_VERSION 2
+#define MUX_REV_VERSION 2
+
 // Chunk object.
 typedef struct WebPChunk WebPChunk;
 struct WebPChunk {
   uint32_t        tag_;
   int             owner_;  // True if *data_ memory is owned internally.
-                           // VP8X, Loop, and other internally created chunks
-                           // like frame/tile are always owned.
+                           // VP8X, ANIM, and other internally created chunks
+                           // like ANMF/FRGM are always owned.
   WebPData        data_;
   WebPChunk*      next_;
 };
 
-// MuxImage object. Store a full webp image (including frame/tile chunk, alpha
+// MuxImage object. Store a full WebP image (including ANMF/FRGM chunk, ALPH
 // chunk and VP8/VP8L chunk),
 typedef struct WebPMuxImage WebPMuxImage;
 struct WebPMuxImage {
-  WebPChunk*  header_;      // Corresponds to WEBP_CHUNK_FRAME/WEBP_CHUNK_TILE.
+  WebPChunk*  header_;      // Corresponds to WEBP_CHUNK_ANMF/WEBP_CHUNK_FRGM.
   WebPChunk*  alpha_;       // Corresponds to WEBP_CHUNK_ALPHA.
   WebPChunk*  img_;         // Corresponds to WEBP_CHUNK_IMAGE.
+  WebPChunk*  unknown_;     // Corresponds to WEBP_CHUNK_UNKNOWN.
+  int         width_;
+  int         height_;
+  int         has_alpha_;   // Through ALPH chunk or as part of VP8L.
   int         is_partial_;  // True if only some of the chunks are filled.
   WebPMuxImage* next_;
 };
@@ -51,11 +60,14 @@ struct WebPMuxImage {
 struct WebPMux {
   WebPMuxImage*   images_;
   WebPChunk*      iccp_;
-  WebPChunk*      meta_;
-  WebPChunk*      loop_;
+  WebPChunk*      exif_;
+  WebPChunk*      xmp_;
+  WebPChunk*      anim_;
   WebPChunk*      vp8x_;
 
-  WebPChunk*  unknown_;
+  WebPChunk*      unknown_;
+  int             canvas_width_;
+  int             canvas_height_;
 };
 
 // CHUNK_INDEX enum: used for indexing within 'kChunks' (defined below) only.
@@ -65,13 +77,14 @@ struct WebPMux {
 typedef enum {
   IDX_VP8X = 0,
   IDX_ICCP,
-  IDX_LOOP,
-  IDX_FRAME,
-  IDX_TILE,
+  IDX_ANIM,
+  IDX_ANMF,
+  IDX_FRGM,
   IDX_ALPHA,
   IDX_VP8,
   IDX_VP8L,
-  IDX_META,
+  IDX_EXIF,
+  IDX_XMP,
   IDX_UNKNOWN,
 
   IDX_NIL,
@@ -80,8 +93,6 @@ typedef enum {
 
 #define NIL_TAG 0x00000000u  // To signal void chunk.
 
-#define MKFOURCC(a, b, c, d) ((uint32_t)(a) | (b) << 8 | (c) << 16 | (d) << 24)
-
 typedef struct {
   uint32_t      tag;
   WebPChunkId   id;
@@ -91,55 +102,23 @@ typedef struct {
 extern const ChunkInfo kChunks[IDX_LAST_CHUNK];
 
 //------------------------------------------------------------------------------
-// Helper functions.
-
-// Read 16, 24 or 32 bits stored in little-endian order.
-static WEBP_INLINE int GetLE16(const uint8_t* const data) {
-  return (int)(data[0] << 0) | (data[1] << 8);
-}
-
-static WEBP_INLINE int GetLE24(const uint8_t* const data) {
-  return GetLE16(data) | (data[2] << 16);
-}
-
-static WEBP_INLINE uint32_t GetLE32(const uint8_t* const data) {
-  return (uint32_t)GetLE16(data) | (GetLE16(data + 2) << 16);
-}
-
-// Store 16, 24 or 32 bits in little-endian order.
-static WEBP_INLINE void PutLE16(uint8_t* const data, int val) {
-  assert(val < (1 << 16));
-  data[0] = (val >> 0);
-  data[1] = (val >> 8);
-}
-
-static WEBP_INLINE void PutLE24(uint8_t* const data, int val) {
-  assert(val < (1 << 24));
-  PutLE16(data, val & 0xffff);
-  data[2] = (val >> 16);
-}
-
-static WEBP_INLINE void PutLE32(uint8_t* const data, uint32_t val) {
-  PutLE16(data, (int)(val & 0xffff));
-  PutLE16(data + 2, (int)(val >> 16));
-}
-
-static WEBP_INLINE size_t SizeWithPadding(size_t chunk_size) {
-  return CHUNK_HEADER_SIZE + ((chunk_size + 1) & ~1U);
-}
-
-//------------------------------------------------------------------------------
 // Chunk object management.
 
 // Initialize.
 void ChunkInit(WebPChunk* const chunk);
 
-// Get chunk index from chunk tag. Returns IDX_NIL if not found.
+// Get chunk index from chunk tag. Returns IDX_UNKNOWN if not found.
 CHUNK_INDEX ChunkGetIndexFromTag(uint32_t tag);
 
-// Get chunk id from chunk tag. Returns WEBP_CHUNK_NIL if not found.
+// Get chunk id from chunk tag. Returns WEBP_CHUNK_UNKNOWN if not found.
 WebPChunkId ChunkGetIdFromTag(uint32_t tag);
 
+// Convert a fourcc string to a tag.
+uint32_t ChunkGetTagFromFourCC(const char fourcc[4]);
+
+// Get chunk index from fourcc. Returns IDX_UNKNOWN if given fourcc is unknown.
+CHUNK_INDEX ChunkGetIndexFromFourCC(const char fourcc[4]);
+
 // Search for nth chunk with given 'tag' in the chunk list.
 // nth = 0 means "last of the list".
 WebPChunk* ChunkSearchList(WebPChunk* first, uint32_t nth, uint32_t tag);
@@ -150,7 +129,8 @@ WebPMuxError ChunkAssignData(WebPChunk* chunk, const WebPData* const data,
 
 // Sets 'chunk' at nth position in the 'chunk_list'.
 // nth = 0 has the special meaning "last of the list".
-WebPMuxError ChunkSetNth(const WebPChunk* chunk, WebPChunk** chunk_list,
+// On success ownership is transferred from 'chunk' to the 'chunk_list'.
+WebPMuxError ChunkSetNth(WebPChunk* chunk, WebPChunk** chunk_list,
                          uint32_t nth);
 
 // Releases chunk and returns chunk->next_.
@@ -159,23 +139,27 @@ WebPChunk* ChunkRelease(WebPChunk* const chunk);
 // Deletes given chunk & returns chunk->next_.
 WebPChunk* ChunkDelete(WebPChunk* const chunk);
 
+// Deletes all chunks in the given chunk list.
+void ChunkListDelete(WebPChunk** const chunk_list);
+
+// Returns size of the chunk including chunk header and padding byte (if any).
+static WEBP_INLINE size_t SizeWithPadding(size_t chunk_size) {
+  return CHUNK_HEADER_SIZE + ((chunk_size + 1) & ~1U);
+}
+
 // Size of a chunk including header and padding.
 static WEBP_INLINE size_t ChunkDiskSize(const WebPChunk* chunk) {
-  const size_t data_size = chunk->data_.size_;
+  const size_t data_size = chunk->data_.size;
   assert(data_size < MAX_CHUNK_PAYLOAD);
   return SizeWithPadding(data_size);
 }
 
 // Total size of a list of chunks.
-size_t ChunksListDiskSize(const WebPChunk* chunk_list);
+size_t ChunkListDiskSize(const WebPChunk* chunk_list);
 
 // Write out the given list of chunks into 'dst'.
 uint8_t* ChunkListEmit(const WebPChunk* chunk_list, uint8_t* dst);
 
-// Get the width & height of image stored in 'image_chunk'.
-WebPMuxError MuxGetImageWidthHeight(const WebPChunk* const image_chunk,
-                                    int* const width, int* const height);
-
 //------------------------------------------------------------------------------
 // MuxImage object management.
 
@@ -189,82 +173,59 @@ WebPMuxImage* MuxImageRelease(WebPMuxImage* const wpi);
 // 'wpi' can be NULL.
 WebPMuxImage* MuxImageDelete(WebPMuxImage* const wpi);
 
-// Delete all images in 'wpi_list'.
-void MuxImageDeleteAll(WebPMuxImage** const wpi_list);
-
 // Count number of images matching the given tag id in the 'wpi_list'.
+// If id == WEBP_CHUNK_NIL, all images will be matched.
 int MuxImageCount(const WebPMuxImage* wpi_list, WebPChunkId id);
 
+// Update width/height/has_alpha info from chunks within wpi.
+// Also remove ALPH chunk if not needed.
+int MuxImageFinalize(WebPMuxImage* const wpi);
+
 // Check if given ID corresponds to an image related chunk.
 static WEBP_INLINE int IsWPI(WebPChunkId id) {
   switch (id) {
-    case WEBP_CHUNK_FRAME:
-    case WEBP_CHUNK_TILE:
+    case WEBP_CHUNK_ANMF:
+    case WEBP_CHUNK_FRGM:
     case WEBP_CHUNK_ALPHA:
     case WEBP_CHUNK_IMAGE:  return 1;
     default:        return 0;
   }
 }
 
-// Get a reference to appropriate chunk list within an image given chunk tag.
-static WEBP_INLINE WebPChunk** MuxImageGetListFromId(
-    const WebPMuxImage* const wpi, WebPChunkId id) {
-  assert(wpi != NULL);
-  switch (id) {
-    case WEBP_CHUNK_FRAME:
-    case WEBP_CHUNK_TILE:  return (WebPChunk**)&wpi->header_;
-    case WEBP_CHUNK_ALPHA: return (WebPChunk**)&wpi->alpha_;
-    case WEBP_CHUNK_IMAGE: return (WebPChunk**)&wpi->img_;
-    default: return NULL;
-  }
-}
-
 // Pushes 'wpi' at the end of 'wpi_list'.
 WebPMuxError MuxImagePush(const WebPMuxImage* wpi, WebPMuxImage** wpi_list);
 
-// Delete nth image in the image list with given tag id.
-WebPMuxError MuxImageDeleteNth(WebPMuxImage** wpi_list, uint32_t nth,
-                               WebPChunkId id);
+// Delete nth image in the image list.
+WebPMuxError MuxImageDeleteNth(WebPMuxImage** wpi_list, uint32_t nth);
 
-// Get nth image in the image list with given tag id.
+// Get nth image in the image list.
 WebPMuxError MuxImageGetNth(const WebPMuxImage** wpi_list, uint32_t nth,
-                            WebPChunkId id, WebPMuxImage** wpi);
+                            WebPMuxImage** wpi);
 
 // Total size of the given image.
 size_t MuxImageDiskSize(const WebPMuxImage* const wpi);
 
-// Total size of a list of images.
-size_t MuxImageListDiskSize(const WebPMuxImage* wpi_list);
-
 // Write out the given image into 'dst'.
 uint8_t* MuxImageEmit(const WebPMuxImage* const wpi, uint8_t* dst);
 
-// Write out the given list of images into 'dst'.
-uint8_t* MuxImageListEmit(const WebPMuxImage* wpi_list, uint8_t* dst);
-
 //------------------------------------------------------------------------------
 // Helper methods for mux.
 
-// Checks if the given image list contains at least one lossless image.
-int MuxHasLosslessImages(const WebPMuxImage* images);
+// Checks if the given image list contains at least one image with alpha.
+int MuxHasAlpha(const WebPMuxImage* images);
 
 // Write out RIFF header into 'data', given total data size 'size'.
 uint8_t* MuxEmitRiffHeader(uint8_t* const data, size_t size);
 
 // Returns the list where chunk with given ID is to be inserted in mux.
-// Return value is NULL if this chunk should be inserted in mux->images_ list
-// or if 'id' is not known.
 WebPChunk** MuxGetChunkListFromId(const WebPMux* mux, WebPChunkId id);
 
-// Validates that the given mux has a single image.
-WebPMuxError MuxValidateForImage(const WebPMux* const mux);
-
 // Validates the given mux object.
 WebPMuxError MuxValidate(const WebPMux* const mux);
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/mux/muxinternal.c b/drivers/webp/mux/muxinternal.c
index 6c3c4fe60a..4babbe82fc 100644
--- a/drivers/webp/mux/muxinternal.c
+++ b/drivers/webp/mux/muxinternal.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Internal objects and utils for mux.
@@ -12,29 +14,33 @@
 
 #include <assert.h>
 #include "./muxi.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "../utils/utils.h"
 
 #define UNDEFINED_CHUNK_SIZE (-1)
 
 const ChunkInfo kChunks[] = {
   { MKFOURCC('V', 'P', '8', 'X'),  WEBP_CHUNK_VP8X,    VP8X_CHUNK_SIZE },
   { MKFOURCC('I', 'C', 'C', 'P'),  WEBP_CHUNK_ICCP,    UNDEFINED_CHUNK_SIZE },
-  { MKFOURCC('L', 'O', 'O', 'P'),  WEBP_CHUNK_LOOP,    LOOP_CHUNK_SIZE },
-  { MKFOURCC('F', 'R', 'M', ' '),  WEBP_CHUNK_FRAME,   FRAME_CHUNK_SIZE },
-  { MKFOURCC('T', 'I', 'L', 'E'),  WEBP_CHUNK_TILE,    TILE_CHUNK_SIZE },
+  { MKFOURCC('A', 'N', 'I', 'M'),  WEBP_CHUNK_ANIM,    ANIM_CHUNK_SIZE },
+  { MKFOURCC('A', 'N', 'M', 'F'),  WEBP_CHUNK_ANMF,    ANMF_CHUNK_SIZE },
+  { MKFOURCC('F', 'R', 'G', 'M'),  WEBP_CHUNK_FRGM,    FRGM_CHUNK_SIZE },
   { MKFOURCC('A', 'L', 'P', 'H'),  WEBP_CHUNK_ALPHA,   UNDEFINED_CHUNK_SIZE },
   { MKFOURCC('V', 'P', '8', ' '),  WEBP_CHUNK_IMAGE,   UNDEFINED_CHUNK_SIZE },
   { MKFOURCC('V', 'P', '8', 'L'),  WEBP_CHUNK_IMAGE,   UNDEFINED_CHUNK_SIZE },
-  { MKFOURCC('M', 'E', 'T', 'A'),  WEBP_CHUNK_META,    UNDEFINED_CHUNK_SIZE },
-  { MKFOURCC('U', 'N', 'K', 'N'),  WEBP_CHUNK_UNKNOWN, UNDEFINED_CHUNK_SIZE },
+  { MKFOURCC('E', 'X', 'I', 'F'),  WEBP_CHUNK_EXIF,    UNDEFINED_CHUNK_SIZE },
+  { MKFOURCC('X', 'M', 'P', ' '),  WEBP_CHUNK_XMP,     UNDEFINED_CHUNK_SIZE },
+  { NIL_TAG,                       WEBP_CHUNK_UNKNOWN, UNDEFINED_CHUNK_SIZE },
 
-  { NIL_TAG,                    WEBP_CHUNK_NIL,     UNDEFINED_CHUNK_SIZE }
+  { NIL_TAG,                       WEBP_CHUNK_NIL,     UNDEFINED_CHUNK_SIZE }
 };
 
 //------------------------------------------------------------------------------
+
+int WebPGetMuxVersion(void) {
+  return (MUX_MAJ_VERSION << 16) | (MUX_MIN_VERSION << 8) | MUX_REV_VERSION;
+}
+
+//------------------------------------------------------------------------------
 // Life of a chunk object.
 
 void ChunkInit(WebPChunk* const chunk) {
@@ -60,9 +66,9 @@ WebPChunk* ChunkRelease(WebPChunk* const chunk) {
 CHUNK_INDEX ChunkGetIndexFromTag(uint32_t tag) {
   int i;
   for (i = 0; kChunks[i].tag != NIL_TAG; ++i) {
-    if (tag == kChunks[i].tag) return i;
+    if (tag == kChunks[i].tag) return (CHUNK_INDEX)i;
   }
-  return IDX_NIL;
+  return IDX_UNKNOWN;
 }
 
 WebPChunkId ChunkGetIdFromTag(uint32_t tag) {
@@ -70,7 +76,16 @@ WebPChunkId ChunkGetIdFromTag(uint32_t tag) {
   for (i = 0; kChunks[i].tag != NIL_TAG; ++i) {
     if (tag == kChunks[i].tag) return kChunks[i].id;
   }
-  return WEBP_CHUNK_NIL;
+  return WEBP_CHUNK_UNKNOWN;
+}
+
+uint32_t ChunkGetTagFromFourCC(const char fourcc[4]) {
+  return MKFOURCC(fourcc[0], fourcc[1], fourcc[2], fourcc[3]);
+}
+
+CHUNK_INDEX ChunkGetIndexFromFourCC(const char fourcc[4]) {
+  const uint32_t tag = ChunkGetTagFromFourCC(fourcc);
+  return ChunkGetIndexFromTag(tag);
 }
 
 //------------------------------------------------------------------------------
@@ -78,7 +93,7 @@ WebPChunkId ChunkGetIdFromTag(uint32_t tag) {
 
 // Returns next chunk in the chunk list with the given tag.
 static WebPChunk* ChunkSearchNextInList(WebPChunk* chunk, uint32_t tag) {
-  while (chunk && chunk->tag_ != tag) {
+  while (chunk != NULL && chunk->tag_ != tag) {
     chunk = chunk->next_;
   }
   return chunk;
@@ -87,7 +102,7 @@ static WebPChunk* ChunkSearchNextInList(WebPChunk* chunk, uint32_t tag) {
 WebPChunk* ChunkSearchList(WebPChunk* first, uint32_t nth, uint32_t tag) {
   uint32_t iter = nth;
   first = ChunkSearchNextInList(first, tag);
-  if (!first) return NULL;
+  if (first == NULL) return NULL;
 
   while (--iter != 0) {
     WebPChunk* next_chunk = ChunkSearchNextInList(first->next_, tag);
@@ -99,14 +114,14 @@ WebPChunk* ChunkSearchList(WebPChunk* first, uint32_t nth, uint32_t tag) {
 
 // Outputs a pointer to 'prev_chunk->next_',
 //   where 'prev_chunk' is the pointer to the chunk at position (nth - 1).
-// Returns 1 if nth chunk was found, 0 otherwise.
+// Returns true if nth chunk was found.
 static int ChunkSearchListToSet(WebPChunk** chunk_list, uint32_t nth,
                                 WebPChunk*** const location) {
   uint32_t count = 0;
-  assert(chunk_list);
+  assert(chunk_list != NULL);
   *location = chunk_list;
 
-  while (*chunk_list) {
+  while (*chunk_list != NULL) {
     WebPChunk* const cur_chunk = *chunk_list;
     ++count;
     if (count == nth) return 1;  // Found.
@@ -124,34 +139,25 @@ static int ChunkSearchListToSet(WebPChunk** chunk_list, uint32_t nth,
 WebPMuxError ChunkAssignData(WebPChunk* chunk, const WebPData* const data,
                              int copy_data, uint32_t tag) {
   // For internally allocated chunks, always copy data & make it owner of data.
-  if (tag == kChunks[IDX_VP8X].tag || tag == kChunks[IDX_LOOP].tag) {
+  if (tag == kChunks[IDX_VP8X].tag || tag == kChunks[IDX_ANIM].tag) {
     copy_data = 1;
   }
 
   ChunkRelease(chunk);
 
   if (data != NULL) {
-    if (copy_data) {
-      // Copy data.
-      chunk->data_.bytes_ = (uint8_t*)malloc(data->size_);
-      if (chunk->data_.bytes_ == NULL) return WEBP_MUX_MEMORY_ERROR;
-      memcpy((uint8_t*)chunk->data_.bytes_, data->bytes_, data->size_);
-      chunk->data_.size_ = data->size_;
-
-      // Chunk is owner of data.
-      chunk->owner_ = 1;
-    } else {
-      // Don't copy data.
+    if (copy_data) {        // Copy data.
+      if (!WebPDataCopy(data, &chunk->data_)) return WEBP_MUX_MEMORY_ERROR;
+      chunk->owner_ = 1;    // Chunk is owner of data.
+    } else {                // Don't copy data.
       chunk->data_ = *data;
     }
   }
-
   chunk->tag_ = tag;
-
   return WEBP_MUX_OK;
 }
 
-WebPMuxError ChunkSetNth(const WebPChunk* chunk, WebPChunk** chunk_list,
+WebPMuxError ChunkSetNth(WebPChunk* chunk, WebPChunk** chunk_list,
                          uint32_t nth) {
   WebPChunk* new_chunk;
 
@@ -159,9 +165,10 @@ WebPMuxError ChunkSetNth(const WebPChunk* chunk, WebPChunk** chunk_list,
     return WEBP_MUX_NOT_FOUND;
   }
 
-  new_chunk = (WebPChunk*)malloc(sizeof(*new_chunk));
+  new_chunk = (WebPChunk*)WebPSafeMalloc(1ULL, sizeof(*new_chunk));
   if (new_chunk == NULL) return WEBP_MUX_MEMORY_ERROR;
   *new_chunk = *chunk;
+  chunk->owner_ = 0;
   new_chunk->next_ = *chunk_list;
   *chunk_list = new_chunk;
   return WEBP_MUX_OK;
@@ -172,70 +179,47 @@ WebPMuxError ChunkSetNth(const WebPChunk* chunk, WebPChunk** chunk_list,
 
 WebPChunk* ChunkDelete(WebPChunk* const chunk) {
   WebPChunk* const next = ChunkRelease(chunk);
-  free(chunk);
+  WebPSafeFree(chunk);
   return next;
 }
 
-//------------------------------------------------------------------------------
-// Chunk serialization methods.
-
-size_t ChunksListDiskSize(const WebPChunk* chunk_list) {
-  size_t size = 0;
-  while (chunk_list) {
-    size += ChunkDiskSize(chunk_list);
-    chunk_list = chunk_list->next_;
+void ChunkListDelete(WebPChunk** const chunk_list) {
+  while (*chunk_list != NULL) {
+    *chunk_list = ChunkDelete(*chunk_list);
   }
-  return size;
 }
 
+//------------------------------------------------------------------------------
+// Chunk serialization methods.
+
 static uint8_t* ChunkEmit(const WebPChunk* const chunk, uint8_t* dst) {
-  const size_t chunk_size = chunk->data_.size_;
+  const size_t chunk_size = chunk->data_.size;
   assert(chunk);
   assert(chunk->tag_ != NIL_TAG);
   PutLE32(dst + 0, chunk->tag_);
   PutLE32(dst + TAG_SIZE, (uint32_t)chunk_size);
   assert(chunk_size == (uint32_t)chunk_size);
-  memcpy(dst + CHUNK_HEADER_SIZE, chunk->data_.bytes_, chunk_size);
+  memcpy(dst + CHUNK_HEADER_SIZE, chunk->data_.bytes, chunk_size);
   if (chunk_size & 1)
     dst[CHUNK_HEADER_SIZE + chunk_size] = 0;  // Add padding.
   return dst + ChunkDiskSize(chunk);
 }
 
 uint8_t* ChunkListEmit(const WebPChunk* chunk_list, uint8_t* dst) {
-  while (chunk_list) {
+  while (chunk_list != NULL) {
     dst = ChunkEmit(chunk_list, dst);
     chunk_list = chunk_list->next_;
   }
   return dst;
 }
 
-//------------------------------------------------------------------------------
-// Manipulation of a WebPData object.
-
-void WebPDataInit(WebPData* webp_data) {
-  if (webp_data != NULL) {
-    memset(webp_data, 0, sizeof(*webp_data));
-  }
-}
-
-void WebPDataClear(WebPData* webp_data) {
-  if (webp_data != NULL) {
-    free((void*)webp_data->bytes_);
-    WebPDataInit(webp_data);
-  }
-}
-
-int WebPDataCopy(const WebPData* src, WebPData* dst) {
-  if (src == NULL || dst == NULL) return 0;
-
-  WebPDataInit(dst);
-  if (src->bytes_ != NULL && src->size_ != 0) {
-    dst->bytes_ = (uint8_t*)malloc(src->size_);
-    if (dst->bytes_ == NULL) return 0;
-    memcpy((void*)dst->bytes_, src->bytes_, src->size_);
-    dst->size_ = src->size_;
+size_t ChunkListDiskSize(const WebPChunk* chunk_list) {
+  size_t size = 0;
+  while (chunk_list != NULL) {
+    size += ChunkDiskSize(chunk_list);
+    chunk_list = chunk_list->next_;
   }
-  return 1;
+  return size;
 }
 
 //------------------------------------------------------------------------------
@@ -252,6 +236,7 @@ WebPMuxImage* MuxImageRelease(WebPMuxImage* const wpi) {
   ChunkDelete(wpi->header_);
   ChunkDelete(wpi->alpha_);
   ChunkDelete(wpi->img_);
+  ChunkListDelete(&wpi->unknown_);
 
   next = wpi->next_;
   MuxImageInit(wpi);
@@ -261,14 +246,31 @@ WebPMuxImage* MuxImageRelease(WebPMuxImage* const wpi) {
 //------------------------------------------------------------------------------
 // MuxImage search methods.
 
