GODOT IS OPEN SOURCE

94 files changed, 34751 insertions, 0 deletions

diff --git a/drivers/webp/SCsub b/drivers/webp/SCsub
new file mode 100644
index 0000000000..6d9707677f
--- /dev/null
+++ b/drivers/webp/SCsub
@@ -0,0 +1,71 @@
+Import('env')
+
+
+webp_sources = [
+	"webp/mux/muxread.c",\
+	"webp/mux/muxedit.c",\
+	"webp/demux/demux.c",\
+	"webp/mux/muxinternal.c",\
+	"webp/enc/alpha.c",\
+	"webp/enc/tree.c",\
+	"webp/enc/webpenc.c",\
+	"webp/enc/cost.c",\
+	"webp/enc/quant.c",\
+	"webp/enc/token.c",\
+	"webp/enc/iterator.c",\
+	"webp/enc/syntax.c",\
+	"webp/enc/frame.c",\
+	"webp/enc/picture.c",\
+	"webp/enc/analysis.c",\
+	"webp/enc/backward_references.c",\
+	"webp/enc/filter.c",\
+	"webp/enc/vp8l.c",\
+	"webp/enc/layer.c",\
+	"webp/enc/histogram.c",\
+	"webp/enc/config.c",\
+	"webp/dec/alpha.c",\
+	"webp/dec/vp8.c",\
+	"webp/dec/tree.c",\
+	"webp/dec/webp.c",\
+	"webp/dec/idec.c",\
+	"webp/dec/quant.c",\
+	"webp/dec/frame.c",\
+	"webp/dec/buffer.c",\
+	"webp/dec/io.c",\
+	"webp/dec/vp8l.c",\
+	"webp/dec/layer.c",
+	"webp/dsp/dec_neon.c",
+	"webp/dsp/upsampling_sse2.c",\
+	"webp/dsp/dec_sse2.c",\
+	"webp/dsp/enc_neon.c",\
+	"webp/dsp/dec.c",\
+	"webp/dsp/upsampling.c",\
+	"webp/dsp/enc_sse2.c",\
+	"webp/dsp/enc.c",\
+	"webp/dsp/cpu.c",\
+	"webp/dsp/lossless.c",\
+	"webp/dsp/upsampling_neon.c",\
+	"webp/dsp/yuv.c",\
+	"webp/utils/bit_reader.c",\
+	"webp/utils/thread.c",\
+	"webp/utils/alpha_processing.c",\
+	"webp/utils/random.c",\
+	"webp/utils/quant_levels.c",\
+	"webp/utils/huffman.c",\
+	"webp/utils/filters.c",\
+	"webp/utils/rescaler.c",\
+	"webp/utils/quant_levels_dec.c",\
+	"webp/utils/color_cache.c",\
+	"webp/utils/utils.c",\
+	"webp/utils/huffman_encode.c",\
+	"webp/utils/bit_writer.c",\
+	"webp/image_loader_webp.cpp",\
+]
+
+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
new file mode 100644
index 0000000000..93729a035f
--- /dev/null
+++ b/drivers/webp/dec/alpha.c
@@ -0,0 +1,161 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Alpha-plane decompression.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+#include "./alphai.h"
+#include "./vp8i.h"
+#include "./vp8li.h"
+#include "../utils/quant_levels_dec.h"
+#include "../webp/format_constants.h"
+
+//------------------------------------------------------------------------------
+// ALPHDecoder object.
+
+ALPHDecoder* ALPHNew(void) {
+  ALPHDecoder* const dec = (ALPHDecoder*)calloc(1, sizeof(*dec));
+  return dec;
+}
+
+void ALPHDelete(ALPHDecoder* const dec) {
+  if (dec != NULL) {
+    VP8LDelete(dec->vp8l_dec_);
+    dec->vp8l_dec_ = NULL;
+    free(dec);
+  }
+}
+
+//------------------------------------------------------------------------------
+// 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;
+  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);
+  assert(data != NULL && output != NULL);
+
+  dec->width_ = width;
+  dec->height_ = height;
+
+  if (data_size <= ALPHA_HEADER_LEN) {
+    return 0;
+  }
+
+  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 (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 (dec->method_ == ALPHA_NO_COMPRESSION) {
+    const size_t alpha_decoded_size = dec->width_ * dec->height_;
+    ok = (alpha_data_size >= alpha_decoded_size);
+  } else {
+    assert(dec->method_ == ALPHA_LOSSLESS_COMPRESSION);
+    ok = VP8LDecodeAlphaHeader(dec, alpha_data, alpha_data_size, output);
+  }
+  return ok;
+}
+
+// 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;
+    }
+  }
+
+  if (unfilter_func != NULL) {
+    unfilter_func(width, height, width, row, num_rows, output);
+  }
+
+  if (alph_dec->pre_processing_ == ALPHA_PREPROCESSED_LEVELS) {
+    if (!DequantizeLevels(output, width, height, row, num_rows)) {
+      return 0;
+    }
+  }
+
+  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 width = dec->pic_hdr_.width_;
+  const int height = dec->pic_hdr_.height_;
+
+  if (row < 0 || num_rows <= 0 || row + num_rows > height) {
+    return NULL;    // sanity check.
+  }
+
+  if (row == 0) {
+    // 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 (!dec->is_alpha_decoded_) {
+    int ok = 0;
+    assert(dec->alph_dec_ != NULL);
+    ok = ALPHDecode(dec, row, num_rows);
+    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 * width;
+}
+
diff --git a/drivers/webp/dec/alphai.h b/drivers/webp/dec/alphai.h
new file mode 100644
index 0000000000..5fa230ca82
--- /dev/null
+++ b/drivers/webp/dec/alphai.h
@@ -0,0 +1,55 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Alpha decoder: internal header.
+//
+// Author: Urvang (urvang@google.com)
+
+#ifndef WEBP_DEC_ALPHAI_H_
+#define WEBP_DEC_ALPHAI_H_
+
+#include "./webpi.h"
+#include "../utils/filters.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP8LDecoder;  // Defined in dec/vp8li.h.
+
+typedef struct ALPHDecoder ALPHDecoder;
+struct ALPHDecoder {
+  int width_;
+  int height_;
+  int method_;
+  WEBP_FILTER_TYPE filter_;
+  int pre_processing_;
+  struct VP8LDecoder* vp8l_dec_;
+  VP8Io io_;
+  int use_8b_decode;  // Although alpha channel requires only 1 byte per
+                      // pixel, sometimes VP8LDecoder may need to allocate
+                      // 4 bytes per pixel internally during decode.
+};
+
+//------------------------------------------------------------------------------
+// internal functions. Not public.
+
+// Allocates a new alpha decoder instance.
+ALPHDecoder* ALPHNew(void);
+
+// Clears and deallocates an alpha decoder instance.
+void ALPHDelete(ALPHDecoder* const dec);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_DEC_ALPHAI_H_ */
diff --git a/drivers/webp/dec/buffer.c b/drivers/webp/dec/buffer.c
new file mode 100644
index 0000000000..1e852efe74
--- /dev/null
+++ b/drivers/webp/dec/buffer.c
@@ -0,0 +1,210 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Everything about WebPDecBuffer
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+
+#include "./vp8i.h"
+#include "./webpi.h"
+#include "../utils/utils.h"
+
+//------------------------------------------------------------------------------
+// WebPDecBuffer
+
+// Number of bytes per pixel for the different color-spaces.
+static const int kModeBpp[MODE_LAST] = {
+  3, 4, 3, 4, 4, 2, 2,
+  4, 4, 4, 2,    // pre-multiplied modes
+  1, 1 };
+
+// Check that webp_csp_mode is within the bounds of WEBP_CSP_MODE.
+// Convert to an integer to handle both the unsigned/signed enum cases
+// without the need for casting to remove type limit warnings.
+static int IsValidColorspace(int webp_csp_mode) {
+  return (webp_csp_mode >= MODE_RGB && webp_csp_mode < MODE_LAST);
+}
+
+static VP8StatusCode CheckDecBuffer(const WebPDecBuffer* const buffer) {
+  int ok = 1;
+  const WEBP_CSP_MODE mode = buffer->colorspace;
+  const int width = buffer->width;
+  const int height = buffer->height;
+  if (!IsValidColorspace(mode)) {
+    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;
+    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 &= (buf->y != NULL);
+    ok &= (buf->u != NULL);
+    ok &= (buf->v != NULL);
+    if (mode == MODE_YUVA) {
+      ok &= (buf->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;
+    ok &= (size <= buf->size);
+    ok &= (buf->stride >= width * kModeBpp[mode]);
+    ok &= (buf->rgba != NULL);
+  }
+  return ok ? VP8_STATUS_OK : VP8_STATUS_INVALID_PARAM;
+}
+
+static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
+  const int w = buffer->width;
+  const int h = buffer->height;
+  const WEBP_CSP_MODE mode = buffer->colorspace;
+
+  if (w <= 0 || h <= 0 || !IsValidColorspace(mode)) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+
+  if (!buffer->is_external_memory && buffer->private_memory == NULL) {
+    uint8_t* output;
+    int uv_stride = 0, a_stride = 0;
+    uint64_t uv_size = 0, a_size = 0, total_size;
+    // We need memory and it hasn't been allocated yet.
+    // => initialize output buffer, now that dimensions are known.
+    const int stride = w * kModeBpp[mode];
+    const uint64_t size = (uint64_t)stride * h;
+
+    if (!WebPIsRGBMode(mode)) {
+      uv_stride = (w + 1) / 2;
+      uv_size = (uint64_t)uv_stride * ((h + 1) / 2);
+      if (mode == MODE_YUVA) {
+        a_stride = w;
+        a_size = (uint64_t)a_stride * h;
+      }
+    }
+    total_size = size + 2 * uv_size + a_size;
+
+    // Security/sanity checks
+    output = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*output));
+    if (output == NULL) {
+      return VP8_STATUS_OUT_OF_MEMORY;
+    }
+    buffer->private_memory = output;
+
+    if (!WebPIsRGBMode(mode)) {   // YUVA initialization
+      WebPYUVABuffer* const buf = &buffer->u.YUVA;
+      buf->y = output;
+      buf->y_stride = stride;
+      buf->y_size = (size_t)size;
+      buf->u = output + size;
+      buf->u_stride = uv_stride;
+      buf->u_size = (size_t)uv_size;
+      buf->v = output + size + uv_size;
+      buf->v_stride = uv_stride;
+      buf->v_size = (size_t)uv_size;
+      if (mode == MODE_YUVA) {
+        buf->a = output + size + 2 * uv_size;
+      }
+      buf->a_size = (size_t)a_size;
+      buf->a_stride = a_stride;
+    } else {  // RGBA initialization
+      WebPRGBABuffer* const buf = &buffer->u.RGBA;
+      buf->rgba = output;
+      buf->stride = stride;
+      buf->size = (size_t)size;
+    }
+  }
+  return CheckDecBuffer(buffer);
+}
+
+VP8StatusCode WebPAllocateDecBuffer(int w, int h,
+                                    const WebPDecoderOptions* const options,
+                                    WebPDecBuffer* const out) {
+  if (out == NULL || w <= 0 || h <= 0) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  if (options != NULL) {    // First, apply options if there is any.
+    if (options->use_cropping) {
+      const int cw = options->crop_width;
+      const int ch = options->crop_height;
+      const int x = options->crop_left & ~1;
+      const int y = options->crop_top & ~1;
+      if (x < 0 || y < 0 || cw <= 0 || ch <= 0 || x + cw > w || y + ch > h) {
+        return VP8_STATUS_INVALID_PARAM;   // out of frame boundary.
+      }
+      w = cw;
+      h = ch;
+    }
+    if (options->use_scaling) {
+      if (options->scaled_width <= 0 || options->scaled_height <= 0) {
+        return VP8_STATUS_INVALID_PARAM;
+      }
+      w = options->scaled_width;
+      h = options->scaled_height;
+    }
+  }
+  out->width = w;
+  out->height = h;
+
+  // Then, allocate buffer for real
+  return AllocateBuffer(out);
+}
+
+//------------------------------------------------------------------------------
+// constructors / destructors
+
+int WebPInitDecBufferInternal(WebPDecBuffer* buffer, int version) {
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_DECODER_ABI_VERSION)) {
+    return 0;  // version mismatch
+  }
+  if (buffer == NULL) return 0;
+  memset(buffer, 0, sizeof(*buffer));
+  return 1;
+}
+
+void WebPFreeDecBuffer(WebPDecBuffer* buffer) {
+  if (buffer != NULL) {
+    if (!buffer->is_external_memory)
+      free(buffer->private_memory);
+    buffer->private_memory = NULL;
+  }
+}
+
+void WebPCopyDecBuffer(const WebPDecBuffer* const src,
+                       WebPDecBuffer* const dst) {
+  if (src != NULL && dst != NULL) {
+    *dst = *src;
+    if (src->private_memory != NULL) {
+      dst->is_external_memory = 1;   // dst buffer doesn't own the memory.
+      dst->private_memory = NULL;
+    }
+  }
+}
+
+// Copy and transfer ownership from src to dst (beware of parameter order!)
+void WebPGrabDecBuffer(WebPDecBuffer* const src, WebPDecBuffer* const dst) {
+  if (src != NULL && dst != NULL) {
+    *dst = *src;
+    if (src->private_memory != NULL) {
+      src->is_external_memory = 1;   // src relinquishes ownership
+      src->private_memory = NULL;
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/dec/decode_vp8.h b/drivers/webp/dec/decode_vp8.h
new file mode 100644
index 0000000000..b9337bbec0
--- /dev/null
+++ b/drivers/webp/dec/decode_vp8.h
@@ -0,0 +1,185 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//  Low-level API for VP8 decoder
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_WEBP_DECODE_VP8_H_
+#define WEBP_WEBP_DECODE_VP8_H_
+
+#include "../webp/decode.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// Lower-level API
+//
+// These functions provide fine-grained control of the decoding process.
+// The call flow should resemble:
+//
+//   VP8Io io;
+//   VP8InitIo(&io);
+//   io.data = data;
+//   io.data_size = size;
+//   /* customize io's functions (setup()/put()/teardown()) if needed. */
+//
+//   VP8Decoder* dec = VP8New();
+//   bool ok = VP8Decode(dec);
+//   if (!ok) printf("Error: %s\n", VP8StatusMessage(dec));
+//   VP8Delete(dec);
+//   return ok;
+
+// Input / Output
+typedef struct VP8Io VP8Io;
+typedef int (*VP8IoPutHook)(const VP8Io* io);
+typedef int (*VP8IoSetupHook)(VP8Io* io);
+typedef void (*VP8IoTeardownHook)(const VP8Io* io);
+
+struct VP8Io {
+  // set by VP8GetHeaders()
+  int width, height;         // picture dimensions, in pixels (invariable).
+                             // These are the original, uncropped dimensions.
+                             // The actual area passed to put() is stored
+                             // in mb_w / mb_h fields.
+
+  // set before calling put()
+  int mb_y;                  // position of the current rows (in pixels)
+  int mb_w;                  // number of columns in the sample
+  int mb_h;                  // number of rows in the sample
+  const uint8_t* y, *u, *v;  // rows to copy (in yuv420 format)
+  int y_stride;              // row stride for luma
+  int uv_stride;             // row stride for chroma
+
+  void* opaque;              // user data
+
+  // called when fresh samples are available. Currently, samples are in
+  // YUV420 format, and can be up to width x 24 in size (depending on the
+  // in-loop filtering level, e.g.). Should return false in case of error
+  // or abort request. The actual size of the area to update is mb_w x mb_h
+  // in size, taking cropping into account.
+  VP8IoPutHook put;
+
+  // called just before starting to decode the blocks.
+  // Must return false in case of setup error, true otherwise. If false is
+  // returned, teardown() will NOT be called. But if the setup succeeded
+  // and true is returned, then teardown() will always be called afterward.
+  VP8IoSetupHook setup;
+
+  // Called just after block decoding is finished (or when an error occurred
+  // during put()). Is NOT called if setup() failed.
+  VP8IoTeardownHook teardown;
+
+  // this is a recommendation for the user-side yuv->rgb converter. This flag
+  // is set when calling setup() hook and can be overwritten by it. It then
+  // can be taken into consideration during the put() method.
+  int fancy_upsampling;
+
+  // Input buffer.
+  size_t data_size;
+  const uint8_t* data;
+
+  // If true, in-loop filtering will not be performed even if present in the
+  // bitstream. Switching off filtering may speed up decoding at the expense
+  // of more visible blocking. Note that output will also be non-compliant
+  // with the VP8 specifications.
+  int bypass_filtering;
+
+  // Cropping parameters.
+  int use_cropping;
+  int crop_left, crop_right, crop_top, crop_bottom;
+
+  // Scaling parameters.
+  int use_scaling;
+  int scaled_width, scaled_height;
+
+  // If non NULL, pointer to the alpha data (if present) corresponding to the
+  // start of the current row (That is: it is pre-offset by mb_y and takes
+  // cropping into account).
+  const uint8_t* a;
+};
+
+// Internal, version-checked, entry point
+int VP8InitIoInternal(VP8Io* const, int);
+
+// Set the custom IO function pointers and user-data. The setter for IO hooks
+// should be called before initiating incremental decoding. Returns true if
+// WebPIDecoder object is successfully modified, false otherwise.
+int WebPISetIOHooks(WebPIDecoder* const idec,
+                    VP8IoPutHook put,
+                    VP8IoSetupHook setup,
+                    VP8IoTeardownHook teardown,
+                    void* user_data);
+
+// Main decoding object. This is an opaque structure.
+typedef struct VP8Decoder VP8Decoder;
+
+// Create a new decoder object.
+VP8Decoder* VP8New(void);
+
+// Must be called to make sure 'io' is initialized properly.
+// Returns false in case of version mismatch. Upon such failure, no other
+// decoding function should be called (VP8Decode, VP8GetHeaders, ...)
+static WEBP_INLINE int VP8InitIo(VP8Io* const io) {
+  return VP8InitIoInternal(io, WEBP_DECODER_ABI_VERSION);
+}
+
+// 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.
+// Returns false in case of error.
+int VP8Decode(VP8Decoder* const dec, VP8Io* const io);
+
+// Return current status of the decoder:
+VP8StatusCode VP8Status(VP8Decoder* const dec);
+
+// return readable string corresponding to the last status.
+const char* VP8StatusMessage(VP8Decoder* const dec);
+
+// Resets the decoder in its initial state, reclaiming memory.
+// Not a mandatory call between calls to VP8Decode().
+void VP8Clear(VP8Decoder* const dec);
+
+// Destroy the decoder object.
+void VP8Delete(VP8Decoder* const dec);
+
+//------------------------------------------------------------------------------
+// Miscellaneous VP8/VP8L bitstream probing functions.
+
+// Returns true if the next 3 bytes in data contain the VP8 signature.
+WEBP_EXTERN(int) VP8CheckSignature(const uint8_t* const data, size_t data_size);
+
+// Validates the VP8 data-header and retrieves basic header information viz
+// width and height. Returns 0 in case of formatting error. *width/*height
+// can be passed NULL.
+WEBP_EXTERN(int) VP8GetInfo(
+    const uint8_t* data,
+    size_t data_size,    // data available so far
+    size_t chunk_size,   // total data size expected in the chunk
+    int* const width, int* const height);
+
+// Returns true if the next byte(s) in data is a VP8L signature.
+WEBP_EXTERN(int) VP8LCheckSignature(const uint8_t* const data, size_t size);
+
+// Validates the VP8L data-header and retrieves basic header information viz
+// width, height and alpha. Returns 0 in case of formatting error.
+// width/height/has_alpha can be passed NULL.
+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);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_WEBP_DECODE_VP8_H_ */
diff --git a/drivers/webp/dec/frame.c b/drivers/webp/dec/frame.c
new file mode 100644
index 0000000000..e1eea94ebe
--- /dev/null
+++ b/drivers/webp/dec/frame.c
@@ -0,0 +1,818 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Frame-reconstruction function. Memory allocation.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+#include "./vp8i.h"
+#include "../utils/utils.h"
+
+#define ALIGN_MASK (32 - 1)
+
+static void ReconstructRow(const VP8Decoder* const dec,
+                           const VP8ThreadContext* ctx);  // TODO(skal): remove
+
+//------------------------------------------------------------------------------
+// Filtering
+
+// kFilterExtraRows[] = How many extra lines are needed on the MB boundary
+// for caching, given a filtering level.
+// Simple filter:  up to 2 luma samples are read and 1 is written.
+// Complex filter: up to 4 luma samples are read and 3 are written. Same for
+//                 U/V, so it's 8 samples total (because of the 2x upsampling).
+static const uint8_t kFilterExtraRows[3] = { 0, 2, 8 };
+
+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_;
+  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 = 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);
+    }
+    if (f_info->f_inner_) {
+      VP8SimpleHFilter16i(y_dst, y_bps, limit);
+    }
+    if (mb_y > 0) {
+      VP8SimpleVFilter16(y_dst, y_bps, limit + 4);
+    }
+    if (f_info->f_inner_) {
+      VP8SimpleVFilter16i(y_dst, y_bps, limit);
+    }
+  } else {    // complex
+    const int uv_bps = dec->cache_uv_stride_;
+    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);
+    }
+    if (f_info->f_inner_) {
+      VP8HFilter16i(y_dst, y_bps, limit, ilevel, hev_thresh);
+      VP8HFilter8i(u_dst, v_dst, uv_bps, limit, ilevel, hev_thresh);
+    }
+    if (mb_y > 0) {
+      VP8VFilter16(y_dst, y_bps, limit + 4, ilevel, hev_thresh);
+      VP8VFilter8(u_dst, v_dst, uv_bps, limit + 4, ilevel, hev_thresh);
+    }
+    if (f_info->f_inner_) {
+      VP8VFilter16i(y_dst, y_bps, limit, ilevel, hev_thresh);
+      VP8VFilter8i(u_dst, v_dst, uv_bps, limit, ilevel, hev_thresh);
+    }
+  }
+}
+
+// Filter the decoded macroblock row (if needed)
+static void FilterRow(const VP8Decoder* const dec) {
+  int mb_x;
+  const int mb_y = dec->thread_ctx_.mb_y_;
+  assert(dec->thread_ctx_.filter_row_);
+  for (mb_x = dec->tl_mb_x_; mb_x < dec->br_mb_x_; ++mb_x) {
+    DoFilter(dec, mb_x, mb_y);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Precompute the filtering strength for each segment and each i4x4/i16x16 mode.
+
+static void PrecomputeFilterStrengths(VP8Decoder* const dec) {
+  if (dec->filter_type_ > 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_) {
+          // TODO(skal): only CURRENT is handled for now.
+          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;
+      }
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+// 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 (all_amp != 0) {
+        VP8InitRandom(&dec->dithering_rg_, 1.0f);
+        dec->dither_ = 1;
+      }
+    }
+  }
+}
+
+// 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;
+  }
+}
+
+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_);
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+// This function is called after a row of macroblocks is finished decoding.
+// It also takes into account the following restrictions:
+//  * In case of in-loop filtering, we must hold off sending some of the bottom
+//    pixels as they are yet unfiltered. They will be when the next macroblock
+//    row is decoded. Meanwhile, we must preserve them by rotating them in the
+//    cache area. This doesn't hold for the very bottom row of the uncropped
+//    picture of course.
+//  * we must clip the remaining pixels against the cropping area. The VP8Io
+//    struct must have the following fields set correctly before calling put():
+
+#define MACROBLOCK_VPOS(mb_y)  ((mb_y) * 16)    // vertical position of a MB
+
+// Finalize and transmit a complete row. Return false in case of user-abort.
+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 = 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 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 (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;
+      io->v = vdst;
+    } else {
+      io->y = dec->cache_y_ + y_offset;
+      io->u = dec->cache_u_ + uv_offset;
+      io->v = dec->cache_v_ + uv_offset;
+    }
+
+    if (!is_last_row) {
+      y_end -= extra_y_rows;
+    }
+    if (y_end > io->crop_bottom) {
+      y_end = io->crop_bottom;    // make sure we don't overflow on last row.