+// Get a reference to appropriate chunk list within an image given chunk tag.
+static WebPChunk** GetChunkListFromId(const WebPMuxImage* const wpi,
+                                      WebPChunkId id) {
+  assert(wpi != NULL);
+  switch (id) {
+    case WEBP_CHUNK_ANMF:
+    case WEBP_CHUNK_FRGM:  return (WebPChunk**)&wpi->header_;
+    case WEBP_CHUNK_ALPHA: return (WebPChunk**)&wpi->alpha_;
+    case WEBP_CHUNK_IMAGE: return (WebPChunk**)&wpi->img_;
+    default: return NULL;
+  }
+}
+
 int MuxImageCount(const WebPMuxImage* wpi_list, WebPChunkId id) {
   int count = 0;
   const WebPMuxImage* current;
   for (current = wpi_list; current != NULL; current = current->next_) {
-    const WebPChunk* const wpi_chunk = *MuxImageGetListFromId(current, id);
-    if (wpi_chunk != NULL) {
-      const WebPChunkId wpi_chunk_id = ChunkGetIdFromTag(wpi_chunk->tag_);
-      if (wpi_chunk_id == id) ++count;
+    if (id == WEBP_CHUNK_NIL) {
+      ++count;  // Special case: count all images.
+    } else {
+      const WebPChunk* const wpi_chunk = *GetChunkListFromId(current, id);
+      if (wpi_chunk != NULL) {
+        const WebPChunkId wpi_chunk_id = ChunkGetIdFromTag(wpi_chunk->tag_);
+        if (wpi_chunk_id == id) ++count;  // Count images with a matching 'id'.
+      }
     }
   }
   return count;
@@ -276,34 +278,22 @@ int MuxImageCount(const WebPMuxImage* wpi_list, WebPChunkId id) {
 
 // Outputs a pointer to 'prev_wpi->next_',
 //   where 'prev_wpi' is the pointer to the image at position (nth - 1).
-// Returns 1 if nth image with given id was found, 0 otherwise.
+// Returns true if nth image was found.
 static int SearchImageToGetOrDelete(WebPMuxImage** wpi_list, uint32_t nth,
-                                    WebPChunkId id,
                                     WebPMuxImage*** const location) {
   uint32_t count = 0;
   assert(wpi_list);
   *location = wpi_list;
 
-  // Search makes sense only for the following.
-  assert(id == WEBP_CHUNK_FRAME || id == WEBP_CHUNK_TILE ||
-         id == WEBP_CHUNK_IMAGE);
-  assert(id != WEBP_CHUNK_IMAGE || nth == 1);
-
   if (nth == 0) {
-    nth = MuxImageCount(*wpi_list, id);
+    nth = MuxImageCount(*wpi_list, WEBP_CHUNK_NIL);
     if (nth == 0) return 0;  // Not found.
   }
 
-  while (*wpi_list) {
+  while (*wpi_list != NULL) {
     WebPMuxImage* const cur_wpi = *wpi_list;
-    const WebPChunk* const wpi_chunk = *MuxImageGetListFromId(cur_wpi, id);
-    if (wpi_chunk != NULL) {
-      const WebPChunkId wpi_chunk_id = ChunkGetIdFromTag(wpi_chunk->tag_);
-      if (wpi_chunk_id == id) {
-        ++count;
-        if (count == nth) return 1;  // Found.
-      }
-    }
+    ++count;
+    if (count == nth) return 1;  // Found.
     wpi_list = &cur_wpi->next_;
     *location = wpi_list;
   }
@@ -322,7 +312,7 @@ WebPMuxError MuxImagePush(const WebPMuxImage* wpi, WebPMuxImage** wpi_list) {
     wpi_list = &cur_wpi->next_;
   }
 
-  new_wpi = (WebPMuxImage*)malloc(sizeof(*new_wpi));
+  new_wpi = (WebPMuxImage*)WebPSafeMalloc(1ULL, sizeof(*new_wpi));
   if (new_wpi == NULL) return WEBP_MUX_MEMORY_ERROR;
   *new_wpi = *wpi;
   new_wpi->next_ = NULL;
@@ -341,20 +331,13 @@ WebPMuxError MuxImagePush(const WebPMuxImage* wpi, WebPMuxImage** wpi_list) {
 WebPMuxImage* MuxImageDelete(WebPMuxImage* const wpi) {
   // Delete the components of wpi. If wpi is NULL this is a noop.
   WebPMuxImage* const next = MuxImageRelease(wpi);
-  free(wpi);
+  WebPSafeFree(wpi);
   return next;
 }
 
-void MuxImageDeleteAll(WebPMuxImage** const wpi_list) {
-  while (*wpi_list) {
-    *wpi_list = MuxImageDelete(*wpi_list);
-  }
-}
-
-WebPMuxError MuxImageDeleteNth(WebPMuxImage** wpi_list, uint32_t nth,
-                               WebPChunkId id) {
+WebPMuxError MuxImageDeleteNth(WebPMuxImage** wpi_list, uint32_t nth) {
   assert(wpi_list);
-  if (!SearchImageToGetOrDelete(wpi_list, nth, id, &wpi_list)) {
+  if (!SearchImageToGetOrDelete(wpi_list, nth, &wpi_list)) {
     return WEBP_MUX_NOT_FOUND;
   }
   *wpi_list = MuxImageDelete(*wpi_list);
@@ -365,10 +348,10 @@ WebPMuxError MuxImageDeleteNth(WebPMuxImage** wpi_list, uint32_t nth,
 // MuxImage reader methods.
 
 WebPMuxError MuxImageGetNth(const WebPMuxImage** wpi_list, uint32_t nth,
-                            WebPChunkId id, WebPMuxImage** wpi) {
+                            WebPMuxImage** wpi) {
   assert(wpi_list);
   assert(wpi);
-  if (!SearchImageToGetOrDelete((WebPMuxImage**)wpi_list, nth, id,
+  if (!SearchImageToGetOrDelete((WebPMuxImage**)wpi_list, nth,
                                 (WebPMuxImage***)&wpi_list)) {
     return WEBP_MUX_NOT_FOUND;
   }
@@ -385,47 +368,48 @@ size_t MuxImageDiskSize(const WebPMuxImage* const wpi) {
   if (wpi->header_ != NULL) size += ChunkDiskSize(wpi->header_);
   if (wpi->alpha_ != NULL) size += ChunkDiskSize(wpi->alpha_);
   if (wpi->img_ != NULL) size += ChunkDiskSize(wpi->img_);
+  if (wpi->unknown_ != NULL) size += ChunkListDiskSize(wpi->unknown_);
   return size;
 }
 
-size_t MuxImageListDiskSize(const WebPMuxImage* wpi_list) {
-  size_t size = 0;
-  while (wpi_list) {
-    size += MuxImageDiskSize(wpi_list);
-    wpi_list = wpi_list->next_;
+// Special case as ANMF/FRGM chunk encapsulates other image chunks.
+static uint8_t* ChunkEmitSpecial(const WebPChunk* const header,
+                                 size_t total_size, uint8_t* dst) {
+  const size_t header_size = header->data_.size;
+  const size_t offset_to_next = total_size - CHUNK_HEADER_SIZE;
+  assert(header->tag_ == kChunks[IDX_ANMF].tag ||
+         header->tag_ == kChunks[IDX_FRGM].tag);
+  PutLE32(dst + 0, header->tag_);
+  PutLE32(dst + TAG_SIZE, (uint32_t)offset_to_next);
+  assert(header_size == (uint32_t)header_size);
+  memcpy(dst + CHUNK_HEADER_SIZE, header->data_.bytes, header_size);
+  if (header_size & 1) {
+    dst[CHUNK_HEADER_SIZE + header_size] = 0;  // Add padding.
   }
-  return size;
+  return dst + ChunkDiskSize(header);
 }
 
 uint8_t* MuxImageEmit(const WebPMuxImage* const wpi, uint8_t* dst) {
   // Ordering of chunks to be emitted is strictly as follows:
-  // 1. Frame/Tile chunk (if present).
-  // 2. Alpha chunk (if present).
+  // 1. ANMF/FRGM chunk (if present).
+  // 2. ALPH chunk (if present).
   // 3. VP8/VP8L chunk.
   assert(wpi);
-  if (wpi->header_ != NULL) dst = ChunkEmit(wpi->header_, dst);
+  if (wpi->header_ != NULL) {
+    dst = ChunkEmitSpecial(wpi->header_, MuxImageDiskSize(wpi), dst);
+  }
   if (wpi->alpha_ != NULL) dst = ChunkEmit(wpi->alpha_, dst);
   if (wpi->img_ != NULL) dst = ChunkEmit(wpi->img_, dst);
-  return dst;
-}
-
-uint8_t* MuxImageListEmit(const WebPMuxImage* wpi_list, uint8_t* dst) {
-  while (wpi_list) {
-    dst = MuxImageEmit(wpi_list, dst);
-    wpi_list = wpi_list->next_;
-  }
+  if (wpi->unknown_ != NULL) dst = ChunkListEmit(wpi->unknown_, dst);
   return dst;
 }
 
 //------------------------------------------------------------------------------
 // Helper methods for mux.
 
-int MuxHasLosslessImages(const WebPMuxImage* images) {
+int MuxHasAlpha(const WebPMuxImage* images) {
   while (images != NULL) {
-    assert(images->img_ != NULL);
-    if (images->img_->tag_ == kChunks[IDX_VP8L].tag) {
-      return 1;
-    }
+    if (images->has_alpha_) return 1;
     images = images->next_;
   }
   return 0;
@@ -441,30 +425,13 @@ uint8_t* MuxEmitRiffHeader(uint8_t* const data, size_t size) {
 
 WebPChunk** MuxGetChunkListFromId(const WebPMux* mux, WebPChunkId id) {
   assert(mux != NULL);
-  switch(id) {
+  switch (id) {
     case WEBP_CHUNK_VP8X:    return (WebPChunk**)&mux->vp8x_;
     case WEBP_CHUNK_ICCP:    return (WebPChunk**)&mux->iccp_;
-    case WEBP_CHUNK_LOOP:    return (WebPChunk**)&mux->loop_;
-    case WEBP_CHUNK_META:    return (WebPChunk**)&mux->meta_;
-    case WEBP_CHUNK_UNKNOWN: return (WebPChunk**)&mux->unknown_;
-    default: return NULL;
-  }
-}
-
-WebPMuxError MuxValidateForImage(const WebPMux* const mux) {
-  const int num_images = MuxImageCount(mux->images_, WEBP_CHUNK_IMAGE);
-  const int num_frames = MuxImageCount(mux->images_, WEBP_CHUNK_FRAME);
-  const int num_tiles  = MuxImageCount(mux->images_, WEBP_CHUNK_TILE);
-
-  if (num_images == 0) {
-    // No images in mux.
-    return WEBP_MUX_NOT_FOUND;
-  } else if (num_images == 1 && num_frames == 0 && num_tiles == 0) {
-    // Valid case (single image).
-    return WEBP_MUX_OK;
-  } else {
-    // Frame/Tile case OR an invalid mux.
-    return WEBP_MUX_INVALID_ARGUMENT;
+    case WEBP_CHUNK_ANIM:    return (WebPChunk**)&mux->anim_;
+    case WEBP_CHUNK_EXIF:    return (WebPChunk**)&mux->exif_;
+    case WEBP_CHUNK_XMP:     return (WebPChunk**)&mux->xmp_;
+    default:                 return (WebPChunk**)&mux->unknown_;
   }
 }
 
@@ -480,7 +447,7 @@ static int IsNotCompatible(int feature, int num_items) {
 // On success returns WEBP_MUX_OK and stores the chunk count in *num.
 static WebPMuxError ValidateChunk(const WebPMux* const mux, CHUNK_INDEX idx,
                                   WebPFeatureFlags feature,
-                                  WebPFeatureFlags vp8x_flags,
+                                  uint32_t vp8x_flags,
                                   int max, int* num) {
   const WebPMuxError err =
       WebPMuxNumChunks(mux, kChunks[idx].id, num);
@@ -494,10 +461,11 @@ static WebPMuxError ValidateChunk(const WebPMux* const mux, CHUNK_INDEX idx,
 
 WebPMuxError MuxValidate(const WebPMux* const mux) {
   int num_iccp;
-  int num_meta;
-  int num_loop_chunks;
+  int num_exif;
+  int num_xmp;
+  int num_anim;
   int num_frames;
-  int num_tiles;
+  int num_fragments;
   int num_vp8x;
   int num_images;
   int num_alpha;
@@ -517,29 +485,33 @@ WebPMuxError MuxValidate(const WebPMux* const mux) {
   err = ValidateChunk(mux, IDX_ICCP, ICCP_FLAG, flags, 1, &num_iccp);
   if (err != WEBP_MUX_OK) return err;
 
+  // At most one EXIF metadata.
+  err = ValidateChunk(mux, IDX_EXIF, EXIF_FLAG, flags, 1, &num_exif);
+  if (err != WEBP_MUX_OK) return err;
+
   // At most one XMP metadata.
-  err = ValidateChunk(mux, IDX_META, META_FLAG, flags, 1, &num_meta);
+  err = ValidateChunk(mux, IDX_XMP, XMP_FLAG, flags, 1, &num_xmp);
   if (err != WEBP_MUX_OK) return err;
 
-  // Animation: ANIMATION_FLAG, loop chunk and frame chunk(s) are consistent.
-  // At most one loop chunk.
-  err = ValidateChunk(mux, IDX_LOOP, NO_FLAG, flags, 1, &num_loop_chunks);
+  // Animation: ANIMATION_FLAG, ANIM chunk and ANMF chunk(s) are consistent.
+  // At most one ANIM chunk.
+  err = ValidateChunk(mux, IDX_ANIM, NO_FLAG, flags, 1, &num_anim);
   if (err != WEBP_MUX_OK) return err;
-  err = ValidateChunk(mux, IDX_FRAME, NO_FLAG, flags, -1, &num_frames);
+  err = ValidateChunk(mux, IDX_ANMF, NO_FLAG, flags, -1, &num_frames);
   if (err != WEBP_MUX_OK) return err;
 
   {
     const int has_animation = !!(flags & ANIMATION_FLAG);
-    if (has_animation && (num_loop_chunks == 0 || num_frames == 0)) {
+    if (has_animation && (num_anim == 0 || num_frames == 0)) {
       return WEBP_MUX_INVALID_ARGUMENT;
     }
-    if (!has_animation && (num_loop_chunks == 1 || num_frames > 0)) {
+    if (!has_animation && (num_anim == 1 || num_frames > 0)) {
       return WEBP_MUX_INVALID_ARGUMENT;
     }
   }
 
-  // Tiling: TILE_FLAG and tile chunk(s) are consistent.
-  err = ValidateChunk(mux, IDX_TILE, TILE_FLAG, flags, -1, &num_tiles);
+  // Fragmentation: FRAGMENTS_FLAG and FRGM chunk(s) are consistent.
+  err = ValidateChunk(mux, IDX_FRGM, FRAGMENTS_FLAG, flags, -1, &num_fragments);
   if (err != WEBP_MUX_OK) return err;
 
   // Verify either VP8X chunk is present OR there is only one elem in
@@ -551,16 +523,22 @@ WebPMuxError MuxValidate(const WebPMux* const mux) {
   if (num_vp8x == 0 && num_images != 1) return WEBP_MUX_INVALID_ARGUMENT;
 
   // ALPHA_FLAG & alpha chunk(s) are consistent.
-  if (num_vp8x > 0 && MuxHasLosslessImages(mux->images_)) {
-    // Special case: we have a VP8X chunk as well as some lossless images.
-    if (!(flags & ALPHA_FLAG)) return WEBP_MUX_INVALID_ARGUMENT;
-  } else {
-    err = ValidateChunk(mux, IDX_ALPHA, ALPHA_FLAG, flags, -1, &num_alpha);
-    if (err != WEBP_MUX_OK) return err;
+  if (MuxHasAlpha(mux->images_)) {
+    if (num_vp8x > 0) {
+      // VP8X chunk is present, so it should contain ALPHA_FLAG.
+      if (!(flags & ALPHA_FLAG)) return WEBP_MUX_INVALID_ARGUMENT;
+    } else {
+      // VP8X chunk is not present, so ALPH chunks should NOT be present either.
+      err = WebPMuxNumChunks(mux, WEBP_CHUNK_ALPHA, &num_alpha);
+      if (err != WEBP_MUX_OK) return err;
+      if (num_alpha > 0) return WEBP_MUX_INVALID_ARGUMENT;
+    }
+  } else {  // Mux doesn't need alpha. So, ALPHA_FLAG should NOT be present.
+    if (flags & ALPHA_FLAG) return WEBP_MUX_INVALID_ARGUMENT;
   }
 
-  // num_tiles & num_images are consistent.
-  if (num_tiles > 0 && num_images != num_tiles) {
+  // num_fragments & num_images are consistent.
+  if (num_fragments > 0 && num_images != num_fragments) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
 
@@ -571,6 +549,3 @@ WebPMuxError MuxValidate(const WebPMux* const mux) {
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/mux/muxread.c b/drivers/webp/mux/muxread.c
index 21c3cfbaeb..8957a1e46e 100644
--- a/drivers/webp/mux/muxread.c
+++ b/drivers/webp/mux/muxread.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Read APIs for mux.
@@ -12,10 +14,7 @@
 
 #include <assert.h>
 #include "./muxi.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 // Helper method(s).
@@ -41,8 +40,9 @@ static WebPMuxError MuxGet(const WebPMux* const mux, CHUNK_INDEX idx,
 
   SWITCH_ID_LIST(IDX_VP8X, mux->vp8x_);
   SWITCH_ID_LIST(IDX_ICCP, mux->iccp_);
-  SWITCH_ID_LIST(IDX_LOOP, mux->loop_);
-  SWITCH_ID_LIST(IDX_META, mux->meta_);
+  SWITCH_ID_LIST(IDX_ANIM, mux->anim_);
+  SWITCH_ID_LIST(IDX_EXIF, mux->exif_);
+  SWITCH_ID_LIST(IDX_XMP, mux->xmp_);
   SWITCH_ID_LIST(IDX_UNKNOWN, mux->unknown_);
   return WEBP_MUX_NOT_FOUND;
 }
@@ -50,15 +50,14 @@ static WebPMuxError MuxGet(const WebPMux* const mux, CHUNK_INDEX idx,
 
 // Fill the chunk with the given data (includes chunk header bytes), after some
 // verifications.
-static WebPMuxError ChunkVerifyAndAssignData(WebPChunk* chunk,
-                                             const uint8_t* data,
-                                             size_t data_size, size_t riff_size,
-                                             int copy_data) {
+static WebPMuxError ChunkVerifyAndAssign(WebPChunk* chunk,
+                                         const uint8_t* data, size_t data_size,
+                                         size_t riff_size, int copy_data) {
   uint32_t chunk_size;
   WebPData chunk_data;
 
   // Sanity checks.
-  if (data_size < TAG_SIZE) return WEBP_MUX_NOT_ENOUGH_DATA;
+  if (data_size < CHUNK_HEADER_SIZE) return WEBP_MUX_NOT_ENOUGH_DATA;
   chunk_size = GetLE32(data + TAG_SIZE);
 
   {
@@ -68,11 +67,103 @@ static WebPMuxError ChunkVerifyAndAssignData(WebPChunk* chunk,
   }
 
   // Data assignment.
-  chunk_data.bytes_ = data + CHUNK_HEADER_SIZE;
-  chunk_data.size_ = chunk_size;
+  chunk_data.bytes = data + CHUNK_HEADER_SIZE;
+  chunk_data.size = chunk_size;
   return ChunkAssignData(chunk, &chunk_data, copy_data, GetLE32(data + 0));
 }
 
+int MuxImageFinalize(WebPMuxImage* const wpi) {
+  const WebPChunk* const img = wpi->img_;
+  const WebPData* const image = &img->data_;
+  const int is_lossless = (img->tag_ == kChunks[IDX_VP8L].tag);
+  int w, h;
+  int vp8l_has_alpha = 0;
+  const int ok = is_lossless ?
+      VP8LGetInfo(image->bytes, image->size, &w, &h, &vp8l_has_alpha) :
+      VP8GetInfo(image->bytes, image->size, image->size, &w, &h);
+  assert(img != NULL);
+  if (ok) {
+    // Ignore ALPH chunk accompanying VP8L.
+    if (is_lossless && (wpi->alpha_ != NULL)) {
+      ChunkDelete(wpi->alpha_);
+      wpi->alpha_ = NULL;
+    }
+    wpi->width_ = w;
+    wpi->height_ = h;
+    wpi->has_alpha_ = vp8l_has_alpha || (wpi->alpha_ != NULL);
+  }
+  return ok;
+}
+
+static int MuxImageParse(const WebPChunk* const chunk, int copy_data,
+                         WebPMuxImage* const wpi) {
+  const uint8_t* bytes = chunk->data_.bytes;
+  size_t size = chunk->data_.size;
+  const uint8_t* const last = bytes + size;
+  WebPChunk subchunk;
+  size_t subchunk_size;
+  ChunkInit(&subchunk);
+
+  assert(chunk->tag_ == kChunks[IDX_ANMF].tag ||
+         chunk->tag_ == kChunks[IDX_FRGM].tag);
+  assert(!wpi->is_partial_);
+
+  // ANMF/FRGM.
+  {
+    const size_t hdr_size = (chunk->tag_ == kChunks[IDX_ANMF].tag) ?
+        ANMF_CHUNK_SIZE : FRGM_CHUNK_SIZE;
+    const WebPData temp = { bytes, hdr_size };
+    // Each of ANMF and FRGM chunk contain a header at the beginning. So, its
+    // size should at least be 'hdr_size'.
+    if (size < hdr_size) goto Fail;
+    ChunkAssignData(&subchunk, &temp, copy_data, chunk->tag_);
+  }
+  ChunkSetNth(&subchunk, &wpi->header_, 1);
+  wpi->is_partial_ = 1;  // Waiting for ALPH and/or VP8/VP8L chunks.
+
+  // Rest of the chunks.
+  subchunk_size = ChunkDiskSize(&subchunk) - CHUNK_HEADER_SIZE;
+  bytes += subchunk_size;
+  size -= subchunk_size;
+
+  while (bytes != last) {
+    ChunkInit(&subchunk);
+    if (ChunkVerifyAndAssign(&subchunk, bytes, size, size,
+                             copy_data) != WEBP_MUX_OK) {
+      goto Fail;
+    }
+    switch (ChunkGetIdFromTag(subchunk.tag_)) {
+      case WEBP_CHUNK_ALPHA:
+        if (wpi->alpha_ != NULL) goto Fail;  // Consecutive ALPH chunks.
+        if (ChunkSetNth(&subchunk, &wpi->alpha_, 1) != WEBP_MUX_OK) goto Fail;
+        wpi->is_partial_ = 1;  // Waiting for a VP8 chunk.
+        break;
+      case WEBP_CHUNK_IMAGE:
+        if (ChunkSetNth(&subchunk, &wpi->img_, 1) != WEBP_MUX_OK) goto Fail;
+        if (!MuxImageFinalize(wpi)) goto Fail;
+        wpi->is_partial_ = 0;  // wpi is completely filled.
+        break;
+      case WEBP_CHUNK_UNKNOWN:
+        if (wpi->is_partial_) goto Fail;  // Encountered an unknown chunk
+                                          // before some image chunks.
+        if (ChunkSetNth(&subchunk, &wpi->unknown_, 0) != WEBP_MUX_OK) goto Fail;
+        break;
+      default:
+        goto Fail;
+        break;
+    }
+    subchunk_size = ChunkDiskSize(&subchunk);
+    bytes += subchunk_size;
+    size -= subchunk_size;
+  }
+  if (wpi->is_partial_) goto Fail;
+  return 1;
+
+ Fail:
+  ChunkRelease(&subchunk);
+  return 0;
+}
+
 //------------------------------------------------------------------------------
 // Create a mux object from WebP-RIFF data.
 
@@ -94,8 +185,8 @@ WebPMux* WebPMuxCreateInternal(const WebPData* bitstream, int copy_data,
   }
   if (bitstream == NULL) return NULL;
 
-  data = bitstream->bytes_;
-  size = bitstream->size_;
+  data = bitstream->bytes;
+  size = bitstream->size;
 
   if (data == NULL) return NULL;
   if (size < RIFF_HEADER_SIZE) return NULL;
@@ -129,48 +220,55 @@ WebPMux* WebPMuxCreateInternal(const WebPData* bitstream, int copy_data,
   data += RIFF_HEADER_SIZE;
   size -= RIFF_HEADER_SIZE;
 
-  wpi = (WebPMuxImage*)malloc(sizeof(*wpi));
+  wpi = (WebPMuxImage*)WebPSafeMalloc(1ULL, sizeof(*wpi));
   if (wpi == NULL) goto Err;
   MuxImageInit(wpi);
 
   // Loop over chunks.
   while (data != end) {
+    size_t data_size;
     WebPChunkId id;
-    WebPMuxError err;
-
-    err = ChunkVerifyAndAssignData(&chunk, data, size, riff_size, copy_data);
-    if (err != WEBP_MUX_OK) goto Err;
-
+    WebPChunk** chunk_list;
+    if (ChunkVerifyAndAssign(&chunk, data, size, riff_size,
+                             copy_data) != WEBP_MUX_OK) {
+      goto Err;
+    }
+    data_size = ChunkDiskSize(&chunk);
     id = ChunkGetIdFromTag(chunk.tag_);
-
-    if (IsWPI(id)) {  // An image chunk (frame/tile/alpha/vp8).
-      WebPChunk** wpi_chunk_ptr =
-          MuxImageGetListFromId(wpi, id);  // Image chunk to set.
-      assert(wpi_chunk_ptr != NULL);
-      if (*wpi_chunk_ptr != NULL) goto Err;  // Consecutive alpha chunks or
-                                             // consecutive frame/tile chunks.
-      if (ChunkSetNth(&chunk, wpi_chunk_ptr, 1) != WEBP_MUX_OK) goto Err;
-      if (id == WEBP_CHUNK_IMAGE) {
+    switch (id) {
+      case WEBP_CHUNK_ALPHA:
+        if (wpi->alpha_ != NULL) goto Err;  // Consecutive ALPH chunks.
+        if (ChunkSetNth(&chunk, &wpi->alpha_, 1) != WEBP_MUX_OK) goto Err;
+        wpi->is_partial_ = 1;  // Waiting for a VP8 chunk.
+        break;
+      case WEBP_CHUNK_IMAGE:
+        if (ChunkSetNth(&chunk, &wpi->img_, 1) != WEBP_MUX_OK) goto Err;
+        if (!MuxImageFinalize(wpi)) goto Err;
         wpi->is_partial_ = 0;  // wpi is completely filled.
+ PushImage:
         // Add this to mux->images_ list.
         if (MuxImagePush(wpi, &mux->images_) != WEBP_MUX_OK) goto Err;
         MuxImageInit(wpi);  // Reset for reading next image.
-      } else {
-        wpi->is_partial_ = 1;  // wpi is only partially filled.
-      }
-    } else {  // A non-image chunk.
-      WebPChunk** chunk_list;
-      if (wpi->is_partial_) goto Err;  // Encountered a non-image chunk before
-                                       // getting all chunks of an image.
-      chunk_list = MuxGetChunkListFromId(mux, id);  // List to add this chunk.
-      if (chunk_list == NULL) chunk_list = &mux->unknown_;
-      if (ChunkSetNth(&chunk, chunk_list, 0) != WEBP_MUX_OK) goto Err;
-    }
-    {
-      const size_t data_size = ChunkDiskSize(&chunk);
-      data += data_size;
-      size -= data_size;
+        break;
+      case WEBP_CHUNK_ANMF:
+        if (wpi->is_partial_) goto Err;  // Previous wpi is still incomplete.
+        if (!MuxImageParse(&chunk, copy_data, wpi)) goto Err;
+        ChunkRelease(&chunk);
+        goto PushImage;
+        break;
+      default:  // A non-image chunk.
+        if (wpi->is_partial_) goto Err;  // Encountered a non-image chunk before
+                                         // getting all chunks of an image.
+        chunk_list = MuxGetChunkListFromId(mux, id);  // List to add this chunk.
+        if (ChunkSetNth(&chunk, chunk_list, 0) != WEBP_MUX_OK) goto Err;
+        if (id == WEBP_CHUNK_VP8X) {  // grab global specs
+          mux->canvas_width_ = GetLE24(data + 12) + 1;
+          mux->canvas_height_ = GetLE24(data + 15) + 1;
+        }
+        break;
     }
+    data += data_size;
+    size -= data_size;
     ChunkInit(&chunk);
   }
 
@@ -190,31 +288,74 @@ WebPMux* WebPMuxCreateInternal(const WebPData* bitstream, int copy_data,
 //------------------------------------------------------------------------------
 // Get API(s).
 