+    }
+    io->a = NULL;
+    if (dec->alpha_data_ != NULL && y_start < y_end) {
+      // 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,
+                           "Could not decode alpha data.");
+      }
+    }
+    if (y_start < io->crop_top) {
+      const int delta_y = io->crop_top - y_start;
+      y_start = io->crop_top;
+      assert(!(delta_y & 1));
+      io->y += dec->cache_y_stride_ * delta_y;
+      io->u += dec->cache_uv_stride_ * (delta_y >> 1);
+      io->v += dec->cache_uv_stride_ * (delta_y >> 1);
+      if (io->a != NULL) {
+        io->a += io->width * delta_y;
+      }
+    }
+    if (y_start < y_end) {
+      io->y += io->crop_left;
+      io->u += io->crop_left >> 1;
+      io->v += io->crop_left >> 1;
+      if (io->a != NULL) {
+        io->a += io->crop_left;
+      }
+      io->mb_y = y_start - io->crop_top;
+      io->mb_w = io->crop_right - io->crop_left;
+      io->mb_h = y_end - y_start;
+      ok = io->put(io);
+    }
+  }
+  // rotate top samples if needed
+  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);
+    }
+  }
+
+  return ok;
+}
+
+#undef MACROBLOCK_VPOS
+
+//------------------------------------------------------------------------------
+
+int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io) {
+  int ok = 1;
+  VP8ThreadContext* const ctx = &dec->thread_ctx_;
+  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_ = filter_row;
+    ReconstructRow(dec, ctx);
+    ok = FinishRow(dec, io);
+  } else {
+    WebPWorker* const worker = &dec->worker_;
+    // Finish previous job *before* updating context
+    ok &= WebPWorkerSync(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_ = 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
+      if (++dec->cache_id_ == dec->num_caches_) {
+        dec->cache_id_ = 0;
+      }
+    }
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+// Finish setting up the decoding parameter once user's setup() is called.
+
+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() no matter what.
+  if (io->setup != NULL && !io->setup(io)) {
+    VP8SetError(dec, VP8_STATUS_USER_ABORT, "Frame setup failed");
+    return dec->status_;
+  }
+
+  // Disable filtering per user request
+  if (io->bypass_filtering) {
+    dec->filter_type_ = 0;
+  }
+  // TODO(skal): filter type / strength / sharpness forcing
+
+  // 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 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.
+  // Means: there's a dependency chain that goes all the way up to the
+  // top-left corner of the picture (MB #0). We must filter all the previous
+  // macroblocks.
+  // TODO(skal): add an 'approximate_decoding' option, that won't produce
+  // a 1:1 bit-exactness for complex filtering?
+  {
+    const int extra_pixels = kFilterExtraRows[dec->filter_type_];
+    if (dec->filter_type_ == 2) {
+      // For complex filter, we need to preserve the dependency chain.
+      dec->tl_mb_x_ = 0;
+      dec->tl_mb_y_ = 0;
+    } else {
+      // For simple filter, we can filter only the cropped region.
+      // We include 'extra_pixels' on the other side of the boundary, since
+      // vertical or horizontal filtering of the previous macroblock can
+      // modify some abutting pixels.
+      dec->tl_mb_x_ = (io->crop_left - extra_pixels) >> 4;
+      dec->tl_mb_y_ = (io->crop_top - extra_pixels) >> 4;
+      if (dec->tl_mb_x_ < 0) dec->tl_mb_x_ = 0;
+      if (dec->tl_mb_y_ < 0) dec->tl_mb_y_ = 0;
+    }
+    // We need some 'extra' pixels on the right/bottom.
+    dec->br_mb_y_ = (io->crop_bottom + 15 + extra_pixels) >> 4;
+    dec->br_mb_x_ = (io->crop_right + 15 + extra_pixels) >> 4;
+    if (dec->br_mb_x_ > dec->mb_w_) {
+      dec->br_mb_x_ = dec->mb_w_;
+    }
+    if (dec->br_mb_y_ > dec->mb_h_) {
+      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->mt_method_ > 0) {
+    ok = WebPWorkerSync(&dec->worker_);
+  }
+
+  if (io->teardown != NULL) {
+    io->teardown(io);
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+// For multi-threaded decoding we need to use 3 rows of 16 pixels as delay line.
+//
+// Reason is: the deblocking filter cannot deblock the bottom horizontal edges
+// immediately, and needs to wait for first few rows of the next macroblock to
+// be decoded. Hence, deblocking is lagging behind by 4 or 8 pixels (depending
+// on strength).
+// With two threads, the vertical positions of the rows being decoded are:
+// Decode:  [ 0..15][16..31][32..47][48..63][64..79][...
+// Deblock:         [ 0..11][12..27][28..43][44..59][...
+// If we use two threads and two caches of 16 pixels, the sequence would be:
+// Decode:  [ 0..15][16..31][ 0..15!!][16..31][ 0..15][...
+// Deblock:         [ 0..11][12..27!!][-4..11][12..27][...
+// The problem occurs during row [12..15!!] that both the decoding and
+// deblocking threads are writing simultaneously.
+// With 3 cache lines, one get a safe write pattern:
+// Decode:  [ 0..15][16..31][32..47][ 0..15][16..31][32..47][0..
+// Deblock:         [ 0..11][12..27][28..43][-4..11][12..27][28...
+// Note that multi-threaded output _without_ deblocking can make use of two
+// cache lines of 16 pixels only, since there's no lagging behind. The decoding
+// and output process have non-concurrent writing:
+// Decode:  [ 0..15][16..31][ 0..15][16..31][...
+// io->put:         [ 0..15][16..31][ 0..15][...
+
+#define MT_CACHE_LINES 3
+#define ST_CACHE_LINES 1   // 1 cache row only for single-threaded case
+
+// Initialize multi/single-thread worker
+static int InitThreadContext(VP8Decoder* const dec) {
+  dec->cache_id_ = 0;
+  if (dec->mt_method_ > 0) {
+    WebPWorker* const worker = &dec->worker_;
+    if (!WebPWorkerReset(worker)) {
+      return VP8SetError(dec, VP8_STATUS_OUT_OF_MEMORY,
+                         "thread initialization failed.");
+    }
+    worker->data1 = dec;
+    worker->data2 = (void*)&dec->thread_ctx_.io_;
+    worker->hook = (WebPWorkerHook)FinishRow;
+    dec->num_caches_ =
+      (dec->filter_type_ > 0) ? MT_CACHE_LINES : MT_CACHE_LINES - 1;
+  } else {
+    dec->num_caches_ = ST_CACHE_LINES;
+  }
+  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->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
+
+//------------------------------------------------------------------------------
+// Memory setup
+
+static int AllocateMemory(VP8Decoder* const dec) {
+  const int num_caches = dec->num_caches_;
+  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 = 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->mt_method_ > 0 ? 2 : 1) * sizeof(VP8FInfo)
+        : 0;
+  const size_t yuv_size = YUV_SIZE * sizeof(*dec->yuv_b_);
+  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;
+  // alpha_size is the only one that scales as width x height.
+  const uint64_t alpha_size = (dec->alpha_data_ != NULL) ?
+      (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 + mb_data_size
+                        + cache_size + alpha_size + ALIGN_MASK;
+  uint8_t* mem;
+
+  if (needed != (size_t)needed) return 0;  // check for overflow
+  if (needed > dec->mem_size_) {
+    free(dec->mem_);
+    dec->mem_size_ = 0;
+    dec->mem_ = WebPSafeMalloc(needed, sizeof(uint8_t));
+    if (dec->mem_ == NULL) {
+      return VP8SetError(dec, VP8_STATUS_OUT_OF_MEMORY,
+                         "no memory during frame initialization.");
+    }
+    // down-cast is ok, thanks to WebPSafeAlloc() above.
+    dec->mem_size_ = (size_t)needed;
+  }
+
+  mem = (uint8_t*)dec->mem_;
+  dec->intra_t_ = (uint8_t*)mem;
+  mem += intra_pred_mode_size;
+
+  dec->yuv_t_ = (VP8TopSamples*)mem;
+  mem += top_size;
+
+  dec->mb_info_ = ((VP8MB*)mem) + 1;
+  mem += mb_info_size;
+
+  dec->f_info_ = f_info_size ? (VP8FInfo*)mem : NULL;
+  mem += f_info_size;
+  dec->thread_ctx_.id_ = 0;
+  dec->thread_ctx_.f_info_ = dec->f_info_;
+  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);
+  dec->yuv_b_ = (uint8_t*)mem;
+  mem += yuv_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;
+  {
+    const int extra_rows = kFilterExtraRows[dec->filter_type_];
+    const int extra_y = extra_rows * dec->cache_y_stride_;
+    const int extra_uv = (extra_rows / 2) * dec->cache_uv_stride_;
+    dec->cache_y_ = ((uint8_t*)mem) + extra_y;
+    dec->cache_u_ = dec->cache_y_
+                  + 16 * num_caches * dec->cache_y_stride_ + extra_uv;
+    dec->cache_v_ = dec->cache_u_
+                  + 8 * num_caches * dec->cache_uv_stride_ + extra_uv;
+    dec->cache_id_ = 0;
+  }
+  mem += cache_size;
+
+  // 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/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);
+
+  return 1;
+}
+
+static void InitIo(VP8Decoder* const dec, VP8Io* io) {
+  // prepare 'io'
+  io->mb_y = 0;
+  io->y = dec->cache_y_;
+  io->u = dec->cache_u_;
+  io->v = dec->cache_v_;
+  io->y_stride = dec->cache_y_stride_;
+  io->uv_stride = dec->cache_uv_stride_;
+  io->a = NULL;
+}
+
+int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
+  if (!InitThreadContext(dec)) return 0;  // call first. Sets dec->num_caches_.
+  if (!AllocateMemory(dec)) return 0;
+  InitIo(dec, io);
+  VP8DspInit();  // Init critical function pointers and look-up tables.
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// 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;
+}
+
+static void Copy32b(uint8_t* dst, uint8_t* 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;
+  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]);
+      }
+    } else {
+      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;
+      }
+    }
+    {
+      // 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);
+      } else if (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 (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);
+      }
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/dec/idec.c b/drivers/webp/dec/idec.c
new file mode 100644
index 0000000000..40d5ff6e0d
--- /dev/null
+++ b/drivers/webp/dec/idec.c
@@ -0,0 +1,849 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Incremental decoding
+//
+// Author: somnath@google.com (Somnath Banerjee)
+
+#include <assert.h>
+#include <string.h>
+#include <stdlib.h>
+
+#include "./alphai.h"
+#include "./webpi.h"
+#include "./vp8i.h"
+#include "../utils/utils.h"
+
+// In append mode, buffer allocations increase as multiples of this value.
+// Needs to be a power of 2.
+#define CHUNK_SIZE 4096
+#define MAX_MB_SIZE 4096
+
+//------------------------------------------------------------------------------
+// Data structures for memory and states
+
+// 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_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,
+  STATE_VP8L_DATA,
+  STATE_DONE,
+  STATE_ERROR
+} DecState;
+
+// Operating state for the MemBuffer
+typedef enum {
+  MEM_MODE_NONE = 0,
+  MEM_MODE_APPEND,
+  MEM_MODE_MAP
+} MemBufferMode;
+
+// storage for partition #0 and partial data (in a rolling fashion)
+typedef struct {
+  MemBufferMode mode_;  // Operation mode
+  size_t start_;        // start location of the data to be decoded
+  size_t end_;          // end location
+  size_t buf_size_;     // size of the allocated buffer
+  uint8_t* buf_;        // We don't own this buffer in case WebPIUpdate()
+
+  size_t part0_size_;         // size of partition #0
+  const uint8_t* part0_buf_;  // buffer to store partition #0
+} MemBuffer;
+
+struct WebPIDecoder {
+  DecState state_;         // current decoding state
+  WebPDecParams params_;   // Params to store output info
+  int is_lossless_;        // for down-casting 'dec_'.
+  void* dec_;              // either a VP8Decoder or a VP8LDecoder instance
+  VP8Io io_;
+
+  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.
+};
+
+// 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_;
+  // note: for VP8, setting up idec->io_ is only really needed at the beginning
+  // of the decoding, till partition #0 is complete.
+  idec->io_.data = new_base;
+  idec->io_.data_size = MemDataSize(mem);
+
+  if (idec->dec_ != NULL) {
+    if (!idec->is_lossless_) {
+      VP8Decoder* const dec = (VP8Decoder*)idec->dec_;
+      const int last_part = dec->num_parts_ - 1;
+      if (offset != 0) {
+        int p;
+        for (p = 0; p <= last_part; ++p) {
+          RemapBitReader(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);
+        }
+      }
+      assert(last_part >= 0);
+      dec->parts_[last_part].buf_end_ = mem->buf_ + mem->end_;
+      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.
+          }
+        }
+      }
+    } else {    // Resize lossless bitreader
+      VP8LDecoder* const dec = (VP8LDecoder*)idec->dec_;
+      VP8LBitReaderSetBuffer(&dec->br_, new_base, MemDataSize(mem));
+    }
+  }
+}
+
+// Appends data to the end of MemBuffer->buf_. It expands the allocated memory
+// 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 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
+    // allows for a chunk should be considered a smoke smell.
+    return 0;
+  }
+
+  if (mem->end_ + data_size > mem->buf_size_) {  // Need some free memory
+    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_);
+    mem->buf_ = new_buf;
+    mem->buf_size_ = (size_t)extra_size;
+    mem->start_ = new_mem_start;
+    mem->end_ = current_size;
+  }
+
+  memcpy(mem->buf_ + mem->end_, data, data_size);
+  mem->end_ += data_size;
+  assert(mem->end_ <= mem->buf_size_);
+
+  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_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!
+
+  mem->buf_ = (uint8_t*)data;
+  mem->end_ = mem->buf_size_ = data_size;
+
+  DoRemap(idec, mem->buf_ + mem->start_ - old_start);
+  return 1;
+}
+
+static void InitMemBuffer(MemBuffer* const mem) {
+  mem->mode_       = MEM_MODE_NONE;
+  mem->buf_        = NULL;
+  mem->buf_size_   = 0;
+  mem->part0_buf_  = NULL;
+  mem->part0_size_ = 0;
+}
+
+static void ClearMemBuffer(MemBuffer* const mem) {
+  assert(mem);
+  if (mem->mode_ == MEM_MODE_APPEND) {
+    free(mem->buf_);
+    free((void*)mem->part0_buf_);
+  }
+}
+
+static int CheckMemBufferMode(MemBuffer* const mem, MemBufferMode expected) {
+  if (mem->mode_ == MEM_MODE_NONE) {
+    mem->mode_ = expected;    // switch to the expected mode
+  } else if (mem->mode_ != expected) {
+    return 0;         // we mixed the modes => error
+  }
+  assert(mem->mode_ == expected);   // mode is ok
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// 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->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_;
+  *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);
+}
+
+//------------------------------------------------------------------------------
+
+static VP8StatusCode IDecError(WebPIDecoder* const idec, VP8StatusCode error) {
+  if (idec->state_ == STATE_VP8_DATA) {
+    VP8Io* const io = &idec->io_;
+    if (io->teardown != NULL) {
+      io->teardown(io);
+    }
+  }
+  idec->state_ = STATE_ERROR;
+  return error;
+}
+
+static void ChangeState(WebPIDecoder* const idec, DecState new_state,
+                        size_t consumed_bytes) {
+  MemBuffer* const mem = &idec->mem_;
+  idec->state_ = new_state;
+  mem->start_ += consumed_bytes;
+  assert(mem->start_ <= mem->end_);
+  idec->io_.data = mem->buf_ + mem->start_;
+  idec->io_.data_size = MemDataSize(mem);
+}
+
+// Headers
+static VP8StatusCode DecodeWebPHeaders(WebPIDecoder* const idec) {
+  MemBuffer* const mem = &idec->mem_;
+  const uint8_t* data = mem->buf_ + mem->start_;
+  size_t curr_size = MemDataSize(mem);
+  VP8StatusCode status;
+  WebPHeaderStructure headers;
+
+  headers.data = data;
+  headers.data_size = curr_size;
+  status = WebPParseHeaders(&headers);
+  if (status == VP8_STATUS_NOT_ENOUGH_DATA) {
+    return VP8_STATUS_SUSPENDED;  // We haven't found a VP8 chunk yet.
+  } else if (status != VP8_STATUS_OK) {
+    return IDecError(idec, status);
+  }
+
+  idec->chunk_size_ = headers.compressed_size;
+  idec->is_lossless_ = headers.is_lossless;
+  if (!idec->is_lossless_) {
+    VP8Decoder* const dec = VP8New();
+    if (dec == NULL) {
+      return VP8_STATUS_OUT_OF_MEMORY;
+    }
+    idec->dec_ = dec;
+    dec->alpha_data_ = headers.alpha_data;
+    dec->alpha_data_size_ = headers.alpha_data_size;
+    ChangeState(idec, STATE_VP8_HEADER, headers.offset);
+  } else {
+    VP8LDecoder* const dec = VP8LNew();
+    if (dec == NULL) {
+      return VP8_STATUS_OUT_OF_MEMORY;
+    }
+    idec->dec_ = dec;
+    ChangeState(idec, STATE_VP8L_HEADER, headers.offset);
+  }
+  return VP8_STATUS_OK;
+}
+
+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_, &width, &height)) {
+    return IDecError(idec, VP8_STATUS_BITSTREAM_ERROR);
+  }
+
+  bits = data[0] | (data[1] << 8) | (data[2] << 16);
+  idec->mem_.part0_size_ = (bits >> 5) + VP8_FRAME_HEADER_SIZE;
+
+  idec->io_.data = data;
+  idec->io_.data_size = curr_size;
+  idec->state_ = STATE_VP8_PARTS0;
+  return VP8_STATUS_OK;
+}
+
+// Partition #0
+static int CopyParts0Data(WebPIDecoder* const idec) {
+  VP8Decoder* const dec = (VP8Decoder*)idec->dec_;
+  VP8BitReader* const br = &dec->br_;
+  const size_t psize = 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
+  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);
+    if (part0_buf == NULL) {
+      return 0;
+    }
+    memcpy(part0_buf, br->buf_, psize);
+    mem->part0_buf_ = part0_buf;
+    br->buf_ = part0_buf;
+    br->buf_end_ = part0_buf + psize;
+  } else {
+    // Else: just keep pointers to the partition #0's data in dec_->br_.
+  }
+  mem->start_ += psize;
+  return 1;
+}
+
+static VP8StatusCode DecodePartition0(WebPIDecoder* const idec) {
+  VP8Decoder* const dec = (VP8Decoder*)idec->dec_;
+  VP8Io* const io = &idec->io_;
+  const WebPDecParams* const params = &idec->params_;
+  WebPDecBuffer* const output = params->output;
+
+  // Wait till we have enough data for the whole partition #0
+  if (MemDataSize(&idec->mem_) < idec->mem_.part0_size_) {
+    return VP8_STATUS_SUSPENDED;
+  }
+
+  if (!VP8GetHeaders(dec, io)) {
+    const VP8StatusCode status = dec->status_;
+    if (status == VP8_STATUS_SUSPENDED ||
+        status == VP8_STATUS_NOT_ENOUGH_DATA) {
+      // treating NOT_ENOUGH_DATA as SUSPENDED state
+      return VP8_STATUS_SUSPENDED;
+    }
+    return IDecError(idec, status);
+  }
+
+  // Allocate/Verify output buffer now
+  dec->status_ = WebPAllocateDecBuffer(io->width, io->height, params->options,
+                                       output);
+  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);
+  }
+
+  // Finish setting up the decoding parameters. Will call io->setup().
+  if (VP8EnterCritical(dec, io) != VP8_STATUS_OK) {
+    return IDecError(idec, dec->status_);
+  }
+
+  // Note: past this point, teardown() must always be called
+  // in case of error.
+  idec->state_ = STATE_VP8_DATA;
+  // Allocate memory and prepare everything.
+  if (!VP8InitFrame(dec, io)) {
+    return IDecError(idec, dec->status_);
+  }
+  return VP8_STATUS_OK;
+}
+
+// Remaining partitions
+static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
+  VP8Decoder* const dec = (VP8Decoder*)idec->dec_;
+  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)];
+    for (; dec->mb_x_ < dec->mb_w_; ++dec->mb_x_) {
+      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);
+        }
+        return VP8_STATUS_SUSPENDED;
+      }
+      // 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);
+    }
+  }
+  // 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;
+}
+
+static VP8StatusCode ErrorStatusLossless(WebPIDecoder* const idec,
+                                         VP8StatusCode status) {
+  if (status == VP8_STATUS_SUSPENDED || status == VP8_STATUS_NOT_ENOUGH_DATA) {
+    return VP8_STATUS_SUSPENDED;
+  }
+  return IDecError(idec, status);
+}
+
+static VP8StatusCode DecodeVP8LHeader(WebPIDecoder* const idec) {
+  VP8Io* const io = &idec->io_;
+  VP8LDecoder* const dec = (VP8LDecoder*)idec->dec_;
+  const WebPDecParams* const params = &idec->params_;
+  WebPDecBuffer* const output = params->output;
+  size_t curr_size = MemDataSize(&idec->mem_);
+  assert(idec->is_lossless_);
+
+  // Wait until there's enough data for decoding header.
+  if (curr_size < (idec->chunk_size_ >> 3)) {
+    return VP8_STATUS_SUSPENDED;
+  }
+  if (!VP8LDecodeHeader(dec, io)) {
+    return ErrorStatusLossless(idec, dec->status_);
+  }
+  // Allocate/verify output buffer now.
+  dec->status_ = WebPAllocateDecBuffer(io->width, io->height, params->options,
+                                       output);
+  if (dec->status_ != VP8_STATUS_OK) {
+    return IDecError(idec, dec->status_);
+  }
+
+  idec->state_ = STATE_VP8L_DATA;
+  return VP8_STATUS_OK;
+}
+
+static VP8StatusCode DecodeVP8LData(WebPIDecoder* const idec) {
+  VP8LDecoder* const dec = (VP8LDecoder*)idec->dec_;
+  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;
+  }
+
+  if (!VP8LDecodeImage(dec)) {
+    return ErrorStatusLossless(idec, dec->status_);
+  }
+
+  idec->state_ = STATE_DONE;
+
+  return VP8_STATUS_OK;
+}
+
+  // Main decoding loop
+static VP8StatusCode IDecode(WebPIDecoder* idec) {
+  VP8StatusCode status = VP8_STATUS_SUSPENDED;
+
+  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_HEADER) {
+    status = DecodeVP8FrameHeader(idec);
+  }
+  if (idec->state_ == STATE_VP8_PARTS0) {
+    status = DecodePartition0(idec);
+  }
+  if (idec->state_ == STATE_VP8_DATA) {
+    status = DecodeRemaining(idec);
+  }
+  if (idec->state_ == STATE_VP8L_HEADER) {
+    status = DecodeVP8LHeader(idec);
+  }
+  if (idec->state_ == STATE_VP8L_DATA) {
+    status = DecodeVP8LData(idec);
+  }
+  return status;
+}
+
+//------------------------------------------------------------------------------
+// Public functions
+
+WebPIDecoder* WebPINewDecoder(WebPDecBuffer* output_buffer) {
+  WebPIDecoder* idec = (WebPIDecoder*)calloc(1, sizeof(*idec));
+  if (idec == NULL) {
+    return NULL;
+  }
+
+  idec->state_ = STATE_WEBP_HEADER;
+  idec->chunk_size_ = 0;
+
+  InitMemBuffer(&idec->mem_);
+  WebPInitDecBuffer(&idec->output_);
+  VP8InitIo(&idec->io_);
+
+  WebPResetDecParams(&idec->params_);
+  idec->params_.output = (output_buffer != NULL) ? output_buffer
+                                                 : &idec->output_;
+  WebPInitCustomIo(&idec->params_, &idec->io_);  // Plug the I/O functions.
+
+  return idec;
+}
+
+WebPIDecoder* WebPIDecode(const uint8_t* data, size_t data_size,
+                          WebPDecoderConfig* config) {
+  WebPIDecoder* idec;
+
+  // Parse the bitstream's features, if requested:
+  if (data != NULL && data_size > 0 && config != NULL) {
+    if (WebPGetFeatures(data, data_size, &config->input) != VP8_STATUS_OK) {
+      return NULL;
+    }
+  }
+  // Create an instance of the incremental decoder
+  idec = WebPINewDecoder(config ? &config->output : NULL);
+  if (idec == NULL) {
+    return NULL;
+  }
+  // Finish initialization
+  if (config != NULL) {
+    idec->params_.options = &config->options;
+  }
+  return idec;
+}
+
+void WebPIDelete(WebPIDecoder* idec) {
+  if (idec == NULL) return;
+  if (idec->dec_ != NULL) {
+    if (!idec->is_lossless_) {
+      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((VP8LDecoder*)idec->dec_);
+    }
+  }
+  ClearMemBuffer(&idec->mem_);
+  WebPFreeDecBuffer(&idec->output_);
+  free(idec);
+}
+
+//------------------------------------------------------------------------------
+// Wrapper toward WebPINewDecoder
+
+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 = 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;
+  return idec;
+}
+
+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) {
+  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 = 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;
+  idec->output_.u.YUVA.u = u;
+  idec->output_.u.YUVA.u_stride = u_stride;
+  idec->output_.u.YUVA.u_size = u_size;
+  idec->output_.u.YUVA.v = v;
+  idec->output_.u.YUVA.v_stride = v_stride;
+  idec->output_.u.YUVA.v_size = v_size;
+  idec->output_.u.YUVA.a = a;
+  idec->output_.u.YUVA.a_stride = a_stride;
+  idec->output_.u.YUVA.a_size = a_size;
+  return idec;
+}
+
+WebPIDecoder* WebPINewYUV(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) {
+  return WebPINewYUVA(luma, luma_size, luma_stride,
+                      u, u_size, u_stride,
+                      v, v_size, v_stride,
+                      NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+
+static VP8StatusCode IDecCheckStatus(const WebPIDecoder* const idec) {
+  assert(idec);
+  if (idec->state_ == STATE_ERROR) {
+    return VP8_STATUS_BITSTREAM_ERROR;
+  }
+  if (idec->state_ == STATE_DONE) {
+    return VP8_STATUS_OK;
+  }
+  return VP8_STATUS_SUSPENDED;
+}
+
+VP8StatusCode WebPIAppend(WebPIDecoder* idec,
+                          const uint8_t* data, size_t data_size) {
+  VP8StatusCode status;
+  if (idec == NULL || data == NULL) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  status = IDecCheckStatus(idec);
+  if (status != VP8_STATUS_SUSPENDED) {
+    return status;
+  }
+  // Check mixed calls between RemapMemBuffer and AppendToMemBuffer.