-WebPMuxError WebPMuxGetFeatures(const WebPMux* mux, uint32_t* flags) {
-  WebPData data;
-  WebPMuxError err;
+// Validates that the given mux has a single image.
+static WebPMuxError ValidateForSingleImage(const WebPMux* const mux) {
+  const int num_images = MuxImageCount(mux->images_, WEBP_CHUNK_IMAGE);
+  const int num_frames = MuxImageCount(mux->images_, WEBP_CHUNK_ANMF);
+  const int num_fragments = MuxImageCount(mux->images_, WEBP_CHUNK_FRGM);
+
+  if (num_images == 0) {
+    // No images in mux.
+    return WEBP_MUX_NOT_FOUND;
+  } else if (num_images == 1 && num_frames == 0 && num_fragments == 0) {
+    // Valid case (single image).
+    return WEBP_MUX_OK;
+  } else {
+    // Frame/Fragment case OR an invalid mux.
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+}
 
-  if (mux == NULL || flags == NULL) return WEBP_MUX_INVALID_ARGUMENT;
-  *flags = 0;
+// Get the canvas width, height and flags after validating that VP8X/VP8/VP8L
+// chunk and canvas size are valid.
+static WebPMuxError MuxGetCanvasInfo(const WebPMux* const mux,
+                                     int* width, int* height, uint32_t* flags) {
+  int w, h;
+  uint32_t f = 0;
+  WebPData data;
+  assert(mux != NULL);
 
   // Check if VP8X chunk is present.
-  err = MuxGet(mux, IDX_VP8X, 1, &data);
-  if (err == WEBP_MUX_NOT_FOUND) {
-    // Check if VP8/VP8L chunk is present.
-    err = WebPMuxGetImage(mux, &data);
-    WebPDataClear(&data);
-    return err;
-  } else if (err != WEBP_MUX_OK) {
-    return err;
+  if (MuxGet(mux, IDX_VP8X, 1, &data) == WEBP_MUX_OK) {
+    if (data.size < VP8X_CHUNK_SIZE) return WEBP_MUX_BAD_DATA;
+    f = GetLE32(data.bytes + 0);
+    w = GetLE24(data.bytes + 4) + 1;
+    h = GetLE24(data.bytes + 7) + 1;
+  } else {
+    const WebPMuxImage* const wpi = mux->images_;
+    // Grab user-forced canvas size as default.
+    w = mux->canvas_width_;
+    h = mux->canvas_height_;
+    if (w == 0 && h == 0 && ValidateForSingleImage(mux) == WEBP_MUX_OK) {
+      // single image and not forced canvas size => use dimension of first frame
+      assert(wpi != NULL);
+      w = wpi->width_;
+      h = wpi->height_;
+    }
+    if (wpi != NULL) {
+      if (wpi->has_alpha_) f |= ALPHA_FLAG;
+    }
   }
+  if (w * (uint64_t)h >= MAX_IMAGE_AREA) return WEBP_MUX_BAD_DATA;
 
-  if (data.size_ < CHUNK_SIZE_BYTES) return WEBP_MUX_BAD_DATA;
-
-  // All OK. Fill up flags.
-  *flags = GetLE32(data.bytes_);
+  if (width != NULL) *width = w;
+  if (height != NULL) *height = h;
+  if (flags != NULL) *flags = f;
   return WEBP_MUX_OK;
 }
 
+WebPMuxError WebPMuxGetCanvasSize(const WebPMux* mux, int* width, int* height) {
+  if (mux == NULL || width == NULL || height == NULL) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+  return MuxGetCanvasInfo(mux, width, height, NULL);
+}
+
+WebPMuxError WebPMuxGetFeatures(const WebPMux* mux, uint32_t* flags) {
+  if (mux == NULL || flags == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+  return MuxGetCanvasInfo(mux, NULL, NULL, flags);
+}
+
 static uint8_t* EmitVP8XChunk(uint8_t* const dst, int width,
                               int height, uint32_t flags) {
   const size_t vp8x_size = CHUNK_HEADER_SIZE + VP8X_CHUNK_SIZE;
@@ -230,7 +371,7 @@ static uint8_t* EmitVP8XChunk(uint8_t* const dst, int width,
 }
 
 // Assemble a single image WebP bitstream from 'wpi'.
-static WebPMuxError SynthesizeBitstream(WebPMuxImage* const wpi,
+static WebPMuxError SynthesizeBitstream(const WebPMuxImage* const wpi,
                                         WebPData* const bitstream) {
   uint8_t* dst;
 
@@ -238,25 +379,17 @@ static WebPMuxError SynthesizeBitstream(WebPMuxImage* const wpi,
   const int need_vp8x = (wpi->alpha_ != NULL);
   const size_t vp8x_size = need_vp8x ? CHUNK_HEADER_SIZE + VP8X_CHUNK_SIZE : 0;
   const size_t alpha_size = need_vp8x ? ChunkDiskSize(wpi->alpha_) : 0;
-  // Note: No need to output FRM/TILE chunk for a single image.
+  // Note: No need to output ANMF/FRGM chunk for a single image.
   const size_t size = RIFF_HEADER_SIZE + vp8x_size + alpha_size +
                       ChunkDiskSize(wpi->img_);
-  uint8_t* const data = (uint8_t*)malloc(size);
+  uint8_t* const data = (uint8_t*)WebPSafeMalloc(1ULL, size);
   if (data == NULL) return WEBP_MUX_MEMORY_ERROR;
 
   // Main RIFF header.
   dst = MuxEmitRiffHeader(data, size);
 
   if (need_vp8x) {
-    int w, h;
-    WebPMuxError err;
-    assert(wpi->img_ != NULL);
-    err = MuxGetImageWidthHeight(wpi->img_, &w, &h);
-    if (err != WEBP_MUX_OK) {
-      free(data);
-      return err;
-    }
-    dst = EmitVP8XChunk(dst, w, h, ALPHA_FLAG);  // VP8X.
+    dst = EmitVP8XChunk(dst, wpi->width_, wpi->height_, ALPHA_FLAG);  // VP8X.
     dst = ChunkListEmit(wpi->alpha_, dst);       // ALPH.
   }
 
@@ -265,107 +398,115 @@ static WebPMuxError SynthesizeBitstream(WebPMuxImage* const wpi,
   assert(dst == data + size);
 
   // Output.
-  bitstream->bytes_ = data;
-  bitstream->size_ = size;
+  bitstream->bytes = data;
+  bitstream->size = size;
   return WEBP_MUX_OK;
 }
 
-WebPMuxError WebPMuxGetImage(const WebPMux* mux, WebPData* bitstream) {
-  WebPMuxError err;
-  WebPMuxImage* wpi = NULL;
-
-  if (mux == NULL || bitstream == NULL) {
+WebPMuxError WebPMuxGetChunk(const WebPMux* mux, const char fourcc[4],
+                             WebPData* chunk_data) {
+  CHUNK_INDEX idx;
+  if (mux == NULL || fourcc == NULL || chunk_data == NULL) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
-
-  err = MuxValidateForImage(mux);
-  if (err != WEBP_MUX_OK) return err;
-
-  // All well. Get the image.
-  err = MuxImageGetNth((const WebPMuxImage**)&mux->images_, 1, WEBP_CHUNK_IMAGE,
-                       &wpi);
-  assert(err == WEBP_MUX_OK);  // Already tested above.
-
-  return SynthesizeBitstream(wpi, bitstream);
-}
-
-WebPMuxError WebPMuxGetMetadata(const WebPMux* mux, WebPData* metadata) {
-  if (mux == NULL || metadata == NULL) return WEBP_MUX_INVALID_ARGUMENT;
-  return MuxGet(mux, IDX_META, 1, metadata);
+  idx = ChunkGetIndexFromFourCC(fourcc);
+  if (IsWPI(kChunks[idx].id)) {     // An image chunk.
+    return WEBP_MUX_INVALID_ARGUMENT;
+  } else if (idx != IDX_UNKNOWN) {  // A known chunk type.
+    return MuxGet(mux, idx, 1, chunk_data);
+  } else {                          // An unknown chunk type.
+    const WebPChunk* const chunk =
+        ChunkSearchList(mux->unknown_, 1, ChunkGetTagFromFourCC(fourcc));
+    if (chunk == NULL) return WEBP_MUX_NOT_FOUND;
+    *chunk_data = chunk->data_;
+    return WEBP_MUX_OK;
+  }
 }
 
-WebPMuxError WebPMuxGetColorProfile(const WebPMux* mux,
-                                    WebPData* color_profile) {
-  if (mux == NULL || color_profile == NULL) return WEBP_MUX_INVALID_ARGUMENT;
-  return MuxGet(mux, IDX_ICCP, 1, color_profile);
+static WebPMuxError MuxGetImageInternal(const WebPMuxImage* const wpi,
+                                        WebPMuxFrameInfo* const info) {
+  // Set some defaults for unrelated fields.
+  info->x_offset = 0;
+  info->y_offset = 0;
+  info->duration = 1;
+  info->dispose_method = WEBP_MUX_DISPOSE_NONE;
+  info->blend_method = WEBP_MUX_BLEND;
+  // Extract data for related fields.
+  info->id = ChunkGetIdFromTag(wpi->img_->tag_);
+  return SynthesizeBitstream(wpi, &info->bitstream);
 }
 
-WebPMuxError WebPMuxGetLoopCount(const WebPMux* mux, int* loop_count) {
-  WebPData image;
-  WebPMuxError err;
-
-  if (mux == NULL || loop_count == NULL) return WEBP_MUX_INVALID_ARGUMENT;
-
-  err = MuxGet(mux, IDX_LOOP, 1, &image);
-  if (err != WEBP_MUX_OK) return err;
-  if (image.size_ < kChunks[WEBP_CHUNK_LOOP].size) return WEBP_MUX_BAD_DATA;
-  *loop_count = GetLE16(image.bytes_);
-
-  return WEBP_MUX_OK;
+static WebPMuxError MuxGetFrameFragmentInternal(const WebPMuxImage* const wpi,
+                                                WebPMuxFrameInfo* const frame) {
+  const int is_frame = (wpi->header_->tag_ == kChunks[IDX_ANMF].tag);
+  const CHUNK_INDEX idx = is_frame ? IDX_ANMF : IDX_FRGM;
+  const WebPData* frame_frgm_data;
+  if (!is_frame) return WEBP_MUX_INVALID_ARGUMENT;
+  assert(wpi->header_ != NULL);  // Already checked by WebPMuxGetFrame().
+  // Get frame/fragment chunk.
+  frame_frgm_data = &wpi->header_->data_;
+  if (frame_frgm_data->size < kChunks[idx].size) return WEBP_MUX_BAD_DATA;
+  // Extract info.
+  frame->x_offset = 2 * GetLE24(frame_frgm_data->bytes + 0);
+  frame->y_offset = 2 * GetLE24(frame_frgm_data->bytes + 3);
+  if (is_frame) {
+    const uint8_t bits = frame_frgm_data->bytes[15];
+    frame->duration = GetLE24(frame_frgm_data->bytes + 12);
+    frame->dispose_method =
+        (bits & 1) ? WEBP_MUX_DISPOSE_BACKGROUND : WEBP_MUX_DISPOSE_NONE;
+    frame->blend_method = (bits & 2) ? WEBP_MUX_NO_BLEND : WEBP_MUX_BLEND;
+  } else {  // Defaults for unused values.
+    frame->duration = 1;
+    frame->dispose_method = WEBP_MUX_DISPOSE_NONE;
+    frame->blend_method = WEBP_MUX_BLEND;
+  }
+  frame->id = ChunkGetIdFromTag(wpi->header_->tag_);
+  return SynthesizeBitstream(wpi, &frame->bitstream);
 }
 
-static WebPMuxError MuxGetFrameTileInternal(
-    const WebPMux* const mux, uint32_t nth, WebPData* const bitstream,
-    int* const x_offset, int* const y_offset, int* const duration,
-    uint32_t tag) {
-  const WebPData* frame_tile_data;
+WebPMuxError WebPMuxGetFrame(
+    const WebPMux* mux, uint32_t nth, WebPMuxFrameInfo* frame) {
   WebPMuxError err;
   WebPMuxImage* wpi;
 
-  const int is_frame = (tag == kChunks[WEBP_CHUNK_FRAME].tag) ? 1 : 0;
-  const CHUNK_INDEX idx = is_frame ? IDX_FRAME : IDX_TILE;
-  const WebPChunkId id = kChunks[idx].id;
-
-  if (mux == NULL || bitstream == NULL ||
-      x_offset == NULL || y_offset == NULL || (is_frame && duration == NULL)) {
+  // Sanity checks.
+  if (mux == NULL || frame == NULL) {
     return WEBP_MUX_INVALID_ARGUMENT;
   }
 
   // Get the nth WebPMuxImage.
-  err = MuxImageGetNth((const WebPMuxImage**)&mux->images_, nth, id, &wpi);
+  err = MuxImageGetNth((const WebPMuxImage**)&mux->images_, nth, &wpi);
   if (err != WEBP_MUX_OK) return err;
 
-  // Get frame chunk.
-  assert(wpi->header_ != NULL);  // As MuxImageGetNth() already checked header_.
-  frame_tile_data = &wpi->header_->data_;
+  // Get frame info.
+  if (wpi->header_ == NULL) {
+    return MuxGetImageInternal(wpi, frame);
+  } else {
+    return MuxGetFrameFragmentInternal(wpi, frame);
+  }
+}
 
-  if (frame_tile_data->size_ < kChunks[idx].size) return WEBP_MUX_BAD_DATA;
-  *x_offset = 2 * GetLE24(frame_tile_data->bytes_ + 0);
-  *y_offset = 2 * GetLE24(frame_tile_data->bytes_ + 3);
-  if (is_frame) *duration = 1 + GetLE24(frame_tile_data->bytes_ + 12);
+WebPMuxError WebPMuxGetAnimationParams(const WebPMux* mux,
+                                       WebPMuxAnimParams* params) {
+  WebPData anim;
+  WebPMuxError err;
 
-  return SynthesizeBitstream(wpi, bitstream);
-}
+  if (mux == NULL || params == NULL) return WEBP_MUX_INVALID_ARGUMENT;
 
-WebPMuxError WebPMuxGetFrame(const WebPMux* mux, uint32_t nth,
-                             WebPData* bitstream,
-                             int* x_offset, int* y_offset, int* duration) {
-  return MuxGetFrameTileInternal(mux, nth, bitstream, x_offset, y_offset,
-                                 duration, kChunks[IDX_FRAME].tag);
-}
+  err = MuxGet(mux, IDX_ANIM, 1, &anim);
+  if (err != WEBP_MUX_OK) return err;
+  if (anim.size < kChunks[WEBP_CHUNK_ANIM].size) return WEBP_MUX_BAD_DATA;
+  params->bgcolor = GetLE32(anim.bytes);
+  params->loop_count = GetLE16(anim.bytes + 4);
 
-WebPMuxError WebPMuxGetTile(const WebPMux* mux, uint32_t nth,
-                            WebPData* bitstream,
-                            int* x_offset, int* y_offset) {
-  return MuxGetFrameTileInternal(mux, nth, bitstream, x_offset, y_offset, NULL,
-                                 kChunks[IDX_TILE].tag);
+  return WEBP_MUX_OK;
 }
 
 // Get chunk index from chunk id. Returns IDX_NIL if not found.
 static CHUNK_INDEX ChunkGetIndexFromId(WebPChunkId id) {
   int i;
   for (i = 0; kChunks[i].id != WEBP_CHUNK_NIL; ++i) {
-    if (id == kChunks[i].id) return i;
+    if (id == kChunks[i].id) return (CHUNK_INDEX)i;
   }
   return IDX_NIL;
 }
@@ -393,19 +534,11 @@ WebPMuxError WebPMuxNumChunks(const WebPMux* mux,
     *num_elements = MuxImageCount(mux->images_, id);
   } else {
     WebPChunk* const* chunk_list = MuxGetChunkListFromId(mux, id);
-    if (chunk_list == NULL) {
-      *num_elements = 0;
-    } else {
-      const CHUNK_INDEX idx = ChunkGetIndexFromId(id);
-      *num_elements = CountChunks(*chunk_list, kChunks[idx].tag);
-    }
+    const CHUNK_INDEX idx = ChunkGetIndexFromId(id);
+    *num_elements = CountChunks(*chunk_list, kChunks[idx].tag);
   }
 
   return WEBP_MUX_OK;
 }
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/types.h b/drivers/webp/types.h
index 3e27190bef..98fff35a11 100644
--- a/drivers/webp/types.h
+++ b/drivers/webp/types.h
@@ -1,8 +1,10 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 //  Common types
@@ -16,10 +18,11 @@
 
 #ifndef _MSC_VER
 #include <inttypes.h>
-#ifdef __STRICT_ANSI__
-#define WEBP_INLINE
-#else  /* __STRICT_ANSI__ */
+#if defined(__cplusplus) || !defined(__STRICT_ANSI__) || \
+    (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L)
 #define WEBP_INLINE inline
+#else
+#define WEBP_INLINE
 #endif
 #else
 typedef signed   char int8_t;
@@ -36,7 +39,11 @@ typedef long long int int64_t;
 #ifndef WEBP_EXTERN
 // This explicitly marks library functions and allows for changing the
 // signature for e.g., Windows DLL builds.
-#define WEBP_EXTERN(type) extern type
+# if defined(__GNUC__) && __GNUC__ >= 4
+#  define WEBP_EXTERN(type) extern __attribute__ ((visibility ("default"))) type
+# else
+#  define WEBP_EXTERN(type) extern type
+# endif  /* __GNUC__ >= 4 */
 #endif  /* WEBP_EXTERN */
 
 // Macro to check ABI compatibility (same major revision number)
diff --git a/drivers/webp/utils/bit_reader.c b/drivers/webp/utils/bit_reader.c
index 1afb1db890..cd265321bb 100644
--- a/drivers/webp/utils/bit_reader.c
+++ b/drivers/webp/utils/bit_reader.c
@@ -1,36 +1,54 @@
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
-// Boolean decoder
+// Boolean decoder non-inlined methods
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
-#include "./bit_reader.h"
-
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
+#ifdef HAVE_CONFIG_H
+#include "../webp/config.h"
 #endif
 
-#define MK(X) (((bit_t)(X) << (BITS)) | (MASK))
+#include "./bit_reader_inl.h"
 
 //------------------------------------------------------------------------------
 // VP8BitReader
 
+void VP8BitReaderSetBuffer(VP8BitReader* const br,
+                           const uint8_t* const start,
+                           size_t size) {
+  br->buf_     = start;
+  br->buf_end_ = start + size;
+  br->buf_max_ =
+      (size >= sizeof(lbit_t)) ? start + size - sizeof(lbit_t) + 1
+                               : start;
+}
+
 void VP8InitBitReader(VP8BitReader* const br,
-                      const uint8_t* const start, const uint8_t* const end) {
+                      const uint8_t* const start, size_t size) {
   assert(br != NULL);
   assert(start != NULL);
-  assert(start <= end);
-  br->range_   = MK(255 - 1);
-  br->buf_     = start;
-  br->buf_end_ = end;
+  assert(size < (1u << 31));   // limit ensured by format and upstream checks
+  br->range_   = 255 - 1;
   br->value_   = 0;
-  br->missing_ = 8;   // to load the very first 8bits
+  br->bits_    = -8;   // to load the very first 8bits
   br->eof_     = 0;
+  VP8BitReaderSetBuffer(br, start, size);
+  VP8LoadNewBytes(br);
+}
+
+void VP8RemapBitReader(VP8BitReader* const br, ptrdiff_t offset) {
+  if (br->buf_ != NULL) {
+    br->buf_ += offset;
+    br->buf_end_ += offset;
+    br->buf_max_ += offset;
+  }
 }
 
 const uint8_t kVP8Log2Range[128] = {
@@ -45,36 +63,38 @@ const uint8_t kVP8Log2Range[128] = {
   0
 };
 
-// range = (range << kVP8Log2Range[range]) + trailing 1's
-const bit_t kVP8NewRange[128] = {
-  MK(127), MK(127), MK(191), MK(127), MK(159), MK(191), MK(223), MK(127),
-  MK(143), MK(159), MK(175), MK(191), MK(207), MK(223), MK(239), MK(127),
-  MK(135), MK(143), MK(151), MK(159), MK(167), MK(175), MK(183), MK(191),
-  MK(199), MK(207), MK(215), MK(223), MK(231), MK(239), MK(247), MK(127),
-  MK(131), MK(135), MK(139), MK(143), MK(147), MK(151), MK(155), MK(159),
-  MK(163), MK(167), MK(171), MK(175), MK(179), MK(183), MK(187), MK(191),
-  MK(195), MK(199), MK(203), MK(207), MK(211), MK(215), MK(219), MK(223),
-  MK(227), MK(231), MK(235), MK(239), MK(243), MK(247), MK(251), MK(127),
-  MK(129), MK(131), MK(133), MK(135), MK(137), MK(139), MK(141), MK(143),
-  MK(145), MK(147), MK(149), MK(151), MK(153), MK(155), MK(157), MK(159),
-  MK(161), MK(163), MK(165), MK(167), MK(169), MK(171), MK(173), MK(175),
-  MK(177), MK(179), MK(181), MK(183), MK(185), MK(187), MK(189), MK(191),
-  MK(193), MK(195), MK(197), MK(199), MK(201), MK(203), MK(205), MK(207),
-  MK(209), MK(211), MK(213), MK(215), MK(217), MK(219), MK(221), MK(223),
-  MK(225), MK(227), MK(229), MK(231), MK(233), MK(235), MK(237), MK(239),
-  MK(241), MK(243), MK(245), MK(247), MK(249), MK(251), MK(253), MK(127)
+// range = ((range - 1) << kVP8Log2Range[range]) + 1
+const uint8_t kVP8NewRange[128] = {
+  127, 127, 191, 127, 159, 191, 223, 127,
+  143, 159, 175, 191, 207, 223, 239, 127,
+  135, 143, 151, 159, 167, 175, 183, 191,
+  199, 207, 215, 223, 231, 239, 247, 127,
+  131, 135, 139, 143, 147, 151, 155, 159,
+  163, 167, 171, 175, 179, 183, 187, 191,
+  195, 199, 203, 207, 211, 215, 219, 223,
+  227, 231, 235, 239, 243, 247, 251, 127,
+  129, 131, 133, 135, 137, 139, 141, 143,
+  145, 147, 149, 151, 153, 155, 157, 159,
+  161, 163, 165, 167, 169, 171, 173, 175,
+  177, 179, 181, 183, 185, 187, 189, 191,
+  193, 195, 197, 199, 201, 203, 205, 207,
+  209, 211, 213, 215, 217, 219, 221, 223,
+  225, 227, 229, 231, 233, 235, 237, 239,
+  241, 243, 245, 247, 249, 251, 253, 127
 };
 
-#undef MK
-
 void VP8LoadFinalBytes(VP8BitReader* const br) {
   assert(br != NULL && br->buf_ != NULL);
   // Only read 8bits at a time
   if (br->buf_ < br->buf_end_) {
-    br->value_ |= (bit_t)(*br->buf_++) << ((BITS) - 8 + br->missing_);
-    br->missing_ -= 8;
-  } else {
+    br->bits_ += 8;
+    br->value_ = (bit_t)(*br->buf_++) | (br->value_ << 8);
+  } else if (!br->eof_) {
+    br->value_ <<= 8;
+    br->bits_ += 8;
     br->eof_ = 1;
+  } else {
+    br->bits_ = 0;  // This is to avoid undefined behaviour with shifts.
   }
 }
 
@@ -97,32 +117,47 @@ int32_t VP8GetSignedValue(VP8BitReader* const br, int bits) {
 //------------------------------------------------------------------------------
 // VP8LBitReader
 
-#define MAX_NUM_BIT_READ 25
+#define VP8L_LOG8_WBITS 4  // Number of bytes needed to store VP8L_WBITS bits.
 
-static const uint32_t kBitMask[MAX_NUM_BIT_READ] = {
-  0, 1, 3, 7, 15, 31, 63, 127, 255, 511, 1023, 2047, 4095, 8191, 16383, 32767,
-  65535, 131071, 262143, 524287, 1048575, 2097151, 4194303, 8388607, 16777215
+#if !defined(WEBP_FORCE_ALIGNED) && \
+    (defined(__arm__) || defined(_M_ARM) || defined(__aarch64__) || \
+     defined(__i386__) || defined(_M_IX86) || \
+     defined(__x86_64__) || defined(_M_X64))
+#define VP8L_USE_UNALIGNED_LOAD
+#endif
+
+static const uint32_t kBitMask[VP8L_MAX_NUM_BIT_READ + 1] = {
+  0,
+  0x000001, 0x000003, 0x000007, 0x00000f,
+  0x00001f, 0x00003f, 0x00007f, 0x0000ff,
+  0x0001ff, 0x0003ff, 0x0007ff, 0x000fff,
+  0x001fff, 0x003fff, 0x007fff, 0x00ffff,
+  0x01ffff, 0x03ffff, 0x07ffff, 0x0fffff,
+  0x1fffff, 0x3fffff, 0x7fffff, 0xffffff
 };
 
-void VP8LInitBitReader(VP8LBitReader* const br,
-                       const uint8_t* const start,
+void VP8LInitBitReader(VP8LBitReader* const br, const uint8_t* const start,
                        size_t length) {
   size_t i;
+  vp8l_val_t value = 0;
   assert(br != NULL);
   assert(start != NULL);
   assert(length < 0xfffffff8u);   // can't happen with a RIFF chunk.
 
-  br->buf_ = start;
   br->len_ = length;
   br->val_ = 0;
-  br->pos_ = 0;
   br->bit_pos_ = 0;
   br->eos_ = 0;
-  br->error_ = 0;
-  for (i = 0; i < sizeof(br->val_) && i < br->len_; ++i) {
-    br->val_ |= ((uint64_t)br->buf_[br->pos_]) << (8 * i);
-    ++br->pos_;
+
+  if (length > sizeof(br->val_)) {
+    length = sizeof(br->val_);
+  }
+  for (i = 0; i < length; ++i) {
+    value |= (vp8l_val_t)start[i] << (8 * i);
   }
+  br->val_ = value;
+  br->pos_ = length;
+  br->buf_ = start;
 }
 
 void VP8LBitReaderSetBuffer(VP8LBitReader* const br,
@@ -130,100 +165,62 @@ void VP8LBitReaderSetBuffer(VP8LBitReader* const br,
   assert(br != NULL);
   assert(buf != NULL);
   assert(len < 0xfffffff8u);   // can't happen with a RIFF chunk.
-  br->eos_ = (br->pos_ >= len);
   br->buf_ = buf;
   br->len_ = len;
+  // pos_ > len_ should be considered a param error.
+  br->eos_ = (br->pos_ > br->len_) || VP8LIsEndOfStream(br);
+}
+
+static void VP8LSetEndOfStream(VP8LBitReader* const br) {
+  br->eos_ = 1;
+  br->bit_pos_ = 0;  // To avoid undefined behaviour with shifts.
 }
 
+// If not at EOS, reload up to VP8L_LBITS byte-by-byte
 static void ShiftBytes(VP8LBitReader* const br) {
   while (br->bit_pos_ >= 8 && br->pos_ < br->len_) {
     br->val_ >>= 8;
-    br->val_ |= ((uint64_t)br->buf_[br->pos_]) << 56;
+    br->val_ |= ((vp8l_val_t)br->buf_[br->pos_]) << (VP8L_LBITS - 8);
     ++br->pos_;
     br->bit_pos_ -= 8;
   }
-}
-
-void VP8LFillBitWindow(VP8LBitReader* const br) {
-  if (br->bit_pos_ >= 32) {
-#if defined(__x86_64__) || defined(_M_X64)
-    if (br->pos_ + 8 < br->len_) {
-      br->val_ >>= 32;
-      // The expression below needs a little-endian arch to work correctly.
-      // This gives a large speedup for decoding speed.
-      br->val_ |= *(const uint64_t *)(br->buf_ + br->pos_) << 32;
-      br->pos_ += 4;
-      br->bit_pos_ -= 32;
-    } else {
-      // Slow path.
-      ShiftBytes(br);
-    }
-#else
-    // Always the slow path.
-    ShiftBytes(br);
-#endif
-  }
-  if (br->pos_ == br->len_ && br->bit_pos_ == 64) {
-    br->eos_ = 1;
+  if (VP8LIsEndOfStream(br)) {
+    VP8LSetEndOfStream(br);
   }
 }
 
-uint32_t VP8LReadOneBit(VP8LBitReader* const br) {
-  const uint32_t val = (br->val_ >> br->bit_pos_) & 1;
-  // Flag an error at end_of_stream.
-  if (!br->eos_) {
-    ++br->bit_pos_;
-    if (br->bit_pos_ >= 32) {
-      ShiftBytes(br);
-    }
-    // After this last bit is read, check if eos needs to be flagged.
-    if (br->pos_ == br->len_ && br->bit_pos_ == 64) {
-      br->eos_ = 1;
-    }
-  } else {
-    br->error_ = 1;
+void VP8LDoFillBitWindow(VP8LBitReader* const br) {
+  assert(br->bit_pos_ >= VP8L_WBITS);
+  // TODO(jzern): given the fixed read size it may be possible to force
+  //              alignment in this block.
+#if defined(VP8L_USE_UNALIGNED_LOAD)
+  if (br->pos_ + sizeof(br->val_) < br->len_) {
+    br->val_ >>= VP8L_WBITS;
+    br->bit_pos_ -= VP8L_WBITS;
+    // The expression below needs a little-endian arch to work correctly.
+    // This gives a large speedup for decoding speed.
+    br->val_ |= (vp8l_val_t)*(const uint32_t*)(br->buf_ + br->pos_) <<
+                (VP8L_LBITS - VP8L_WBITS);
+    br->pos_ += VP8L_LOG8_WBITS;
+    return;
   }
-  return val;
+#endif
+  ShiftBytes(br);       // Slow path.
 }
 
 uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits) {
-  uint32_t val = 0;
   assert(n_bits >= 0);
   // Flag an error if end_of_stream or n_bits is more than allowed limit.
-  if (!br->eos_ && n_bits < MAX_NUM_BIT_READ) {
-    // If this read is going to cross the read buffer, set the eos flag.
-    if (br->pos_ == br->len_) {
-      if ((br->bit_pos_ + n_bits) >= 64) {
-        br->eos_ = 1;
-        if ((br->bit_pos_ + n_bits) > 64) return val;
-      }
-    }
-    val = (br->val_ >> br->bit_pos_) & kBitMask[n_bits];
-    br->bit_pos_ += n_bits;
-    if (br->bit_pos_ >= 40) {
-      if (br->pos_ + 5 < br->len_) {
-        br->val_ >>= 40;
-        br->val_ |=
-            (((uint64_t)br->buf_[br->pos_ + 0]) << 24) |
-            (((uint64_t)br->buf_[br->pos_ + 1]) << 32) |
-            (((uint64_t)br->buf_[br->pos_ + 2]) << 40) |
-            (((uint64_t)br->buf_[br->pos_ + 3]) << 48) |
-            (((uint64_t)br->buf_[br->pos_ + 4]) << 56);
-        br->pos_ += 5;
-        br->bit_pos_ -= 40;
-      }
-      if (br->bit_pos_ >= 8) {
-        ShiftBytes(br);
-      }
-    }
+  if (!br->eos_ && n_bits <= VP8L_MAX_NUM_BIT_READ) {
+    const uint32_t val = VP8LPrefetchBits(br) & kBitMask[n_bits];
+    const int new_bits = br->bit_pos_ + n_bits;
+    br->bit_pos_ = new_bits;
+    ShiftBytes(br);
+    return val;
   } else {
-    br->error_ = 1;
+    VP8LSetEndOfStream(br);
+    return 0;
   }
-  return val;
 }
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/utils/bit_reader.h b/drivers/webp/utils/bit_reader.h
index 43cd948fd4..0fc62d33b7 100644
--- a/drivers/webp/utils/bit_reader.h
+++ b/drivers/webp/utils/bit_reader.h
@@ -1,9 +1,10 @@
-//
 // Copyright 2010 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Boolean decoder
@@ -18,44 +19,75 @@
 #ifdef _MSC_VER
 #include <stdlib.h>  // _byteswap_ulong
 #endif
-#include <string.h>  // For memcpy
-#include "../types.h"
+#include "../webp/types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-#define BITS 32     // can be 32, 16 or 8
-#define MASK ((((bit_t)1) << (BITS)) - 1)
-#if (BITS == 32)
-typedef uint64_t bit_t;   // natural register type
-typedef uint32_t lbit_t;  // natural type for memory I/O
-#elif (BITS == 16)
-typedef uint32_t bit_t;
-typedef uint16_t lbit_t;
+// The Boolean decoder needs to maintain infinite precision on the value_ field.
+// However, since range_ is only 8bit, we only need an active window of 8 bits
+// for value_. Left bits (MSB) gets zeroed and shifted away when value_ falls
+// below 128, range_ is updated, and fresh bits read from the bitstream are
+// brought in as LSB. To avoid reading the fresh bits one by one (slow), we
+// cache BITS of them ahead. The total of (BITS + 8) bits must fit into a
+// natural register (with type bit_t). To fetch BITS bits from bitstream we
+// use a type lbit_t.
+//
+// BITS can be any multiple of 8 from 8 to 56 (inclusive).
+// Pick values that fit natural register size.
+
+#if defined(__i386__) || defined(_M_IX86)      // x86 32bit
+#define BITS 24
+#elif defined(__x86_64__) || defined(_M_X64)   // x86 64bit
+#define BITS 56
+#elif defined(__arm__) || defined(_M_ARM)      // ARM
+#define BITS 24
+#elif defined(__mips__)                        // MIPS
+#define BITS 24
+#else                                          // reasonable default
+#define BITS 24  // TODO(skal): test aarch64 and find the proper BITS value.
+#endif
+
+//------------------------------------------------------------------------------
+// Derived types and constants:
+//   bit_t = natural register type for storing 'value_' (which is BITS+8 bits)
+//   range_t = register for 'range_' (which is 8bits only)
+
+#if (BITS > 24)
+typedef uint64_t bit_t;
 #else
 typedef uint32_t bit_t;
-typedef uint8_t lbit_t;
 #endif
 
+typedef uint32_t range_t;
+
 //------------------------------------------------------------------------------
-// Bitreader and code-tree reader
+// Bitreader
 
 typedef struct VP8BitReader VP8BitReader;
 struct VP8BitReader {
+  // boolean decoder  (keep the field ordering as is!)
+  bit_t value_;               // current value
+  range_t range_;             // current range minus 1. In [127, 254] interval.
+  int bits_;                  // number of valid bits left
+  // read buffer
   const uint8_t* buf_;        // next byte to be read
   const uint8_t* buf_end_;    // end of read buffer
+  const uint8_t* buf_max_;    // max packed-read position on buffer
   int eof_;                   // true if input is exhausted
-
-  // boolean decoder
-  bit_t range_;            // current range minus 1. In [127, 254] interval.
-  bit_t value_;            // current value
-  int missing_;            // number of missing bits in value_ (8bit)
 };
 
 // Initialize the bit reader and the boolean decoder.
 void VP8InitBitReader(VP8BitReader* const br,
-                      const uint8_t* const start, const uint8_t* const end);
+                      const uint8_t* const start, size_t size);
+// Sets the working read buffer.
+void VP8BitReaderSetBuffer(VP8BitReader* const br,
+                           const uint8_t* const start, size_t size);
+
+// Update internal pointers to displace the byte buffer by the
+// relative offset 'offset'.
+void VP8RemapBitReader(VP8BitReader* const br, ptrdiff_t offset);
 
 // return the next value made of 'num_bits' bits
 uint32_t VP8GetValue(VP8BitReader* const br, int num_bits);
@@ -66,100 +98,31 @@ static WEBP_INLINE uint32_t VP8Get(VP8BitReader* const br) {
 // return the next value with sign-extension.
 int32_t VP8GetSignedValue(VP8BitReader* const br, int num_bits);
 
-// Read a bit with proba 'prob'. Speed-critical function!
-extern const uint8_t kVP8Log2Range[128];
-extern const bit_t kVP8NewRange[128];
-
-void VP8LoadFinalBytes(VP8BitReader* const br);    // special case for the tail
-
-static WEBP_INLINE void VP8LoadNewBytes(VP8BitReader* const br) {
-  assert(br && br->buf_);
-  // Read 'BITS' bits at a time if possible.
-  if (br->buf_ + sizeof(lbit_t) <= br->buf_end_) {
-    // convert memory type to register type (with some zero'ing!)
-    bit_t bits;
-    lbit_t in_bits = *(lbit_t*)br->buf_;
-    br->buf_ += (BITS) >> 3;
-#if !defined(__BIG_ENDIAN__)
-#if (BITS == 32)
-#if defined(__i386__) || defined(__x86_64__)
-    __asm__ volatile("bswap %k0" : "=r"(in_bits) : "0"(in_bits));
-    bits = (bit_t)in_bits;   // 32b -> 64b zero-extension
-#elif defined(_MSC_VER)
-    bits = _byteswap_ulong(in_bits);
-#else
-    bits = (bit_t)(in_bits >> 24) | ((in_bits >> 8) & 0xff00)
-         | ((in_bits << 8) & 0xff0000)  | (in_bits << 24);
-#endif  // x86
-#elif (BITS == 16)
-    // gcc will recognize a 'rorw $8, ...' here:
-    bits = (bit_t)(in_bits >> 8) | ((in_bits & 0xff) << 8);
-#endif
-#else    // LITTLE_ENDIAN
-    bits = (bit_t)in_bits;
-#endif
-    br->value_ |= bits << br->missing_;
-    br->missing_ -= (BITS);
-  } else {
-    VP8LoadFinalBytes(br);    // no need to be inlined
-  }
-}
+// bit_reader_inl.h will implement the following methods:
+//   static WEBP_INLINE int VP8GetBit(VP8BitReader* const br, int prob)
+//   static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v)
+// and should be included by the .c files that actually need them.
+// This is to avoid recompiling the whole library whenever this file is touched,
+// and also allowing platform-specific ad-hoc hacks.
 
-static WEBP_INLINE int VP8BitUpdate(VP8BitReader* const br, bit_t split) {
-  const bit_t value_split = split | (MASK);
-  if (br->missing_ > 0) {  // Make sure we have a least BITS bits in 'value_'
-    VP8LoadNewBytes(br);
-  }
-  if (br->value_ > value_split) {
-    br->range_ -= value_split + 1;
-    br->value_ -= value_split + 1;
-    return 1;
-  } else {
-    br->range_ = value_split;
-    return 0;
-  }
-}
-
-static WEBP_INLINE void VP8Shift(VP8BitReader* const br) {
-  // range_ is in [0..127] interval here.
-  const int idx = br->range_ >> (BITS);
-  const int shift = kVP8Log2Range[idx];
-  br->range_ = kVP8NewRange[idx];
-  br->value_ <<= shift;
-  br->missing_ += shift;
-}
-
-static WEBP_INLINE int VP8GetBit(VP8BitReader* const br, int prob) {
-  // It's important to avoid generating a 64bit x 64bit multiply here.
-  // We just need an 8b x 8b after all.
-  const bit_t split =
-      (bit_t)((uint32_t)(br->range_ >> (BITS)) * prob) << ((BITS) - 8);
-  const int bit = VP8BitUpdate(br, split);
-  if (br->range_ <= (((bit_t)0x7e << (BITS)) | (MASK))) {
-    VP8Shift(br);
-  }
-  return bit;
-}
+// -----------------------------------------------------------------------------
+// Bitreader for lossless format
 
-static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v) {
-  const bit_t split = (br->range_ >> 1);
-  const int bit = VP8BitUpdate(br, split);
-  VP8Shift(br);
-  return bit ? -v : v;
-}
+// maximum number of bits (inclusive) the bit-reader can handle:
+#define VP8L_MAX_NUM_BIT_READ 24
 
+#define VP8L_LBITS 64  // Number of bits prefetched (= bit-size of vp8l_val_t).
+#define VP8L_WBITS 32  // Minimum number of bytes ready after VP8LFillBitWindow.
 
-// -----------------------------------------------------------------------------
-// Bitreader
+typedef uint64_t vp8l_val_t;  // right now, this bit-reader can only use 64bit.
 
 typedef struct {
-  uint64_t       val_;
-  const uint8_t* buf_;
-  size_t         len_;
-  size_t         pos_;
-  int            bit_pos_;
-  int            eos_;
-  int            error_;
+  vp8l_val_t     val_;        // pre-fetched bits
+  const uint8_t* buf_;        // input byte buffer
+  size_t         len_;        // buffer length
+  size_t         pos_;        // byte position in buf_
+  int            bit_pos_;    // current bit-reading position in val_
+  int            eos_;        // true if a bit was read past the end of buffer
 } VP8LBitReader;
 
 void VP8LInitBitReader(VP8LBitReader* const br,
@@ -170,28 +133,39 @@ void VP8LInitBitReader(VP8LBitReader* const br,
 void VP8LBitReaderSetBuffer(VP8LBitReader* const br,
                             const uint8_t* const buffer, size_t length);
 
-// Reads the specified number of bits from Read Buffer.
-// Flags an error in case end_of_stream or n_bits is more than allowed limit.
-// Flags eos if this read attempt is going to cross the read buffer.
+// Reads the specified number of bits from read buffer.
+// Flags an error in case end_of_stream or n_bits is more than the allowed limit
+// of VP8L_MAX_NUM_BIT_READ (inclusive).
+// Flags eos_ if this read attempt is going to cross the read buffer.
 uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits);
 
-// Reads one bit from Read Buffer. Flags an error in case end_of_stream.
-// Flags eos after reading last bit from the buffer.
-uint32_t VP8LReadOneBit(VP8LBitReader* const br);
-
-// VP8LReadOneBitUnsafe is faster than VP8LReadOneBit, but it can be called only
-// 32 times after the last VP8LFillBitWindow. Any subsequent calls
-// (without VP8LFillBitWindow) will return invalid data.
-static WEBP_INLINE uint32_t VP8LReadOneBitUnsafe(VP8LBitReader* const br) {
-  const uint32_t val = (br->val_ >> br->bit_pos_) & 1;
-  ++br->bit_pos_;
-  return val;
+// Return the prefetched bits, so they can be looked up.
+static WEBP_INLINE uint32_t VP8LPrefetchBits(VP8LBitReader* const br) {
+  return (uint32_t)(br->val_ >> (br->bit_pos_ & (VP8L_LBITS - 1)));
+}
+
+// Returns true if there was an attempt at reading bit past the end of
+// the buffer. Doesn't set br->eos_ flag.
+static WEBP_INLINE int VP8LIsEndOfStream(const VP8LBitReader* const br) {
+  assert(br->pos_ <= br->len_);
+  return br->eos_ || ((br->pos_ == br->len_) && (br->bit_pos_ > VP8L_LBITS));
 }
 
-// Advances the Read buffer by 4 bytes to make room for reading next 32 bits.
-void VP8LFillBitWindow(VP8LBitReader* const br);
+// For jumping over a number of bits in the bit stream when accessed with
+// VP8LPrefetchBits and VP8LFillBitWindow.
+static WEBP_INLINE void VP8LSetBitPos(VP8LBitReader* const br, int val) {
+  br->bit_pos_ = val;
+  br->eos_ = VP8LIsEndOfStream(br);
+}
+
+// Advances the read buffer by 4 bytes to make room for reading next 32 bits.
+// Speed critical, but infrequent part of the code can be non-inlined.
+extern void VP8LDoFillBitWindow(VP8LBitReader* const br);
+static WEBP_INLINE void VP8LFillBitWindow(VP8LBitReader* const br) {
+  if (br->bit_pos_ >= VP8L_WBITS) VP8LDoFillBitWindow(br);
+}
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/utils/bit_writer.c b/drivers/webp/utils/bit_writer.c
index 671159cacd..064428691b 100644
--- a/drivers/webp/utils/bit_writer.c
+++ b/drivers/webp/utils/bit_writer.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Bit writing and boolean coder
@@ -13,11 +15,10 @@
 #include <assert.h>
 #include <string.h>   // for memcpy()
 #include <stdlib.h>
-#include "./bit_writer.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+#include "./bit_writer.h"
+#include "./endian_inl.h"
+#include "./utils.h"
 
 //------------------------------------------------------------------------------
 // VP8BitWriter
@@ -36,19 +37,22 @@ static int BitWriterResize(VP8BitWriter* const bw, size_t extra_size) {
   new_size = 2 * bw->max_pos_;
   if (new_size < needed_size) new_size = needed_size;
   if (new_size < 1024) new_size = 1024;
-  new_buf = (uint8_t*)malloc(new_size);
+  new_buf = (uint8_t*)WebPSafeMalloc(1ULL, new_size);
   if (new_buf == NULL) {
     bw->error_ = 1;
     return 0;
   }
-  memcpy(new_buf, bw->buf_, bw->pos_);
-  free(bw->buf_);
+  if (bw->pos_ > 0) {
+    assert(bw->buf_ != NULL);
+    memcpy(new_buf, bw->buf_, bw->pos_);
+  }
+  WebPSafeFree(bw->buf_);
   bw->buf_ = new_buf;
   bw->max_pos_ = new_size;
   return 1;
 }
 
-static void kFlush(VP8BitWriter* const bw) {
+static void Flush(VP8BitWriter* const bw) {
   const int s = 8 + bw->nb_bits_;
   const int32_t bits = bw->value_ >> s;
   assert(bw->nb_bits_ >= 0);
@@ -114,7 +118,7 @@ int VP8PutBit(VP8BitWriter* const bw, int bit, int prob) {
     bw->range_ = kNewRange[bw->range_];
     bw->value_ <<= shift;
     bw->nb_bits_ += shift;
-    if (bw->nb_bits_ > 0) kFlush(bw);
+    if (bw->nb_bits_ > 0) Flush(bw);
   }
   return bit;
 }
@@ -131,24 +135,25 @@ int VP8PutBitUniform(VP8BitWriter* const bw, int bit) {
     bw->range_ = kNewRange[bw->range_];
     bw->value_ <<= 1;
     bw->nb_bits_ += 1;
-    if (bw->nb_bits_ > 0) kFlush(bw);
+    if (bw->nb_bits_ > 0) Flush(bw);
   }
   return bit;
 }
 
-void VP8PutValue(VP8BitWriter* const bw, int value, int nb_bits) {
-  int mask;
-  for (mask = 1 << (nb_bits - 1); mask; mask >>= 1)
+void VP8PutBits(VP8BitWriter* const bw, uint32_t value, int nb_bits) {
+  uint32_t mask;
+  assert(nb_bits > 0 && nb_bits < 32);
+  for (mask = 1u << (nb_bits - 1); mask; mask >>= 1)
     VP8PutBitUniform(bw, value & mask);
 }
 
-void VP8PutSignedValue(VP8BitWriter* const bw, int value, int nb_bits) {
+void VP8PutSignedBits(VP8BitWriter* const bw, int value, int nb_bits) {
   if (!VP8PutBitUniform(bw, value != 0))
     return;
   if (value < 0) {
-    VP8PutValue(bw, ((-value) << 1) | 1, nb_bits + 1);
+    VP8PutBits(bw, ((-value) << 1) | 1, nb_bits + 1);
   } else {
-    VP8PutValue(bw, value << 1, nb_bits + 1);
+    VP8PutBits(bw, value << 1, nb_bits + 1);
   }
 }
 
@@ -167,16 +172,16 @@ int VP8BitWriterInit(VP8BitWriter* const bw, size_t expected_size) {
 }
 
 uint8_t* VP8BitWriterFinish(VP8BitWriter* const bw) {
-  VP8PutValue(bw, 0, 9 - bw->nb_bits_);
+  VP8PutBits(bw, 0, 9 - bw->nb_bits_);
   bw->nb_bits_ = 0;   // pad with zeroes
-  kFlush(bw);
+  Flush(bw);
   return bw->buf_;
 }
 
 int VP8BitWriterAppend(VP8BitWriter* const bw,
                        const uint8_t* data, size_t size) {
-  assert(data);
-  if (bw->nb_bits_ != -8) return 0;   // kFlush() must have been called
+  assert(data != NULL);
+  if (bw->nb_bits_ != -8) return 0;   // Flush() must have been called
   if (!BitWriterResize(bw, size)) return 0;
   memcpy(bw->buf_ + bw->pos_, data, size);
   bw->pos_ += size;
@@ -184,8 +189,8 @@ int VP8BitWriterAppend(VP8BitWriter* const bw,
 }
 
 void VP8BitWriterWipeOut(VP8BitWriter* const bw) {
-  if (bw) {
-    free(bw->buf_);
+  if (bw != NULL) {
+    WebPSafeFree(bw->buf_);
     memset(bw, 0, sizeof(*bw));
   }
 }
@@ -193,32 +198,39 @@ void VP8BitWriterWipeOut(VP8BitWriter* const bw) {
 //------------------------------------------------------------------------------
 // VP8LBitWriter
 
+// This is the minimum amount of size the memory buffer is guaranteed to grow
+// when extra space is needed.
+#define MIN_EXTRA_SIZE  (32768ULL)
+
 // Returns 1 on success.
 static int VP8LBitWriterResize(VP8LBitWriter* const bw, size_t extra_size) {
   uint8_t* allocated_buf;
   size_t allocated_size;
-  const size_t current_size = VP8LBitWriterNumBytes(bw);
+  const size_t max_bytes = bw->end_ - bw->buf_;
+  const size_t current_size = bw->cur_ - bw->buf_;
   const uint64_t size_required_64b = (uint64_t)current_size + extra_size;
   const size_t size_required = (size_t)size_required_64b;
   if (size_required != size_required_64b) {
     bw->error_ = 1;
     return 0;
   }
-  if (bw->max_bytes_ > 0 && size_required <= bw->max_bytes_) return 1;
-  allocated_size = (3 * bw->max_bytes_) >> 1;
+  if (max_bytes > 0 && size_required <= max_bytes) return 1;
+  allocated_size = (3 * max_bytes) >> 1;
   if (allocated_size < size_required) allocated_size = size_required;
   // make allocated size multiple of 1k
   allocated_size = (((allocated_size >> 10) + 1) << 10);
-  allocated_buf = (uint8_t*)malloc(allocated_size);
+  allocated_buf = (uint8_t*)WebPSafeMalloc(1ULL, allocated_size);
   if (allocated_buf == NULL) {
     bw->error_ = 1;
     return 0;
   }
-  memcpy(allocated_buf, bw->buf_, current_size);
-  free(bw->buf_);
+  if (current_size > 0) {
+    memcpy(allocated_buf, bw->buf_, current_size);
+  }
+  WebPSafeFree(bw->buf_);
   bw->buf_ = allocated_buf;
-  bw->max_bytes_ = allocated_size;
-  memset(allocated_buf + current_size, 0, allocated_size - current_size);
+  bw->cur_ = bw->buf_ + current_size;
+  bw->end_ = bw->buf_ + allocated_size;
   return 1;
 }
 