+  if (!CheckMemBufferMode(&idec->mem_, MEM_MODE_APPEND)) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  // Append data to memory buffer
+  if (!AppendToMemBuffer(idec, data, data_size)) {
+    return VP8_STATUS_OUT_OF_MEMORY;
+  }
+  return IDecode(idec);
+}
+
+VP8StatusCode WebPIUpdate(WebPIDecoder* idec,
+                          const uint8_t* data, size_t data_size) {
+  VP8StatusCode status;
+  if (idec == NULL || data == NULL) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  status = IDecCheckStatus(idec);
+  if (status != VP8_STATUS_SUSPENDED) {
+    return status;
+  }
+  // Check mixed calls between RemapMemBuffer and AppendToMemBuffer.
+  if (!CheckMemBufferMode(&idec->mem_, MEM_MODE_MAP)) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  // Make the memory buffer point to the new buffer
+  if (!RemapMemBuffer(idec, data, data_size)) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  return IDecode(idec);
+}
+
+//------------------------------------------------------------------------------
+
+static const WebPDecBuffer* GetOutputBuffer(const WebPIDecoder* const idec) {
+  if (idec == NULL || idec->dec_ == NULL) {
+    return NULL;
+  }
+  if (idec->state_ <= STATE_VP8_PARTS0) {
+    return NULL;
+  }
+  return idec->params_.output;
+}
+
+const WebPDecBuffer* WebPIDecodedArea(const WebPIDecoder* idec,
+                                      int* left, int* top,
+                                      int* width, int* height) {
+  const WebPDecBuffer* const src = GetOutputBuffer(idec);
+  if (left != NULL) *left = 0;
+  if (top != NULL) *top = 0;
+  // TODO(skal): later include handling of rotations.
+  if (src) {
+    if (width != NULL) *width = src->width;
+    if (height != NULL) *height = idec->params_.last_y;
+  } else {
+    if (width != NULL) *width = 0;
+    if (height != NULL) *height = 0;
+  }
+  return src;
+}
+
+uint8_t* WebPIDecGetRGB(const WebPIDecoder* idec, int* last_y,
+                        int* width, int* height, int* stride) {
+  const WebPDecBuffer* const src = GetOutputBuffer(idec);
+  if (src == NULL) return NULL;
+  if (src->colorspace >= MODE_YUV) {
+    return NULL;
+  }
+
+  if (last_y != NULL) *last_y = idec->params_.last_y;
+  if (width != NULL) *width = src->width;
+  if (height != NULL) *height = src->height;
+  if (stride != NULL) *stride = src->u.RGBA.stride;
+
+  return src->u.RGBA.rgba;
+}
+
+uint8_t* WebPIDecGetYUVA(const WebPIDecoder* idec, int* last_y,
+                         uint8_t** u, uint8_t** v, uint8_t** a,
+                         int* width, int* height,
+                         int* stride, int* uv_stride, int* a_stride) {
+  const WebPDecBuffer* const src = GetOutputBuffer(idec);
+  if (src == NULL) return NULL;
+  if (src->colorspace < MODE_YUV) {
+    return NULL;
+  }
+
+  if (last_y != NULL) *last_y = idec->params_.last_y;
+  if (u != NULL) *u = src->u.YUVA.u;
+  if (v != NULL) *v = src->u.YUVA.v;
+  if (a != NULL) *a = src->u.YUVA.a;
+  if (width != NULL) *width = src->width;
+  if (height != NULL) *height = src->height;
+  if (stride != NULL) *stride = src->u.YUVA.y_stride;
+  if (uv_stride != NULL) *uv_stride = src->u.YUVA.u_stride;
+  if (a_stride != NULL) *a_stride = src->u.YUVA.a_stride;
+
+  return src->u.YUVA.y;
+}
+
+int WebPISetIOHooks(WebPIDecoder* const idec,
+                    VP8IoPutHook put,
+                    VP8IoSetupHook setup,
+                    VP8IoTeardownHook teardown,
+                    void* user_data) {
+  if (idec == NULL || idec->state_ > STATE_WEBP_HEADER) {
+    return 0;
+  }
+
+  idec->io_.put = put;
+  idec->io_.setup = setup;
+  idec->io_.teardown = teardown;
+  idec->io_.opaque = user_data;
+
+  return 1;
+}
+
diff --git a/drivers/webp/dec/io.c b/drivers/webp/dec/io.c
new file mode 100644
index 0000000000..1ba376ed27
--- /dev/null
+++ b/drivers/webp/dec/io.c
@@ -0,0 +1,628 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// functions for sample output.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include "../dec/vp8i.h"
+#include "./webpi.h"
+#include "../dsp/dsp.h"
+#include "../dsp/yuv.h"
+
+//------------------------------------------------------------------------------
+// Main YUV<->RGB conversion functions
+
+static int EmitYUV(const VP8Io* const io, WebPDecParams* const p) {
+  WebPDecBuffer* output = p->output;
+  const WebPYUVABuffer* const buf = &output->u.YUVA;
+  uint8_t* const y_dst = buf->y + io->mb_y * buf->y_stride;
+  uint8_t* const u_dst = buf->u + (io->mb_y >> 1) * buf->u_stride;
+  uint8_t* const v_dst = buf->v + (io->mb_y >> 1) * buf->v_stride;
+  const int mb_w = io->mb_w;
+  const int mb_h = io->mb_h;
+  const int uv_w = (mb_w + 1) / 2;
+  const int uv_h = (mb_h + 1) / 2;
+  int j;
+  for (j = 0; j < mb_h; ++j) {
+    memcpy(y_dst + j * buf->y_stride, io->y + j * io->y_stride, mb_w);
+  }
+  for (j = 0; j < uv_h; ++j) {
+    memcpy(u_dst + j * buf->u_stride, io->u + j * io->uv_stride, uv_w);
+    memcpy(v_dst + j * buf->v_stride, io->v + j * io->uv_stride, uv_w);
+  }
+  return io->mb_h;
+}
+
+// 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);
+  }
+  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
+static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
+  int num_lines_out = io->mb_h;   // a priori guess
+  const WebPRGBABuffer* const buf = &p->output->u.RGBA;
+  uint8_t* dst = buf->rgba + io->mb_y * buf->stride;
+  WebPUpsampleLinePairFunc upsample = WebPUpsamplers[p->output->colorspace];
+  const uint8_t* cur_y = io->y;
+  const uint8_t* cur_u = io->u;
+  const uint8_t* cur_v = io->v;
+  const uint8_t* top_u = p->tmp_u;
+  const uint8_t* top_v = p->tmp_v;
+  int y = io->mb_y;
+  const int y_end = io->mb_y + io->mb_h;
+  const int mb_w = io->mb_w;
+  const int uv_w = (mb_w + 1) / 2;
+
+  if (y == 0) {
+    // First line is special cased. We mirror the u/v samples at boundary.
+    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,
+             dst - buf->stride, dst, mb_w);
+    ++num_lines_out;
+  }
+  // Loop over each output pairs of row.
+  for (; y + 2 < y_end; y += 2) {
+    top_u = cur_u;
+    top_v = cur_v;
+    cur_u += io->uv_stride;
+    cur_v += io->uv_stride;
+    dst += 2 * buf->stride;
+    cur_y += 2 * io->y_stride;
+    upsample(cur_y - io->y_stride, cur_y,
+             top_u, top_v, cur_u, cur_v,
+             dst - buf->stride, dst, mb_w);
+  }
+  // move to last row
+  cur_y += io->y_stride;
+  if (io->crop_top + y_end < io->crop_bottom) {
+    // Save the unfinished samples for next call (as we're not done yet).
+    memcpy(p->tmp_y, cur_y, mb_w * sizeof(*p->tmp_y));
+    memcpy(p->tmp_u, cur_u, uv_w * sizeof(*p->tmp_u));
+    memcpy(p->tmp_v, cur_v, uv_w * sizeof(*p->tmp_v));
+    // The fancy upsampler leaves a row unfinished behind
+    // (except for the very last row)
+    num_lines_out--;
+  } else {
+    // Process the very last row of even-sized picture
+    if (!(y_end & 1)) {
+      upsample(cur_y, NULL, cur_u, cur_v, cur_u, cur_v,
+               dst + buf->stride, NULL, mb_w);
+    }
+  }
+  return num_lines_out;
+}
+
+#endif    /* FANCY_UPSAMPLING */
+
+//------------------------------------------------------------------------------
+
+static int EmitAlphaYUV(const VP8Io* const io, WebPDecParams* const p) {
+  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;
+
+  if (alpha != NULL) {
+    for (j = 0; j < mb_h; ++j) {
+      memcpy(dst, alpha, mb_w * sizeof(*dst));
+      alpha += io->width;
+      dst += buf->a_stride;
+    }
+  } else if (buf->a != NULL) {
+    // the user requested alpha, but there is none, set it to opaque.
+    for (j = 0; j < mb_h; ++j) {
+      memset(dst, 0xff, mb_w * sizeof(*dst));
+      dst += buf->a_stride;
+    }
+  }
+  return 0;
+}
+
+static int GetAlphaSourceRow(const VP8Io* const io,
+                             const uint8_t** alpha, int* const num_rows) {
+  int start_y = io->mb_y;
+  *num_rows = io->mb_h;
+
+  // Compensate for the 1-line delay of the fancy upscaler.
+  // This is similar to EmitFancyRGB().
+  if (io->fancy_upsampling) {
+    if (start_y == 0) {
+      // We don't process the last row yet. It'll be done during the next call.
+      --*num_rows;
+    } else {
+      --start_y;
+      // Fortunately, *alpha data is persistent, so we can go back
+      // one row and finish alpha blending, now that the fancy upscaler
+      // completed the YUV->RGB interpolation.
+      *alpha -= io->width;
+    }
+    if (io->crop_top + io->mb_y + io->mb_h == io->crop_bottom) {
+      // If it's the very last call, we process all the remaining rows!
+      *num_rows = io->crop_bottom - io->crop_top - start_y;
+    }
+  }
+  return start_y;
+}
+
+static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
+  const uint8_t* alpha = io->a;
+  if (alpha != NULL) {
+    const int mb_w = io->mb_w;
+    const WEBP_CSP_MODE colorspace = p->output->colorspace;
+    const int alpha_first =
+        (colorspace == MODE_ARGB || colorspace == MODE_Argb);
+    const WebPRGBABuffer* const buf = &p->output->u.RGBA;
+    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)) {
+      WebPApplyAlphaMultiply(base_rgba, alpha_first,
+                             mb_w, num_rows, buf->stride);
+    }
+  }
+  return 0;
+}
+
+static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
+  const uint8_t* alpha = io->a;
+  if (alpha != NULL) {
+    const int mb_w = io->mb_w;
+    const WEBP_CSP_MODE colorspace = p->output->colorspace;
+    const WebPRGBABuffer* const buf = &p->output->u.RGBA;
+    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* alpha_dst = base_rgba + 1;
+    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).
+        const uint32_t alpha_value = alpha[i] >> 4;
+        alpha_dst[2 * i] = (alpha_dst[2 * i] & 0xf0) | alpha_value;
+        alpha_mask &= alpha_value;
+      }
+      alpha += io->width;
+      alpha_dst += buf->stride;
+    }
+    if (alpha_mask != 0x0f && WebPIsPremultipliedMode(colorspace)) {
+      WebPApplyAlphaMultiply4444(base_rgba, mb_w, num_rows, buf->stride);
+    }
+  }
+  return 0;
+}
+
+//------------------------------------------------------------------------------
+// YUV rescaling (no final RGB conversion needed)
+
+static int Rescale(const uint8_t* src, int src_stride,
+                   int new_lines, WebPRescaler* const wrk) {
+  int num_lines_out = 0;
+  while (new_lines > 0) {    // import new contributions of source rows.
+    const int lines_in = WebPRescalerImport(wrk, new_lines, src, src_stride);
+    src += lines_in * src_stride;
+    new_lines -= lines_in;
+    num_lines_out += WebPRescalerExport(wrk);    // emit output row(s)
+  }
+  return num_lines_out;
+}
+
+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);
+  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) {
+  if (io->a != NULL) {
+    Rescale(io->a, io->width, io->mb_h, &p->scaler_a);
+  }
+  return 0;
+}
+
+static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
+  const int has_alpha = WebPIsAlphaMode(p->output->colorspace);
+  const WebPYUVABuffer* const buf = &p->output->u.YUVA;
+  const int out_width  = io->scaled_width;
+  const int out_height = io->scaled_height;
+  const int uv_out_width  = (out_width + 1) >> 1;
+  const int uv_out_height = (out_height + 1) >> 1;
+  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 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;
+
+  tmp_size = work_size + 2 * uv_work_size;
+  if (has_alpha) {
+    tmp_size += work_size;
+  }
+  p->memory = calloc(1, tmp_size * sizeof(*work));
+  if (p->memory == NULL) {
+    return 0;   // memory error
+  }
+  work = (int32_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;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// RGBA rescaling
+
+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;
+  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(p->scaler_u.y_accum == p->scaler_v.y_accum);
+    WebPRescalerExportRow(&p->scaler_y);
+    WebPRescalerExportRow(&p->scaler_u);
+    WebPRescalerExportRow(&p->scaler_v);
+    convert(p->scaler_y.dst, p->scaler_u.dst, p->scaler_v.dst,
+            dst, p->scaler_y.dst_width);
+    dst += buf->stride;
+    ++num_lines_out;
+  }
+  return num_lines_out;
+}
+
+static int EmitRescaledRGB(const VP8Io* const io, WebPDecParams* const p) {
+  const int mb_h = io->mb_h;
+  const int uv_mb_h = (mb_h + 1) >> 1;
+  int j = 0, uv_j = 0;
+  int num_lines_out = 0;
+  while (j < mb_h) {
+    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);
+  }
+  return num_lines_out;
+}
+
+static int ExportAlpha(WebPDecParams* const p, int y_pos) {
+  const WebPRGBABuffer* const buf = &p->output->u.RGBA;
+  uint8_t* const base_rgba = buf->rgba + (p->last_y + 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;
+  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);
+    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;
+    }
+    dst += buf->stride;
+    ++num_lines_out;
+  }
+  if (is_premult_alpha && alpha_mask != 0xff) {
+    WebPApplyAlphaMultiply(base_rgba, alpha_first,
+                           width, num_lines_out, buf->stride);
+  }
+  return num_lines_out;
+}
+
+static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
+  const WebPRGBABuffer* const buf = &p->output->u.RGBA;
+  uint8_t* const base_rgba = buf->rgba + (p->last_y + y_pos) * buf->stride;
+  uint8_t* alpha_dst = base_rgba + 1;
+  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)) {
+    int i;
+    assert(p->last_y + 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).
+      const uint32_t alpha_value = p->scaler_a.dst[i] >> 4;
+      alpha_dst[2 * i] = (alpha_dst[2 * i] & 0xf0) | alpha_value;
+      alpha_mask &= alpha_value;
+    }
+    alpha_dst += buf->stride;
+    ++num_lines_out;
+  }
+  if (is_premult_alpha && alpha_mask != 0x0f) {
+    WebPApplyAlphaMultiply4444(base_rgba, width, num_lines_out, buf->stride);
+  }
+  return num_lines_out;
+}
+
+static int EmitRescaledAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
+  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);
+    }
+  }
+  return 0;
+}
+
+static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
+  const int has_alpha = WebPIsAlphaMode(p->output->colorspace);
+  const int out_width  = io->scaled_width;
+  const int out_height = io->scaled_height;
+  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
+  uint8_t* tmp;   // tmp storage for scaled YUV444 samples before RGB conversion
+  size_t tmp_size1, tmp_size2;
+
+  tmp_size1 = 3 * work_size;
+  tmp_size2 = 3 * out_width;
+  if (has_alpha) {
+    tmp_size1 += work_size;
+    tmp_size2 += out_width;
+  }
+  p->memory = calloc(1, tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp));
+  if (p->memory == NULL) {
+    return 0;   // memory error
+  }
+  work = (int32_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;
+
+  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 ||
+        p->output->colorspace == MODE_rgbA_4444) {
+      p->emit_alpha_row = ExportAlphaRGBA4444;
+    } else {
+      p->emit_alpha_row = ExportAlpha;
+    }
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Default custom functions
+
+static int CustomSetup(VP8Io* io) {
+  WebPDecParams* const p = (WebPDecParams*)io->opaque;
+  const WEBP_CSP_MODE colorspace = p->output->colorspace;
+  const int is_rgb = WebPIsRGBMode(colorspace);
+  const int is_alpha = WebPIsAlphaMode(colorspace);
+
+  p->memory = NULL;
+  p->emit = NULL;
+  p->emit_alpha = NULL;
+  p->emit_alpha_row = NULL;
+  if (!WebPIoInitFromOptions(p->options, io, is_alpha ? MODE_YUV : MODE_YUVA)) {
+    return 0;
+  }
+
+  if (io->use_scaling) {
+    const int ok = is_rgb ? InitRGBRescaler(io, p) : InitYUVRescaler(io, p);
+    if (!ok) {
+      return 0;    // memory error
+    }
+  } else {
+    if (is_rgb) {
+      p->emit = EmitSampledRGB;   // default
+#ifdef FANCY_UPSAMPLING
+      if (io->fancy_upsampling) {
+        const int uv_width = (io->mb_w + 1) >> 1;
+        p->memory = malloc(io->mb_w + 2 * uv_width);
+        if (p->memory == NULL) {
+          return 0;   // memory error.
+        }
+        p->tmp_y = (uint8_t*)p->memory;
+        p->tmp_u = p->tmp_y + io->mb_w;
+        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) {
+    VP8YUVInit();
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+static int CustomPut(const VP8Io* io) {
+  WebPDecParams* const p = (WebPDecParams*)io->opaque;
+  const int mb_w = io->mb_w;
+  const int mb_h = io->mb_h;
+  int num_lines_out;
+  assert(!(io->mb_y & 1));
+
+  if (mb_w <= 0 || mb_h <= 0) {
+    return 0;
+  }
+  num_lines_out = p->emit(io, p);
+  if (p->emit_alpha != NULL) {
+    p->emit_alpha(io, p);
+  }
+  p->last_y += num_lines_out;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+static void CustomTeardown(const VP8Io* io) {
+  WebPDecParams* const p = (WebPDecParams*)io->opaque;
+  free(p->memory);
+  p->memory = NULL;
+}
+
+//------------------------------------------------------------------------------
+// Main entry point
+
+void WebPInitCustomIo(WebPDecParams* const params, VP8Io* const io) {
+  io->put      = CustomPut;
+  io->setup    = CustomSetup;
+  io->teardown = CustomTeardown;
+  io->opaque   = params;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/dec/layer.c b/drivers/webp/dec/layer.c
new file mode 100644
index 0000000000..dacb9e23cd
--- /dev/null
+++ b/drivers/webp/dec/layer.c
@@ -0,0 +1,30 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Enhancement layer (for YUV444/422)
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include "./vp8i.h"
+
+//------------------------------------------------------------------------------
+
+int VP8DecodeLayer(VP8Decoder* const dec) {
+  assert(dec);
+  assert(dec->layer_data_size_ > 0);
+  (void)dec;
+
+  // TODO: handle enhancement layer here.
+
+  return 1;
+}
+
diff --git a/drivers/webp/dec/quant.c b/drivers/webp/dec/quant.c
new file mode 100644
index 0000000000..5b648f942c
--- /dev/null
+++ b/drivers/webp/dec/quant.c
@@ -0,0 +1,110 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Quantizer initialization
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./vp8i.h"
+
+static WEBP_INLINE int clip(int v, int M) {
+  return v < 0 ? 0 : v > M ? M : v;
+}
+
+// Paragraph 14.1
+static const uint8_t kDcTable[128] = {
+  4,     5,   6,   7,   8,   9,  10,  10,
+  11,   12,  13,  14,  15,  16,  17,  17,
+  18,   19,  20,  20,  21,  21,  22,  22,
+  23,   23,  24,  25,  25,  26,  27,  28,
+  29,   30,  31,  32,  33,  34,  35,  36,
+  37,   37,  38,  39,  40,  41,  42,  43,
+  44,   45,  46,  46,  47,  48,  49,  50,
+  51,   52,  53,  54,  55,  56,  57,  58,
+  59,   60,  61,  62,  63,  64,  65,  66,
+  67,   68,  69,  70,  71,  72,  73,  74,
+  75,   76,  76,  77,  78,  79,  80,  81,
+  82,   83,  84,  85,  86,  87,  88,  89,
+  91,   93,  95,  96,  98, 100, 101, 102,
+  104, 106, 108, 110, 112, 114, 116, 118,
+  122, 124, 126, 128, 130, 132, 134, 136,
+  138, 140, 143, 145, 148, 151, 154, 157
+};
+
+static const uint16_t kAcTable[128] = {
+  4,     5,   6,   7,   8,   9,  10,  11,
+  12,   13,  14,  15,  16,  17,  18,  19,
+  20,   21,  22,  23,  24,  25,  26,  27,
+  28,   29,  30,  31,  32,  33,  34,  35,
+  36,   37,  38,  39,  40,  41,  42,  43,
+  44,   45,  46,  47,  48,  49,  50,  51,
+  52,   53,  54,  55,  56,  57,  58,  60,
+  62,   64,  66,  68,  70,  72,  74,  76,
+  78,   80,  82,  84,  86,  88,  90,  92,
+  94,   96,  98, 100, 102, 104, 106, 108,
+  110, 112, 114, 116, 119, 122, 125, 128,
+  131, 134, 137, 140, 143, 146, 149, 152,
+  155, 158, 161, 164, 167, 170, 173, 177,
+  181, 185, 189, 193, 197, 201, 205, 209,
+  213, 217, 221, 225, 229, 234, 239, 245,
+  249, 254, 259, 264, 269, 274, 279, 284
+};
+
+//------------------------------------------------------------------------------
+// Paragraph 9.6
+
+void VP8ParseQuant(VP8Decoder* const dec) {
+  VP8BitReader* const br = &dec->br_;
+  const int base_q0 = VP8GetValue(br, 7);
+  const int dqy1_dc = VP8Get(br) ? VP8GetSignedValue(br, 4) : 0;
+  const int dqy2_dc = VP8Get(br) ? VP8GetSignedValue(br, 4) : 0;
+  const int dqy2_ac = VP8Get(br) ? VP8GetSignedValue(br, 4) : 0;
+  const int dquv_dc = VP8Get(br) ? VP8GetSignedValue(br, 4) : 0;
+  const int dquv_ac = VP8Get(br) ? VP8GetSignedValue(br, 4) : 0;
+
+  const VP8SegmentHeader* const hdr = &dec->segment_hdr_;
+  int i;
+
+  for (i = 0; i < NUM_MB_SEGMENTS; ++i) {
+    int q;
+    if (hdr->use_segment_) {
+      q = hdr->quantizer_[i];
+      if (!hdr->absolute_delta_) {
+        q += base_q0;
+      }
+    } else {
+      if (i > 0) {
+        dec->dqm_[i] = dec->dqm_[0];
+        continue;
+      } else {
+        q = base_q0;
+      }
+    }
+    {
+      VP8QuantMatrix* const m = &dec->dqm_[i];
+      m->y1_mat_[0] = kDcTable[clip(q + dqy1_dc, 127)];
+      m->y1_mat_[1] = kAcTable[clip(q + 0,       127)];
+
+      m->y2_mat_[0] = kDcTable[clip(q + dqy2_dc, 127)] * 2;
+      // For all x in [0..284], x*155/100 is bitwise equal to (x*101581) >> 16.
+      // The smallest precision for that is '(x*6349) >> 12' but 16 is a good
+      // word size.
+      m->y2_mat_[1] = (kAcTable[clip(q + dqy2_ac, 127)] * 101581) >> 16;
+      if (m->y2_mat_[1] < 8) m->y2_mat_[1] = 8;
+
+      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
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/dec/tree.c b/drivers/webp/dec/tree.c
new file mode 100644
index 0000000000..bf9b7c551a
--- /dev/null
+++ b/drivers/webp/dec/tree.c
@@ -0,0 +1,494 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Coding trees and probas
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "vp8i.h"
+
+#define USE_GENERIC_TREE
+
+#ifdef USE_GENERIC_TREE
+static const int8_t kYModesIntra4[18] = {
+  -B_DC_PRED, 1,
+    -B_TM_PRED, 2,
+      -B_VE_PRED, 3,
+        4, 6,
+          -B_HE_PRED, 5,
+            -B_RD_PRED, -B_VR_PRED,
+        -B_LD_PRED, 7,
+          -B_VL_PRED, 8,
+            -B_HD_PRED, -B_HU_PRED
+};
+#endif
+
+//------------------------------------------------------------------------------
+// Default probabilities
+
+// Paragraph 13.5
+static const uint8_t
+  CoeffsProba0[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS] = {
+  { { { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 253, 136, 254, 255, 228, 219, 128, 128, 128, 128, 128 },
+      { 189, 129, 242, 255, 227, 213, 255, 219, 128, 128, 128 },
+      { 106, 126, 227, 252, 214, 209, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 98, 248, 255, 236, 226, 255, 255, 128, 128, 128 },
+      { 181, 133, 238, 254, 221, 234, 255, 154, 128, 128, 128 },
+      { 78, 134, 202, 247, 198, 180, 255, 219, 128, 128, 128 },
+    },
+    { { 1, 185, 249, 255, 243, 255, 128, 128, 128, 128, 128 },
+      { 184, 150, 247, 255, 236, 224, 128, 128, 128, 128, 128 },
+      { 77, 110, 216, 255, 236, 230, 128, 128, 128, 128, 128 },
+    },
+    { { 1, 101, 251, 255, 241, 255, 128, 128, 128, 128, 128 },
+      { 170, 139, 241, 252, 236, 209, 255, 255, 128, 128, 128 },
+      { 37, 116, 196, 243, 228, 255, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 204, 254, 255, 245, 255, 128, 128, 128, 128, 128 },
+      { 207, 160, 250, 255, 238, 128, 128, 128, 128, 128, 128 },
+      { 102, 103, 231, 255, 211, 171, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 152, 252, 255, 240, 255, 128, 128, 128, 128, 128 },
+      { 177, 135, 243, 255, 234, 225, 128, 128, 128, 128, 128 },
+      { 80, 129, 211, 255, 194, 224, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 246, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 255, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 198, 35, 237, 223, 193, 187, 162, 160, 145, 155, 62 },
+      { 131, 45, 198, 221, 172, 176, 220, 157, 252, 221, 1 },
+      { 68, 47, 146, 208, 149, 167, 221, 162, 255, 223, 128 }
+    },
+    { { 1, 149, 241, 255, 221, 224, 255, 255, 128, 128, 128 },
+      { 184, 141, 234, 253, 222, 220, 255, 199, 128, 128, 128 },
+      { 81, 99, 181, 242, 176, 190, 249, 202, 255, 255, 128 }
+    },
+    { { 1, 129, 232, 253, 214, 197, 242, 196, 255, 255, 128 },
+      { 99, 121, 210, 250, 201, 198, 255, 202, 128, 128, 128 },
+      { 23, 91, 163, 242, 170, 187, 247, 210, 255, 255, 128 }
+    },
+    { { 1, 200, 246, 255, 234, 255, 128, 128, 128, 128, 128 },
+      { 109, 178, 241, 255, 231, 245, 255, 255, 128, 128, 128 },
+      { 44, 130, 201, 253, 205, 192, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 132, 239, 251, 219, 209, 255, 165, 128, 128, 128 },
+      { 94, 136, 225, 251, 218, 190, 255, 255, 128, 128, 128 },
+      { 22, 100, 174, 245, 186, 161, 255, 199, 128, 128, 128 }
+    },
+    { { 1, 182, 249, 255, 232, 235, 128, 128, 128, 128, 128 },
+      { 124, 143, 241, 255, 227, 234, 128, 128, 128, 128, 128 },
+      { 35, 77, 181, 251, 193, 211, 255, 205, 128, 128, 128 }
+    },
+    { { 1, 157, 247, 255, 236, 231, 255, 255, 128, 128, 128 },
+      { 121, 141, 235, 255, 225, 227, 255, 255, 128, 128, 128 },
+      { 45, 99, 188, 251, 195, 217, 255, 224, 128, 128, 128 }
+    },
+    { { 1, 1, 251, 255, 213, 255, 128, 128, 128, 128, 128 },
+      { 203, 1, 248, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 137, 1, 177, 255, 224, 255, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 253, 9, 248, 251, 207, 208, 255, 192, 128, 128, 128 },
+      { 175, 13, 224, 243, 193, 185, 249, 198, 255, 255, 128 },
+      { 73, 17, 171, 221, 161, 179, 236, 167, 255, 234, 128 }
+    },
+    { { 1, 95, 247, 253, 212, 183, 255, 255, 128, 128, 128 },
+      { 239, 90, 244, 250, 211, 209, 255, 255, 128, 128, 128 },
+      { 155, 77, 195, 248, 188, 195, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 24, 239, 251, 218, 219, 255, 205, 128, 128, 128 },
+      { 201, 51, 219, 255, 196, 186, 128, 128, 128, 128, 128 },
+      { 69, 46, 190, 239, 201, 218, 255, 228, 128, 128, 128 }
+    },
+    { { 1, 191, 251, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 223, 165, 249, 255, 213, 255, 128, 128, 128, 128, 128 },
+      { 141, 124, 248, 255, 255, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 16, 248, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 190, 36, 230, 255, 236, 255, 128, 128, 128, 128, 128 },
+      { 149, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 226, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 247, 192, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 240, 128, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 134, 252, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 213, 62, 250, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 55, 93, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 202, 24, 213, 235, 186, 191, 220, 160, 240, 175, 255 },
+      { 126, 38, 182, 232, 169, 184, 228, 174, 255, 187, 128 },
+      { 61, 46, 138, 219, 151, 178, 240, 170, 255, 216, 128 }
+    },
+    { { 1, 112, 230, 250, 199, 191, 247, 159, 255, 255, 128 },
+      { 166, 109, 228, 252, 211, 215, 255, 174, 128, 128, 128 },
+      { 39, 77, 162, 232, 172, 180, 245, 178, 255, 255, 128 }
+    },
+    { { 1, 52, 220, 246, 198, 199, 249, 220, 255, 255, 128 },
+      { 124, 74, 191, 243, 183, 193, 250, 221, 255, 255, 128 },
+      { 24, 71, 130, 219, 154, 170, 243, 182, 255, 255, 128 }
+    },
+    { { 1, 182, 225, 249, 219, 240, 255, 224, 128, 128, 128 },
+      { 149, 150, 226, 252, 216, 205, 255, 171, 128, 128, 128 },
+      { 28, 108, 170, 242, 183, 194, 254, 223, 255, 255, 128 }
+    },
+    { { 1, 81, 230, 252, 204, 203, 255, 192, 128, 128, 128 },
+      { 123, 102, 209, 247, 188, 196, 255, 233, 128, 128, 128 },
+      { 20, 95, 153, 243, 164, 173, 255, 203, 128, 128, 128 }
+    },
+    { { 1, 222, 248, 255, 216, 213, 128, 128, 128, 128, 128 },
+      { 168, 175, 246, 252, 235, 205, 255, 255, 128, 128, 128 },
+      { 47, 116, 215, 255, 211, 212, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 121, 236, 253, 212, 214, 255, 255, 128, 128, 128 },
+      { 141, 84, 213, 252, 201, 202, 255, 219, 128, 128, 128 },
+      { 42, 80, 160, 240, 162, 185, 255, 205, 128, 128, 128 }
+    },
+    { { 1, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 244, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 238, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  }
+};
+
+// Paragraph 11.5
+static const uint8_t kBModesProba[NUM_BMODES][NUM_BMODES][NUM_BMODES - 1] = {
+  { { 231, 120, 48, 89, 115, 113, 120, 152, 112 },
+    { 152, 179, 64, 126, 170, 118, 46, 70, 95 },
+    { 175, 69, 143, 80, 85, 82, 72, 155, 103 },
+    { 56, 58, 10, 171, 218, 189, 17, 13, 152 },
+    { 114, 26, 17, 163, 44, 195, 21, 10, 173 },
+    { 121, 24, 80, 195, 26, 62, 44, 64, 85 },
+    { 144, 71, 10, 38, 171, 213, 144, 34, 26 },
+    { 170, 46, 55, 19, 136, 160, 33, 206, 71 },
+    { 63, 20, 8, 114, 114, 208, 12, 9, 226 },
+    { 81, 40, 11, 96, 182, 84, 29, 16, 36 } },
+  { { 134, 183, 89, 137, 98, 101, 106, 165, 148 },
+    { 72, 187, 100, 130, 157, 111, 32, 75, 80 },
+    { 66, 102, 167, 99, 74, 62, 40, 234, 128 },
+    { 41, 53, 9, 178, 241, 141, 26, 8, 107 },
+    { 74, 43, 26, 146, 73, 166, 49, 23, 157 },
+    { 65, 38, 105, 160, 51, 52, 31, 115, 128 },
+    { 104, 79, 12, 27, 217, 255, 87, 17, 7 },
+    { 87, 68, 71, 44, 114, 51, 15, 186, 23 },
+    { 47, 41, 14, 110, 182, 183, 21, 17, 194 },
+    { 66, 45, 25, 102, 197, 189, 23, 18, 22 } },
+  { { 88, 88, 147, 150, 42, 46, 45, 196, 205 },
+    { 43, 97, 183, 117, 85, 38, 35, 179, 61 },
+    { 39, 53, 200, 87, 26, 21, 43, 232, 171 },
+    { 56, 34, 51, 104, 114, 102, 29, 93, 77 },
+    { 39, 28, 85, 171, 58, 165, 90, 98, 64 },
+    { 34, 22, 116, 206, 23, 34, 43, 166, 73 },
+    { 107, 54, 32, 26, 51, 1, 81, 43, 31 },
+    { 68, 25, 106, 22, 64, 171, 36, 225, 114 },
+    { 34, 19, 21, 102, 132, 188, 16, 76, 124 },
+    { 62, 18, 78, 95, 85, 57, 50, 48, 51 } },
+  { { 193, 101, 35, 159, 215, 111, 89, 46, 111 },
+    { 60, 148, 31, 172, 219, 228, 21, 18, 111 },
+    { 112, 113, 77, 85, 179, 255, 38, 120, 114 },
+    { 40, 42, 1, 196, 245, 209, 10, 25, 109 },
+    { 88, 43, 29, 140, 166, 213, 37, 43, 154 },
+    { 61, 63, 30, 155, 67, 45, 68, 1, 209 },
+    { 100, 80, 8, 43, 154, 1, 51, 26, 71 },
+    { 142, 78, 78, 16, 255, 128, 34, 197, 171 },
+    { 41, 40, 5, 102, 211, 183, 4, 1, 221 },
+    { 51, 50, 17, 168, 209, 192, 23, 25, 82 } },
+  { { 138, 31, 36, 171, 27, 166, 38, 44, 229 },
+    { 67, 87, 58, 169, 82, 115, 26, 59, 179 },
+    { 63, 59, 90, 180, 59, 166, 93, 73, 154 },
+    { 40, 40, 21, 116, 143, 209, 34, 39, 175 },
+    { 47, 15, 16, 183, 34, 223, 49, 45, 183 },
+    { 46, 17, 33, 183, 6, 98, 15, 32, 183 },
+    { 57, 46, 22, 24, 128, 1, 54, 17, 37 },
+    { 65, 32, 73, 115, 28, 128, 23, 128, 205 },
+    { 40, 3, 9, 115, 51, 192, 18, 6, 223 },
+    { 87, 37, 9, 115, 59, 77, 64, 21, 47 } },
+  { { 104, 55, 44, 218, 9, 54, 53, 130, 226 },
+    { 64, 90, 70, 205, 40, 41, 23, 26, 57 },
+    { 54, 57, 112, 184, 5, 41, 38, 166, 213 },
+    { 30, 34, 26, 133, 152, 116, 10, 32, 134 },
+    { 39, 19, 53, 221, 26, 114, 32, 73, 255 },
+    { 31, 9, 65, 234, 2, 15, 1, 118, 73 },
+    { 75, 32, 12, 51, 192, 255, 160, 43, 51 },
+    { 88, 31, 35, 67, 102, 85, 55, 186, 85 },
+    { 56, 21, 23, 111, 59, 205, 45, 37, 192 },
+    { 55, 38, 70, 124, 73, 102, 1, 34, 98 } },
+  { { 125, 98, 42, 88, 104, 85, 117, 175, 82 },
+    { 95, 84, 53, 89, 128, 100, 113, 101, 45 },
+    { 75, 79, 123, 47, 51, 128, 81, 171, 1 },
+    { 57, 17, 5, 71, 102, 57, 53, 41, 49 },
+    { 38, 33, 13, 121, 57, 73, 26, 1, 85 },
+    { 41, 10, 67, 138, 77, 110, 90, 47, 114 },
+    { 115, 21, 2, 10, 102, 255, 166, 23, 6 },
+    { 101, 29, 16, 10, 85, 128, 101, 196, 26 },
+    { 57, 18, 10, 102, 102, 213, 34, 20, 43 },
+    { 117, 20, 15, 36, 163, 128, 68, 1, 26 } },
+  { { 102, 61, 71, 37, 34, 53, 31, 243, 192 },
+    { 69, 60, 71, 38, 73, 119, 28, 222, 37 },
+    { 68, 45, 128, 34, 1, 47, 11, 245, 171 },
+    { 62, 17, 19, 70, 146, 85, 55, 62, 70 },
+    { 37, 43, 37, 154, 100, 163, 85, 160, 1 },
+    { 63, 9, 92, 136, 28, 64, 32, 201, 85 },
+    { 75, 15, 9, 9, 64, 255, 184, 119, 16 },
+    { 86, 6, 28, 5, 64, 255, 25, 248, 1 },
+    { 56, 8, 17, 132, 137, 255, 55, 116, 128 },
+    { 58, 15, 20, 82, 135, 57, 26, 121, 40 } },
+  { { 164, 50, 31, 137, 154, 133, 25, 35, 218 },
+    { 51, 103, 44, 131, 131, 123, 31, 6, 158 },
+    { 86, 40, 64, 135, 148, 224, 45, 183, 128 },
+    { 22, 26, 17, 131, 240, 154, 14, 1, 209 },
+    { 45, 16, 21, 91, 64, 222, 7, 1, 197 },
+    { 56, 21, 39, 155, 60, 138, 23, 102, 213 },
+    { 83, 12, 13, 54, 192, 255, 68, 47, 28 },
+    { 85, 26, 85, 85, 128, 128, 32, 146, 171 },
+    { 18, 11, 7, 63, 144, 171, 4, 4, 246 },
+    { 35, 27, 10, 146, 174, 171, 12, 26, 128 } },
+  { { 190, 80, 35, 99, 180, 80, 126, 54, 45 },
+    { 85, 126, 47, 87, 176, 51, 41, 20, 32 },
+    { 101, 75, 128, 139, 118, 146, 116, 128, 85 },
+    { 56, 41, 15, 176, 236, 85, 37, 9, 62 },
+    { 71, 30, 17, 119, 118, 255, 17, 18, 138 },
+    { 101, 38, 60, 138, 55, 70, 43, 26, 142 },
+    { 146, 36, 19, 30, 171, 255, 97, 27, 20 },
+    { 138, 45, 61, 62, 219, 1, 81, 188, 64 },
+    { 32, 41, 20, 117, 151, 142, 20, 21, 163 },
+    { 112, 19, 12, 61, 195, 128, 48, 4, 24 } }
+};
+
+void VP8ResetProba(VP8Proba* const proba) {
+  memset(proba->segments_, 255u, sizeof(proba->segments_));
+  // proba->bands_[][] is initialized later
+}
+
+void VP8ParseIntraMode(VP8BitReader* const br, VP8Decoder* const dec) {
+  uint8_t* const top = dec->intra_t_ + 4 * dec->mb_x_;
+  uint8_t* const left = dec->intra_l_;
+  VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+
+  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);
+    block->imodes_[0] = ymode;
+    memset(top, ymode, 4 * sizeof(*top));
+    memset(left, ymode, 4 * sizeof(*left));
+  } else {
+    uint8_t* modes = block->imodes_;
+    int y;
+    for (y = 0; y < 4; ++y) {
+      int ymode = left[y];
+      int x;
+      for (x = 0; x < 4; ++x) {
+        const uint8_t* const prob = kBModesProba[top[x]][ymode];
+#ifdef USE_GENERIC_TREE
+        // Generic tree-parsing
+        int i = kYModesIntra4[VP8GetBit(br, prob[0])];
+        while (i > 0) {
+          i = kYModesIntra4[2 * i + VP8GetBit(br, prob[i])];
+        }
+        ymode = -i;
+#else
+        // Hardcoded tree parsing
+        ymode = !VP8GetBit(br, prob[0]) ? B_DC_PRED :
+                  !VP8GetBit(br, prob[1]) ? B_TM_PRED :
+                    !VP8GetBit(br, prob[2]) ? B_VE_PRED :
+                      !VP8GetBit(br, prob[3]) ?
+                        (!VP8GetBit(br, prob[4]) ? B_HE_PRED :
+                          (!VP8GetBit(br, prob[5]) ? B_RD_PRED : B_VR_PRED)) :
+                        (!VP8GetBit(br, prob[6]) ? B_LD_PRED :
+                          (!VP8GetBit(br, prob[7]) ? B_VL_PRED :
+                            (!VP8GetBit(br, prob[8]) ? B_HD_PRED : B_HU_PRED)));
+#endif    // USE_GENERIC_TREE
+        top[x] = ymode;
+      }
+      memcpy(modes, top, 4 * sizeof(*top));
+      modes += 4;
+      left[y] = ymode;
+    }
+  }
+  // Hardcoded UVMode decision tree
+  block->uvmode_ = !VP8GetBit(br, 142) ? DC_PRED
+                 : !VP8GetBit(br, 114) ? V_PRED
+                 : VP8GetBit(br, 183) ? TM_PRED : H_PRED;
+}
+
+//------------------------------------------------------------------------------
+// Paragraph 13
+
+static const uint8_t
+    CoeffsUpdateProba[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS] = {
+  { { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 176, 246, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 223, 241, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 244, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 234, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 246, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 239, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 253, 255, 254, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 254, 255, 254, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 217, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 225, 252, 241, 253, 255, 255, 254, 255, 255, 255, 255 },
+      { 234, 250, 241, 250, 253, 255, 253, 254, 255, 255, 255 }
+    },
+    { { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 223, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 238, 253, 254, 254, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 247, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 186, 251, 250, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 234, 251, 244, 254, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 251, 243, 253, 254, 255, 254, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 236, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 253, 253, 254, 254, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 248, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 254, 252, 254, 255, 255, 255, 255, 255, 255, 255 },
+      { 248, 254, 249, 253, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 246, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 254, 251, 254, 254, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 248, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 254, 254, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 245, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 251, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 252, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  }
+};
+
+// Paragraph 9.9
+void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec) {
+  VP8Proba* const proba = &dec->proba_;
+  int t, b, c, p;
+  for (t = 0; t < NUM_TYPES; ++t) {
+    for (b = 0; b < NUM_BANDS; ++b) {
+      for (c = 0; c < NUM_CTX; ++c) {
+        for (p = 0; p < NUM_PROBAS; ++p) {
+          const 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;
+        }
+      }
+    }
+  }
+  dec->use_skip_proba_ = VP8Get(br);
+  if (dec->use_skip_proba_) {
+    dec->skip_p_ = VP8GetValue(br, 8);
+  }
+}
+
diff --git a/drivers/webp/dec/vp8.c b/drivers/webp/dec/vp8.c
new file mode 100644
index 0000000000..bfd0e8f9d3
--- /dev/null
+++ b/drivers/webp/dec/vp8.c
@@ -0,0 +1,713 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// main entry for the decoder
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+
+#include "./alphai.h"
+#include "./vp8i.h"
+#include "./vp8li.h"
+#include "./webpi.h"
+#include "../utils/bit_reader.h"
+
+//------------------------------------------------------------------------------
+
+int WebPGetDecoderVersion(void) {
+  return (DEC_MAJ_VERSION << 16) | (DEC_MIN_VERSION << 8) | DEC_REV_VERSION;
+}
+
+//------------------------------------------------------------------------------
+// VP8Decoder
+
+static void SetOk(VP8Decoder* const dec) {
+  dec->status_ = VP8_STATUS_OK;
+  dec->error_msg_ = "OK";
+}
+
+int VP8InitIoInternal(VP8Io* const io, int version) {
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_DECODER_ABI_VERSION)) {
+    return 0;  // mismatch error
+  }
+  if (io != NULL) {
+    memset(io, 0, sizeof(*io));
+  }
+  return 1;
+}
+
+VP8Decoder* VP8New(void) {
+  VP8Decoder* const dec = (VP8Decoder*)calloc(1, sizeof(*dec));
+  if (dec != NULL) {
+    SetOk(dec);
+    WebPWorkerInit(&dec->worker_);
+    dec->ready_ = 0;
+    dec->num_parts_ = 1;
+  }
+  return dec;
+}
+
+VP8StatusCode VP8Status(VP8Decoder* const dec) {
+  if (!dec) return VP8_STATUS_INVALID_PARAM;
+  return dec->status_;
+}
+
+const char* VP8StatusMessage(VP8Decoder* const dec) {
+  if (dec == NULL) return "no object";
+  if (!dec->error_msg_) return "OK";
+  return dec->error_msg_;
+}
+
+void VP8Delete(VP8Decoder* const dec) {
+  if (dec != NULL) {
+    VP8Clear(dec);
+    free(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.
+  if (dec->status_ == VP8_STATUS_OK) {
+    dec->status_ = error;
+    dec->error_msg_ = msg;
+    dec->ready_ = 0;
+  }
+  return 0;
+}
+
+//------------------------------------------------------------------------------
+
+int VP8CheckSignature(const uint8_t* const data, size_t data_size) {
+  return (data_size >= 3 &&
+          data[0] == 0x9d && data[1] == 0x01 && data[2] == 0x2a);
+}
+
+int VP8GetInfo(const uint8_t* data, size_t data_size, size_t chunk_size,
+               int* const width, int* const height) {
+  if (data == NULL || data_size < VP8_FRAME_HEADER_SIZE) {
+    return 0;         // not enough data
+  }
+  // check signature
+  if (!VP8CheckSignature(data + 3, data_size - 3)) {
+    return 0;         // Wrong signature.
+  } else {
+    const uint32_t bits = data[0] | (data[1] << 8) | (data[2] << 16);
+    const int key_frame = !(bits & 1);
+    const int w = ((data[7] << 8) | data[6]) & 0x3fff;
+    const int h = ((data[9] << 8) | data[8]) & 0x3fff;
+
+    if (!key_frame) {   // Not a keyframe.
+      return 0;
+    }
+
+    if (((bits >> 1) & 7) > 3) {
+      return 0;         // unknown profile
+    }
+    if (!((bits >> 4) & 1)) {
+      return 0;         // first frame is invisible!
+    }
+    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;
+    }
+    if (height) {
+      *height = h;
+    }
+
+    return 1;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Header parsing
+
+static void ResetSegmentHeader(VP8SegmentHeader* const hdr) {
+  assert(hdr != NULL);
+  hdr->use_segment_ = 0;
+  hdr->update_map_ = 0;
+  hdr->absolute_delta_ = 1;
+  memset(hdr->quantizer_, 0, sizeof(hdr->quantizer_));
+  memset(hdr->filter_strength_, 0, sizeof(hdr->filter_strength_));
+}
+
+// Paragraph 9.3
+static int ParseSegmentHeader(VP8BitReader* br,
+                              VP8SegmentHeader* hdr, VP8Proba* proba) {
+  assert(br != NULL);
+  assert(hdr != NULL);
+  hdr->use_segment_ = VP8Get(br);
+  if (hdr->use_segment_) {
+    hdr->update_map_ = VP8Get(br);
+    if (VP8Get(br)) {   // update data
+      int s;
+      hdr->absolute_delta_ = VP8Get(br);
+      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+        hdr->quantizer_[s] = VP8Get(br) ? VP8GetSignedValue(br, 7) : 0;
+      }
+      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+        hdr->filter_strength_[s] = VP8Get(br) ? VP8GetSignedValue(br, 6) : 0;
+      }
+    }
+    if (hdr->update_map_) {
+      int s;
+      for (s = 0; s < MB_FEATURE_TREE_PROBS; ++s) {
+        proba->segments_[s] = VP8Get(br) ? VP8GetValue(br, 8) : 255u;
+      }
+    }
+  } else {
+    hdr->update_map_ = 0;
+  }
+  return !br->eof_;
+}
+
+// Paragraph 9.5
+// This function returns VP8_STATUS_SUSPENDED if we don't have all the
+// necessary data in 'buf'.
+// This case is not necessarily an error (for incremental decoding).