@@ -227,58 +239,81 @@ int VP8LBitWriterInit(VP8LBitWriter* const bw, size_t expected_size) {
   return VP8LBitWriterResize(bw, expected_size);
 }
 
-void VP8LBitWriterDestroy(VP8LBitWriter* const bw) {
+void VP8LBitWriterWipeOut(VP8LBitWriter* const bw) {
   if (bw != NULL) {
-    free(bw->buf_);
+    WebPSafeFree(bw->buf_);
     memset(bw, 0, sizeof(*bw));
   }
 }
 
-void VP8LWriteBits(VP8LBitWriter* const bw, int n_bits, uint32_t bits) {
-  if (n_bits < 1) return;
-#if !defined(__BIG_ENDIAN__)
-  // Technically, this branch of the code can write up to 25 bits at a time,
-  // but in prefix encoding, the maximum number of bits written is 18 at a time.
-  {
-    uint8_t* const p = &bw->buf_[bw->bit_pos_ >> 3];
-    uint32_t v = *(const uint32_t*)p;
-    v |= bits << (bw->bit_pos_ & 7);
-    *(uint32_t*)p = v;
-    bw->bit_pos_ += n_bits;
+void VP8LPutBitsFlushBits(VP8LBitWriter* const bw) {
+  // If needed, make some room by flushing some bits out.
+  if (bw->cur_ + VP8L_WRITER_BYTES > bw->end_) {
+    const uint64_t extra_size = (bw->end_ - bw->buf_) + MIN_EXTRA_SIZE;
+    if (extra_size != (size_t)extra_size ||
+        !VP8LBitWriterResize(bw, (size_t)extra_size)) {
+      bw->cur_ = bw->buf_;
+      bw->error_ = 1;
+      return;
+    }
   }
-#else  // BIG_ENDIAN
-  {
-    uint8_t* p = &bw->buf_[bw->bit_pos_ >> 3];
-    const int bits_reserved_in_first_byte = bw->bit_pos_ & 7;
-    const int bits_left_to_write = n_bits - 8 + bits_reserved_in_first_byte;
-    // implicit & 0xff is assumed for uint8_t arithmetics
-    *p++ |= bits << bits_reserved_in_first_byte;
-    bits >>= 8 - bits_reserved_in_first_byte;
-    if (bits_left_to_write >= 1) {
-      *p++ = bits;
-      bits >>= 8;
-      if (bits_left_to_write >= 9) {
-        *p++ = bits;
-        bits >>= 8;
+  *(vp8l_wtype_t*)bw->cur_ = (vp8l_wtype_t)WSWAP((vp8l_wtype_t)bw->bits_);
+  bw->cur_ += VP8L_WRITER_BYTES;
+  bw->bits_ >>= VP8L_WRITER_BITS;
+  bw->used_ -= VP8L_WRITER_BITS;
+}
+
+void VP8LPutBitsInternal(VP8LBitWriter* const bw, uint32_t bits, int n_bits) {
+  assert(n_bits <= 32);
+  // That's the max we can handle:
+  assert(sizeof(vp8l_wtype_t) == 2);
+  if (n_bits > 0) {
+    vp8l_atype_t lbits = bw->bits_;
+    int used = bw->used_;
+    // Special case of overflow handling for 32bit accumulator (2-steps flush).
+#if VP8L_WRITER_BITS == 16
+    if (used + n_bits >= VP8L_WRITER_MAX_BITS) {
+      // Fill up all the VP8L_WRITER_MAX_BITS so it can be flushed out below.
+      const int shift = VP8L_WRITER_MAX_BITS - used;
+      lbits |= (vp8l_atype_t)bits << used;
+      used = VP8L_WRITER_MAX_BITS;
+      n_bits -= shift;
+      bits >>= shift;
+      assert(n_bits <= VP8L_WRITER_MAX_BITS);
+    }
+#endif
+    // If needed, make some room by flushing some bits out.
+    while (used >= VP8L_WRITER_BITS) {
+      if (bw->cur_ + VP8L_WRITER_BYTES > bw->end_) {
+        const uint64_t extra_size = (bw->end_ - bw->buf_) + MIN_EXTRA_SIZE;
+        if (extra_size != (size_t)extra_size ||
+            !VP8LBitWriterResize(bw, (size_t)extra_size)) {
+          bw->cur_ = bw->buf_;
+          bw->error_ = 1;
+          return;
+        }
       }
+      *(vp8l_wtype_t*)bw->cur_ = (vp8l_wtype_t)WSWAP((vp8l_wtype_t)lbits);
+      bw->cur_ += VP8L_WRITER_BYTES;
+      lbits >>= VP8L_WRITER_BITS;
+      used -= VP8L_WRITER_BITS;
     }
-    assert(n_bits <= 25);
-    *p = bits;
-    bw->bit_pos_ += n_bits;
+    bw->bits_ = lbits | ((vp8l_atype_t)bits << used);
+    bw->used_ = used + n_bits;
   }
-#endif
-  if ((bw->bit_pos_ >> 3) > (bw->max_bytes_ - 8)) {
-    const uint64_t extra_size = 32768ULL + bw->max_bytes_;
-    if (extra_size != (size_t)extra_size ||
-        !VP8LBitWriterResize(bw, (size_t)extra_size)) {
-      bw->bit_pos_ = 0;
-      bw->error_ = 1;
+}
+
+uint8_t* VP8LBitWriterFinish(VP8LBitWriter* const bw) {
+  // flush leftover bits
+  if (VP8LBitWriterResize(bw, (bw->used_ + 7) >> 3)) {
+    while (bw->used_ > 0) {
+      *bw->cur_++ = (uint8_t)bw->bits_;
+      bw->bits_ >>= 8;
+      bw->used_ -= 8;
     }
+    bw->used_ = 0;
   }
+  return bw->buf_;
 }
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/utils/bit_writer.h b/drivers/webp/utils/bit_writer.h
index 57f39b11b1..ef360d1dc6 100644
--- a/drivers/webp/utils/bit_writer.h
+++ b/drivers/webp/utils/bit_writer.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Bit writing and boolean coder
@@ -12,9 +14,9 @@
 #ifndef WEBP_UTILS_BIT_WRITER_H_
 #define WEBP_UTILS_BIT_WRITER_H_
 
-#include "../types.h"
+#include "../webp/types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -43,8 +45,8 @@ void VP8BitWriterWipeOut(VP8BitWriter* const bw);
 
 int VP8PutBit(VP8BitWriter* const bw, int bit, int prob);
 int VP8PutBitUniform(VP8BitWriter* const bw, int bit);
-void VP8PutValue(VP8BitWriter* const bw, int value, int nb_bits);
-void VP8PutSignedValue(VP8BitWriter* const bw, int value, int nb_bits);
+void VP8PutBits(VP8BitWriter* const bw, uint32_t value, int nb_bits);
+void VP8PutSignedBits(VP8BitWriter* const bw, int value, int nb_bits);
 
 // Appends some bytes to the internal buffer. Data is copied.
 int VP8BitWriterAppend(VP8BitWriter* const bw,
@@ -66,57 +68,77 @@ static WEBP_INLINE size_t VP8BitWriterSize(const VP8BitWriter* const bw) {
 
 //------------------------------------------------------------------------------
 // VP8LBitWriter
-// TODO(vikasa): VP8LBitWriter is copied as-is from lossless code. There's scope
-// of re-using VP8BitWriter. Will evaluate once basic lossless encoder is
-// implemented.
 
-typedef struct {
-  uint8_t* buf_;
-  size_t bit_pos_;
-  size_t max_bytes_;
+#if defined(__x86_64__) || defined(_M_X64)   // 64bit
+typedef uint64_t vp8l_atype_t;   // accumulator type
+typedef uint32_t vp8l_wtype_t;   // writing type
+#define WSWAP HToLE32
+#define VP8L_WRITER_BYTES    4   // sizeof(vp8l_wtype_t)
+#define VP8L_WRITER_BITS     32  // 8 * sizeof(vp8l_wtype_t)
+#define VP8L_WRITER_MAX_BITS 64  // 8 * sizeof(vp8l_atype_t)
+#else
+typedef uint32_t vp8l_atype_t;
+typedef uint16_t vp8l_wtype_t;
+#define WSWAP HToLE16
+#define VP8L_WRITER_BYTES    2
+#define VP8L_WRITER_BITS     16
+#define VP8L_WRITER_MAX_BITS 32
+#endif
 
-  // After all bits are written, the caller must observe the state of
-  // error_. A value of 1 indicates that a memory allocation failure
-  // has happened during bit writing. A value of 0 indicates successful
+typedef struct {
+  vp8l_atype_t bits_;   // bit accumulator
+  int          used_;   // number of bits used in accumulator
+  uint8_t*     buf_;    // start of buffer
+  uint8_t*     cur_;    // current write position
+  uint8_t*     end_;    // end of buffer
+
+  // After all bits are written (VP8LBitWriterFinish()), the caller must observe
+  // the state of error_. A value of 1 indicates that a memory allocation
+  // failure has happened during bit writing. A value of 0 indicates successful
   // writing of bits.
   int error_;
 } VP8LBitWriter;
 
 static WEBP_INLINE size_t VP8LBitWriterNumBytes(VP8LBitWriter* const bw) {
-  return (bw->bit_pos_ + 7) >> 3;
-}
-
-static WEBP_INLINE uint8_t* VP8LBitWriterFinish(VP8LBitWriter* const bw) {
-  return bw->buf_;
+  return (bw->cur_ - bw->buf_) + ((bw->used_ + 7) >> 3);
 }
 
-// Returns 0 in case of memory allocation error.
+// Returns false in case of memory allocation error.
 int VP8LBitWriterInit(VP8LBitWriter* const bw, size_t expected_size);
+// Finalize the bitstream coding. Returns a pointer to the internal buffer.
+uint8_t* VP8LBitWriterFinish(VP8LBitWriter* const bw);
+// Release any pending memory and zeroes the object.
+void VP8LBitWriterWipeOut(VP8LBitWriter* const bw);
 
-void VP8LBitWriterDestroy(VP8LBitWriter* const bw);
+// Internal function for VP8LPutBits flushing 32 bits from the written state.
+void VP8LPutBitsFlushBits(VP8LBitWriter* const bw);
 
-// This function writes bits into bytes in increasing addresses, and within
-// a byte least-significant-bit first.
-//
-// The function can write up to 16 bits in one go with WriteBits
-// Example: let's assume that 3 bits (Rs below) have been written already:
-//
-// BYTE-0     BYTE+1       BYTE+2
-//
-// 0000 0RRR    0000 0000    0000 0000
-//
-// Now, we could write 5 or less bits in MSB by just sifting by 3
-// and OR'ing to BYTE-0.
-//
-// For n bits, we take the last 5 bytes, OR that with high bits in BYTE-0,
-// and locate the rest in BYTE+1 and BYTE+2.
-//
+// PutBits internal function used in the 16 bit vp8l_wtype_t case.
+void VP8LPutBitsInternal(VP8LBitWriter* const bw, uint32_t bits, int n_bits);
+
+// This function writes bits into bytes in increasing addresses (little endian),
+// and within a byte least-significant-bit first.
+// This function can write up to 32 bits in one go, but VP8LBitReader can only
+// read 24 bits max (VP8L_MAX_NUM_BIT_READ).
 // VP8LBitWriter's error_ flag is set in case of  memory allocation error.
-void VP8LWriteBits(VP8LBitWriter* const bw, int n_bits, uint32_t bits);
+static WEBP_INLINE void VP8LPutBits(VP8LBitWriter* const bw,
+                                    uint32_t bits, int n_bits) {
+  if (sizeof(vp8l_wtype_t) == 4) {
+    if (n_bits > 0) {
+      if (bw->used_ >= 32) {
+        VP8LPutBitsFlushBits(bw);
+      }
+      bw->bits_ |= (vp8l_atype_t)bits << bw->used_;
+      bw->used_ += n_bits;
+    }
+  } else {
+    VP8LPutBitsInternal(bw, bits, n_bits);
+  }
+}
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/utils/color_cache.c b/drivers/webp/utils/color_cache.c
index 560f81db10..f9ff4b5451 100644
--- a/drivers/webp/utils/color_cache.c
+++ b/drivers/webp/utils/color_cache.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Color Cache for WebP Lossless
@@ -11,13 +13,10 @@
 
 #include <assert.h>
 #include <stdlib.h>
+#include <string.h>
 #include "./color_cache.h"
 #include "../utils/utils.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
 // VP8LColorCache.
 
@@ -29,16 +28,22 @@ int VP8LColorCacheInit(VP8LColorCache* const cc, int hash_bits) {
                                           sizeof(*cc->colors_));
   if (cc->colors_ == NULL) return 0;
   cc->hash_shift_ = 32 - hash_bits;
+  cc->hash_bits_ = hash_bits;
   return 1;
 }
 
 void VP8LColorCacheClear(VP8LColorCache* const cc) {
   if (cc != NULL) {
-    free(cc->colors_);
+    WebPSafeFree(cc->colors_);
     cc->colors_ = NULL;
   }
 }
 
-#if defined(__cplusplus) || defined(c_plusplus)
+void VP8LColorCacheCopy(const VP8LColorCache* const src,
+                        VP8LColorCache* const dst) {
+  assert(src != NULL);
+  assert(dst != NULL);
+  assert(src->hash_bits_ == dst->hash_bits_);
+  memcpy(dst->colors_, src->colors_,
+         ((size_t)1u << dst->hash_bits_) * sizeof(*dst->colors_));
 }
-#endif
diff --git a/drivers/webp/utils/color_cache.h b/drivers/webp/utils/color_cache.h
index da5e260195..a9a9f64270 100644
--- a/drivers/webp/utils/color_cache.h
+++ b/drivers/webp/utils/color_cache.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Color Cache for WebP Lossless
@@ -13,26 +15,33 @@
 #ifndef WEBP_UTILS_COLOR_CACHE_H_
 #define WEBP_UTILS_COLOR_CACHE_H_
 
-#include "../types.h"
+#include "../webp/types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
 // Main color cache struct.
 typedef struct {
   uint32_t *colors_;  // color entries
-  int hash_shift_;    // Hash shift: 32 - hash_bits.
+  int hash_shift_;    // Hash shift: 32 - hash_bits_.
+  int hash_bits_;
 } VP8LColorCache;
 
 static const uint32_t kHashMul = 0x1e35a7bd;
 
 static WEBP_INLINE uint32_t VP8LColorCacheLookup(
     const VP8LColorCache* const cc, uint32_t key) {
-  assert(key <= (~0U >> cc->hash_shift_));
+  assert((key >> cc->hash_bits_) == 0u);
   return cc->colors_[key];
 }
 
+static WEBP_INLINE void VP8LColorCacheSet(const VP8LColorCache* const cc,
+                                          uint32_t key, uint32_t argb) {
+  assert((key >> cc->hash_bits_) == 0u);
+  cc->colors_[key] = argb;
+}
+
 static WEBP_INLINE void VP8LColorCacheInsert(const VP8LColorCache* const cc,
                                              uint32_t argb) {
   const uint32_t key = (kHashMul * argb) >> cc->hash_shift_;
@@ -47,7 +56,7 @@ static WEBP_INLINE int VP8LColorCacheGetIndex(const VP8LColorCache* const cc,
 static WEBP_INLINE int VP8LColorCacheContains(const VP8LColorCache* const cc,
                                               uint32_t argb) {
   const uint32_t key = (kHashMul * argb) >> cc->hash_shift_;
-  return cc->colors_[key] == argb;
+  return (cc->colors_[key] == argb);
 }
 
 //------------------------------------------------------------------------------
@@ -56,12 +65,15 @@ static WEBP_INLINE int VP8LColorCacheContains(const VP8LColorCache* const cc,
 // Returns false in case of memory error.
 int VP8LColorCacheInit(VP8LColorCache* const color_cache, int hash_bits);
 
+void VP8LColorCacheCopy(const VP8LColorCache* const src,
+                        VP8LColorCache* const dst);
+
 // Delete the memory associated to color cache.
 void VP8LColorCacheClear(VP8LColorCache* const color_cache);
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }
 #endif
 
diff --git a/drivers/webp/utils/filters.c b/drivers/webp/utils/filters.c
index 08f52a3d20..15543b1271 100644
--- a/drivers/webp/utils/filters.c
+++ b/drivers/webp/utils/filters.c
@@ -1,171 +1,37 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
-// Spatial prediction using various filters
+// filter estimation
 //
 // Author: Urvang (urvang@google.com)
 
 #include "./filters.h"
-#include <assert.h>
 #include <stdlib.h>
 #include <string.h>
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-//------------------------------------------------------------------------------
-// Helpful macro.
-
-# define SANITY_CHECK(in, out)                              \
-  assert(in != NULL);                                       \
-  assert(out != NULL);                                      \
-  assert(width > 0);                                        \
-  assert(height > 0);                                       \
-  assert(bpp > 0);                                          \
-  assert(stride >= width * bpp);
-
-static WEBP_INLINE void PredictLine(const uint8_t* src, const uint8_t* pred,
-                                    uint8_t* dst, int length, int inverse) {
-  int i;
-  if (inverse) {
-    for (i = 0; i < length; ++i) dst[i] = src[i] + pred[i];
-  } else {
-    for (i = 0; i < length; ++i) dst[i] = src[i] - pred[i];
-  }
-}
-
-//------------------------------------------------------------------------------
-// Horizontal filter.
-
-static WEBP_INLINE void DoHorizontalFilter(const uint8_t* in,
-    int width, int height, int bpp, int stride, int inverse, uint8_t* out) {
-  int h;
-  const uint8_t* preds = (inverse ? out : in);
-  SANITY_CHECK(in, out);
-
-  // Filter line-by-line.
-  for (h = 0; h < height; ++h) {
-    // Leftmost pixel is predicted from above (except for topmost scanline).
-    if (h == 0) {
-      memcpy((void*)out, (const void*)in, bpp);
-    } else {
-      PredictLine(in, preds - stride, out, bpp, inverse);
-    }
-    PredictLine(in + bpp, preds, out + bpp, bpp * (width - 1), inverse);
-    preds += stride;
-    in += stride;
-    out += stride;
-  }
-}
-
-static void HorizontalFilter(const uint8_t* data, int width, int height,
-                             int bpp, int stride, uint8_t* filtered_data) {
-  DoHorizontalFilter(data, width, height, bpp, stride, 0, filtered_data);
-}
-
-static void HorizontalUnfilter(const uint8_t* data, int width, int height,
-                               int bpp, int stride, uint8_t* recon_data) {
-  DoHorizontalFilter(data, width, height, bpp, stride, 1, recon_data);
-}
-
-//------------------------------------------------------------------------------
-// Vertical filter.
-
-static WEBP_INLINE void DoVerticalFilter(const uint8_t* in,
-    int width, int height, int bpp, int stride, int inverse, uint8_t* out) {
-  int h;
-  const uint8_t* preds = (inverse ? out : in);
-  SANITY_CHECK(in, out);
-
-  // Very first top-left pixel is copied.
-  memcpy((void*)out, (const void*)in, bpp);
-  // Rest of top scan-line is left-predicted.
-  PredictLine(in + bpp, preds, out + bpp, bpp * (width - 1), inverse);
-
-  // Filter line-by-line.
-  for (h = 1; h < height; ++h) {
-    in += stride;
-    out += stride;
-    PredictLine(in, preds, out, bpp * width, inverse);
-    preds += stride;
-  }
-}
-
-static void VerticalFilter(const uint8_t* data, int width, int height,
-                           int bpp, int stride, uint8_t* filtered_data) {
-  DoVerticalFilter(data, width, height, bpp, stride, 0, filtered_data);
-}
-
-static void VerticalUnfilter(const uint8_t* data, int width, int height,
-                             int bpp, int stride, uint8_t* recon_data) {
-  DoVerticalFilter(data, width, height, bpp, stride, 1, recon_data);
-}
+// -----------------------------------------------------------------------------
+// Quick estimate of a potentially interesting filter mode to try.
 
-//------------------------------------------------------------------------------
-// Gradient filter.
+#define SMAX 16
+#define SDIFF(a, b) (abs((a) - (b)) >> 4)   // Scoring diff, in [0..SMAX)
 
 static WEBP_INLINE int GradientPredictor(uint8_t a, uint8_t b, uint8_t c) {
   const int g = a + b - c;
-  return (g < 0) ? 0 : (g > 255) ? 255 : g;
-}
-
-static WEBP_INLINE
-void DoGradientFilter(const uint8_t* in, int width, int height,
-                      int bpp, int stride, int inverse, uint8_t* out) {
-  const uint8_t* preds = (inverse ? out : in);
-  int h;
-  SANITY_CHECK(in, out);
-
-  // left prediction for top scan-line
-  memcpy((void*)out, (const void*)in, bpp);
-  PredictLine(in + bpp, preds, out + bpp, bpp * (width - 1), inverse);
-
-  // Filter line-by-line.
-  for (h = 1; h < height; ++h) {
-    int w;
-    preds += stride;
-    in += stride;
-    out += stride;
-    // leftmost pixel: predict from above.
-    PredictLine(in, preds - stride, out, bpp, inverse);
-    for (w = bpp; w < width * bpp; ++w) {
-      const int pred = GradientPredictor(preds[w - bpp],
-                                         preds[w - stride],
-                                         preds[w - stride - bpp]);
-      out[w] = in[w] + (inverse ? pred : -pred);
-    }
-  }
+  return ((g & ~0xff) == 0) ? g : (g < 0) ? 0 : 255;  // clip to 8bit
 }
 
-static void GradientFilter(const uint8_t* data, int width, int height,
-                           int bpp, int stride, uint8_t* filtered_data) {
-  DoGradientFilter(data, width, height, bpp, stride, 0, filtered_data);
-}
-
-static void GradientUnfilter(const uint8_t* data, int width, int height,
-                             int bpp, int stride, uint8_t* recon_data) {
-  DoGradientFilter(data, width, height, bpp, stride, 1, recon_data);
-}
-
-#undef SANITY_CHECK
-
-// -----------------------------------------------------------------------------
-// Quick estimate of a potentially interesting filter mode to try, in addition
-// to the default NONE.
-
-#define SMAX 16
-#define SDIFF(a, b) (abs((a) - (b)) >> 4)   // Scoring diff, in [0..SMAX)
-
-WEBP_FILTER_TYPE EstimateBestFilter(const uint8_t* data,
-                                    int width, int height, int stride) {
+WEBP_FILTER_TYPE WebPEstimateBestFilter(const uint8_t* data,
+                                        int width, int height, int stride) {
   int i, j;
   int bins[WEBP_FILTER_LAST][SMAX];
   memset(bins, 0, sizeof(bins));
+
   // We only sample every other pixels. That's enough.
   for (j = 2; j < height - 1; j += 2) {
     const uint8_t* const p = data + j * stride;
@@ -185,7 +51,8 @@ WEBP_FILTER_TYPE EstimateBestFilter(const uint8_t* data,
     }
   }
   {
-    WEBP_FILTER_TYPE filter, best_filter = WEBP_FILTER_NONE;
+    int filter;
+    WEBP_FILTER_TYPE best_filter = WEBP_FILTER_NONE;
     int best_score = 0x7fffffff;
     for (filter = WEBP_FILTER_NONE; filter < WEBP_FILTER_LAST; ++filter) {
       int score = 0;
@@ -196,7 +63,7 @@ WEBP_FILTER_TYPE EstimateBestFilter(const uint8_t* data,
       }
       if (score < best_score) {
         best_score = score;
-        best_filter = filter;
+        best_filter = (WEBP_FILTER_TYPE)filter;
       }
     }
     return best_filter;
@@ -207,23 +74,3 @@ WEBP_FILTER_TYPE EstimateBestFilter(const uint8_t* data,
 #undef SDIFF
 
 //------------------------------------------------------------------------------
-
-const WebPFilterFunc WebPFilters[WEBP_FILTER_LAST] = {
-  NULL,              // WEBP_FILTER_NONE
-  HorizontalFilter,  // WEBP_FILTER_HORIZONTAL
-  VerticalFilter,    // WEBP_FILTER_VERTICAL
-  GradientFilter     // WEBP_FILTER_GRADIENT
-};
-
-const WebPFilterFunc WebPUnfilters[WEBP_FILTER_LAST] = {
-  NULL,                // WEBP_FILTER_NONE
-  HorizontalUnfilter,  // WEBP_FILTER_HORIZONTAL
-  VerticalUnfilter,    // WEBP_FILTER_VERTICAL
-  GradientUnfilter     // WEBP_FILTER_GRADIENT
-};
-
-//------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/utils/filters.h b/drivers/webp/utils/filters.h
index db886be29a..088b132fc5 100644
--- a/drivers/webp/utils/filters.h
+++ b/drivers/webp/utils/filters.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Spatial prediction using various filters
@@ -12,42 +14,18 @@
 #ifndef WEBP_UTILS_FILTERS_H_
 #define WEBP_UTILS_FILTERS_H_
 