+// Still, no bitreader is ever initialized to make it possible to read
+// unavailable memory.
+// If we don't even have the partitions' sizes, than VP8_STATUS_NOT_ENOUGH_DATA
+// is returned, and this is an unrecoverable error.
+// If the partitions were positioned ok, VP8_STATUS_OK is returned.
+static VP8StatusCode ParsePartitions(VP8Decoder* const dec,
+                                     const uint8_t* buf, size_t size) {
+  VP8BitReader* const br = &dec->br_;
+  const uint8_t* sz = buf;
+  const uint8_t* buf_end = buf + size;
+  const uint8_t* part_start;
+  int last_part;
+  int 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) {
+    // we can't even read the sizes with sz[]! That's a failure.
+    return VP8_STATUS_NOT_ENOUGH_DATA;
+  }
+  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;
+    sz += 3;
+  }
+  VP8InitBitReader(dec->parts_ + last_part, part_start, buf_end);
+  return (part_start < buf_end) ? VP8_STATUS_OK :
+           VP8_STATUS_SUSPENDED;   // Init is ok, but there's not enough data
+}
+
+// Paragraph 9.4
+static int ParseFilterHeader(VP8BitReader* br, VP8Decoder* const dec) {
+  VP8FilterHeader* const hdr = &dec->filter_hdr_;
+  hdr->simple_    = VP8Get(br);
+  hdr->level_     = VP8GetValue(br, 6);
+  hdr->sharpness_ = VP8GetValue(br, 3);
+  hdr->use_lf_delta_ = VP8Get(br);
+  if (hdr->use_lf_delta_) {
+    if (VP8Get(br)) {   // update lf-delta?
+      int i;
+      for (i = 0; i < NUM_REF_LF_DELTAS; ++i) {
+        if (VP8Get(br)) {
+          hdr->ref_lf_delta_[i] = VP8GetSignedValue(br, 6);
+        }
+      }
+      for (i = 0; i < NUM_MODE_LF_DELTAS; ++i) {
+        if (VP8Get(br)) {
+          hdr->mode_lf_delta_[i] = VP8GetSignedValue(br, 6);
+        }
+      }
+    }
+  }
+  dec->filter_type_ = (hdr->level_ == 0) ? 0 : hdr->simple_ ? 1 : 2;
+  return !br->eof_;
+}
+
+// Topmost call
+int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
+  const uint8_t* buf;
+  size_t buf_size;
+  VP8FrameHeader* frm_hdr;
+  VP8PictureHeader* pic_hdr;
+  VP8BitReader* br;
+  VP8StatusCode status;
+
+  if (dec == NULL) {
+    return 0;
+  }
+  SetOk(dec);
+  if (io == NULL) {
+    return VP8SetError(dec, VP8_STATUS_INVALID_PARAM,
+                       "null VP8Io passed to VP8GetHeaders()");
+  }
+  buf = io->data;
+  buf_size = io->data_size;
+  if (buf_size < 4) {
+    return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
+                       "Truncated header.");
+  }
+
+  // Paragraph 9.1
+  {
+    const uint32_t bits = buf[0] | (buf[1] << 8) | (buf[2] << 16);
+    frm_hdr = &dec->frm_hdr_;
+    frm_hdr->key_frame_ = !(bits & 1);
+    frm_hdr->profile_ = (bits >> 1) & 7;
+    frm_hdr->show_ = (bits >> 4) & 1;
+    frm_hdr->partition_length_ = (bits >> 5);
+    if (frm_hdr->profile_ > 3)
+      return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
+                         "Incorrect keyframe parameters.");
+    if (!frm_hdr->show_)
+      return VP8SetError(dec, VP8_STATUS_UNSUPPORTED_FEATURE,
+                         "Frame not displayable.");
+    buf += 3;
+    buf_size -= 3;
+  }
+
+  pic_hdr = &dec->pic_hdr_;
+  if (frm_hdr->key_frame_) {
+    // Paragraph 9.2
+    if (buf_size < 7) {
+      return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
+                         "cannot parse picture header");
+    }
+    if (!VP8CheckSignature(buf, buf_size)) {
+      return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
+                         "Bad code word");
+    }
+    pic_hdr->width_ = ((buf[4] << 8) | buf[3]) & 0x3fff;
+    pic_hdr->xscale_ = buf[4] >> 6;   // ratio: 1, 5/4 5/3 or 2
+    pic_hdr->height_ = ((buf[6] << 8) | buf[5]) & 0x3fff;
+    pic_hdr->yscale_ = buf[6] >> 6;
+    buf += 7;
+    buf_size -= 7;
+
+    dec->mb_w_ = (pic_hdr->width_ + 15) >> 4;
+    dec->mb_h_ = (pic_hdr->height_ + 15) >> 4;
+    // Setup default output area (can be later modified during io->setup())
+    io->width = pic_hdr->width_;
+    io->height = pic_hdr->height_;
+    io->use_scaling  = 0;
+    io->use_cropping = 0;
+    io->crop_top  = 0;
+    io->crop_left = 0;
+    io->crop_right  = io->width;
+    io->crop_bottom = io->height;
+    io->mb_w = io->width;   // sanity check
+    io->mb_h = io->height;  // ditto
+
+    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_
+  // to read this partition (and this partition only).
+  if (frm_hdr->partition_length_ > buf_size) {
+    return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
+                       "bad partition length");
+  }
+
+  br = &dec->br_;
+  VP8InitBitReader(br, buf, buf + frm_hdr->partition_length_);
+  buf += frm_hdr->partition_length_;
+  buf_size -= frm_hdr->partition_length_;
+
+  if (frm_hdr->key_frame_) {
+    pic_hdr->colorspace_ = VP8Get(br);
+    pic_hdr->clamp_type_ = VP8Get(br);
+  }
+  if (!ParseSegmentHeader(br, &dec->segment_hdr_, &dec->proba_)) {
+    return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
+                       "cannot parse segment header");
+  }
+  // Filter specs
+  if (!ParseFilterHeader(br, dec)) {
+    return VP8SetError(dec, VP8_STATUS_BITSTREAM_ERROR,
+                       "cannot parse filter header");
+  }
+  status = ParsePartitions(dec, buf, buf_size);
+  if (status != VP8_STATUS_OK) {
+    return VP8SetError(dec, status, "cannot parse partitions");
+  }
+
+  // quantizer change
+  VP8ParseQuant(dec);
+
+  // Frame buffer marking
+  if (!frm_hdr->key_frame_) {
+    return VP8SetError(dec, VP8_STATUS_UNSUPPORTED_FEATURE,
+                       "Not a key frame.");
+  }
+
+  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;
+}
+
+//------------------------------------------------------------------------------
+// Residual decoding (Paragraph 13.2 / 13.3)
+
+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
+};
+
+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 };
+static const uint8_t kCat6[] =
+  { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129, 0 };
+static const uint8_t* const kCat3456[] = { kCat3, kCat4, kCat5, kCat6 };
+static const uint8_t kZigzag[16] = {
+  0, 1, 4, 8,  5, 2, 3, 6,  9, 12, 13, 10,  7, 11, 14, 15
+};
+
+// 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
+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].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[kBands[++n]].probas_[0];
+      if (n == 16) return 16;
+    }
+    {        // non zero coeff
+      const VP8ProbaArray* const p_ctx = &prob[kBands[n + 1]].probas_[0];
+      int v;
+      if (!VP8GetBit(br, p[2])) {
+        v = 1;
+        p = p_ctx[1];
+      } else {
+        v = GetLargeValue(br, p);
+        p = p_ctx[2];
+      }
+      out[kZigzag[n]] = VP8GetSigned(br, v) * dq[n > 0];
+    }
+  }
+  return 16;
+}
+
+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) {
+  VP8BandProbas (* const bands)[NUM_BANDS] = dec->proba_.bands_;
+  const VP8BandProbas* ac_proba;
+  const VP8QuantMatrix* const q = &dec->dqm_[dec->segment_];
+  VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+  int16_t* dst = block->coeffs_;
+  VP8MB* const left_mb = dec->mb_info_ - 1;
+  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;
+
+  memset(dst, 0, 384 * sizeof(*dst));
+  if (!block->is_i4x4_) {    // parse DC
+    int16_t dc[16] = { 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_proba = bands[0];
+  } else {
+    first = 0;
+    ac_proba = bands[3];
+  }
+
+  tnz = mb->nz_ & 0x0f;
+  lnz = left_mb->nz_ & 0x0f;
+  for (y = 0; y < 4; ++y) {
+    int l = lnz & 1;
+    uint32_t nz_coeffs = 0;
+    for (x = 0; x < 4; ++x) {
+      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;
+    }
+    tnz >>= 4;
+    lnz = (lnz >> 1) | (l << 7);
+    non_zero_y = (non_zero_y << 8) | nz_coeffs;
+  }
+  out_t_nz = tnz;
+  out_l_nz = lnz >> 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 & 1;
+      for (x = 0; x < 2; ++x) {
+        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;
+      }
+      tnz >>= 2;
+      lnz = (lnz >> 1) | (l << 5);
+    }
+    // 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;
+  }
+  mb->nz_ = out_t_nz;
+  left_mb->nz_ = out_l_nz;
+
+  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
+}
+
+//------------------------------------------------------------------------------
+// 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 mb = dec->mb_info_ + dec->mb_x_;
+  VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+  int skip;
+
+  // 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]);
+  }
+  skip = dec->use_skip_proba_ ? VP8GetBit(br, dec->skip_p_) : 0;
+
+  VP8ParseIntraMode(br, dec);
+  if (br->eof_) {
+    return 0;
+  }
+
+  if (!skip) {
+    skip = ParseResiduals(dec, mb, token_br);
+  } else {
+    left->nz_ = mb->nz_ = 0;
+    if (!block->is_i4x4_) {
+      left->nz_dc_ = mb->nz_dc_ = 0;
+    }
+    block->non_zero_y_ = 0;
+    block->non_zero_uv_ = 0;
+  }
+
+  if (dec->filter_type_ > 0) {  // store filter info
+    VP8FInfo* const finfo = dec->f_info_ + dec->mb_x_;
+    *finfo = dec->fstrengths_[dec->segment_][block->is_i4x4_];
+    finfo->f_inner_ |= !skip;
+  }
+
+  return !token_br->eof_;
+}
+
+void VP8InitScanline(VP8Decoder* const dec) {
+  VP8MB* const left = dec->mb_info_ - 1;
+  left->nz_ = 0;
+  left->nz_dc_ = 0;
+  memset(dec->intra_l_, B_DC_PRED, sizeof(dec->intra_l_));
+  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)];
+    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.");
+      }
+    }
+    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->mt_method_ > 0) {
+    if (!WebPWorkerSync(&dec->worker_)) return 0;
+  }
+
+  // Finish
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  if (dec->layer_data_size_ > 0) {
+    if (!VP8DecodeLayer(dec)) {
+      return 0;
+    }
+  }
+#endif
+
+  return 1;
+}
+
+// Main entry point
+int VP8Decode(VP8Decoder* const dec, VP8Io* const io) {
+  int ok = 0;
+  if (dec == NULL) {
+    return 0;
+  }
+  if (io == NULL) {
+    return VP8SetError(dec, VP8_STATUS_INVALID_PARAM,
+                       "NULL VP8Io parameter in VP8Decode().");
+  }
+
+  if (!dec->ready_) {
+    if (!VP8GetHeaders(dec, io)) {
+      return 0;
+    }
+  }
+  assert(dec->ready_);
+
+  // Finish setting up the decoding parameter. Will call io->setup().
+  ok = (VP8EnterCritical(dec, io) == VP8_STATUS_OK);
+  if (ok) {   // good to go.
+    // Will allocate memory and prepare everything.
+    if (ok) ok = VP8InitFrame(dec, io);
+
+    // Main decoding loop
+    if (ok) ok = ParseFrame(dec, io);
+
+    // Exit.
+    ok &= VP8ExitCritical(dec, io);
+  }
+
+  if (!ok) {
+    VP8Clear(dec);
+    return 0;
+  }
+
+  dec->ready_ = 0;
+  return ok;
+}
+
+void VP8Clear(VP8Decoder* const dec) {
+  if (dec == NULL) {
+    return;
+  }
+  if (dec->mt_method_ > 0) {
+    WebPWorkerEnd(&dec->worker_);
+  }
+  ALPHDelete(dec->alph_dec_);
+  dec->alph_dec_ = NULL;
+  free(dec->mem_);
+  dec->mem_ = NULL;
+  dec->mem_size_ = 0;
+  memset(&dec->br_, 0, sizeof(dec->br_));
+  dec->ready_ = 0;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/dec/vp8i.h b/drivers/webp/dec/vp8i.h
new file mode 100644
index 0000000000..3f4cf297d9
--- /dev/null
+++ b/drivers/webp/dec/vp8i.h
@@ -0,0 +1,359 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// VP8 decoder: internal header.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_DEC_VP8I_H_
+#define WEBP_DEC_VP8I_H_
+
+#include <string.h>     // for memcpy()
+#include "./vp8li.h"
+#include "../utils/bit_reader.h"
+#include "../utils/random.h"
+#include "../utils/thread.h"
+#include "../dsp/dsp.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// Various defines and enums
+
+// version numbers
+#define DEC_MAJ_VERSION 0
+#define DEC_MIN_VERSION 4
+#define DEC_REV_VERSION 0
+
+// 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.
+// 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
+// top-left samples (from previously decoded blocks), along with four
+// extra top-right samples for luma (intra4x4 prediction only).
+// One possible layout is, using 32 * (17 + 9) bytes:
+//
+//   .+------   <- only 1 pixel high
+//   .|yyyyt.
+//   .|yyyyt.
+//   .|yyyyt.
+//   .|yyyy..
+//   .+--.+--   <- only 1 pixel high
+//   .|uu.|vv
+//   .|uu.|vv
+//
+// Every character is a 4x4 block, with legend:
+//  '.' = unused
+//  '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
+
+typedef struct {
+  uint8_t key_frame_;
+  uint8_t profile_;
+  uint8_t show_;
+  uint32_t partition_length_;
+} VP8FrameHeader;
+
+typedef struct {
+  uint16_t width_;
+  uint16_t height_;
+  uint8_t xscale_;
+  uint8_t yscale_;
+  uint8_t colorspace_;   // 0 = YCbCr
+  uint8_t clamp_type_;
+} VP8PictureHeader;
+
+// segment features
+typedef struct {
+  int use_segment_;
+  int update_map_;        // whether to update the segment map or not
+  int absolute_delta_;    // absolute or delta values for quantizer and filter
+  int8_t quantizer_[NUM_MB_SEGMENTS];        // quantization changes
+  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
+  VP8BandProbas bands_[NUM_TYPES][NUM_BANDS];
+} VP8Proba;
+
+// Filter parameters
+typedef struct {
+  int simple_;                  // 0=complex, 1=simple
+  int level_;                   // [0..63]
+  int sharpness_;               // [0..7]
+  int use_lf_delta_;
+  int ref_lf_delta_[NUM_REF_LF_DELTAS];
+  int mode_lf_delta_[NUM_MODE_LF_DELTAS];
+} VP8FilterHeader;
+
+//------------------------------------------------------------------------------
+// Informations about the macroblocks.
+
+typedef struct {  // filter specs
+  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 {  // 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_*)
+} 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 (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
+
+struct VP8Decoder {
+  VP8StatusCode status_;
+  int ready_;     // true if ready to decode a picture with VP8Decode()
+  const char* error_msg_;  // set when status_ is not OK.
+
+  // Main data source
+  VP8BitReader br_;
+
+  // headers
+  VP8FrameHeader   frm_hdr_;
+  VP8PictureHeader pic_hdr_;
+  VP8FilterHeader  filter_hdr_;
+  VP8SegmentHeader segment_hdr_;
+
+  // Worker
+  WebPWorker worker_;
+  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
+
+  // dimension, in macroblock units.
+  int mb_w_, mb_h_;
+
+  // Macroblock to process/filter, depending on cropping and filter_type.
+  int tl_mb_x_, tl_mb_y_;  // top-left MB that must be in-loop filtered
+  int br_mb_x_, br_mb_y_;  // last bottom-right MB that must be decoded
+
+  // number of partitions.
+  int num_parts_;
+  // per-partition boolean decoders.
+  VP8BitReader parts_[MAX_NUM_PARTITIONS];
+
+  // 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];
+
+  // probabilities
+  VP8Proba proba_;
+  int use_skip_proba_;
+  uint8_t skip_p_;
+
+  // 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 segment_;       // segment of the currently parsed block
+  VP8TopSamples* yuv_t_;  // top y/u/v 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_;
+  int cache_uv_stride_;
+
+  // main memory chunk for the above data. Persistent.
+  void* mem_;
+  size_t mem_size_;
+
+  // Per macroblock non-persistent infos.
+  int mb_x_, mb_y_;       // current position, in macroblock units
+  VP8MBData* mb_data_;    // parsed reconstruction data
+
+  // Filtering side-info
+  int filter_type_;                          // 0=off, 1=simple, 2=complex
+  VP8FInfo fstrengths_[NUM_MB_SEGMENTS][2];  // precalculated per-segment/type
+
+  // Alpha
+  struct ALPHDecoder* alph_dec_;  // alpha-plane decoder object
+  const uint8_t* alpha_data_;     // compressed alpha data (if present)
+  size_t alpha_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.
+
+  // extensions
+  int layer_colorspace_;
+  const uint8_t* layer_data_;   // compressed layer data (if present)
+  size_t layer_data_size_;
+};
+
+//------------------------------------------------------------------------------
+// internal functions. Not public.
+
+// in vp8.c
+int VP8SetError(VP8Decoder* const dec,
+                VP8StatusCode error, const char* const msg);
+
+// in tree.c
+void VP8ResetProba(VP8Proba* const proba);
+void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec);
+void VP8ParseIntraMode(VP8BitReader* const br,  VP8Decoder* const dec);
+
+// in quant.c
+void VP8ParseQuant(VP8Decoder* const dec);
+
+// in frame.c
+int VP8InitFrame(VP8Decoder* const dec, VP8Io* 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
+// if ok, otherwise sets and returns the error status on *dec.
+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);
+// 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);
+// 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.
+int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br);
+
+// in alpha.c
+const uint8_t* VP8DecompressAlphaRows(VP8Decoder* const dec,
+                                      int row, int num_rows);
+
+// in layer.c
+int VP8DecodeLayer(VP8Decoder* const dec);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_DEC_VP8I_H_ */
diff --git a/drivers/webp/dec/vp8l.c b/drivers/webp/dec/vp8l.c
new file mode 100644
index 0000000000..ea0254d7a8
--- /dev/null
+++ b/drivers/webp/dec/vp8l.c
@@ -0,0 +1,1380 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// main entry for the decoder
+//
+// 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/lossless.h"
+#include "../dsp/yuv.h"
+#include "../utils/alpha_processing.h"
+#include "../utils/huffman.h"
+#include "../utils/utils.h"
+
+#define NUM_ARGB_CACHE_ROWS          16
+
+static const int kCodeLengthLiterals = 16;
+static const int kCodeLengthRepeatCode = 16;
+static const int kCodeLengthExtraBits[3] = { 2, 3, 7 };
+static const int kCodeLengthRepeatOffsets[3] = { 3, 3, 11 };
+
+// -----------------------------------------------------------------------------
+//  Five Huffman codes are used at each meta code:
+//  1. green + length prefix codes + color cache codes,
+//  2. alpha,
+//  3. red,
+//  4. blue, and,
+//  5. distance prefix codes.
+typedef enum {
+  GREEN = 0,
+  RED   = 1,
+  BLUE  = 2,
+  ALPHA = 3,
+  DIST  = 4
+} HuffIndex;
+
+static const uint16_t kAlphabetSize[HUFFMAN_CODES_PER_META_CODE] = {
+  NUM_LITERAL_CODES + NUM_LENGTH_CODES,
+  NUM_LITERAL_CODES, NUM_LITERAL_CODES, NUM_LITERAL_CODES,
+  NUM_DISTANCE_CODES
+};
+
+
+#define NUM_CODE_LENGTH_CODES       19
+static const uint8_t kCodeLengthCodeOrder[NUM_CODE_LENGTH_CODES] = {
+  17, 18, 0, 1, 2, 3, 4, 5, 16, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
+};
+
+#define CODE_TO_PLANE_CODES        120
+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
+};
+
+static int DecodeImageStream(int xsize, int ysize,
+                             int is_level0,
+                             VP8LDecoder* const dec,
+                             uint32_t** const decoded_data);
+
+//------------------------------------------------------------------------------
+
+int VP8LCheckSignature(const uint8_t* const data, size_t size) {
+  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) {
+  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);
+  if (VP8LReadBits(br, VP8L_VERSION_BITS) != 0) return 0;
+  return 1;
+}
+
+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;
+    VP8LInitBitReader(&br, data, data_size);
+    if (!ReadImageInfo(&br, &w, &h, &a)) {
+      return 0;
+    }
+    if (width != NULL) *width = w;
+    if (height != NULL) *height = h;
+    if (has_alpha != NULL) *has_alpha = a;
+    return 1;
+  }
+}
+
+//------------------------------------------------------------------------------
+
+static WEBP_INLINE int GetCopyDistance(int distance_symbol,
+                                       VP8LBitReader* const br) {
+  int extra_bits, offset;
+  if (distance_symbol < 4) {
+    return distance_symbol + 1;
+  }
+  extra_bits = (distance_symbol - 2) >> 1;
+  offset = (2 + (distance_symbol & 1)) << extra_bits;
+  return offset + VP8LReadBits(br, extra_bits) + 1;
+}
+
+static WEBP_INLINE int GetCopyLength(int length_symbol,
+                                     VP8LBitReader* const br) {
+  // Length and distance prefixes are encoded the same way.
+  return GetCopyDistance(length_symbol, br);
+}
+
+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 = 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;  // 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 ReadSymbol, in order to pre-fetch enough bits.
+static WEBP_INLINE int ReadSymbol(const HuffmanTree* tree,
+                                  VP8LBitReader* const br) {
+  const HuffmanTreeNode* node = tree->root_;
+  uint32_t bits = VP8LPrefetchBits(br);
+  int bitpos = br->bit_pos_;
+  // Check if we find the bit combination from the Huffman lookup table.
+  const int lut_ix = bits & (HUFF_LUT - 1);
+  const int lut_bits = tree->lut_bits_[lut_ix];
+  if (lut_bits <= HUFF_LUT_BITS) {
+    VP8LSetBitPos(br, bitpos + lut_bits);
+    return tree->lut_symbol_[lut_ix];
+  }
+  node += tree->lut_jump_[lut_ix];
+  bitpos += HUFF_LUT_BITS;
+  bits >>= HUFF_LUT_BITS;
+
+  // Decode the value from a binary tree.
+  assert(node != NULL);
+  do {
+    node = HuffmanTreeNextNode(node, bits & 1);
+    bits >>= 1;
+    ++bitpos;
+  } while (HuffmanTreeNodeIsNotLeaf(node));
+  VP8LSetBitPos(br, bitpos);
+  return node->symbol_;
+}
+
+static int ReadHuffmanCodeLengths(
+    VP8LDecoder* const dec, const int* const code_length_code_lengths,
+    int num_symbols, int* const code_lengths) {
+  int ok = 0;
+  VP8LBitReader* const br = &dec->br_;
+  int symbol;
+  int max_symbol;
+  int prev_code_len = DEFAULT_CODE_LENGTH;
+  HuffmanTree tree;
+
+  if (!HuffmanTreeBuildImplicit(&tree, code_length_code_lengths,
+                                NUM_CODE_LENGTH_CODES)) {
+    dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+    return 0;
+  }
+
+  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 {
+    max_symbol = num_symbols;
+  }
+
+  symbol = 0;
+  while (symbol < num_symbols) {
+    int code_len;
+    if (max_symbol-- == 0) break;
+    VP8LFillBitWindow(br);
+    code_len = ReadSymbol(&tree, br);
+    if (code_len < kCodeLengthLiterals) {
+      code_lengths[symbol++] = code_len;
+      if (code_len != 0) prev_code_len = code_len;
+    } else {
+      const int use_prev = (code_len == kCodeLengthRepeatCode);
+      const int slot = code_len - kCodeLengthLiterals;
+      const int extra_bits = kCodeLengthExtraBits[slot];
+      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;
+        while (repeat-- > 0) code_lengths[symbol++] = length;
+      }
+    }
+  }
+  ok = 1;
+
+ End:
+  HuffmanTreeRelease(&tree);
+  return ok;
+}
+
+static int ReadHuffmanCode(int alphabet_size, VP8LDecoder* const dec,
+                           HuffmanTree* const tree) {
+  int ok = 0;
+  VP8LBitReader* const br = &dec->br_;
+  const int simple_code = VP8LReadBits(br, 1);
+
+  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;
+    // 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;
+    }
+    ok = HuffmanTreeBuildExplicit(tree, code_lengths, codes, symbols,
+                                  alphabet_size, num_symbols);
+  } 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;
+    if (num_codes > NUM_CODE_LENGTH_CODES) {
+      dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+      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) {
+    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);
+  }
+}
+
+static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
+                            int color_cache_bits, int allow_recursion) {
+  int i, j;
+  VP8LBitReader* const br = &dec->br_;
+  VP8LMetadata* const hdr = &dec->hdr_;
+  uint32_t* huffman_image = NULL;
+  HTreeGroup* htree_groups = NULL;
+  int num_htree_groups = 1;
+
+  if (allow_recursion && VP8LReadBits(br, 1)) {
+    // use meta Huffman codes.