-#include "../types.h"
+#include "../webp/types.h"
+#include "../dsp/dsp.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-// Filters.
-typedef enum {
-  WEBP_FILTER_NONE = 0,
-  WEBP_FILTER_HORIZONTAL,
-  WEBP_FILTER_VERTICAL,
-  WEBP_FILTER_GRADIENT,
-  WEBP_FILTER_LAST = WEBP_FILTER_GRADIENT + 1,  // end marker
-  WEBP_FILTER_BEST,
-  WEBP_FILTER_FAST
-} WEBP_FILTER_TYPE;
-
-typedef void (*WebPFilterFunc)(const uint8_t* in, int width, int height,
-                               int bpp, int stride, uint8_t* out);
-
-// Filter the given data using the given predictor.
-// 'in' corresponds to a 2-dimensional pixel array of size (stride * height)
-// in raster order.
-// 'bpp' is number of bytes per pixel, and
-// 'stride' is number of bytes per scan line (with possible padding).
-// 'out' should be pre-allocated.
-extern const WebPFilterFunc WebPFilters[WEBP_FILTER_LAST];
-
-// Reconstruct the original data from the given filtered data.
-extern const WebPFilterFunc WebPUnfilters[WEBP_FILTER_LAST];
-
 // Fast estimate of a potentially good filter.
-extern WEBP_FILTER_TYPE EstimateBestFilter(const uint8_t* data,
-                                           int width, int height, int stride);
+WEBP_FILTER_TYPE WebPEstimateBestFilter(const uint8_t* data,
+                                        int width, int height, int stride);
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/utils/huffman.c b/drivers/webp/utils/huffman.c
index 1cc1cfd355..d57376aa6b 100644
--- a/drivers/webp/utils/huffman.c
+++ b/drivers/webp/utils/huffman.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Utilities for building and looking up Huffman trees.
@@ -11,228 +13,193 @@
 
 #include <assert.h>
 #include <stdlib.h>
+#include <string.h>
 #include "./huffman.h"
 #include "../utils/utils.h"
-#include "../format_constants.h"
+#include "../webp/format_constants.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+// Huffman data read via DecodeImageStream is represented in two (red and green)
+// bytes.
+#define MAX_HTREE_GROUPS    0x10000
 
-#define NON_EXISTENT_SYMBOL (-1)
-
-static void TreeNodeInit(HuffmanTreeNode* const node) {
-  node->children_ = -1;   // means: 'unassigned so far'
-}
-
-static int NodeIsEmpty(const HuffmanTreeNode* const node) {
-  return (node->children_ < 0);
+HTreeGroup* VP8LHtreeGroupsNew(int num_htree_groups) {
+  HTreeGroup* const htree_groups =
+      (HTreeGroup*)WebPSafeMalloc(num_htree_groups, sizeof(*htree_groups));
+  if (htree_groups == NULL) {
+    return NULL;
+  }
+  assert(num_htree_groups <= MAX_HTREE_GROUPS);
+  return htree_groups;
 }
 
-static int IsFull(const HuffmanTree* const tree) {
-  return (tree->num_nodes_ == tree->max_nodes_);
+void VP8LHtreeGroupsFree(HTreeGroup* const htree_groups) {
+  if (htree_groups != NULL) {
+    WebPSafeFree(htree_groups);
+  }
 }
 
-static void AssignChildren(HuffmanTree* const tree,
-                           HuffmanTreeNode* const node) {
-  HuffmanTreeNode* const children = tree->root_ + tree->num_nodes_;
-  node->children_ = (int)(children - node);
-  assert(children - node == (int)(children - node));
-  tree->num_nodes_ += 2;
-  TreeNodeInit(children + 0);
-  TreeNodeInit(children + 1);
+// Returns reverse(reverse(key, len) + 1, len), where reverse(key, len) is the
+// bit-wise reversal of the len least significant bits of key.
+static WEBP_INLINE uint32_t GetNextKey(uint32_t key, int len) {
+  uint32_t step = 1 << (len - 1);
+  while (key & step) {
+    step >>= 1;
+  }
+  return (key & (step - 1)) + step;
 }
 
-static int TreeInit(HuffmanTree* const tree, int num_leaves) {
-  assert(tree != NULL);
-  if (num_leaves == 0) return 0;
-  // We allocate maximum possible nodes in the tree at once.
-  // Note that a Huffman tree is a full binary tree; and in a full binary tree
-  // with L leaves, the total number of nodes N = 2 * L - 1.
-  tree->max_nodes_ = 2 * num_leaves - 1;
-  tree->root_ = (HuffmanTreeNode*)WebPSafeMalloc((uint64_t)tree->max_nodes_,
-                                                 sizeof(*tree->root_));
-  if (tree->root_ == NULL) return 0;
-  TreeNodeInit(tree->root_);  // Initialize root.
-  tree->num_nodes_ = 1;
-  return 1;
+// Stores code in table[0], table[step], table[2*step], ..., table[end].
+// Assumes that end is an integer multiple of step.
+static WEBP_INLINE void ReplicateValue(HuffmanCode* table,
+                                       int step, int end,
+                                       HuffmanCode code) {
+  assert(end % step == 0);
+  do {
+    end -= step;
+    table[end] = code;
+  } while (end > 0);
 }
 
-void HuffmanTreeRelease(HuffmanTree* const tree) {
-  if (tree != NULL) {
-    free(tree->root_);
-    tree->root_ = NULL;
-    tree->max_nodes_ = 0;
-    tree->num_nodes_ = 0;
+// Returns the table width of the next 2nd level table. count is the histogram
+// of bit lengths for the remaining symbols, len is the code length of the next
+// processed symbol
+static WEBP_INLINE int NextTableBitSize(const int* const count,
+                                        int len, int root_bits) {
+  int left = 1 << (len - root_bits);
+  while (len < MAX_ALLOWED_CODE_LENGTH) {
+    left -= count[len];
+    if (left <= 0) break;
+    ++len;
+    left <<= 1;
   }
+  return len - root_bits;
 }
 
-int HuffmanCodeLengthsToCodes(const int* const code_lengths,
-                              int code_lengths_size, int* const huff_codes) {
-  int symbol;
-  int code_len;
-  int code_length_hist[MAX_ALLOWED_CODE_LENGTH + 1] = { 0 };
-  int curr_code;
-  int next_codes[MAX_ALLOWED_CODE_LENGTH + 1] = { 0 };
-  int max_code_length = 0;
-
+int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
+                          const int code_lengths[], int code_lengths_size) {
+  HuffmanCode* table = root_table;  // next available space in table
+  int total_size = 1 << root_bits;  // total size root table + 2nd level table
+  int* sorted = NULL;               // symbols sorted by code length
+  int len;                          // current code length
+  int symbol;                       // symbol index in original or sorted table
+  // number of codes of each length:
+  int count[MAX_ALLOWED_CODE_LENGTH + 1] = { 0 };
+  // offsets in sorted table for each length:
+  int offset[MAX_ALLOWED_CODE_LENGTH + 1];
+
+  assert(code_lengths_size != 0);
   assert(code_lengths != NULL);
-  assert(code_lengths_size > 0);
-  assert(huff_codes != NULL);
+  assert(root_table != NULL);
+  assert(root_bits > 0);
 
-  // Calculate max code length.
+  // Build histogram of code lengths.
   for (symbol = 0; symbol < code_lengths_size; ++symbol) {
-    if (code_lengths[symbol] > max_code_length) {
-      max_code_length = code_lengths[symbol];
+    if (code_lengths[symbol] > MAX_ALLOWED_CODE_LENGTH) {
+      return 0;
     }
+    ++count[code_lengths[symbol]];
   }
-  if (max_code_length > MAX_ALLOWED_CODE_LENGTH) return 0;
 
-  // Calculate code length histogram.
-  for (symbol = 0; symbol < code_lengths_size; ++symbol) {
-    ++code_length_hist[code_lengths[symbol]];
-  }
-  code_length_hist[0] = 0;
-
-  // Calculate the initial values of 'next_codes' for each code length.
-  // next_codes[code_len] denotes the code to be assigned to the next symbol
-  // of code length 'code_len'.
-  curr_code = 0;
-  next_codes[0] = -1;  // Unused, as code length = 0 implies code doesn't exist.
-  for (code_len = 1; code_len <= max_code_length; ++code_len) {
-    curr_code = (curr_code + code_length_hist[code_len - 1]) << 1;
-    next_codes[code_len] = curr_code;
+  // Error, all code lengths are zeros.
+  if (count[0] == code_lengths_size) {
+    return 0;
   }
 
-  // Get symbols.
-  for (symbol = 0; symbol < code_lengths_size; ++symbol) {
-    if (code_lengths[symbol] > 0) {
-      huff_codes[symbol] = next_codes[code_lengths[symbol]]++;
-    } else {
-      huff_codes[symbol] = NON_EXISTENT_SYMBOL;
-    }
-  }
-  return 1;
-}
-
-static int TreeAddSymbol(HuffmanTree* const tree,
-                         int symbol, int code, int code_length) {
-  HuffmanTreeNode* node = tree->root_;
-  const HuffmanTreeNode* const max_node = tree->root_ + tree->max_nodes_;
-  while (code_length-- > 0) {
-    if (node >= max_node) {
+  // Generate offsets into sorted symbol table by code length.
+  offset[1] = 0;
+  for (len = 1; len < MAX_ALLOWED_CODE_LENGTH; ++len) {
+    if (count[len] > (1 << len)) {
       return 0;
     }
-    if (NodeIsEmpty(node)) {
-      if (IsFull(tree)) return 0;    // error: too many symbols.
-      AssignChildren(tree, node);
-    } else if (HuffmanTreeNodeIsLeaf(node)) {
-      return 0;  // leaf is already occupied.
-    }
-    node += node->children_ + ((code >> code_length) & 1);
-  }
-  if (NodeIsEmpty(node)) {
-    node->children_ = 0;      // turn newly created node into a leaf.
-  } else if (!HuffmanTreeNodeIsLeaf(node)) {
-    return 0;   // trying to assign a symbol to already used code.
+    offset[len + 1] = offset[len] + count[len];
   }
-  node->symbol_ = symbol;  // Add symbol in this node.
-  return 1;
-}
 
-int HuffmanTreeBuildImplicit(HuffmanTree* const tree,
-                             const int* const code_lengths,
-                             int code_lengths_size) {
-  int symbol;
-  int num_symbols = 0;
-  int root_symbol = 0;
-
-  assert(tree != NULL);
-  assert(code_lengths != NULL);
+  sorted = (int*)WebPSafeMalloc(code_lengths_size, sizeof(*sorted));
+  if (sorted == NULL) {
+    return 0;
+  }
 
-  // Find out number of symbols and the root symbol.
+  // Sort symbols by length, by symbol order within each length.
   for (symbol = 0; symbol < code_lengths_size; ++symbol) {
+    const int symbol_code_length = code_lengths[symbol];
     if (code_lengths[symbol] > 0) {
-      // Note: code length = 0 indicates non-existent symbol.
-      ++num_symbols;
-      root_symbol = symbol;
+      sorted[offset[symbol_code_length]++] = symbol;
     }
   }
 
-  // Initialize the tree. Will fail for num_symbols = 0
-  if (!TreeInit(tree, num_symbols)) return 0;
-
-  // Build tree.
-  if (num_symbols == 1) {  // Trivial case.
-    const int max_symbol = code_lengths_size;
-    if (root_symbol < 0 || root_symbol >= max_symbol) {
-      HuffmanTreeRelease(tree);
-      return 0;
-    }
-    return TreeAddSymbol(tree, root_symbol, 0, 0);
-  } else {  // Normal case.
-    int ok = 0;
-
-    // Get Huffman codes from the code lengths.
-    int* const codes =
-        (int*)WebPSafeMalloc((uint64_t)code_lengths_size, sizeof(*codes));
-    if (codes == NULL) goto End;
+  // Special case code with only one value.
+  if (offset[MAX_ALLOWED_CODE_LENGTH] == 1) {
+    HuffmanCode code;
+    code.bits = 0;
+    code.value = (uint16_t)sorted[0];
+    ReplicateValue(table, 1, total_size, code);
+    WebPSafeFree(sorted);
+    return total_size;
+  }
 
-    if (!HuffmanCodeLengthsToCodes(code_lengths, code_lengths_size, codes)) {
-      goto End;
+  {
+    int step;              // step size to replicate values in current table
+    uint32_t low = -1;     // low bits for current root entry
+    uint32_t mask = total_size - 1;    // mask for low bits
+    uint32_t key = 0;      // reversed prefix code
+    int num_nodes = 1;     // number of Huffman tree nodes
+    int num_open = 1;      // number of open branches in current tree level
+    int table_bits = root_bits;        // key length of current table
+    int table_size = 1 << table_bits;  // size of current table
+    symbol = 0;
+    // Fill in root table.
+    for (len = 1, step = 2; len <= root_bits; ++len, step <<= 1) {
+      num_open <<= 1;
+      num_nodes += num_open;
+      num_open -= count[len];
+      if (num_open < 0) {
+        WebPSafeFree(sorted);
+        return 0;
+      }
+      for (; count[len] > 0; --count[len]) {
+        HuffmanCode code;
+        code.bits = (uint8_t)len;
+        code.value = (uint16_t)sorted[symbol++];
+        ReplicateValue(&table[key], step, table_size, code);
+        key = GetNextKey(key, len);
+      }
     }
 
-    // Add symbols one-by-one.
-    for (symbol = 0; symbol < code_lengths_size; ++symbol) {
-      if (code_lengths[symbol] > 0) {
-        if (!TreeAddSymbol(tree, symbol, codes[symbol], code_lengths[symbol])) {
-          goto End;
+    // Fill in 2nd level tables and add pointers to root table.
+    for (len = root_bits + 1, step = 2; len <= MAX_ALLOWED_CODE_LENGTH;
+         ++len, step <<= 1) {
+      num_open <<= 1;
+      num_nodes += num_open;
+      num_open -= count[len];
+      if (num_open < 0) {
+        WebPSafeFree(sorted);
+        return 0;
+      }
+      for (; count[len] > 0; --count[len]) {
+        HuffmanCode code;
+        if ((key & mask) != low) {
+          table += table_size;
+          table_bits = NextTableBitSize(count, len, root_bits);
+          table_size = 1 << table_bits;
+          total_size += table_size;
+          low = key & mask;
+          root_table[low].bits = (uint8_t)(table_bits + root_bits);
+          root_table[low].value = (uint16_t)((table - root_table) - low);
         }
+        code.bits = (uint8_t)(len - root_bits);
+        code.value = (uint16_t)sorted[symbol++];
+        ReplicateValue(&table[key >> root_bits], step, table_size, code);
+        key = GetNextKey(key, len);
       }
     }
-    ok = 1;
- End:
-    free(codes);
-    ok = ok && IsFull(tree);
-    if (!ok) HuffmanTreeRelease(tree);
-    return ok;
-  }
-}
-
-int HuffmanTreeBuildExplicit(HuffmanTree* const tree,
-                             const int* const code_lengths,
-                             const int* const codes,
-                             const int* const symbols, int max_symbol,
-                             int num_symbols) {
-  int ok = 0;
-  int i;
-
-  assert(tree != NULL);
-  assert(code_lengths != NULL);
-  assert(codes != NULL);
-  assert(symbols != NULL);
-
-  // Initialize the tree. Will fail if num_symbols = 0.
-  if (!TreeInit(tree, num_symbols)) return 0;
 
-  // Add symbols one-by-one.
-  for (i = 0; i < num_symbols; ++i) {
-    if (codes[i] != NON_EXISTENT_SYMBOL) {
-      if (symbols[i] < 0 || symbols[i] >= max_symbol) {
-        goto End;
-      }
-      if (!TreeAddSymbol(tree, symbols[i], codes[i], code_lengths[i])) {
-        goto End;
-      }
+    // Check if tree is full.
+    if (num_nodes != 2 * offset[MAX_ALLOWED_CODE_LENGTH] - 1) {
+      WebPSafeFree(sorted);
+      return 0;
     }
   }
-  ok = 1;
- End:
-  ok = ok && IsFull(tree);
-  if (!ok) HuffmanTreeRelease(tree);
-  return ok;
-}
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
+  WebPSafeFree(sorted);
+  return total_size;
+}
diff --git a/drivers/webp/utils/huffman.h b/drivers/webp/utils/huffman.h
index f16447e649..c6dd6aaa45 100644
--- a/drivers/webp/utils/huffman.h
+++ b/drivers/webp/utils/huffman.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Utilities for building and looking up Huffman trees.
@@ -13,65 +15,73 @@
 #define WEBP_UTILS_HUFFMAN_H_
 
 #include <assert.h>
-#include "../types.h"
+#include "../webp/format_constants.h"
+#include "../webp/types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-// A node of a Huffman tree.
-typedef struct {
-  int symbol_;
-  int children_;  // delta offset to both children (contiguous) or 0 if leaf.
-} HuffmanTreeNode;
+#define HUFFMAN_TABLE_BITS      8
+#define HUFFMAN_TABLE_MASK      ((1 << HUFFMAN_TABLE_BITS) - 1)
+
+#define LENGTHS_TABLE_BITS      7
+#define LENGTHS_TABLE_MASK      ((1 << LENGTHS_TABLE_BITS) - 1)
 
-// Huffman Tree.
-typedef struct HuffmanTree HuffmanTree;
-struct HuffmanTree {
-  HuffmanTreeNode* root_;   // all the nodes, starting at root.
-  int max_nodes_;           // max number of nodes
-  int num_nodes_;           // number of currently occupied nodes
-};
 
-// Returns true if the given node is a leaf of the Huffman tree.
-static WEBP_INLINE int HuffmanTreeNodeIsLeaf(
-    const HuffmanTreeNode* const node) {
-  return (node->children_ == 0);
-}
+// Huffman lookup table entry
+typedef struct {
+  uint8_t bits;     // number of bits used for this symbol
+  uint16_t value;   // symbol value or table offset
+} HuffmanCode;
 
-// Go down one level. Most critical function. 'right_child' must be 0 or 1.
-static WEBP_INLINE const HuffmanTreeNode* HuffmanTreeNextNode(
-    const HuffmanTreeNode* node, int right_child) {
-  return node + node->children_ + right_child;
-}
+// long version for holding 32b values
+typedef struct {
+  int bits;         // number of bits used for this symbol,
+                    // or an impossible value if not a literal code.
+  uint32_t value;   // 32b packed ARGB value if literal,
+                    // or non-literal symbol otherwise
+} HuffmanCode32;
 
-// Releases the nodes of the Huffman tree.
-// Note: It does NOT free 'tree' itself.
-void HuffmanTreeRelease(HuffmanTree* const tree);
+#define HUFFMAN_PACKED_BITS 6
+#define HUFFMAN_PACKED_TABLE_SIZE (1u << HUFFMAN_PACKED_BITS)
 
-// Builds Huffman tree assuming code lengths are implicitly in symbol order.
-// Returns false in case of error (invalid tree or memory error).
-int HuffmanTreeBuildImplicit(HuffmanTree* const tree,
-                             const int* const code_lengths,
-                             int code_lengths_size);
+// Huffman table group.
+// Includes special handling for the following cases:
+//  - is_trivial_literal: one common literal base for RED/BLUE/ALPHA (not GREEN)
+//  - is_trivial_code: only 1 code (no bit is read from bitstream)
+//  - use_packed_table: few enough literal symbols, so all the bit codes
+//    can fit into a small look-up table packed_table[]
+// The common literal base, if applicable, is stored in 'literal_arb'.
+typedef struct HTreeGroup HTreeGroup;
+struct HTreeGroup {
+  HuffmanCode* htrees[HUFFMAN_CODES_PER_META_CODE];
+  int      is_trivial_literal;  // True, if huffman trees for Red, Blue & Alpha
+                                // Symbols are trivial (have a single code).
+  uint32_t literal_arb;         // If is_trivial_literal is true, this is the
+                                // ARGB value of the pixel, with Green channel
+                                // being set to zero.
+  int is_trivial_code;          // true if is_trivial_literal with only one code
+  int use_packed_table;         // use packed table below for short literal code
+  // table mapping input bits to a packed values, or escape case to literal code
+  HuffmanCode32 packed_table[HUFFMAN_PACKED_TABLE_SIZE];
+};
 
-// Build a Huffman tree with explicitly given lists of code lengths, codes
-// and symbols. Verifies that all symbols added are smaller than max_symbol.
-// Returns false in case of an invalid symbol, invalid tree or memory error.
-int HuffmanTreeBuildExplicit(HuffmanTree* const tree,
-                             const int* const code_lengths,
-                             const int* const codes,
-                             const int* const symbols, int max_symbol,
-                             int num_symbols);
+// Creates the instance of HTreeGroup with specified number of tree-groups.
+HTreeGroup* VP8LHtreeGroupsNew(int num_htree_groups);
 
-// Utility: converts Huffman code lengths to corresponding Huffman codes.
-// 'huff_codes' should be pre-allocated.
-// Returns false in case of error (memory allocation, invalid codes).
-int HuffmanCodeLengthsToCodes(const int* const code_lengths,
-                              int code_lengths_size, int* const huff_codes);
+// Releases the memory allocated for HTreeGroup.
+void VP8LHtreeGroupsFree(HTreeGroup* const htree_groups);
 
+// Builds Huffman lookup table assuming code lengths are in symbol order.
+// The 'code_lengths' is pre-allocated temporary memory buffer used for creating
+// the huffman table.
+// Returns built table size or 0 in case of error (invalid tree or
+// memory error).
+int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
+                          const int code_lengths[], int code_lengths_size);
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/utils/huffman_encode.c b/drivers/webp/utils/huffman_encode.c
index e172b10a85..6421c2beed 100644
--- a/drivers/webp/utils/huffman_encode.c
+++ b/drivers/webp/utils/huffman_encode.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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)
@@ -14,7 +16,7 @@
 #include <string.h>
 #include "./huffman_encode.h"
 #include "../utils/utils.h"
-#include "../format_constants.h"
+#include "../webp/format_constants.h"
 
 // -----------------------------------------------------------------------------
 // Util function to optimize the symbol map for RLE coding
@@ -25,14 +27,14 @@ static int ValuesShouldBeCollapsedToStrideAverage(int a, int b) {
 }
 
 // Change the population counts in a way that the consequent
-// Hufmann tree compression, especially its RLE-part, give smaller output.
-static int OptimizeHuffmanForRle(int length, int* const counts) {
-  uint8_t* good_for_rle;
+// Huffman tree compression, especially its RLE-part, give smaller output.
+static void OptimizeHuffmanForRle(int length, uint8_t* const good_for_rle,
+                                  uint32_t* const counts) {
   // 1) Let's make the Huffman code more compatible with rle encoding.
   int i;
   for (; length >= 0; --length) {
     if (length == 0) {
-      return 1;  // All zeros.
+      return;  // All zeros.
     }
     if (counts[length - 1] != 0) {
       // Now counts[0..length - 1] does not have trailing zeros.
@@ -41,15 +43,11 @@ static int OptimizeHuffmanForRle(int length, int* const counts) {
   }
   // 2) Let's mark all population counts that already can be encoded
   // with an rle code.
-  good_for_rle = (uint8_t*)calloc(length, 1);
-  if (good_for_rle == NULL) {
-    return 0;
-  }
   {
     // Let's not spoil any of the existing good rle codes.
     // Mark any seq of 0's that is longer as 5 as a good_for_rle.
     // Mark any seq of non-0's that is longer as 7 as a good_for_rle.
-    int symbol = counts[0];
+    uint32_t symbol = counts[0];
     int stride = 0;
     for (i = 0; i < length + 1; ++i) {
       if (i == length || counts[i] != symbol) {
@@ -71,17 +69,17 @@ static int OptimizeHuffmanForRle(int length, int* const counts) {
   }
   // 3) Let's replace those population counts that lead to more rle codes.
   {
-    int stride = 0;
-    int limit = counts[0];
-    int sum = 0;
+    uint32_t stride = 0;
+    uint32_t limit = counts[0];
+    uint32_t sum = 0;
     for (i = 0; i < length + 1; ++i) {
       if (i == length || good_for_rle[i] ||
           (i != 0 && good_for_rle[i - 1]) ||
           !ValuesShouldBeCollapsedToStrideAverage(counts[i], limit)) {
         if (stride >= 4 || (stride >= 3 && sum == 0)) {
-          int k;
+          uint32_t k;
           // The stride must end, collapse what we have, if we have enough (4).
-          int count = (sum + stride / 2) / stride;
+          uint32_t count = (sum + stride / 2) / stride;
           if (count < 1) {
             count = 1;
           }
@@ -117,17 +115,8 @@ static int OptimizeHuffmanForRle(int length, int* const counts) {
       }
     }
   }
-  free(good_for_rle);
-  return 1;
 }
 
-typedef struct {
-  int total_count_;
-  int value_;
-  int pool_index_left_;
-  int pool_index_right_;
-} HuffmanTree;
-
 // A comparer function for two Huffman trees: sorts first by 'total count'
 // (more comes first), and then by 'value' (more comes first).
 static int CompareHuffmanTrees(const void* ptr1, const void* ptr2) {
@@ -138,13 +127,8 @@ static int CompareHuffmanTrees(const void* ptr1, const void* ptr2) {
   } else if (t1->total_count_ < t2->total_count_) {
     return 1;
   } else {
-    if (t1->value_ < t2->value_) {
-      return -1;
-    }
-    if (t1->value_ > t2->value_) {
-      return 1;
-    }
-    return 0;
+    assert(t1->value_ != t2->value_);
+    return (t1->value_ < t2->value_) ? -1 : 1;
   }
 }
 
@@ -178,12 +162,12 @@ static void SetBitDepths(const HuffmanTree* const tree,
 // we are not planning to use this with extremely long blocks.
 //
 // See http://en.wikipedia.org/wiki/Huffman_coding
-static int GenerateOptimalTree(const int* const histogram, int histogram_size,
-                               int tree_depth_limit,
-                               uint8_t* const bit_depths) {
-  int count_min;
+static void GenerateOptimalTree(const uint32_t* const histogram,
+                                int histogram_size,
+                                HuffmanTree* tree, int tree_depth_limit,
+                                uint8_t* const bit_depths) {
+  uint32_t count_min;
   HuffmanTree* tree_pool;
-  HuffmanTree* tree;
   int tree_size_orig = 0;
   int i;
 
@@ -193,12 +177,10 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
     }
   }
 
-  // 3 * tree_size is enough to cover all the nodes representing a
-  // population and all the inserted nodes combining two existing nodes.
-  // The tree pool needs 2 * (tree_size_orig - 1) entities, and the
-  // tree needs exactly tree_size_orig entities.
-  tree = (HuffmanTree*)WebPSafeMalloc(3ULL * tree_size_orig, sizeof(*tree));
-  if (tree == NULL) return 0;
+  if (tree_size_orig == 0) {   // pretty optimal already!
+    return;
+  }
+
   tree_pool = tree + tree_size_orig;
 