+    const int huffman_precision = VP8LReadBits(br, 3) + 2;
+    const int huffman_xsize = VP8LSubSampleSize(xsize, huffman_precision);
+    const int huffman_ysize = VP8LSubSampleSize(ysize, huffman_precision);
+    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 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;
+
+  assert(num_htree_groups <= 0x10000);
+  htree_groups =
+      (HTreeGroup*)WebPSafeCalloc((uint64_t)num_htree_groups,
+                                  sizeof(*htree_groups));
+  if (htree_groups == NULL) {
+    dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  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) {
+      int alphabet_size = kAlphabetSize[j];
+      if (j == 0 && color_cache_bits > 0) {
+        alphabet_size += 1 << color_cache_bits;
+      }
+      if (!ReadHuffmanCode(alphabet_size, dec, htrees + j)) goto Error;
+    }
+  }
+
+  // All OK. Finalize pointers and return.
+  hdr->huffman_image_ = huffman_image;
+  hdr->num_htree_groups_ = num_htree_groups;
+  hdr->htree_groups_ = htree_groups;
+  return 1;
+
+ Error:
+  free(huffman_image);
+  DeleteHtreeGroups(htree_groups, num_htree_groups);
+  return 0;
+}
+
+//------------------------------------------------------------------------------
+// Scaling.
+
+static int AllocateAndInitRescaler(VP8LDecoder* const dec, VP8Io* const io) {
+  const int num_channels = 4;
+  const int in_width = io->mb_w;
+  const int out_width = io->scaled_width;
+  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;
+  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));
+  if (memory == NULL) {
+    dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+    return 0;
+  }
+  assert(dec->rescaler_memory == NULL);
+  dec->rescaler_memory = memory;
+
+  dec->rescaler = (WebPRescaler*)memory;
+  memory += sizeof(*dec->rescaler);
+  work = (int32_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);
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Export to ARGB
+
+// 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) {
+  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;
+  }
+  return num_lines_out;
+}
+
+// Emit scaled rows.
+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;
+  int num_lines_in = 0;
+  int num_lines_out = 0;
+  while (num_lines_in < mb_h) {
+    uint8_t* const row_in = in + num_lines_in * in_stride;
+    uint8_t* const row_out = out + num_lines_out * out_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;
+}
+
+// Emit rows without any scaling.
+static int EmitRows(WEBP_CSP_MODE colorspace,
+                    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;
+  uint8_t* row_out = out;
+  while (lines-- > 0) {
+    VP8LConvertFromBGRA((const uint32_t*)row_in, mb_w, colorspace, row_out);
+    row_in += in_stride;
+    row_out += out_stride;
+  }
+  return mb_h;  // Num rows out == num rows in.
+}
+
+//------------------------------------------------------------------------------
+// Export to YUVA
+
+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,
+                       YUV_HALF);
+    }
+  }
+
+  // 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, YUV_HALF << 2);
+        v[i] = VP8RGBToV(r, g, b, YUV_HALF << 2);
+      } else {             // odd lines: average with previous values
+        const int tmp_u = VP8RGBToU(r, g, b, YUV_HALF << 2);
+        const int tmp_v = VP8RGBToV(r, g, b, YUV_HALF << 2);
+        // 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, YUV_HALF << 2);
+        v[i] = VP8RGBToV(r, g, b, YUV_HALF << 2);
+      } else {             // odd lines (note: we could just skip this)
+        const int tmp_u = VP8RGBToU(r, g, b, YUV_HALF << 2);
+        const int tmp_v = VP8RGBToV(r, g, b, YUV_HALF << 2);
+        u[i] = (u[i] + tmp_u + 1) >> 1;
+        v[i] = (v[i] + tmp_v + 1) >> 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);
+  }
+}
+
+static int ExportYUVA(const VP8LDecoder* const dec, int y_pos) {
+  WebPRescaler* const rescaler = dec->rescaler;
+  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;
+  }
+  return num_lines_out;
+}
+
+static int EmitRescaledRowsYUVA(const VP8LDecoder* const dec,
+                                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 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 uint8_t* in, int in_stride,
+                        int mb_w, int num_rows) {
+  int y_pos = dec->last_out_row_;
+  while (num_rows-- > 0) {
+    ConvertToYUVA((const uint32_t*)in, mb_w, y_pos, dec->output_);
+    in += in_stride;
+    ++y_pos;
+  }
+  return y_pos;
+}
+
+//------------------------------------------------------------------------------
+// Cropping.
+
+// 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 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,
+                         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) {
+    y_end = io->crop_bottom;  // make sure we don't overflow on last row.
+  }
+  if (y_start < io->crop_top) {
+    const int delta = io->crop_top - y_start;
+    y_start = io->crop_top;
+    *in_data += delta * pixel_stride;
+  }
+  if (y_start >= y_end) return 0;  // Crop window is empty.
+
+  *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;
+  io->mb_h = y_end - y_start;
+  return 1;  // Non-empty crop window.
+}
+
+//------------------------------------------------------------------------------
+
+static WEBP_INLINE int GetMetaIndex(
+    const uint32_t* const image, int xsize, int bits, int x, int y) {
+  if (bits == 0) return 0;
+  return image[xsize * (y >> bits) + (x >> bits)];
+}
+
+static WEBP_INLINE HTreeGroup* GetHtreeGroupForPos(VP8LMetadata* const hdr,
+                                                   int x, int y) {
+  const int meta_index = GetMetaIndex(hdr->huffman_image_, hdr->huffman_xsize_,
+                                      hdr->huffman_subsample_bits_, x, y);
+  assert(meta_index < hdr->num_htree_groups_);
+  return hdr->htree_groups_ + meta_index;
+}
+
+//------------------------------------------------------------------------------
+// Main loop, with custom row-processing function
+
+typedef void (*ProcessRowsFunc)(VP8LDecoder* const dec, int row);
+
+static void ApplyInverseTransforms(VP8LDecoder* const dec, int num_rows,
+                                   const uint32_t* const rows) {
+  int n = dec->next_transform_;
+  const int cache_pixs = dec->width_ * num_rows;
+  const int start_row = dec->last_row_;
+  const int end_row = start_row + num_rows;
+  const uint32_t* rows_in = rows;
+  uint32_t* const rows_out = dec->argb_cache_;
+
+  // Inverse transforms.
+  // TODO: most transforms only need to operate on the cropped region only.
+  memcpy(rows_out, rows_in, cache_pixs * sizeof(*rows_out));
+  while (n-- > 0) {
+    VP8LTransform* const transform = &dec->transforms_[n];
+    VP8LInverseTransform(transform, start_row, end_row, rows_in, rows_out);
+    rows_in = rows_out;
+  }
+}
+
+// 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->pixels_ + dec->width_ * dec->last_row_;
+  const int num_rows = row - dec->last_row_;
+
+  if (num_rows <= 0) return;  // Nothing to be done.
+  ApplyInverseTransforms(dec, num_rows, rows);
+
+  // Emit output.
+  {
+    VP8Io* const io = dec->io_;
+    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_;
+      if (output->colorspace < MODE_YUV) {  // 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 ?
+            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_'.
+        dec->last_out_row_ += num_rows_out;
+      } else {                              // convert to YUVA
+        dec->last_out_row_ = io->use_scaling ?
+            EmitRescaledRowsYUVA(dec, rows_data, in_stride, io->mb_h) :
+            EmitRowsYUVA(dec, rows_data, in_stride, io->mb_w, io->mb_h);
+      }
+      assert(dec->last_out_row_ <= output->height);
+    }
+  }
+
+  // Update 'last_row_'.
+  dec->last_row_ = row;
+  assert(dec->last_row_ <= dec->height_);
+}
+
+// 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) {
+    const HuffmanTree* const htrees = hdr->htree_groups_[i].htrees_;
+    if (htrees[RED].num_nodes_ > 1) return 0;
+    if (htrees[BLUE].num_nodes_ > 1) return 0;
+    if (htrees[ALPHA].num_nodes_ > 1) 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;
+}
+
+static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
+                           int width, int height, int last_row) {
+  int ok = 1;
+  int row = dec->last_pixel_ / width;
+  int col = dec->last_pixel_ % width;
+  VP8LBitReader* const br = &dec->br_;
+  VP8LMetadata* const hdr = &dec->hdr_;
+  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 mask = hdr->huffman_mask_;
+  assert(htree_group != NULL);
+  assert(last_row <= height);
+  assert(Is8bOptimizable(hdr));
+
+  while (!br->eos_ && pos < last) {
+    int code;
+    // 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
+      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);
+      dist_code = GetCopyDistance(dist_symbol, br);
+      dist = PlaneCodeToDistance(width, dist_code);
+      if (pos >= dist && end - pos >= length) {
+        int i;
+        for (i = 0; i < length; ++i) data[pos + i] = data[pos + i - dist];
+      } 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;
+    }
+    ok = !br->error_;
+    if (!ok) goto End;
+  }
+  // Process the remaining rows corresponding to last row-block.
+  ExtractPalettedAlphaRows(dec, row);
+
+ End:
+  if (br->error_ || !ok || (br->eos_ && pos < end)) {
+    ok = 0;
+    dec->status_ = br->eos_ ? VP8_STATUS_SUSPENDED
+                            : VP8_STATUS_BITSTREAM_ERROR;
+  } else {
+    dec->last_pixel_ = (int)pos;
+    if (pos == end) dec->state_ = READ_DATA;
+  }
+  return ok;
+}
+
+static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
+                           int width, int height, int last_row,
+                           ProcessRowsFunc process_func) {
+  int ok = 1;
+  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_;
+  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_last <= src_end);
+
+  while (!br->eos_ && src < src_last) {
+    int code;
+    // 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);
+    }
+    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;
+      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:
+      ++src;
+      ++col;
+      if (col >= width) {
+        col = 0;
+        ++row;
+        if ((row % NUM_ARGB_CACHE_ROWS == 0) && (process_func != NULL)) {
+          process_func(dec, row);
+        }
+        if (color_cache != NULL) {
+          while (last_cached < src) {
+            VP8LColorCacheInsert(color_cache, *last_cached++);
+          }
+        }
+      }
+    } 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);
+      dist_code = GetCopyDistance(dist_symbol, br);
+      dist = PlaneCodeToDistance(width, dist_code);
+      if (src - data < (ptrdiff_t)dist || src_end - src < (ptrdiff_t)length) {
+        ok = 0;
+        goto End;
+      } else {
+        int i;
+        for (i = 0; i < length; ++i) src[i] = src[i - dist];
+        src += length;
+      }
+      col += length;
+      while (col >= width) {
+        col -= width;
+        ++row;
+        if ((row % NUM_ARGB_CACHE_ROWS == 0) && (process_func != NULL)) {
+          process_func(dec, row);
+        }
+      }
+      if (src < src_last) {
+        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
+      const int key = code - len_code_limit;
+      assert(color_cache != NULL);
+      while (last_cached < src) {
+        VP8LColorCacheInsert(color_cache, *last_cached++);
+      }
+      *src = VP8LColorCacheLookup(color_cache, key);
+      goto AdvanceByOne;
+    } else {  // Not reached
+      ok = 0;
+      goto End;
+    }
+    ok = !br->error_;
+    if (!ok) goto End;
+  }
+  // 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_SUSPENDED
+                            : VP8_STATUS_BITSTREAM_ERROR;
+  } else {
+    dec->last_pixel_ = (int)(src - data);
+    if (src == src_end) dec->state_ = READ_DATA;
+  }
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+// VP8LTransform
+
+static void ClearTransform(VP8LTransform* const transform) {
+  free(transform->data_);
+  transform->data_ = NULL;
+}
+
+// For security reason, we need to remap the color map to span
+// the total possible bundled values, and not just the num_colors.
+static int ExpandColorMap(int num_colors, VP8LTransform* const transform) {
+  int i;
+  const int final_num_colors = 1 << (8 >> transform->bits_);
+  uint32_t* const new_color_map =
+      (uint32_t*)WebPSafeMalloc((uint64_t)final_num_colors,
+                                sizeof(*new_color_map));
+  if (new_color_map == NULL) {
+    return 0;
+  } else {
+    uint8_t* const data = (uint8_t*)transform->data_;
+    uint8_t* const new_data = (uint8_t*)new_color_map;
+    new_color_map[0] = transform->data_[0];
+    for (i = 4; i < 4 * num_colors; ++i) {
+      // Equivalent to AddPixelEq(), on a byte-basis.
+      new_data[i] = (data[i] + new_data[i - 4]) & 0xff;
+    }
+    for (; i < 4 * final_num_colors; ++i)
+      new_data[i] = 0;  // black tail.
+    free(transform->data_);
+    transform->data_ = new_color_map;
+  }
+  return 1;
+}
+
+static int ReadTransform(int* const xsize, int const* ysize,
+                         VP8LDecoder* const dec) {
+  int ok = 1;
+  VP8LBitReader* const br = &dec->br_;
+  VP8LTransform* transform = &dec->transforms_[dec->next_transform_];
+  const VP8LImageTransformType type =
+      (VP8LImageTransformType)VP8LReadBits(br, 2);
+
+  // Each transform type can only be present once in the stream.
+  if (dec->transforms_seen_ & (1U << type)) {
+    return 0;  // Already there, let's not accept the second same transform.
+  }
+  dec->transforms_seen_ |= (1U << type);
+
+  transform->type_ = type;
+  transform->xsize_ = *xsize;
+  transform->ysize_ = *ysize;
+  transform->data_ = NULL;
+  ++dec->next_transform_;
+  assert(dec->next_transform_ <= NUM_TRANSFORMS);
+
+  switch (type) {
+    case PREDICTOR_TRANSFORM:
+    case CROSS_COLOR_TRANSFORM:
+      transform->bits_ = VP8LReadBits(br, 3) + 2;
+      ok = DecodeImageStream(VP8LSubSampleSize(transform->xsize_,
+                                               transform->bits_),
+                             VP8LSubSampleSize(transform->ysize_,
+                                               transform->bits_),
+                             0, dec, &transform->data_);
+      break;
+    case COLOR_INDEXING_TRANSFORM: {
+       const int num_colors = VP8LReadBits(br, 8) + 1;
+       const int bits = (num_colors > 16) ? 0
+                      : (num_colors > 4) ? 1
+                      : (num_colors > 2) ? 2
+                      : 3;
+       *xsize = VP8LSubSampleSize(transform->xsize_, bits);
+       transform->bits_ = bits;
+       ok = DecodeImageStream(num_colors, 1, 0, dec, &transform->data_);
+       ok = ok && ExpandColorMap(num_colors, transform);
+      break;
+    }
+    case SUBTRACT_GREEN:
+      break;
+    default:
+      assert(0);    // can't happen
+      break;
+  }
+
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+// VP8LMetadata
+
+static void InitMetadata(VP8LMetadata* const hdr) {
+  assert(hdr);
+  memset(hdr, 0, sizeof(*hdr));
+}
+
+static void ClearMetadata(VP8LMetadata* const hdr) {
+  assert(hdr);
+
+  free(hdr->huffman_image_);
+  DeleteHtreeGroups(hdr->htree_groups_, hdr->num_htree_groups_);
+  VP8LColorCacheClear(&hdr->color_cache_);
+  InitMetadata(hdr);
+}
+
+// -----------------------------------------------------------------------------
+// VP8LDecoder
+
+VP8LDecoder* VP8LNew(void) {
+  VP8LDecoder* const dec = (VP8LDecoder*)calloc(1, 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;
+}
+
+void VP8LClear(VP8LDecoder* const dec) {
+  int i;
+  if (dec == NULL) return;
+  ClearMetadata(&dec->hdr_);
+
+  free(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);
+  dec->rescaler_memory = NULL;
+
+  dec->output_ = NULL;   // leave no trace behind
+}
+
+void VP8LDelete(VP8LDecoder* const dec) {
+  if (dec != NULL) {
+    VP8LClear(dec);
+    free(dec);
+  }
+}
+
+static void UpdateDecoder(VP8LDecoder* const dec, int width, int height) {
+  VP8LMetadata* const hdr = &dec->hdr_;
+  const int num_bits = hdr->huffman_subsample_bits_;
+  dec->width_ = width;
+  dec->height_ = height;
+
+  hdr->huffman_xsize_ = VP8LSubSampleSize(width, num_bits);
+  hdr->huffman_mask_ = (num_bits == 0) ? ~0 : (1 << num_bits) - 1;
+}
+
+static int DecodeImageStream(int xsize, int ysize,
+                             int is_level0,
+                             VP8LDecoder* const dec,
+                             uint32_t** const decoded_data) {
+  int ok = 1;
+  int transform_xsize = xsize;
+  int transform_ysize = ysize;
+  VP8LBitReader* const br = &dec->br_;
+  VP8LMetadata* const hdr = &dec->hdr_;
+  uint32_t* data = NULL;
+  int color_cache_bits = 0;
+
+  // Read the transforms (may recurse).
+  if (is_level0) {
+    while (ok && VP8LReadBits(br, 1)) {
+      ok = ReadTransform(&transform_xsize, &transform_ysize, dec);
+    }
+  }
+
+  // Color cache
+  if (ok && VP8LReadBits(br, 1)) {
+    color_cache_bits = VP8LReadBits(br, 4);
+    ok = (color_cache_bits >= 1 && color_cache_bits <= MAX_CACHE_BITS);
+    if (!ok) {
+      dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+      goto End;
+    }
+  }
+
+  // Read the Huffman codes (may recurse).
+  ok = ok && ReadHuffmanCodes(dec, transform_xsize, transform_ysize,
+                              color_cache_bits, is_level0);
+  if (!ok) {
+    dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+    goto End;
+  }
+
+  // Finish setting up the color-cache
+  if (color_cache_bits > 0) {
+    hdr->color_cache_size_ = 1 << color_cache_bits;
+    if (!VP8LColorCacheInit(&hdr->color_cache_, color_cache_bits)) {
+      dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+      ok = 0;
+      goto End;
+    }
+  } else {
+    hdr->color_cache_size_ = 0;
+  }
+  UpdateDecoder(dec, transform_xsize, transform_ysize);
+
+  if (is_level0) {   // level 0 complete
+    dec->state_ = READ_HDR;
+    goto End;
+  }
+
+  {
+    const uint64_t total_size = (uint64_t)transform_xsize * transform_ysize;
+    data = (uint32_t*)WebPSafeMalloc(total_size, sizeof(*data));
+    if (data == NULL) {
+      dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
+      ok = 0;
+      goto End;
+    }
+  }
+
+  // Use the Huffman trees to decode the LZ77 encoded data.
+  ok = DecodeImageData(dec, data, transform_xsize, transform_ysize,
+                       transform_ysize, NULL);
+  ok = ok && !br->error_;
+
+ End:
+
+  if (!ok) {
+    free(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;
+    } else {
+      // We allocate image data in this function only for transforms. At level 0
+      // (that is: not the transforms), we shouldn't have allocated anything.
+      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 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. 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->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->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.
+static void ExtractAlphaRows(VP8LDecoder* const dec, int row) {
+  const int num_rows = row - 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);
+
+  // Extract alpha (which is stored in the green plane).
+  {
+    const int width = dec->io_->width;      // the final width (!= dec->width_)
+    const int cache_pixs = width * num_rows;
+    uint8_t* const dst = (uint8_t*)dec->io_->opaque + width * dec->last_row_;
+    const uint32_t* const src = dec->argb_cache_;
+    int i;
+    for (i = 0; i < cache_pixs; ++i) dst[i] = (src[i] >> 8) & 0xff;
+  }
+  dec->last_row_ = dec->last_out_row_ = row;
+}
+
+int VP8LDecodeAlphaHeader(ALPHDecoder* const alph_dec,
+                          const uint8_t* const data, size_t data_size,
+                          uint8_t* const output) {
+  int ok = 0;
+  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 = 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(alph_dec->width_, alph_dec->height_, 1, dec, NULL)) {
+    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_);
+  }
+
+  if (!ok) goto Err;
+
+  dec->action_ = READ_DATA;
+  return 1;
+
+ Err:
+  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(dec->action_ == READ_DATA);
+  assert(last_row <= dec->height_);
+
+  // 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);
+}
+
+//------------------------------------------------------------------------------
+
+int VP8LDecodeHeader(VP8LDecoder* const dec, VP8Io* const io) {
+  int width, height, has_alpha;
+
+  if (dec == NULL) return 0;
+  if (io == NULL) {
+    dec->status_ = VP8_STATUS_INVALID_PARAM;
+    return 0;
+  }
+
+  dec->io_ = io;
+  dec->status_ = VP8_STATUS_OK;
+  VP8LInitBitReader(&dec->br_, io->data, io->data_size);
+  if (!ReadImageInfo(&dec->br_, &width, &height, &has_alpha)) {
+    dec->status_ = VP8_STATUS_BITSTREAM_ERROR;
+    goto Error;
+  }
+  dec->state_ = READ_DIM;
+  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;
+}
+
+int VP8LDecodeImage(VP8LDecoder* const dec) {
+  VP8Io* io = NULL;
+  WebPDecParams* params = NULL;
+
+  // Sanity checks.
+  if (dec == NULL) return 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 (!AllocateInternalBuffers32b(dec, io->width)) goto Err;
+
+  if (io->use_scaling && !AllocateAndInitRescaler(dec, io)) goto Err;
+
+  // Decode.
+  dec->action_ = READ_DATA;
+  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:
+  VP8LClear(dec);
+  assert(dec->status_ != VP8_STATUS_OK);
+  return 0;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/dec/vp8li.h b/drivers/webp/dec/vp8li.h
new file mode 100644
index 0000000000..afa294db1e
--- /dev/null
+++ b/drivers/webp/dec/vp8li.h
@@ -0,0 +1,137 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Lossless decoder: internal header.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+//         Vikas Arora(vikaas.arora@gmail.com)
+
+#ifndef WEBP_DEC_VP8LI_H_
+#define WEBP_DEC_VP8LI_H_
+
+#include <string.h>     // for memcpy()
+#include "./webpi.h"
+#include "../utils/bit_reader.h"
+#include "../utils/color_cache.h"
+#include "../utils/huffman.h"
+#include "../webp/format_constants.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum {
+  READ_DATA = 0,
+  READ_HDR = 1,
+  READ_DIM = 2
+} VP8LDecodeState;
+
+typedef struct VP8LTransform VP8LTransform;
+struct VP8LTransform {
+  VP8LImageTransformType type_;   // transform type.
+  int                    bits_;   // subsampling bits defining transform window.
+  int                    xsize_;  // transform window X index.
+  int                    ysize_;  // transform window Y index.
+  uint32_t              *data_;   // transform data.
+};
+
+typedef struct {
+  HuffmanTree htrees_[HUFFMAN_CODES_PER_META_CODE];
+} HTreeGroup;
+
+typedef struct {
+  int             color_cache_size_;
+  VP8LColorCache  color_cache_;
+
+  int             huffman_mask_;
+  int             huffman_subsample_bits_;
+  int             huffman_xsize_;
+  uint32_t       *huffman_image_;
+  int             num_htree_groups_;
+  HTreeGroup     *htree_groups_;
+} VP8LMetadata;
+
+typedef struct VP8LDecoder VP8LDecoder;
+struct VP8LDecoder {
+  VP8StatusCode    status_;
+  VP8LDecodeState  action_;
+  VP8LDecodeState  state_;
+  VP8Io           *io_;
+
+  const WebPDecBuffer *output_;    // shortcut to io->opaque->output
+
+  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              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_;
+
+  int              next_transform_;
+  VP8LTransform    transforms_[NUM_TRANSFORMS];
+  // or'd bitset storing the transforms types.
+  uint32_t         transforms_seen_;
+
+  uint8_t         *rescaler_memory;  // Working memory for rescaling work.
+  WebPRescaler    *rescaler;         // Common rescaler for all channels.
+};
+
+//------------------------------------------------------------------------------
+// internal functions. Not public.
+
+struct ALPHDecoder;  // Defined in dec/alphai.h.
+
+// in vp8l.c
+
+// 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);
+
+// Decodes the image header. Returns false in case of error.
+int VP8LDecodeHeader(VP8LDecoder* const dec, VP8Io* const io);
+
+// Decodes an image. It's required to decode the lossless header before calling
+// this function. Returns false in case of error, with updated dec->status_.
+int VP8LDecodeImage(VP8LDecoder* const dec);
+
+// Resets the decoder in its initial state, reclaiming memory.
+// Preserves the dec->status_ value.
+void VP8LClear(VP8LDecoder* const dec);
+
+// Clears and deallocate a lossless decoder instance.
+void VP8LDelete(VP8LDecoder* const dec);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_DEC_VP8LI_H_ */
diff --git a/drivers/webp/dec/webp.c b/drivers/webp/dec/webp.c
new file mode 100644
index 0000000000..fda88bda26
--- /dev/null
+++ b/drivers/webp/dec/webp.c
@@ -0,0 +1,822 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Main decoding functions for WEBP images.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+
+#include "./vp8i.h"
+#include "./vp8li.h"
+#include "./webpi.h"
+#include "../webp/mux_types.h"  // ALPHA_FLAG
+
+//------------------------------------------------------------------------------
+// RIFF layout is:
+//   Offset  tag
+//   0...3   "RIFF" 4-byte tag
+//   4...7   size of image data (including metadata) starting at offset 8
+//   8...11  "WEBP"   our form-type signature
+// The RIFF container (12 bytes) is followed by appropriate chunks:
+//   12..15  "VP8 ": 4-bytes tags, signaling the use of VP8 video format
+//   16..19  size of the raw VP8 image data, starting at offset 20
+//   20....  the VP8 bytes
+// Or,
+//   12..15  "VP8L": 4-bytes tags, signaling the use of VP8L lossless format
+//   16..19  size of the raw VP8L image data, starting at offset 20
+//   20....  the VP8L bytes
+// Or,
+//   12..15  "VP8X": 4-bytes tags, describing the extended-VP8 chunk.
+//   16..19  size of the VP8X chunk starting at offset 20.
+//   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, 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.
+// 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 riff_size) {
+  assert(data != NULL);
+  assert(data_size != NULL);
+  assert(riff_size != NULL);
+
+  *riff_size = 0;  // Default: no RIFF present.
+  if (*data_size >= RIFF_HEADER_SIZE && !memcmp(*data, "RIFF", TAG_SIZE)) {
+    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);
+      // 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;
+      }
+      // We have a RIFF container. Skip it.
+      *riff_size = size;
+      *data += RIFF_HEADER_SIZE;
+      *data_size -= RIFF_HEADER_SIZE;
+    }
+  }
+  return VP8_STATUS_OK;
+}
+
+// Validates the VP8X header and skips over it.
+// Returns VP8_STATUS_BITSTREAM_ERROR for invalid VP8X header,
+//         VP8_STATUS_NOT_ENOUGH_DATA in case of insufficient data, and
+//         VP8_STATUS_OK otherwise.
+// If a VP8X chunk is found, found_vp8x is set to true and *width_ptr,
+// *height_ptr and *flags_ptr are set to the corresponding values extracted
+// from the VP8X chunk.
+static VP8StatusCode ParseVP8X(const uint8_t** const data,
+                               size_t* const data_size,
+                               int* const found_vp8x,
+                               int* const width_ptr, int* const height_ptr,
+                               uint32_t* const flags_ptr) {
+  const uint32_t vp8x_size = CHUNK_HEADER_SIZE + VP8X_CHUNK_SIZE;
+  assert(data != NULL);
+  assert(data_size != NULL);
+  assert(found_vp8x != NULL);
+
+  *found_vp8x = 0;
+
+  if (*data_size < CHUNK_HEADER_SIZE) {
+    return VP8_STATUS_NOT_ENOUGH_DATA;  // Insufficient data.
+  }
+
+  if (!memcmp(*data, "VP8X", TAG_SIZE)) {
+    int width, height;
+    uint32_t flags;
+    const uint32_t chunk_size = get_le32(*data + TAG_SIZE);
+    if (chunk_size != VP8X_CHUNK_SIZE) {
+      return VP8_STATUS_BITSTREAM_ERROR;  // Wrong chunk size.
+    }
+
+    // Verify if enough data is available to validate the VP8X chunk.
+    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);
+    if (width * (uint64_t)height >= MAX_IMAGE_AREA) {
+      return VP8_STATUS_BITSTREAM_ERROR;  // image is too large
+    }
+
+    if (flags_ptr != NULL) *flags_ptr = flags;
+    if (width_ptr != NULL) *width_ptr = width;
+    if (height_ptr != NULL) *height_ptr = height;
+    // Skip over VP8X header bytes.
+    *data += vp8x_size;
+    *data_size -= vp8x_size;
+    *found_vp8x = 1;
+  }
+  return VP8_STATUS_OK;
+}
+
+// Skips to the next VP8/VP8L chunk header in the data given the size of the
+// RIFF chunk 'riff_size'.
+// Returns VP8_STATUS_BITSTREAM_ERROR if any invalid chunk size is encountered,
+//         VP8_STATUS_NOT_ENOUGH_DATA in case of insufficient data, and
+//         VP8_STATUS_OK otherwise.
+// If an alpha chunk is found, *alpha_data and *alpha_size are set
+// appropriately.
+static VP8StatusCode ParseOptionalChunks(const uint8_t** const data,
+                                         size_t* const data_size,
+                                         size_t const riff_size,
+                                         const uint8_t** const alpha_data,
+                                         size_t* const alpha_size) {
+  const uint8_t* buf;
+  size_t buf_size;
+  uint32_t total_size = TAG_SIZE +           // "WEBP".
+                        CHUNK_HEADER_SIZE +  // "VP8Xnnnn".
+                        VP8X_CHUNK_SIZE;     // data.
+  assert(data != NULL);
+  assert(data_size != NULL);
+  buf = *data;
+  buf_size = *data_size;
+
+  assert(alpha_data != NULL);
+  assert(alpha_size != NULL);
+  *alpha_data = NULL;
+  *alpha_size = 0;
+
+  while (1) {
+    uint32_t chunk_size;
+    uint32_t disk_chunk_size;   // chunk_size with padding
+
+    *data = buf;
+    *data_size = buf_size;
+
+    if (buf_size < CHUNK_HEADER_SIZE) {  // Insufficient data.
+      return VP8_STATUS_NOT_ENOUGH_DATA;
+    }
+
+    chunk_size = get_le32(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;
+
+    // Check that total bytes skipped so far does not exceed riff_size.
+    if (riff_size > 0 && (total_size > riff_size)) {
+      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;
+    }
+
+    if (!memcmp(buf, "ALPH", TAG_SIZE)) {         // A valid ALPH header.
+      *alpha_data = buf + CHUNK_HEADER_SIZE;
+      *alpha_size = chunk_size;
+    }
+
+    // We have a full and valid chunk; skip it.
+    buf += disk_chunk_size;
+    buf_size -= disk_chunk_size;
+  }
+}
+
+// Validates the VP8/VP8L Header ("VP8 nnnn" or "VP8L nnnn") and skips over it.
+// Returns VP8_STATUS_BITSTREAM_ERROR for invalid (chunk larger than
+//         riff_size) VP8/VP8L header,
+//         VP8_STATUS_NOT_ENOUGH_DATA in case of insufficient data, and
+//         VP8_STATUS_OK otherwise.
+// If a VP8/VP8L chunk is found, *chunk_size is set to the total number of bytes
+// 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,
+                                    int* const is_lossless) {
+  const uint8_t* const data = *data_ptr;
+  const int is_vp8 = !memcmp(data, "VP8 ", TAG_SIZE);
+  const int is_vp8l = !memcmp(data, "VP8L", TAG_SIZE);
+  const uint32_t minimal_size =
+      TAG_SIZE + CHUNK_HEADER_SIZE;  // "WEBP" + "VP8 nnnn" OR
+                                     // "WEBP" + "VP8Lnnnn"
+  assert(data != NULL);
+  assert(data_size != NULL);
+  assert(chunk_size != NULL);
+  assert(is_lossless != NULL);
+
+  if (*data_size < CHUNK_HEADER_SIZE) {
+    return VP8_STATUS_NOT_ENOUGH_DATA;  // Insufficient data.
+  }
+
+  if (is_vp8 || is_vp8l) {
+    // Bitstream contains VP8/VP8L header.
+    const uint32_t size = get_le32(data + TAG_SIZE);
+    if ((riff_size >= minimal_size) && (size > riff_size - minimal_size)) {
+      return VP8_STATUS_BITSTREAM_ERROR;  // Inconsistent size information.
+    }
+    // Skip over CHUNK_HEADER_SIZE bytes from VP8/VP8L Header.
+    *chunk_size = size;
+    *data_ptr += CHUNK_HEADER_SIZE;
+    *data_size -= CHUNK_HEADER_SIZE;
+    *is_lossless = is_vp8l;
+  } else {
+    // Raw VP8/VP8L bitstream (no header).
+    *is_lossless = VP8LCheckSignature(data, *data_size);
+    *chunk_size = *data_size;
+  }
+
+  return VP8_STATUS_OK;
+}
+
+//------------------------------------------------------------------------------
+
+// Fetch '*width', '*height', '*has_alpha' and fill out 'headers' based on
+// 'data'. All the output parameters may be NULL. If 'headers' is NULL only the
+// minimal amount will be read to fetch the remaining parameters.
+// If 'headers' is non-NULL this function will attempt to locate both alpha
+// data (with or without a VP8X chunk) and the bitstream chunk (VP8/VP8L).
+// Note: The following chunk sequences (before the raw VP8/VP8L data) are
+// considered valid by this function:
+// RIFF + VP8(L)
+// RIFF + VP8X + (optional chunks) + VP8(L)
+// ALPH + VP8 <-- Not a valid WebP format: only allowed for internal purpose.
+// VP8(L)     <-- Not a valid WebP format: only allowed for internal purpose.
+static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
+                                          size_t data_size,
+                                          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;
+
+  VP8StatusCode status;
+  WebPHeaderStructure hdrs;
+
+  if (data == NULL || data_size < RIFF_HEADER_SIZE) {
+    return VP8_STATUS_NOT_ENOUGH_DATA;
+  }
+  memset(&hdrs, 0, sizeof(hdrs));
+  hdrs.data = data;
+  hdrs.data_size = data_size;
+
+  // Skip over RIFF header.
+  status = ParseRIFF(&data, &data_size, &hdrs.riff_size);
+  if (status != VP8_STATUS_OK) {
+    return status;   // Wrong RIFF header / insufficient data.
+  }
+  found_riff = (hdrs.riff_size > 0);
+
+  // Skip over VP8X.
+  {
+    uint32_t flags = 0;
+    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);
+    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) {
+    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) ||
+      (!found_riff && !found_vp8x && !memcmp(data, "ALPH", TAG_SIZE))) {
+    status = ParseOptionalChunks(&data, &data_size, hdrs.riff_size,
+                                 &hdrs.alpha_data, &hdrs.alpha_data_size);
+    if (status != VP8_STATUS_OK) {
+      goto ReturnWidthHeight;  // Invalid chunk size / insufficient data.
+    }
+  }
+
+  // Skip over VP8/VP8L header.
+  status = ParseVP8Header(&data, &data_size, hdrs.riff_size,
+                          &hdrs.compressed_size, &hdrs.is_lossless);
+  if (status != VP8_STATUS_OK) {
+    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) {
+      status = VP8_STATUS_NOT_ENOUGH_DATA;
+      goto ReturnWidthHeight;
+    }
+    // Validates raw VP8 data.
+    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) {
+      status = VP8_STATUS_NOT_ENOUGH_DATA;
+      goto ReturnWidthHeight;
+    }
+    // Validates raw VP8L data.
+    if (!VP8LGetInfo(data, data_size, &image_width, &image_height, has_alpha)) {
+      return VP8_STATUS_BITSTREAM_ERROR;
+    }
+  }
+  // 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;
+    headers->offset = data - headers->data;
+    assert((uint64_t)(data - headers->data) < MAX_CHUNK_PAYLOAD);
+    assert(headers->offset == headers->data_size - data_size);
+  }
+ 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.
+  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 != NULL) {
+    memset(params, 0, sizeof(*params));
+  }
+}
+
+//------------------------------------------------------------------------------
+// "Into" decoding variants
+
+// Main flow
+static VP8StatusCode DecodeInto(const uint8_t* const data, size_t data_size,
+                                WebPDecParams* const params) {
+  VP8StatusCode status;
+  VP8Io io;
+  WebPHeaderStructure headers;
+
+  headers.data = data;
+  headers.data_size = data_size;
+  status = WebPParseHeaders(&headers);   // Process Pre-VP8 chunks.
+  if (status != VP8_STATUS_OK) {
+    return status;
+  }
+
+  assert(params != NULL);
+  VP8InitIo(&io);
+  io.data = headers.data + headers.offset;
+  io.data_size = headers.data_size - headers.offset;
+  WebPInitCustomIo(params, &io);  // Plug the I/O functions.
+
+  if (!headers.is_lossless) {
+    VP8Decoder* const dec = VP8New();
+    if (dec == NULL) {
+      return VP8_STATUS_OUT_OF_MEMORY;
+    }
+    dec->alpha_data_ = headers.alpha_data;
+    dec->alpha_data_size_ = headers.alpha_data_size;
+
+    // Decode bitstream header, update io->width/io->height.
+    if (!VP8GetHeaders(dec, &io)) {
+      status = dec->status_;   // An error occurred. Grab error status.
+    } else {
+      // Allocate/check output buffers.
+      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_;
+        }
+      }
+    }
+    VP8Delete(dec);
+  } else {
+    VP8LDecoder* const dec = VP8LNew();
+    if (dec == NULL) {
+      return VP8_STATUS_OUT_OF_MEMORY;
+    }
+    if (!VP8LDecodeHeader(dec, &io)) {
+      status = dec->status_;   // An error occurred. Grab error status.
+    } else {
+      // Allocate/check output buffers.
+      status = WebPAllocateDecBuffer(io.width, io.height, params->options,
+                                     params->output);
+      if (status == VP8_STATUS_OK) {  // Decode
+        if (!VP8LDecodeImage(dec)) {
+          status = dec->status_;
+        }
+      }
+    }
+    VP8LDelete(dec);
+  }
+
+  if (status != VP8_STATUS_OK) {
+    WebPFreeDecBuffer(params->output);
+  }
+  return status;
+}
+
+// Helpers
+static uint8_t* DecodeIntoRGBABuffer(WEBP_CSP_MODE colorspace,
+                                     const uint8_t* const data,
+                                     size_t data_size,
+                                     uint8_t* const rgba,
+                                     int stride, size_t size) {
+  WebPDecParams params;
+  WebPDecBuffer buf;
+  if (rgba == NULL) {
+    return NULL;
+  }
+  WebPInitDecBuffer(&buf);
+  WebPResetDecParams(&params);
+  params.output = &buf;
+  buf.colorspace    = colorspace;
+  buf.u.RGBA.rgba   = rgba;
+  buf.u.RGBA.stride = stride;
+  buf.u.RGBA.size   = size;
+  buf.is_external_memory = 1;
+  if (DecodeInto(data, data_size, &params) != VP8_STATUS_OK) {
+    return NULL;
+  }
+  return rgba;
+}
+
+uint8_t* WebPDecodeRGBInto(const uint8_t* data, size_t data_size,
+                           uint8_t* output, size_t size, int stride) {
+  return DecodeIntoRGBABuffer(MODE_RGB, data, data_size, output, stride, size);
+}
+
+uint8_t* WebPDecodeRGBAInto(const uint8_t* data, size_t data_size,
+                            uint8_t* output, size_t size, int stride) {
+  return DecodeIntoRGBABuffer(MODE_RGBA, data, data_size, output, stride, size);
+}
+
+uint8_t* WebPDecodeARGBInto(const uint8_t* data, size_t data_size,
+                            uint8_t* output, size_t size, int stride) {
+  return DecodeIntoRGBABuffer(MODE_ARGB, data, data_size, output, stride, size);
+}
+
+uint8_t* WebPDecodeBGRInto(const uint8_t* data, size_t data_size,
+                           uint8_t* output, size_t size, int stride) {
+  return DecodeIntoRGBABuffer(MODE_BGR, data, data_size, output, stride, size);
+}
+
+uint8_t* WebPDecodeBGRAInto(const uint8_t* data, size_t data_size,
+                            uint8_t* output, size_t size, int stride) {
+  return DecodeIntoRGBABuffer(MODE_BGRA, data, data_size, output, stride, size);
+}
+
+uint8_t* WebPDecodeYUVInto(const uint8_t* data, size_t data_size,
+                           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) {
+  WebPDecParams params;
+  WebPDecBuffer output;
+  if (luma == NULL) return NULL;
+  WebPInitDecBuffer(&output);
+  WebPResetDecParams(&params);
+  params.output = &output;
+  output.colorspace      = MODE_YUV;
+  output.u.YUVA.y        = luma;
+  output.u.YUVA.y_stride = luma_stride;
+  output.u.YUVA.y_size   = luma_size;
+  output.u.YUVA.u        = u;
+  output.u.YUVA.u_stride = u_stride;
+  output.u.YUVA.u_size   = u_size;
+  output.u.YUVA.v        = v;
+  output.u.YUVA.v_stride = v_stride;
+  output.u.YUVA.v_size   = v_size;
+  output.is_external_memory = 1;
+  if (DecodeInto(data, data_size, &params) != VP8_STATUS_OK) {
+    return NULL;
+  }
+  return luma;
+}
+
+//------------------------------------------------------------------------------
+
+static uint8_t* Decode(WEBP_CSP_MODE mode, const uint8_t* const data,
+                       size_t data_size, int* const width, int* const height,
+                       WebPDecBuffer* const keep_info) {
+  WebPDecParams params;
+  WebPDecBuffer output;
+
+  WebPInitDecBuffer(&output);
+  WebPResetDecParams(&params);
+  params.output = &output;
+  output.colorspace = mode;
+
+  // Retrieve (and report back) the required dimensions from bitstream.
+  if (!WebPGetInfo(data, data_size, &output.width, &output.height)) {
+    return NULL;
+  }
+  if (width != NULL) *width = output.width;
+  if (height != NULL) *height = output.height;
+
+  // Decode
+  if (DecodeInto(data, data_size, &params) != VP8_STATUS_OK) {
+    return NULL;
+  }
+  if (keep_info != NULL) {    // keep track of the side-info
+    WebPCopyDecBuffer(&output, keep_info);
+  }
+  // return decoded samples (don't clear 'output'!)
+  return WebPIsRGBMode(mode) ? output.u.RGBA.rgba : output.u.YUVA.y;
+}
+
+uint8_t* WebPDecodeRGB(const uint8_t* data, size_t data_size,
+                       int* width, int* height) {
+  return Decode(MODE_RGB, data, data_size, width, height, NULL);
+}
+
+uint8_t* WebPDecodeRGBA(const uint8_t* data, size_t data_size,
+                        int* width, int* height) {
+  return Decode(MODE_RGBA, data, data_size, width, height, NULL);
+}
+
+uint8_t* WebPDecodeARGB(const uint8_t* data, size_t data_size,
+                        int* width, int* height) {
+  return Decode(MODE_ARGB, data, data_size, width, height, NULL);
+}
+
+uint8_t* WebPDecodeBGR(const uint8_t* data, size_t data_size,
+                       int* width, int* height) {
+  return Decode(MODE_BGR, data, data_size, width, height, NULL);
+}
+
+uint8_t* WebPDecodeBGRA(const uint8_t* data, size_t data_size,
+                        int* width, int* height) {
+  return Decode(MODE_BGRA, data, data_size, width, height, NULL);
+}
+
+uint8_t* WebPDecodeYUV(const uint8_t* data, size_t data_size,
+                       int* width, int* height, uint8_t** u, uint8_t** v,
+                       int* stride, int* uv_stride) {
+  WebPDecBuffer output;   // only to preserve the side-infos
+  uint8_t* const out = Decode(MODE_YUV, data, data_size,
+                              width, height, &output);
+
+  if (out != NULL) {
+    const WebPYUVABuffer* const buf = &output.u.YUVA;
+    *u = buf->u;
+    *v = buf->v;
+    *stride = buf->y_stride;
+    *uv_stride = buf->u_stride;
+    assert(buf->u_stride == buf->v_stride);
+  }
+  return out;
+}
+
+static void DefaultFeatures(WebPBitstreamFeatures* const features) {
+  assert(features != NULL);
+  memset(features, 0, sizeof(*features));
+}
+
+static VP8StatusCode GetFeatures(const uint8_t* const data, size_t data_size,
+                                 WebPBitstreamFeatures* const features) {
+  if (features == NULL || data == NULL) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  DefaultFeatures(features);
+
+  // Only parse enough of the data to retrieve the features.
+  return ParseHeadersInternal(data, data_size,
+                              &features->width, &features->height,
+                              &features->has_alpha, &features->has_animation,
+                              &features->format, NULL);
+}
+
+//------------------------------------------------------------------------------
+// WebPGetInfo()
+
+int WebPGetInfo(const uint8_t* data, size_t data_size,
+                int* width, int* height) {
+  WebPBitstreamFeatures features;
+
+  if (GetFeatures(data, data_size, &features) != VP8_STATUS_OK) {
+    return 0;
+  }
+
+  if (width != NULL) {
+    *width  = features.width;
+  }
+  if (height != NULL) {
+    *height = features.height;
+  }
+
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Advance decoding API
+
+int WebPInitDecoderConfigInternal(WebPDecoderConfig* config,
+                                  int version) {
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_DECODER_ABI_VERSION)) {
+    return 0;   // version mismatch
+  }
+  if (config == NULL) {
+    return 0;
+  }
+  memset(config, 0, sizeof(*config));
+  DefaultFeatures(&config->input);
+  WebPInitDecBuffer(&config->output);
+  return 1;
+}
+
+VP8StatusCode WebPGetFeaturesInternal(const uint8_t* data, size_t data_size,
+                                      WebPBitstreamFeatures* features,
+                                      int version) {
+  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;
+  }
+  return GetFeatures(data, data_size, features);
+}
+
+VP8StatusCode WebPDecode(const uint8_t* data, size_t data_size,
+                         WebPDecoderConfig* config) {
+  WebPDecParams params;
+  VP8StatusCode status;
+
+  if (config == NULL) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+
+  status = GetFeatures(data, data_size, &config->input);
+  if (status != VP8_STATUS_OK) {
+    if (status == VP8_STATUS_NOT_ENOUGH_DATA) {
+      return VP8_STATUS_BITSTREAM_ERROR;  // Not-enough-data treated as error.
+    }
+    return status;
+  }
+
+  WebPResetDecParams(&params);
+  params.output = &config->output;
+  params.options = &config->options;
+  status = DecodeInto(data, data_size, &params);
+
+  return status;
+}
+
+//------------------------------------------------------------------------------
+// Cropping and rescaling.
+
+int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
+                          VP8Io* const io, WEBP_CSP_MODE src_colorspace) {
+  const int W = io->width;
+  const int H = io->height;
+  int x = 0, y = 0, w = W, h = H;
+
+  // Cropping
+  io->use_cropping = (options != NULL) && (options->use_cropping > 0);
+  if (io->use_cropping) {
+    w = options->crop_width;
+    h = options->crop_height;
+    x = options->crop_left;
+    y = options->crop_top;
+    if (!WebPIsRGBMode(src_colorspace)) {   // only snap for YUV420 or YUV422
+      x &= ~1;
+      y &= ~1;    // TODO(later): only for YUV420, not YUV422.
+    }
+    if (x < 0 || y < 0 || w <= 0 || h <= 0 || x + w > W || y + h > H) {
+      return 0;  // out of frame boundary error
+    }
+  }
+  io->crop_left   = x;
+  io->crop_top    = y;
+  io->crop_right  = x + w;
+  io->crop_bottom = y + h;
+  io->mb_w = w;
+  io->mb_h = h;
+
+  // Scaling
+  io->use_scaling = (options != NULL) && (options->use_scaling > 0);
+  if (io->use_scaling) {
+    if (options->scaled_width <= 0 || options->scaled_height <= 0) {
+      return 0;
+    }
+    io->scaled_width = options->scaled_width;
+    io->scaled_height = options->scaled_height;
+  }
+
+  // Filter
+  io->bypass_filtering = options && options->bypass_filtering;
+
+  // Fancy upsampler
+#ifdef FANCY_UPSAMPLING
+  io->fancy_upsampling = (options == NULL) || (!options->no_fancy_upsampling);
+#endif
+
+  if (io->use_scaling) {
+    // disable filter (only for large downscaling ratio).
+    io->bypass_filtering = (io->scaled_width < W * 3 / 4) &&
+                           (io->scaled_height < H * 3 / 4);
+    io->fancy_upsampling = 0;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/dec/webpi.h b/drivers/webp/dec/webpi.h
new file mode 100644
index 0000000000..d915f5ef6f
--- /dev/null
+++ b/drivers/webp/dec/webpi.h
@@ -0,0 +1,116 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Internal header: WebP decoding parameters and custom IO on buffer
+//
+// Author: somnath@google.com (Somnath Banerjee)
+
+#ifndef WEBP_DEC_WEBPI_H_
+#define WEBP_DEC_WEBPI_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "../utils/rescaler.h"
+#include "./decode_vp8.h"
+
+//------------------------------------------------------------------------------
+// WebPDecParams: Decoding output parameters. Transient internal object.