   // For block sizes with less than 64k symbols we never need to do a
@@ -214,7 +196,7 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
     int j;
     for (j = 0; j < histogram_size; ++j) {
       if (histogram[j] != 0) {
-        const int count =
+        const uint32_t count =
             (histogram[j] < count_min) ? count_min : histogram[j];
         tree[idx].total_count_ = count;
         tree[idx].value_ = j;
@@ -230,11 +212,11 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
     if (tree_size > 1) {  // Normal case.
       int tree_pool_size = 0;
       while (tree_size > 1) {  // Finish when we have only one root.
-        int count;
+        uint32_t count;
         tree_pool[tree_pool_size++] = tree[tree_size - 1];
         tree_pool[tree_pool_size++] = tree[tree_size - 2];
         count = tree_pool[tree_pool_size - 1].total_count_ +
-            tree_pool[tree_pool_size - 2].total_count_;
+                tree_pool[tree_pool_size - 2].total_count_;
         tree_size -= 2;
         {
           // Search for the insertion point.
@@ -271,8 +253,6 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
       }
     }
   }
-  free(tree);
-  return 1;
 }
 
 // -----------------------------------------------------------------------------
@@ -423,17 +403,15 @@ static void ConvertBitDepthsToSymbols(HuffmanTreeCode* const tree) {
 // -----------------------------------------------------------------------------
 // Main entry point
 
-int VP8LCreateHuffmanTree(int* const histogram, int tree_depth_limit,
-                          HuffmanTreeCode* const tree) {
-  const int num_symbols = tree->num_symbols;
-  if (!OptimizeHuffmanForRle(num_symbols, histogram)) {
-    return 0;
-  }
-  if (!GenerateOptimalTree(histogram, num_symbols,
-                           tree_depth_limit, tree->code_lengths)) {
-    return 0;
-  }
+void VP8LCreateHuffmanTree(uint32_t* const histogram, int tree_depth_limit,
+                           uint8_t* const buf_rle,
+                           HuffmanTree* const huff_tree,
+                           HuffmanTreeCode* const huff_code) {
+  const int num_symbols = huff_code->num_symbols;
+  memset(buf_rle, 0, num_symbols * sizeof(*buf_rle));
+  OptimizeHuffmanForRle(num_symbols, buf_rle, histogram);
+  GenerateOptimalTree(histogram, num_symbols, huff_tree, tree_depth_limit,
+                      huff_code->code_lengths);
   // Create the actual bit codes for the bit lengths.
-  ConvertBitDepthsToSymbols(tree);
-  return 1;
+  ConvertBitDepthsToSymbols(huff_code);
 }
diff --git a/drivers/webp/utils/huffman_encode.h b/drivers/webp/utils/huffman_encode.h
index 7f4aedc102..a157165148 100644
--- a/drivers/webp/utils/huffman_encode.h
+++ b/drivers/webp/utils/huffman_encode.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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)
@@ -12,9 +14,9 @@
 #ifndef WEBP_UTILS_HUFFMAN_ENCODE_H_
 #define WEBP_UTILS_HUFFMAN_ENCODE_H_
 
-#include "../types.h"
+#include "../webp/types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -31,16 +33,27 @@ typedef struct {
   uint16_t* codes;         // Symbol Codes.
 } HuffmanTreeCode;
 
+// Struct to represent the Huffman tree.
+typedef struct {
+  uint32_t total_count_;   // Symbol frequency.
+  int value_;              // Symbol value.
+  int pool_index_left_;    // Index for the left sub-tree.
+  int pool_index_right_;   // Index for the right sub-tree.
+} HuffmanTree;
+
 // Turn the Huffman tree into a token sequence.
 // Returns the number of tokens used.
 int VP8LCreateCompressedHuffmanTree(const HuffmanTreeCode* const tree,
                                     HuffmanTreeToken* tokens, int max_tokens);
 
 // Create an optimized tree, and tokenize it.
-int VP8LCreateHuffmanTree(int* const histogram, int tree_depth_limit,
-                          HuffmanTreeCode* const tree);
+// 'buf_rle' and 'huff_tree' are pre-allocated and the 'tree' is the constructed
+// huffman code tree.
+void VP8LCreateHuffmanTree(uint32_t* const histogram, int tree_depth_limit,
+                           uint8_t* const buf_rle, HuffmanTree* const huff_tree,
+                           HuffmanTreeCode* const tree);
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }
 #endif
 
diff --git a/drivers/webp/utils/quant_levels.c b/drivers/webp/utils/quant_levels.c
index f6884392aa..d7c8aab922 100644
--- a/drivers/webp/utils/quant_levels.c
+++ b/drivers/webp/utils/quant_levels.c
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Quantize levels for specified number of quantization-levels ([2, 256]).
@@ -14,10 +16,6 @@
 
 #include "./quant_levels.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 #define NUM_SYMBOLS     256
 
 #define MAX_ITER  6             // Maximum number of convergence steps.
@@ -140,15 +138,3 @@ int QuantizeLevels(uint8_t* const data, int width, int height,
   return 1;
 }
 
-int DequantizeLevels(uint8_t* const data, int width, int height) {
-  if (data == NULL || width <= 0 || height <= 0) return 0;
-  // TODO(skal): implement gradient smoothing.
-  (void)data;
-  (void)width;
-  (void)height;
-  return 1;
-}
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/utils/quant_levels.h b/drivers/webp/utils/quant_levels.h
index 4f165fd230..1cb5a32cae 100644
--- a/drivers/webp/utils/quant_levels.h
+++ b/drivers/webp/utils/quant_levels.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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 quantization utility
@@ -14,9 +16,9 @@
 
 #include <stdlib.h>
 
-#include "../types.h"
+#include "../webp/types.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -27,12 +29,7 @@ extern "C" {
 int QuantizeLevels(uint8_t* const data, int width, int height, int num_levels,
                    uint64_t* const sse);
 
-// Apply post-processing to input 'data' of size 'width'x'height' assuming
-// that the source was quantized to a reduced number of levels.
-// Returns false in case of error (data is NULL, invalid parameters, ...).
-int DequantizeLevels(uint8_t* const data, int width, int height);
-
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/utils/rescaler.c b/drivers/webp/utils/rescaler.c
index 9825dcbc5f..00c9300bfb 100644
--- a/drivers/webp/utils/rescaler.c
+++ b/drivers/webp/utils/rescaler.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Rescaling functions
@@ -11,124 +13,116 @@
 
 #include <assert.h>
 #include <stdlib.h>
+#include <string.h>
+#include "../dsp/dsp.h"
 #include "./rescaler.h"
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
-#define RFIX 30
-#define MULT_FIX(x,y) (((int64_t)(x) * (y) + (1 << (RFIX - 1))) >> RFIX)
-
 void WebPRescalerInit(WebPRescaler* const wrk, int src_width, int src_height,
-                      uint8_t* const dst, int dst_width, int dst_height,
-                      int dst_stride, int num_channels, int x_add, int x_sub,
-                      int y_add, int y_sub, int32_t* const work) {
+                      uint8_t* const dst,
+                      int dst_width, int dst_height, int dst_stride,
+                      int num_channels, rescaler_t* const work) {
+  const int x_add = src_width, x_sub = dst_width;
+  const int y_add = src_height, y_sub = dst_height;
   wrk->x_expand = (src_width < dst_width);
+  wrk->y_expand = (src_height < dst_height);
   wrk->src_width = src_width;
   wrk->src_height = src_height;
   wrk->dst_width = dst_width;
   wrk->dst_height = dst_height;
+  wrk->src_y = 0;
+  wrk->dst_y = 0;
   wrk->dst = dst;
   wrk->dst_stride = dst_stride;
   wrk->num_channels = num_channels;
+
   // for 'x_expand', we use bilinear interpolation
-  wrk->x_add = wrk->x_expand ? (x_sub - 1) : x_add - x_sub;
+  wrk->x_add = wrk->x_expand ? (x_sub - 1) : x_add;
   wrk->x_sub = wrk->x_expand ? (x_add - 1) : x_sub;
-  wrk->y_accum = y_add;
-  wrk->y_add = y_add;
-  wrk->y_sub = y_sub;
-  wrk->fx_scale = (1 << RFIX) / x_sub;
-  wrk->fy_scale = (1 << RFIX) / y_sub;
-  wrk->fxy_scale = wrk->x_expand ?
-      ((int64_t)dst_height << RFIX) / (x_sub * src_height) :
-      ((int64_t)dst_height << RFIX) / (x_add * src_height);
+  if (!wrk->x_expand) {  // fx_scale is not used otherwise
+    wrk->fx_scale = WEBP_RESCALER_FRAC(1, wrk->x_sub);
+  }
+  // vertical scaling parameters
+  wrk->y_add = wrk->y_expand ? y_add - 1 : y_add;
+  wrk->y_sub = wrk->y_expand ? y_sub - 1 : y_sub;
+  wrk->y_accum = wrk->y_expand ? wrk->y_sub : wrk->y_add;
+  if (!wrk->y_expand) {
+    // this is WEBP_RESCALER_FRAC(dst_height, x_add * y_add) without the cast.
+    const uint64_t ratio =
+        (uint64_t)dst_height * WEBP_RESCALER_ONE / (wrk->x_add * wrk->y_add);
+    if (ratio != (uint32_t)ratio) {
+      // We can't represent the ratio with the current fixed-point precision.
+      // => We special-case fxy_scale = 0, in WebPRescalerExportRow().
+      wrk->fxy_scale = 0;
+    } else {
+      wrk->fxy_scale = (uint32_t)ratio;
+    }
+    wrk->fy_scale = WEBP_RESCALER_FRAC(1, wrk->y_sub);
+  } else {
+    wrk->fy_scale = WEBP_RESCALER_FRAC(1, wrk->x_add);
+    // wrk->fxy_scale is unused here.
+  }
   wrk->irow = work;
   wrk->frow = work + num_channels * dst_width;
-}
+  memset(work, 0, 2 * dst_width * num_channels * sizeof(*work));
 
-void WebPRescalerImportRow(WebPRescaler* const wrk,
-                           const uint8_t* const src, int channel) {
-  const int x_stride = wrk->num_channels;
-  const int x_out_max = wrk->dst_width * wrk->num_channels;
-  int x_in = channel;
-  int x_out;
-  int accum = 0;
-  if (!wrk->x_expand) {
-    int sum = 0;
-    for (x_out = channel; x_out < x_out_max; x_out += x_stride) {
-      accum += wrk->x_add;
-      for (; accum > 0; accum -= wrk->x_sub) {
-        sum += src[x_in];
-        x_in += x_stride;
-      }
-      {        // Emit next horizontal pixel.
-        const int32_t base = src[x_in];
-        const int32_t frac = base * (-accum);
-        x_in += x_stride;
-        wrk->frow[x_out] = (sum + base) * wrk->x_sub - frac;
-        // fresh fractional start for next pixel
-        sum = (int)MULT_FIX(frac, wrk->fx_scale);
-      }
-    }
-  } else {        // simple bilinear interpolation
-    int left = src[channel], right = src[channel];
-    for (x_out = channel; x_out < x_out_max; x_out += x_stride) {
-      if (accum < 0) {
-        left = right;
-        x_in += x_stride;
-        right = src[x_in];
-        accum += wrk->x_add;
-      }
-      wrk->frow[x_out] = right * wrk->x_add + (left - right) * accum;
-      accum -= wrk->x_sub;
-    }
-  }
-  // Accumulate the new row's contribution
-  for (x_out = channel; x_out < x_out_max; x_out += x_stride) {
-    wrk->irow[x_out] += wrk->frow[x_out];
-  }
+  WebPRescalerDspInit();
 }
 
-uint8_t* WebPRescalerExportRow(WebPRescaler* const wrk) {
-  if (wrk->y_accum <= 0) {
-    int x_out;
-    uint8_t* const dst = wrk->dst;
-    int32_t* const irow = wrk->irow;
-    const int32_t* const frow = wrk->frow;
-    const int yscale = wrk->fy_scale * (-wrk->y_accum);
-    const int x_out_max = wrk->dst_width * wrk->num_channels;
+int WebPRescalerGetScaledDimensions(int src_width, int src_height,
+                                    int* const scaled_width,
+                                    int* const scaled_height) {
+  assert(scaled_width != NULL);
+  assert(scaled_height != NULL);
+  {
+    int width = *scaled_width;
+    int height = *scaled_height;
 
-    for (x_out = 0; x_out < x_out_max; ++x_out) {
-      const int frac = (int)MULT_FIX(frow[x_out], yscale);
-      const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale);
-      dst[x_out] = (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
-      irow[x_out] = frac;   // new fractional start
+    // if width is unspecified, scale original proportionally to height ratio.
+    if (width == 0) {
+      width = (src_width * height + src_height / 2) / src_height;
     }
-    wrk->y_accum += wrk->y_add;
-    wrk->dst += wrk->dst_stride;
-    return dst;
-  } else {
-    return NULL;
+    // if height is unspecified, scale original proportionally to width ratio.
+    if (height == 0) {
+      height = (src_height * width + src_width / 2) / src_width;
+    }
+    // Check if the overall dimensions still make sense.
+    if (width <= 0 || height <= 0) {
+      return 0;
+    }
+
+    *scaled_width = width;
+    *scaled_height = height;
+    return 1;
   }
 }
 
-#undef MULT_FIX
-#undef RFIX
-
 //------------------------------------------------------------------------------
 // all-in-one calls
 
+int WebPRescaleNeededLines(const WebPRescaler* const wrk, int max_num_lines) {
+  const int num_lines = (wrk->y_accum + wrk->y_sub - 1) / wrk->y_sub;
+  return (num_lines > max_num_lines) ? max_num_lines : num_lines;
+}
+
 int WebPRescalerImport(WebPRescaler* const wrk, int num_lines,
                        const uint8_t* src, int src_stride) {
   int total_imported = 0;
-  while (total_imported < num_lines && wrk->y_accum > 0) {
-    int channel;
-    for (channel = 0; channel < wrk->num_channels; ++channel) {
-      WebPRescalerImportRow(wrk, src, channel);
+  while (total_imported < num_lines && !WebPRescalerHasPendingOutput(wrk)) {
+    if (wrk->y_expand) {
+      rescaler_t* const tmp = wrk->irow;
+      wrk->irow = wrk->frow;
+      wrk->frow = tmp;
     }
+    WebPRescalerImportRow(wrk, src);
+    if (!wrk->y_expand) {     // Accumulate the contribution of the new row.
+      int x;
+      for (x = 0; x < wrk->num_channels * wrk->dst_width; ++x) {
+        wrk->irow[x] += wrk->frow[x];
+      }
+    }
+    ++wrk->src_y;
     src += src_stride;
     ++total_imported;
     wrk->y_accum -= wrk->y_sub;
@@ -146,7 +140,3 @@ int WebPRescalerExport(WebPRescaler* const rescaler) {
 }
 
 //------------------------------------------------------------------------------
-
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
diff --git a/drivers/webp/utils/rescaler.h b/drivers/webp/utils/rescaler.h
index 9c9133d19b..98b01a76d0 100644
--- a/drivers/webp/utils/rescaler.h
+++ b/drivers/webp/utils/rescaler.h
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Rescaling functions
@@ -12,64 +14,87 @@
 #ifndef WEBP_UTILS_RESCALER_H_
 #define WEBP_UTILS_RESCALER_H_
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 extern "C" {
 #endif
 
-#include "../types.h"
+#include "../webp/types.h"
+
+#define WEBP_RESCALER_RFIX 32   // fixed-point precision for multiplies
+#define WEBP_RESCALER_ONE (1ull << WEBP_RESCALER_RFIX)
+#define WEBP_RESCALER_FRAC(x, y) \
+    ((uint32_t)(((uint64_t)(x) << WEBP_RESCALER_RFIX) / (y)))
 
 // Structure used for on-the-fly rescaling
-typedef struct {
+typedef uint32_t rescaler_t;   // type for side-buffer
+typedef struct WebPRescaler WebPRescaler;
+struct WebPRescaler {
   int x_expand;               // true if we're expanding in the x direction
+  int y_expand;               // true if we're expanding in the y direction
   int num_channels;           // bytes to jump between pixels
-  int fy_scale, fx_scale;     // fixed-point scaling factor
-  int64_t fxy_scale;          // ''
-  // we need hpel-precise add/sub increments, for the downsampled U/V planes.
+  uint32_t fx_scale;          // fixed-point scaling factors
+  uint32_t fy_scale;          // ''
+  uint32_t fxy_scale;         // ''
   int y_accum;                // vertical accumulator
-  int y_add, y_sub;           // vertical increments (add ~= src, sub ~= dst)
-  int x_add, x_sub;           // horizontal increments (add ~= src, sub ~= dst)
+  int y_add, y_sub;           // vertical increments
+  int x_add, x_sub;           // horizontal increments
   int src_width, src_height;  // source dimensions
   int dst_width, dst_height;  // destination dimensions
+  int src_y, dst_y;           // row counters for input and output
   uint8_t* dst;
   int dst_stride;
-  int32_t* irow, *frow;       // work buffer
-} WebPRescaler;
+  rescaler_t* irow, *frow;    // work buffer
+};
 
 // Initialize a rescaler given scratch area 'work' and dimensions of src & dst.
-void WebPRescalerInit(WebPRescaler* const wrk, int src_width, int src_height,
+void WebPRescalerInit(WebPRescaler* const rescaler,
+                      int src_width, int src_height,
                       uint8_t* const dst,
                       int dst_width, int dst_height, int dst_stride,
                       int num_channels,
-                      int x_add, int x_sub,
-                      int y_add, int y_sub,
-                      int32_t* const work);
+                      rescaler_t* const work);
+
+// If either 'scaled_width' or 'scaled_height' (but not both) is 0 the value
+// will be calculated preserving the aspect ratio, otherwise the values are
+// left unmodified. Returns true on success, false if either value is 0 after
+// performing the scaling calculation.
+int WebPRescalerGetScaledDimensions(int src_width, int src_height,
+                                    int* const scaled_width,
+                                    int* const scaled_height);
 
-// Import a row of data and save its contribution in the rescaler.
-// 'channel' denotes the channel number to be imported.
-void WebPRescalerImportRow(WebPRescaler* const rescaler,
-                           const uint8_t* const src, int channel);
+// Returns the number of input lines needed next to produce one output line,
+// considering that the maximum available input lines are 'max_num_lines'.
+int WebPRescaleNeededLines(const WebPRescaler* const rescaler,
+                           int max_num_lines);
 
 // Import multiple rows over all channels, until at least one row is ready to
 // be exported. Returns the actual number of lines that were imported.
 int WebPRescalerImport(WebPRescaler* const rescaler, int num_rows,
                        const uint8_t* src, int src_stride);
 
-// Return true if there is pending output rows ready.
+// Export as many rows as possible. Return the numbers of rows written.
+int WebPRescalerExport(WebPRescaler* const rescaler);
+
+// Return true if input is finished
 static WEBP_INLINE
-int WebPRescalerHasPendingOutput(const WebPRescaler* const rescaler) {
-  return (rescaler->y_accum <= 0);
+int WebPRescalerInputDone(const WebPRescaler* const rescaler) {
+  return (rescaler->src_y >= rescaler->src_height);
+}
+// Return true if output is finished
+static WEBP_INLINE
+int WebPRescalerOutputDone(const WebPRescaler* const rescaler) {
+  return (rescaler->dst_y >= rescaler->dst_height);
 }
 
-// Export one row from rescaler. Returns the pointer where output was written,
-// or NULL if no row was pending.
-uint8_t* WebPRescalerExportRow(WebPRescaler* const wrk);
-
-// Export as many rows as possible. Return the numbers of rows written.
-int WebPRescalerExport(WebPRescaler* const wrk);
+// Return true if there are pending output rows ready.
+static WEBP_INLINE
+int WebPRescalerHasPendingOutput(const WebPRescaler* const rescaler) {
+  return !WebPRescalerOutputDone(rescaler) && (rescaler->y_accum <= 0);
+}
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/utils/thread.c b/drivers/webp/utils/thread.c
index ce89cf9dc7..93f7622797 100644
--- a/drivers/webp/utils/thread.c
+++ b/drivers/webp/utils/thread.c
@@ -1,27 +1,60 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Multi-threaded worker
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
-#ifdef HAVE_CONFIG_H
-#include "config.h"
-#endif
-
 #include <assert.h>
 #include <string.h>   // for memset()
 #include "./thread.h"
+#include "./utils.h"
+
+#ifdef WEBP_USE_THREAD
+
+#if defined(_WIN32)
+
+#include <windows.h>
+typedef HANDLE pthread_t;
+typedef CRITICAL_SECTION pthread_mutex_t;
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
+#if _WIN32_WINNT >= 0x0600  // Windows Vista / Server 2008 or greater
+#define USE_WINDOWS_CONDITION_VARIABLE
+typedef CONDITION_VARIABLE pthread_cond_t;
+#else
+typedef struct {
+  HANDLE waiting_sem_;
+  HANDLE received_sem_;
+  HANDLE signal_event_;
+} pthread_cond_t;
+#endif  // _WIN32_WINNT >= 0x600
+
+#ifndef WINAPI_FAMILY_PARTITION
+#define WINAPI_PARTITION_DESKTOP 1
+#define WINAPI_FAMILY_PARTITION(x) x
 #endif
 
-#ifdef WEBP_USE_THREAD
+#if !WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
+#define USE_CREATE_THREAD
+#endif
+
+#else  // !_WIN32
+
+#include <pthread.h>
+
+#endif  // _WIN32
+
+struct WebPWorkerImpl {
+  pthread_mutex_t mutex_;
+  pthread_cond_t  condition_;
+  pthread_t       thread_;
+};
 
 #if defined(_WIN32)
 
@@ -34,15 +67,29 @@ extern "C" {
 #define THREADFN unsigned int __stdcall
 #define THREAD_RETURN(val) (unsigned int)((DWORD_PTR)val)
 
+#if _WIN32_WINNT >= 0x0501  // Windows XP or greater
+#define WaitForSingleObject(obj, timeout) \
+  WaitForSingleObjectEx(obj, timeout, FALSE /*bAlertable*/)
+#endif
+
 static int pthread_create(pthread_t* const thread, const void* attr,
                           unsigned int (__stdcall *start)(void*), void* arg) {
   (void)attr;
+#ifdef USE_CREATE_THREAD
+  *thread = CreateThread(NULL,   /* lpThreadAttributes */
+                         0,      /* dwStackSize */
+                         start,
+                         arg,
+                         0,      /* dwStackSize */
+                         NULL);  /* lpThreadId */
+#else
   *thread = (pthread_t)_beginthreadex(NULL,   /* void *security */
                                       0,      /* unsigned stack_size */
                                       start,
                                       arg,
                                       0,      /* unsigned initflag */
                                       NULL);  /* unsigned *thrdaddr */
+#endif
   if (*thread == NULL) return 1;
   SetThreadPriority(*thread, THREAD_PRIORITY_ABOVE_NORMAL);
   return 0;
@@ -57,7 +104,11 @@ static int pthread_join(pthread_t thread, void** value_ptr) {
 // Mutex
 static int pthread_mutex_init(pthread_mutex_t* const mutex, void* mutexattr) {
   (void)mutexattr;
+#if _WIN32_WINNT >= 0x0600  // Windows Vista / Server 2008 or greater
+  InitializeCriticalSectionEx(mutex, 0 /*dwSpinCount*/, 0 /*Flags*/);
+#else
   InitializeCriticalSection(mutex);
+#endif
   return 0;
 }
 
@@ -79,14 +130,21 @@ static int pthread_mutex_destroy(pthread_mutex_t* const mutex) {
 // Condition
 static int pthread_cond_destroy(pthread_cond_t* const condition) {
   int ok = 1;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+  (void)condition;
+#else
   ok &= (CloseHandle(condition->waiting_sem_) != 0);
   ok &= (CloseHandle(condition->received_sem_) != 0);
   ok &= (CloseHandle(condition->signal_event_) != 0);
+#endif
   return !ok;
 }
 
 static int pthread_cond_init(pthread_cond_t* const condition, void* cond_attr) {
   (void)cond_attr;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+  InitializeConditionVariable(condition);
+#else
   condition->waiting_sem_ = CreateSemaphore(NULL, 0, 1, NULL);
   condition->received_sem_ = CreateSemaphore(NULL, 0, 1, NULL);
   condition->signal_event_ = CreateEvent(NULL, FALSE, FALSE, NULL);
@@ -96,11 +154,15 @@ static int pthread_cond_init(pthread_cond_t* const condition, void* cond_attr) {
     pthread_cond_destroy(condition);
     return 1;
   }
+#endif
   return 0;
 }
 
 static int pthread_cond_signal(pthread_cond_t* const condition) {
   int ok = 1;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+  WakeConditionVariable(condition);
+#else
   if (WaitForSingleObject(condition->waiting_sem_, 0) == WAIT_OBJECT_0) {
     // a thread is waiting in pthread_cond_wait: allow it to be notified
     ok = SetEvent(condition->signal_event_);
@@ -109,12 +171,16 @@ static int pthread_cond_signal(pthread_cond_t* const condition) {
     ok &= (WaitForSingleObject(condition->received_sem_, INFINITE) !=
            WAIT_OBJECT_0);
   }
+#endif
   return !ok;
 }
 
 static int pthread_cond_wait(pthread_cond_t* const condition,
                              pthread_mutex_t* const mutex) {
   int ok;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+  ok = SleepConditionVariableCS(condition, mutex, INFINITE);
+#else
   // note that there is a consumer available so the signal isn't dropped in
   // pthread_cond_signal
   if (!ReleaseSemaphore(condition->waiting_sem_, 1, NULL))
@@ -125,123 +191,168 @@ static int pthread_cond_wait(pthread_cond_t* const condition,
         WAIT_OBJECT_0);
   ok &= ReleaseSemaphore(condition->received_sem_, 1, NULL);
   pthread_mutex_lock(mutex);
+#endif
   return !ok;
 }
 
-#else  // _WIN32
+#else  // !_WIN32
 # define THREADFN void*
 # define THREAD_RETURN(val) val
-#endif
+#endif  // _WIN32
 
 //------------------------------------------------------------------------------
 
-static THREADFN WebPWorkerThreadLoop(void *ptr) {    // thread loop
+static void Execute(WebPWorker* const worker);  // Forward declaration.
+
+static THREADFN ThreadLoop(void* ptr) {
   WebPWorker* const worker = (WebPWorker*)ptr;
   int done = 0;
   while (!done) {
-    pthread_mutex_lock(&worker->mutex_);
+    pthread_mutex_lock(&worker->impl_->mutex_);
     while (worker->status_ == OK) {   // wait in idling mode
-      pthread_cond_wait(&worker->condition_, &worker->mutex_);
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
     }
     if (worker->status_ == WORK) {
-      if (worker->hook) {
-        worker->had_error |= !worker->hook(worker->data1, worker->data2);
-      }
+      Execute(worker);
       worker->status_ = OK;
     } else if (worker->status_ == NOT_OK) {   // finish the worker
       done = 1;
     }
     // signal to the main thread that we're done (for Sync())
-    pthread_cond_signal(&worker->condition_);
-    pthread_mutex_unlock(&worker->mutex_);
+    pthread_cond_signal(&worker->impl_->condition_);
+    pthread_mutex_unlock(&worker->impl_->mutex_);
   }
   return THREAD_RETURN(NULL);    // Thread is finished
 }
 
 // main thread state control
-static void WebPWorkerChangeState(WebPWorker* const worker,
-                                  WebPWorkerStatus new_status) {
-  // no-op when attempting to change state on a thread that didn't come up
-  if (worker->status_ < OK) return;
-
-  pthread_mutex_lock(&worker->mutex_);
-  // wait for the worker to finish
-  while (worker->status_ != OK) {
-    pthread_cond_wait(&worker->condition_, &worker->mutex_);
-  }
-  // assign new status and release the working thread if needed
-  if (new_status != OK) {
-    worker->status_ = new_status;
-    pthread_cond_signal(&worker->condition_);
+static void ChangeState(WebPWorker* const worker,
+                        WebPWorkerStatus new_status) {
+  // No-op when attempting to change state on a thread that didn't come up.
+  // Checking status_ without acquiring the lock first would result in a data
+  // race.
+  if (worker->impl_ == NULL) return;
+
+  pthread_mutex_lock(&worker->impl_->mutex_);
+  if (worker->status_ >= OK) {
+    // wait for the worker to finish
+    while (worker->status_ != OK) {
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    // assign new status and release the working thread if needed
+    if (new_status != OK) {
+      worker->status_ = new_status;
+      pthread_cond_signal(&worker->impl_->condition_);
+    }
   }
-  pthread_mutex_unlock(&worker->mutex_);
+  pthread_mutex_unlock(&worker->impl_->mutex_);
 }
 
-#endif
+#endif  // WEBP_USE_THREAD
 
 //------------------------------------------------------------------------------
 
-void WebPWorkerInit(WebPWorker* const worker) {
+static void Init(WebPWorker* const worker) {
   memset(worker, 0, sizeof(*worker));
   worker->status_ = NOT_OK;
 }
 
-int WebPWorkerSync(WebPWorker* const worker) {
+static int Sync(WebPWorker* const worker) {
 #ifdef WEBP_USE_THREAD
-  WebPWorkerChangeState(worker, OK);
+  ChangeState(worker, OK);
 #endif
   assert(worker->status_ <= OK);
   return !worker->had_error;
 }
 
-int WebPWorkerReset(WebPWorker* const worker) {
+static int Reset(WebPWorker* const worker) {
   int ok = 1;
   worker->had_error = 0;
   if (worker->status_ < OK) {
 #ifdef WEBP_USE_THREAD
-    if (pthread_mutex_init(&worker->mutex_, NULL) ||
-        pthread_cond_init(&worker->condition_, NULL)) {
+    worker->impl_ = (WebPWorkerImpl*)WebPSafeCalloc(1, sizeof(*worker->impl_));
+    if (worker->impl_ == NULL) {
       return 0;
     }
-    pthread_mutex_lock(&worker->mutex_);
-    ok = !pthread_create(&worker->thread_, NULL, WebPWorkerThreadLoop, worker);
+    if (pthread_mutex_init(&worker->impl_->mutex_, NULL)) {
+      goto Error;
+    }
+    if (pthread_cond_init(&worker->impl_->condition_, NULL)) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      goto Error;
+    }
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    ok = !pthread_create(&worker->impl_->thread_, NULL, ThreadLoop, worker);
     if (ok) worker->status_ = OK;
-    pthread_mutex_unlock(&worker->mutex_);
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+    if (!ok) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      pthread_cond_destroy(&worker->impl_->condition_);
+ Error:
+      WebPSafeFree(worker->impl_);
+      worker->impl_ = NULL;
+      return 0;
+    }
 #else
     worker->status_ = OK;
 #endif
   } else if (worker->status_ > OK) {
-    ok = WebPWorkerSync(worker);
+    ok = Sync(worker);
   }
   assert(!ok || (worker->status_ == OK));
   return ok;
 }
 
-void WebPWorkerLaunch(WebPWorker* const worker) {
+static void Execute(WebPWorker* const worker) {
+  if (worker->hook != NULL) {
+    worker->had_error |= !worker->hook(worker->data1, worker->data2);
+  }
+}
+
+static void Launch(WebPWorker* const worker) {
 #ifdef WEBP_USE_THREAD
-  WebPWorkerChangeState(worker, WORK);
+  ChangeState(worker, WORK);
 #else
-  if (worker->hook)
-    worker->had_error |= !worker->hook(worker->data1, worker->data2);
+  Execute(worker);
 #endif
 }
 
-void WebPWorkerEnd(WebPWorker* const worker) {
-  if (worker->status_ >= OK) {
+static void End(WebPWorker* const worker) {
 #ifdef WEBP_USE_THREAD
-    WebPWorkerChangeState(worker, NOT_OK);
-    pthread_join(worker->thread_, NULL);
-    pthread_mutex_destroy(&worker->mutex_);
-    pthread_cond_destroy(&worker->condition_);
+  if (worker->impl_ != NULL) {
+    ChangeState(worker, NOT_OK);
+    pthread_join(worker->impl_->thread_, NULL);
+    pthread_mutex_destroy(&worker->impl_->mutex_);
+    pthread_cond_destroy(&worker->impl_->condition_);
+    WebPSafeFree(worker->impl_);
+    worker->impl_ = NULL;
+  }
 #else
-    worker->status_ = NOT_OK;
+  worker->status_ = NOT_OK;
+  assert(worker->impl_ == NULL);
 #endif
-  }
   assert(worker->status_ == NOT_OK);
 }
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
+static WebPWorkerInterface g_worker_interface = {
+  Init, Reset, Sync, Launch, Execute, End
+};
+
+int WebPSetWorkerInterface(const WebPWorkerInterface* const winterface) {
+  if (winterface == NULL ||
+      winterface->Init == NULL || winterface->Reset == NULL ||
+      winterface->Sync == NULL || winterface->Launch == NULL ||
+      winterface->Execute == NULL || winterface->End == NULL) {
+    return 0;
+  }
+  g_worker_interface = *winterface;
+  return 1;
+}
+
+const WebPWorkerInterface* WebPGetWorkerInterface(void) {
+  return &g_worker_interface;
+}
+
+//------------------------------------------------------------------------------
diff --git a/drivers/webp/utils/thread.h b/drivers/webp/utils/thread.h
index 3191890b76..8408311855 100644
--- a/drivers/webp/utils/thread.h
+++ b/drivers/webp/utils/thread.h
@@ -1,8 +1,10 @@
 // Copyright 2011 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Multi-threaded worker
@@ -12,29 +14,15 @@
 #ifndef WEBP_UTILS_THREAD_H_
 #define WEBP_UTILS_THREAD_H_
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
+#ifdef HAVE_CONFIG_H
+#include "../webp/config.h"
 #endif
 
-#if WEBP_USE_THREAD
-
-#if defined(_WIN32)
-
-#include <windows.h>
-typedef HANDLE pthread_t;
-typedef CRITICAL_SECTION pthread_mutex_t;
-typedef struct {
-  HANDLE waiting_sem_;
-  HANDLE received_sem_;
-  HANDLE signal_event_;
-} pthread_cond_t;
+#include "../webp/types.h"
 
-#else
-
-#include <pthread.h>
-
-#endif    /* _WIN32 */
-#endif    /* WEBP_USE_THREAD */
+#ifdef __cplusplus
+extern "C" {
+#endif
 
 // State of the worker thread object
 typedef enum {
@@ -47,13 +35,12 @@ typedef enum {
 // arguments (data1 and data2), and should return false in case of error.
 typedef int (*WebPWorkerHook)(void*, void*);
 
-// Synchronize object used to launch job in the worker thread
+// Platform-dependent implementation details for the worker.
+typedef struct WebPWorkerImpl WebPWorkerImpl;
+
+// Synchronization object used to launch job in the worker thread
 typedef struct {
-#if WEBP_USE_THREAD
-  pthread_mutex_t mutex_;
-  pthread_cond_t  condition_;
-  pthread_t       thread_;
-#endif
+  WebPWorkerImpl* impl_;
   WebPWorkerStatus status_;
   WebPWorkerHook hook;    // hook to call
   void* data1;            // first argument passed to 'hook'
@@ -61,25 +48,45 @@ typedef struct {
   int had_error;          // return value of the last call to 'hook'
 } WebPWorker;
 
-// Must be called first, before any other method.
-void WebPWorkerInit(WebPWorker* const worker);
-// Must be called initialize the object and spawn the thread. Re-entrant.
-// Will potentially launch the thread. Returns false in case of error.
-int WebPWorkerReset(WebPWorker* const worker);
-// Make sure the previous work is finished. Returns true if worker->had_error
-// was not set and not error condition was triggered by the working thread.
-int WebPWorkerSync(WebPWorker* const worker);
-// Trigger the thread to call hook() with data1 and data2 argument. These
-// hook/data1/data2 can be changed at any time before calling this function,
-// but not be changed afterward until the next call to WebPWorkerSync().
-void WebPWorkerLaunch(WebPWorker* const worker);
-// Kill the thread and terminate the object. To use the object again, one
-// must call WebPWorkerReset() again.
-void WebPWorkerEnd(WebPWorker* const worker);
+// The interface for all thread-worker related functions. All these functions
+// must be implemented.
+typedef struct {
+  // Must be called first, before any other method.
+  void (*Init)(WebPWorker* const worker);
+  // Must be called to initialize the object and spawn the thread. Re-entrant.
+  // Will potentially launch the thread. Returns false in case of error.
+  int (*Reset)(WebPWorker* const worker);
+  // Makes sure the previous work is finished. Returns true if worker->had_error
+  // was not set and no error condition was triggered by the working thread.
+  int (*Sync)(WebPWorker* const worker);
+  // Triggers the thread to call hook() with data1 and data2 arguments. These
+  // hook/data1/data2 values can be changed at any time before calling this
+  // function, but not be changed afterward until the next call to Sync().
+  void (*Launch)(WebPWorker* const worker);
+  // This function is similar to Launch() except that it calls the
+  // hook directly instead of using a thread. Convenient to bypass the thread
+  // mechanism while still using the WebPWorker structs. Sync() must
+  // still be called afterward (for error reporting).
+  void (*Execute)(WebPWorker* const worker);
+  // Kill the thread and terminate the object. To use the object again, one
+  // must call Reset() again.
+  void (*End)(WebPWorker* const worker);
+} WebPWorkerInterface;
+
+// Install a new set of threading functions, overriding the defaults. This
+// should be done before any workers are started, i.e., before any encoding or
+// decoding takes place. The contents of the interface struct are copied, it
+// is safe to free the corresponding memory after this call. This function is
+// not thread-safe. Return false in case of invalid pointer or methods.
+WEBP_EXTERN(int) WebPSetWorkerInterface(
+    const WebPWorkerInterface* const winterface);
+
+// Retrieve the currently set thread worker interface.
+WEBP_EXTERN(const WebPWorkerInterface*) WebPGetWorkerInterface(void);
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif
 
diff --git a/drivers/webp/utils/utils.c b/drivers/webp/utils/utils.c
index 673b7e284c..d8e30930af 100644
--- a/drivers/webp/utils/utils.c
+++ b/drivers/webp/utils/utils.c
@@ -1,8 +1,10 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Misc. common utility functions
@@ -10,35 +12,228 @@
 // Author: Skal (pascal.massimino@gmail.com)
 
 #include <stdlib.h>
+#include <string.h>  // for memcpy()
+#include "../webp/decode.h"
+#include "../webp/encode.h"
 #include "./utils.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
+// If PRINT_MEM_INFO is defined, extra info (like total memory used, number of
+// alloc/free etc) is printed. For debugging/tuning purpose only (it's slow,
+// and not multi-thread safe!).
+// An interesting alternative is valgrind's 'massif' tool:
+//    http://valgrind.org/docs/manual/ms-manual.html
+// Here is an example command line:
+/*    valgrind --tool=massif --massif-out-file=massif.out \
+               --stacks=yes --alloc-fn=WebPSafeAlloc --alloc-fn=WebPSafeCalloc
+      ms_print massif.out
+*/
+// In addition:
+// * if PRINT_MEM_TRAFFIC is defined, all the details of the malloc/free cycles
+//   are printed.
+// * if MALLOC_FAIL_AT is defined, the global environment variable
+//   $MALLOC_FAIL_AT is used to simulate a memory error when calloc or malloc
+//   is called for the nth time. Example usage:
+//   export MALLOC_FAIL_AT=50 && ./examples/cwebp input.png
+// * if MALLOC_LIMIT is defined, the global environment variable $MALLOC_LIMIT
+//   sets the maximum amount of memory (in bytes) made available to libwebp.
+//   This can be used to emulate environment with very limited memory.
+//   Example: export MALLOC_LIMIT=64000000 && ./examples/dwebp picture.webp
+
+// #define PRINT_MEM_INFO
+// #define PRINT_MEM_TRAFFIC
+// #define MALLOC_FAIL_AT
+// #define MALLOC_LIMIT
 
 //------------------------------------------------------------------------------
 // Checked memory allocation
 
-static int CheckSizeArguments(uint64_t nmemb, size_t size) {
+#if defined(PRINT_MEM_INFO)
+
+#include <stdio.h>
+
+static int num_malloc_calls = 0;
+static int num_calloc_calls = 0;
+static int num_free_calls = 0;
+static int countdown_to_fail = 0;     // 0 = off
+
+typedef struct MemBlock MemBlock;
+struct MemBlock {
+  void* ptr_;
+  size_t size_;
+  MemBlock* next_;
+};
+
+static MemBlock* all_blocks = NULL;
+static size_t total_mem = 0;
+static size_t total_mem_allocated = 0;
+static size_t high_water_mark = 0;
+static size_t mem_limit = 0;
+
+static int exit_registered = 0;
+
+static void PrintMemInfo(void) {
+  fprintf(stderr, "\nMEMORY INFO:\n");
+  fprintf(stderr, "num calls to: malloc = %4d\n", num_malloc_calls);
+  fprintf(stderr, "              calloc = %4d\n", num_calloc_calls);
+  fprintf(stderr, "              free   = %4d\n", num_free_calls);
+  fprintf(stderr, "total_mem: %u\n", (uint32_t)total_mem);
+  fprintf(stderr, "total_mem allocated: %u\n", (uint32_t)total_mem_allocated);
+  fprintf(stderr, "high-water mark: %u\n", (uint32_t)high_water_mark);
+  while (all_blocks != NULL) {
+    MemBlock* b = all_blocks;
+    all_blocks = b->next_;
+    free(b);
+  }
+}
+
+static void Increment(int* const v) {
+  if (!exit_registered) {
+#if defined(MALLOC_FAIL_AT)
+    {
+      const char* const malloc_fail_at_str = getenv("MALLOC_FAIL_AT");
+      if (malloc_fail_at_str != NULL) {
+        countdown_to_fail = atoi(malloc_fail_at_str);
+      }
+    }
+#endif
+#if defined(MALLOC_LIMIT)
+    {
+      const char* const malloc_limit_str = getenv("MALLOC_LIMIT");
+      if (malloc_limit_str != NULL) {
+        mem_limit = atoi(malloc_limit_str);
+      }
+    }
+#endif
+    (void)countdown_to_fail;
+    (void)mem_limit;
+    atexit(PrintMemInfo);
+    exit_registered = 1;
+  }
+  ++*v;
+}
+
+static void AddMem(void* ptr, size_t size) {
+  if (ptr != NULL) {
+    MemBlock* const b = (MemBlock*)malloc(sizeof(*b));
+    if (b == NULL) abort();
+    b->next_ = all_blocks;
+    all_blocks = b;
+    b->ptr_ = ptr;
+    b->size_ = size;
+    total_mem += size;
+    total_mem_allocated += size;
+#if defined(PRINT_MEM_TRAFFIC)
+#if defined(MALLOC_FAIL_AT)
+    fprintf(stderr, "fail-count: %5d [mem=%u]\n",
+            num_malloc_calls + num_calloc_calls, (uint32_t)total_mem);
+#else
+    fprintf(stderr, "Mem: %u (+%u)\n", (uint32_t)total_mem, (uint32_t)size);
+#endif
+#endif
+    if (total_mem > high_water_mark) high_water_mark = total_mem;
+  }
+}
+
+static void SubMem(void* ptr) {
+  if (ptr != NULL) {
+    MemBlock** b = &all_blocks;
+    // Inefficient search, but that's just for debugging.
+    while (*b != NULL && (*b)->ptr_ != ptr) b = &(*b)->next_;
+    if (*b == NULL) {
+      fprintf(stderr, "Invalid pointer free! (%p)\n", ptr);
+      abort();
+    }
+    {
+      MemBlock* const block = *b;
+      *b = block->next_;
+      total_mem -= block->size_;
+#if defined(PRINT_MEM_TRAFFIC)
+      fprintf(stderr, "Mem: %u (-%u)\n",
+              (uint32_t)total_mem, (uint32_t)block->size_);
+#endif
+      free(block);
+    }
+  }
+}
+
+#else
+#define Increment(v) do {} while (0)
+#define AddMem(p, s) do {} while (0)
+#define SubMem(p)    do {} while (0)
+#endif
+
+// Returns 0 in case of overflow of nmemb * size.
+static int CheckSizeArgumentsOverflow(uint64_t nmemb, size_t size) {
   const uint64_t total_size = nmemb * size;
   if (nmemb == 0) return 1;
   if ((uint64_t)size > WEBP_MAX_ALLOCABLE_MEMORY / nmemb) return 0;
   if (total_size != (size_t)total_size) return 0;
+#if defined(PRINT_MEM_INFO) && defined(MALLOC_FAIL_AT)
+  if (countdown_to_fail > 0 && --countdown_to_fail == 0) {
+    return 0;    // fake fail!
+  }
+#endif
+#if defined(MALLOC_LIMIT)
+  if (mem_limit > 0 && total_mem + total_size >= mem_limit) {
+    return 0;   // fake fail!
+  }
+#endif
+
   return 1;
 }
 
 void* WebPSafeMalloc(uint64_t nmemb, size_t size) {
-  if (!CheckSizeArguments(nmemb, size)) return NULL;
-  return malloc((size_t)(nmemb * size));
+  void* ptr;
+  Increment(&num_malloc_calls);
+  if (!CheckSizeArgumentsOverflow(nmemb, size)) return NULL;
+  assert(nmemb * size > 0);
+  ptr = malloc((size_t)(nmemb * size));
+  AddMem(ptr, (size_t)(nmemb * size));
+  return ptr;
 }
 
 void* WebPSafeCalloc(uint64_t nmemb, size_t size) {
-  if (!CheckSizeArguments(nmemb, size)) return NULL;
-  return calloc((size_t)nmemb, size);
+  void* ptr;
+  Increment(&num_calloc_calls);
+  if (!CheckSizeArgumentsOverflow(nmemb, size)) return NULL;
+  assert(nmemb * size > 0);
+  ptr = calloc((size_t)nmemb, size);
+  AddMem(ptr, (size_t)(nmemb * size));
+  return ptr;
+}
+
+void WebPSafeFree(void* const ptr) {
+  if (ptr != NULL) {
+    Increment(&num_free_calls);
+    SubMem(ptr);
+  }
+  free(ptr);
+}
+
+// Public API function.
+void WebPFree(void* ptr) {
+  free(ptr);
 }
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
-}    // extern "C"
-#endif
+void WebPCopyPlane(const uint8_t* src, int src_stride,
+                   uint8_t* dst, int dst_stride, int width, int height) {
+  assert(src != NULL && dst != NULL);
+  assert(src_stride >= width && dst_stride >= width);
+  while (height-- > 0) {
+    memcpy(dst, src, width);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void WebPCopyPixels(const WebPPicture* const src, WebPPicture* const dst) {
+  assert(src != NULL && dst != NULL);
+  assert(src->width == dst->width && src->height == dst->height);
+  assert(src->use_argb && dst->use_argb);
+  WebPCopyPlane((uint8_t*)src->argb, 4 * src->argb_stride, (uint8_t*)dst->argb,
+                4 * dst->argb_stride, 4 * src->width, src->height);
+}
+
+//------------------------------------------------------------------------------
diff --git a/drivers/webp/utils/utils.h b/drivers/webp/utils/utils.h
index 316ac90612..fcdb7e139b 100644
--- a/drivers/webp/utils/utils.h
+++ b/drivers/webp/utils/utils.h
@@ -1,20 +1,25 @@
 // Copyright 2012 Google Inc. All Rights Reserved.
 //
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// 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.
 // -----------------------------------------------------------------------------
 //
 // Misc. common utility functions
 //
-// Author: Skal (pascal.massimino@gmail.com)
+// Authors: Skal (pascal.massimino@gmail.com)
+//          Urvang (urvang@google.com)
 
 #ifndef WEBP_UTILS_UTILS_H_
 #define WEBP_UTILS_UTILS_H_
 
-#include "../types.h"
+#include <assert.h>
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#include "../webp/types.h"
+
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -30,14 +35,107 @@ extern "C" {
 // somewhere (like: malloc(num_pixels * sizeof(*something))). That's why this
 // safe malloc() borrows the signature from calloc(), pointing at the dangerous
 // underlying multiply involved.
-void* WebPSafeMalloc(uint64_t nmemb, size_t size);
+WEBP_EXTERN(void*) WebPSafeMalloc(uint64_t nmemb, size_t size);
 // Note that WebPSafeCalloc() expects the second argument type to be 'size_t'
 // in order to favor the "calloc(num_foo, sizeof(foo))" pattern.
-void* WebPSafeCalloc(uint64_t nmemb, size_t size);
+WEBP_EXTERN(void*) WebPSafeCalloc(uint64_t nmemb, size_t size);
+
+// Companion deallocation function to the above allocations.
+WEBP_EXTERN(void) WebPSafeFree(void* const ptr);
+
+//------------------------------------------------------------------------------
+// Alignment
+
+#define WEBP_ALIGN_CST 31
+#define WEBP_ALIGN(PTR) ((uintptr_t)((PTR) + WEBP_ALIGN_CST) & ~WEBP_ALIGN_CST)
+
+//------------------------------------------------------------------------------
+// Reading/writing data.
+
+// Read 16, 24 or 32 bits stored in little-endian order.
+static WEBP_INLINE int GetLE16(const uint8_t* const data) {
+  return (int)(data[0] << 0) | (data[1] << 8);
+}
+
+static WEBP_INLINE int GetLE24(const uint8_t* const data) {
+  return GetLE16(data) | (data[2] << 16);
+}
+
+static WEBP_INLINE uint32_t GetLE32(const uint8_t* const data) {
+  return GetLE16(data) | ((uint32_t)GetLE16(data + 2) << 16);
+}
+
+// Store 16, 24 or 32 bits in little-endian order.
+static WEBP_INLINE void PutLE16(uint8_t* const data, int val) {
+  assert(val < (1 << 16));
+  data[0] = (val >> 0);
+  data[1] = (val >> 8);
+}
+
+static WEBP_INLINE void PutLE24(uint8_t* const data, int val) {
+  assert(val < (1 << 24));
+  PutLE16(data, val & 0xffff);
+  data[2] = (val >> 16);
+}
+
+static WEBP_INLINE void PutLE32(uint8_t* const data, uint32_t val) {
+  PutLE16(data, (int)(val & 0xffff));
+  PutLE16(data + 2, (int)(val >> 16));
+}
+
+// Returns (int)floor(log2(n)). n must be > 0.
+// use GNU builtins where available.
+#if defined(__GNUC__) && \
+    ((__GNUC__ == 3 && __GNUC_MINOR__ >= 4) || __GNUC__ >= 4)
+static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
+  return 31 ^ __builtin_clz(n);
+}
+#elif defined(_MSC_VER) && _MSC_VER > 1310 && \
+      (defined(_M_X64) || defined(_M_IX86))
+#include <intrin.h>
+#pragma intrinsic(_BitScanReverse)
+
+static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
+  unsigned long first_set_bit;
+  _BitScanReverse(&first_set_bit, n);
+  return first_set_bit;
+}
+#else
+static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
+  int log = 0;
+  uint32_t value = n;
+  int i;
+
+  for (i = 4; i >= 0; --i) {
+    const int shift = (1 << i);
+    const uint32_t x = value >> shift;
+    if (x != 0) {
+      value = x;
+      log += shift;
+    }
+  }
+  return log;
+}
+#endif
+
+//------------------------------------------------------------------------------
+// Pixel copying.
+
+struct WebPPicture;
+
+// Copy width x height pixels from 'src' to 'dst' honoring the strides.
+WEBP_EXTERN(void) WebPCopyPlane(const uint8_t* src, int src_stride,
+                                uint8_t* dst, int dst_stride,
+                                int width, int height);
+
+// Copy ARGB pixels from 'src' to 'dst' honoring strides. 'src' and 'dst' are
+// assumed to be already allocated and using ARGB data.
+WEBP_EXTERN(void) WebPCopyPixels(const struct WebPPicture* const src,
+                                 struct WebPPicture* const dst);
 
 //------------------------------------------------------------------------------
 
-#if defined(__cplusplus) || defined(c_plusplus)
+#ifdef __cplusplus
 }    // extern "C"
 #endif