+
+typedef struct WebPDecParams WebPDecParams;
+typedef int (*OutputFunc)(const VP8Io* const io, WebPDecParams* const p);
+typedef int (*OutputRowFunc)(WebPDecParams* const p, int y_pos);
+
+struct WebPDecParams {
+  WebPDecBuffer* output;             // output buffer.
+  uint8_t* tmp_y, *tmp_u, *tmp_v;    // cache for the fancy upsampler
+                                     // or used for tmp rescaling
+
+  int last_y;                 // coordinate of the line that was last output
+  const WebPDecoderOptions* options;  // if not NULL, use alt decoding features
+  // rescalers
+  WebPRescaler scaler_y, scaler_u, scaler_v, scaler_a;
+  void* memory;                  // overall scratch memory for the output work.
+
+  OutputFunc emit;               // output RGB or YUV samples
+  OutputFunc emit_alpha;         // output alpha channel
+  OutputRowFunc emit_alpha_row;  // output one line of rescaled alpha values
+};
+
+// Should be called first, before any use of the WebPDecParams object.
+void WebPResetDecParams(WebPDecParams* const params);
+
+//------------------------------------------------------------------------------
+// Header parsing helpers
+
+// Structure storing a description of the RIFF headers.
+typedef struct {
+  const uint8_t* data;         // input buffer
+  size_t data_size;            // input buffer size
+  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
+  size_t compressed_size;      // VP8/VP8L compressed data size
+  size_t riff_size;            // size of the riff payload (or 0 if absent)
+  int is_lossless;             // true if a VP8L chunk is present
+} WebPHeaderStructure;
+
+// Skips over all valid chunks prior to the first VP8/VP8L frame header.
+// 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);
+
+//------------------------------------------------------------------------------
+// Misc utils
+
+// Initializes VP8Io with custom setup, io and teardown functions. The default
+// hooks will use the supplied 'params' as io->opaque handle.
+void WebPInitCustomIo(WebPDecParams* const params, VP8Io* const io);
+
+// Setup crop_xxx fields, mb_w and mb_h in io. 'src_colorspace' refers
+// to the *compressed* format, not the output one.
+int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
+                          VP8Io* const io, WEBP_CSP_MODE src_colorspace);
+
+//------------------------------------------------------------------------------
+// Internal functions regarding WebPDecBuffer memory (in buffer.c).
+// Don't really need to be externally visible for now.
+
+// Prepare 'buffer' with the requested initial dimensions width/height.
+// If no external storage is supplied, initializes buffer by allocating output
+// memory and setting up the stride information. Validate the parameters. Return
+// an error code in case of problem (no memory, or invalid stride / size /
+// 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.
+VP8StatusCode WebPAllocateDecBuffer(int width, int height,
+                                    const WebPDecoderOptions* const options,
+                                    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,
+                       WebPDecBuffer* const dst);
+
+// Copy and transfer ownership from src to dst (beware of parameter order!)
+void WebPGrabDecBuffer(WebPDecBuffer* const src, WebPDecBuffer* const dst);
+
+
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_DEC_WEBPI_H_ */
diff --git a/drivers/webp/decode.h b/drivers/webp/decode.h
new file mode 100644
index 0000000000..0c3b62e215
--- /dev/null
+++ b/drivers/webp/decode.h
@@ -0,0 +1,491 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//  Main decoding functions for WebP images.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_WEBP_DECODE_H_
+#define WEBP_WEBP_DECODE_H_
+
+#include "./types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define WEBP_DECODER_ABI_VERSION 0x0203    // 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.
+WEBP_EXTERN(int) WebPGetDecoderVersion(void);
+
+// Retrieve basic header information: width, height.
+// This function will also validate the header and return 0 in
+// case of formatting error.
+// Pointers 'width' and 'height' can be passed NULL if deemed irrelevant.
+WEBP_EXTERN(int) WebPGetInfo(const uint8_t* data, size_t data_size,
+                             int* width, int* height);
+
+// 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().
+// Returns NULL in case of error.
+WEBP_EXTERN(uint8_t*) WebPDecodeRGBA(const uint8_t* data, size_t data_size,
+                                     int* width, int* height);
+
+// Same as WebPDecodeRGBA, but returning A, R, G, B, A, R, G, B... ordered data.
+WEBP_EXTERN(uint8_t*) WebPDecodeARGB(const uint8_t* data, size_t data_size,
+                                     int* width, int* height);
+
+// Same as WebPDecodeRGBA, but returning B, G, R, A, B, G, R, A... ordered data.
+WEBP_EXTERN(uint8_t*) WebPDecodeBGRA(const uint8_t* data, size_t data_size,
+                                     int* width, int* height);
+
+// Same as WebPDecodeRGBA, but returning R, G, B, R, G, B... ordered data.
+// If the bitstream contains transparency, it is ignored.
+WEBP_EXTERN(uint8_t*) WebPDecodeRGB(const uint8_t* data, size_t data_size,
+                                    int* width, int* height);
+
+// Same as WebPDecodeRGB, but returning B, G, R, B, G, R... ordered data.
+WEBP_EXTERN(uint8_t*) WebPDecodeBGR(const uint8_t* data, size_t data_size,
+                                    int* width, int* height);
+
+
+// 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.
+// 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.
+// (*) Also named Y'CbCr. See: http://en.wikipedia.org/wiki/YCbCr
+WEBP_EXTERN(uint8_t*) WebPDecodeYUV(const uint8_t* data, size_t data_size,
+                                    int* width, int* height,
+                                    uint8_t** u, uint8_t** v,
+                                    int* stride, int* uv_stride);
+
+// 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
+// storage is not sufficient (or an error occurred), NULL is returned.
+// Otherwise, output_buffer is returned, for convenience.
+// The parameter 'output_stride' specifies the distance (in bytes)
+// between scanlines. Hence, output_buffer_size is expected to be at least
+// output_stride x picture-height.
+WEBP_EXTERN(uint8_t*) WebPDecodeRGBAInto(
+    const uint8_t* data, size_t data_size,
+    uint8_t* output_buffer, size_t output_buffer_size, int output_stride);
+WEBP_EXTERN(uint8_t*) WebPDecodeARGBInto(
+    const uint8_t* data, size_t data_size,
+    uint8_t* output_buffer, size_t output_buffer_size, int output_stride);
+WEBP_EXTERN(uint8_t*) WebPDecodeBGRAInto(
+    const uint8_t* data, size_t data_size,
+    uint8_t* output_buffer, size_t output_buffer_size, int output_stride);
+
+// RGB and BGR variants. Here too the transparency information, if present,
+// will be dropped and ignored.
+WEBP_EXTERN(uint8_t*) WebPDecodeRGBInto(
+    const uint8_t* data, size_t data_size,
+    uint8_t* output_buffer, size_t output_buffer_size, int output_stride);
+WEBP_EXTERN(uint8_t*) WebPDecodeBGRInto(
+    const uint8_t* data, size_t data_size,
+    uint8_t* output_buffer, size_t output_buffer_size, int output_stride);
+
+// WebPDecodeYUVInto() is a variant of WebPDecodeYUV() that operates directly
+// into pre-allocated luma/chroma plane buffers. This function requires the
+// strides to be passed: one for the luma plane and one for each of the
+// chroma ones. The size of each plane buffer is passed as 'luma_size',
+// 'u_size' and 'v_size' respectively.
+// Pointer to the luma plane ('*luma') is returned or NULL if an error occurred
+// during decoding (or because some buffers were found to be too small).
+WEBP_EXTERN(uint8_t*) WebPDecodeYUVInto(
+    const uint8_t* data, size_t data_size,
+    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);
+
+//------------------------------------------------------------------------------
+// Output colorspaces and buffer
+
+// Colorspaces
+// 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.
+// 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) {
+  return (mode == MODE_rgbA || mode == MODE_bgrA || mode == MODE_Argb ||
+          mode == MODE_rgbA_4444);
+}
+
+static WEBP_INLINE int WebPIsAlphaMode(WEBP_CSP_MODE mode) {
+  return (mode == MODE_RGBA || mode == MODE_BGRA || mode == MODE_ARGB ||
+          mode == MODE_RGBA_4444 || mode == MODE_YUVA ||
+          WebPIsPremultipliedMode(mode));
+}
+
+static WEBP_INLINE int WebPIsRGBMode(WEBP_CSP_MODE mode) {
+  return (mode < MODE_YUV);
+}
+
+//------------------------------------------------------------------------------
+// WebPDecBuffer: Generic structure for describing the output sample buffer.
+
+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.
+};
+
+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
+  int a_stride;               // alpha stride
+  size_t y_size;              // luma plane size
+  size_t u_size, v_size;      // chroma planes size
+  size_t a_size;              // alpha-plane size
+};
+
+// Output buffer
+struct WebPDecBuffer {
+  WEBP_CSP_MODE colorspace;  // Colorspace.
+  int width, height;         // Dimensions.
+  int is_external_memory;    // If true, 'internal_memory' pointer is not used.
+  union {
+    WebPRGBABuffer RGBA;
+    WebPYUVABuffer YUVA;
+  } u;                       // Nameless union of buffer parameters.
+  uint32_t       pad[4];     // padding for later use
+
+  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.
+};
+
+// Internal, version-checked, entry point
+WEBP_EXTERN(int) WebPInitDecBufferInternal(WebPDecBuffer*, int);
+
+// Initialize the structure as empty. Must be called before any other use.
+// Returns false in case of version mismatch
+static WEBP_INLINE int WebPInitDecBuffer(WebPDecBuffer* buffer) {
+  return WebPInitDecBufferInternal(buffer, WEBP_DECODER_ABI_VERSION);
+}
+
+// Free any memory associated with the buffer. Must always be called last.
+// Note: doesn't free the 'buffer' structure itself.
+WEBP_EXTERN(void) WebPFreeDecBuffer(WebPDecBuffer* buffer);
+
+//------------------------------------------------------------------------------
+// Enumeration of the status codes
+
+typedef enum VP8StatusCode {
+  VP8_STATUS_OK = 0,
+  VP8_STATUS_OUT_OF_MEMORY,
+  VP8_STATUS_INVALID_PARAM,
+  VP8_STATUS_BITSTREAM_ERROR,
+  VP8_STATUS_UNSUPPORTED_FEATURE,
+  VP8_STATUS_SUSPENDED,
+  VP8_STATUS_USER_ABORT,
+  VP8_STATUS_NOT_ENOUGH_DATA
+} VP8StatusCode;
+
+//------------------------------------------------------------------------------
+// Incremental decoding
+//
+// This API allows streamlined decoding of partial data.
+// Picture can be incrementally decoded as data become available thanks to the
+// WebPIDecoder object. This object can be left in a SUSPENDED state if the
+// picture is only partially decoded, pending additional input.
+// Code example:
+//
+//   WebPInitDecBuffer(&buffer);
+//   buffer.colorspace = mode;
+//   ...
+//   WebPIDecoder* idec = WebPINewDecoder(&buffer);
+//   while (has_more_data) {
+//     // ... (get additional data)
+//     status = WebPIAppend(idec, new_data, new_data_size);
+//     if (status != VP8_STATUS_SUSPENDED ||
+//       break;
+//     }
+//
+//     // The above call decodes the current available buffer.
+//     // Part of the image can now be refreshed by calling to
+//     // WebPIDecGetRGB()/WebPIDecGetYUVA() etc.
+//   }
+//   WebPIDelete(idec);
+
+// 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);
+
+// This function allocates and initializes an incremental-decoder object, which
+// 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'.
+// 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 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,
+    uint8_t* v, size_t v_size, int v_stride,
+    uint8_t* a, size_t a_size, int a_stride);
+
+// Deprecated version of the above, without the alpha plane.
+// Kept for backward compatibility.
+WEBP_EXTERN(WebPIDecoder*) WebPINewYUV(
+    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);
+
+// Deletes the WebPIDecoder object and associated memory. Must always be called
+// if WebPINewDecoder, WebPINewRGB or WebPINewYUV succeeded.
+WEBP_EXTERN(void) WebPIDelete(WebPIDecoder* idec);
+
+// Copies and decodes the next available data. Returns VP8_STATUS_OK when
+// the image is successfully decoded. Returns VP8_STATUS_SUSPENDED when more
+// data is expected. Returns error in other cases.
+WEBP_EXTERN(VP8StatusCode) WebPIAppend(
+    WebPIDecoder* idec, const uint8_t* data, size_t data_size);
+
+// A variant of the above function to be used when data buffer contains
+// partial data from the beginning. In this case data buffer is not copied
+// to the internal memory.
+// Note that the value of the 'data' pointer can change between calls to
+// WebPIUpdate, for instance when the data buffer is resized to fit larger data.
+WEBP_EXTERN(VP8StatusCode) WebPIUpdate(
+    WebPIDecoder* idec, const uint8_t* data, size_t data_size);
+
+// Returns the RGB/A image decoded so far. Returns NULL if output params
+// are not initialized yet. The RGB/A output type corresponds to the colorspace
+// specified during call to WebPINewDecoder() or WebPINewRGB().
+// *last_y is the index of last decoded row in raster scan order. Some pointers
+// (*last_y, *width etc.) can be NULL if corresponding information is not
+// needed.
+WEBP_EXTERN(uint8_t*) WebPIDecGetRGB(
+    const WebPIDecoder* idec, int* last_y,
+    int* width, int* height, int* stride);
+
+// Same as above function to get a YUVA image. Returns pointer to the luma
+// plane or NULL in case of error. If there is no alpha information
+// the alpha pointer '*a' will be returned NULL.
+WEBP_EXTERN(uint8_t*) WebPIDecGetYUVA(
+    const WebPIDecoder* idec, int* last_y,
+    uint8_t** u, uint8_t** v, uint8_t** a,
+    int* width, int* height, int* stride, int* uv_stride, int* a_stride);
+
+// Deprecated alpha-less version of WebPIDecGetYUVA(): it will ignore the
+// alpha information (if present). Kept for backward compatibility.
+static WEBP_INLINE uint8_t* WebPIDecGetYUV(
+    const WebPIDecoder* idec, int* last_y, uint8_t** u, uint8_t** v,
+    int* width, int* height, int* stride, int* uv_stride) {
+  return WebPIDecGetYUVA(idec, last_y, u, v, NULL, width, height,
+                         stride, uv_stride, NULL);
+}
+
+// Generic call to retrieve information about the displayable area.
+// If non NULL, the left/right/width/height pointers are filled with the visible
+// rectangular area so far.
+// Returns NULL in case the incremental decoder object is in an invalid state.
+// Otherwise returns the pointer to the internal representation. This structure
+// is read-only, tied to WebPIDecoder's lifespan and should not be modified.
+WEBP_EXTERN(const WebPDecBuffer*) WebPIDecodedArea(
+    const WebPIDecoder* idec, int* left, int* top, int* width, int* height);
+
+//------------------------------------------------------------------------------
+// Advanced decoding parametrization
+//
+//  Code sample for using the advanced decoding API
+/*
+     // A) Init a configuration object
+     WebPDecoderConfig config;
+     CHECK(WebPInitDecoderConfig(&config));
+
+     // B) optional: retrieve the bitstream's features.
+     CHECK(WebPGetFeatures(data, data_size, &config.input) == VP8_STATUS_OK);
+
+     // C) Adjust 'config', if needed
+     config.no_fancy_upsampling = 1;
+     config.output.colorspace = MODE_BGRA;
+     // etc.
+
+     // Note that you can also make config.output point to an externally
+     // supplied memory buffer, provided it's big enough to store the decoded
+     // picture. Otherwise, config.output will just be used to allocate memory
+     // and store the decoded picture.
+
+     // D) Decode!
+     CHECK(WebPDecode(data, data_size, &config) == VP8_STATUS_OK);
+
+     // E) Decoded image is now in config.output (and config.output.u.RGBA)
+
+     // F) Reclaim memory allocated in config's object. It's safe to call
+     // this function even if the memory is external and wasn't allocated
+     // by WebPDecode().
+     WebPFreeDecBuffer(&config.output);
+*/
+
+// Features gathered from the bitstream
+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
+
+  // Unused for now:
+  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[2];              // padding for later use
+};
+
+// Internal, version-checked, entry point
+WEBP_EXTERN(VP8StatusCode) WebPGetFeaturesInternal(
+    const uint8_t*, size_t, WebPBitstreamFeatures*, int);
+
+// Retrieve features from the bitstream. The *features structure is filled
+// with information gathered from the bitstream.
+// 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) {
+  return WebPGetFeaturesInternal(data, data_size, features,
+                                 WEBP_DECODER_ABI_VERSION);
+}
+
+// Decoding options
+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_
+  int crop_left, crop_top;            // top-left position for cropping.
+                                      // Will be snapped to even values.
+  int crop_width, crop_height;        // dimension of the cropping area
+  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)
+
+  // Unused for now:
+  int force_rotation;                 // forced rotation (to be applied _last_)
+  int no_enhancement;                 // if true, discard enhancement layer
+  uint32_t pad[5];                    // padding for later use
+};
+
+// Main object storing the configuration for advanced decoding.
+struct WebPDecoderConfig {
+  WebPBitstreamFeatures input;  // Immutable bitstream features (optional)
+  WebPDecBuffer output;         // Output buffer (can point to external mem)
+  WebPDecoderOptions options;   // Decoding options
+};
+
+// Internal, version-checked, entry point
+WEBP_EXTERN(int) WebPInitDecoderConfigInternal(WebPDecoderConfig*, int);
+
+// Initialize the configuration as empty. This function must always be
+// called first, unless WebPGetFeatures() is to be called.
+// Returns false in case of mismatched version.
+static WEBP_INLINE int WebPInitDecoderConfig(WebPDecoderConfig* config) {
+  return WebPInitDecoderConfigInternal(config, WEBP_DECODER_ABI_VERSION);
+}
+
+// Instantiate a new incremental decoder object with the requested
+// configuration. The bitstream can be passed using 'data' and 'data_size'
+// parameter, in which case the features will be parsed and stored into
+// config->input. Otherwise, 'data' can be NULL and no parsing will occur.
+// Note that 'config' can be NULL too, in which case a default configuration
+// is used.
+// The return WebPIDecoder object must always be deleted calling WebPIDelete().
+// Returns NULL in case of error (and config->status will then reflect
+// the error condition).
+WEBP_EXTERN(WebPIDecoder*) WebPIDecode(const uint8_t* data, size_t data_size,
+                                       WebPDecoderConfig* config);
+
+// Non-incremental version. This version decodes the full data at once, taking
+// 'config' into account. Returns decoding status (which should be VP8_STATUS_OK
+// if the decoding was successful).
+WEBP_EXTERN(VP8StatusCode) WebPDecode(const uint8_t* data, size_t data_size,
+                                      WebPDecoderConfig* config);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_WEBP_DECODE_H_ */
diff --git a/drivers/webp/demux.h b/drivers/webp/demux.h
new file mode 100644
index 0000000000..2da3239dd9
--- /dev/null
+++ b/drivers/webp/demux.h
@@ -0,0 +1,224 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Demux API.
+// Enables extraction of image and extended format data from WebP files.
+
+// Code Example: Demuxing WebP data to extract all the frames, ICC profile
+// and EXIF/XMP metadata.
+/*
+  WebPDemuxer* demux = WebPDemux(&webp_data);
+
+  uint32_t width = WebPDemuxGetI(demux, WEBP_FF_CANVAS_WIDTH);
+  uint32_t height = WebPDemuxGetI(demux, WEBP_FF_CANVAS_HEIGHT);
+  // ... (Get information about the features present in the WebP file).
+  uint32_t flags = WebPDemuxGetI(demux, WEBP_FF_FORMAT_FLAGS);
+
+  // ... (Iterate over all frames).
+  WebPIterator iter;
+  if (WebPDemuxGetFrame(demux, 1, &iter)) {
+    do {
+      // ... (Consume 'iter'; e.g. Decode 'iter.fragment' with WebPDecode(),
+      // ... and get other frame properties like width, height, offsets etc.
+      // ... see 'struct WebPIterator' below for more info).
+    } while (WebPDemuxNextFrame(&iter));
+    WebPDemuxReleaseIterator(&iter);
+  }
+
+  // ... (Extract metadata).
+  WebPChunkIterator chunk_iter;
+  if (flags & ICCP_FLAG) WebPDemuxGetChunk(demux, "ICCP", 1, &chunk_iter);
+  // ... (Consume the ICC profile in 'chunk_iter.chunk').
+  WebPDemuxReleaseChunkIterator(&chunk_iter);
+  if (flags & EXIF_FLAG) WebPDemuxGetChunk(demux, "EXIF", 1, &chunk_iter);
+  // ... (Consume the EXIF metadata in 'chunk_iter.chunk').
+  WebPDemuxReleaseChunkIterator(&chunk_iter);
+  if (flags & XMP_FLAG) WebPDemuxGetChunk(demux, "XMP ", 1, &chunk_iter);
+  // ... (Consume the XMP metadata in 'chunk_iter.chunk').
+  WebPDemuxReleaseChunkIterator(&chunk_iter);
+  WebPDemuxDelete(demux);
+*/
+
+#ifndef WEBP_WEBP_DEMUX_H_
+#define WEBP_WEBP_DEMUX_H_
+
+#include "./mux_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define WEBP_DEMUX_ABI_VERSION 0x0101    // 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 WebPDemuxState WebPDemuxState;
+// typedef enum WebPFormatFeature WebPFormatFeature;
+typedef struct WebPDemuxer WebPDemuxer;
+typedef struct WebPIterator WebPIterator;
+typedef struct WebPChunkIterator WebPChunkIterator;
+
+//------------------------------------------------------------------------------
+
+// Returns the version number of the demux library, packed in hexadecimal using
+// 8bits for each of major/minor/revision. E.g: v2.5.7 is 0x020507.
+WEBP_EXTERN(int) WebPGetDemuxVersion(void);
+
+//------------------------------------------------------------------------------
+// Life of a Demux object
+
+typedef enum WebPDemuxState {
+  WEBP_DEMUX_PARSE_ERROR    = -1,  // An error occurred while parsing.
+  WEBP_DEMUX_PARSING_HEADER =  0,  // Not enough data to parse full header.
+  WEBP_DEMUX_PARSED_HEADER  =  1,  // Header parsing complete,
+                                   // data may be available.
+  WEBP_DEMUX_DONE           =  2   // Entire file has been parsed.
+} WebPDemuxState;
+
+// Internal, version-checked, entry point
+WEBP_EXTERN(WebPDemuxer*) WebPDemuxInternal(
+    const WebPData*, int, WebPDemuxState*, int);
+
+// Parses the full WebP file given by 'data'.
+// 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 possibly incomplete WebP file given by 'data'.
+// If 'state' is non-NULL it will be set to indicate the status of the demuxer.
+// Returns NULL in case of error or if there isn't enough data to start parsing;
+// and a WebPDemuxer object on successful parse.
+// Note that WebPDemuxer keeps internal pointers to 'data' memory segment.
+// If this data is volatile, the demuxer object should be deleted (by calling
+// WebPDemuxDelete()) and WebPDemuxPartial() called again on the new data.
+// This is usually an inexpensive operation.
+static WEBP_INLINE WebPDemuxer* WebPDemuxPartial(
+    const WebPData* data, WebPDemuxState* state) {
+  return WebPDemuxInternal(data, 1, state, WEBP_DEMUX_ABI_VERSION);
+}
+
+// Frees memory associated with 'dmux'.
+WEBP_EXTERN(void) WebPDemuxDelete(WebPDemuxer* dmux);
+
+//------------------------------------------------------------------------------
+// Data/information extraction.
+
+typedef enum WebPFormatFeature {
+  WEBP_FF_FORMAT_FLAGS,  // Extended format flags present in the 'VP8X' chunk.
+  WEBP_FF_CANVAS_WIDTH,
+  WEBP_FF_CANVAS_HEIGHT,
+  WEBP_FF_LOOP_COUNT,
+  WEBP_FF_BACKGROUND_COLOR,
+  WEBP_FF_FRAME_COUNT    // Number of frames present in the demux object.
+                         // In case of a partial demux, this is the number of
+                         // frames seen so far, with the last frame possibly
+                         // being partial.
+} WebPFormatFeature;
+
+// 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);
+
+//------------------------------------------------------------------------------
+// Frame iteration.
+
+struct WebPIterator {
+  int frame_num;
+  int num_frames;          // equivalent to WEBP_FF_FRAME_COUNT.
+  int fragment_num;
+  int num_fragments;
+  int x_offset, y_offset;  // offset relative to the canvas.
+  int width, height;       // dimensions of this frame or fragment.
+  int duration;            // display duration in milliseconds.
+  WebPMuxAnimDispose dispose_method;  // dispose method for the frame.
+  int complete;   // true if 'fragment' contains a full frame. partial images
+                  // may still be decoded with the WebP incremental decoder.
+  WebPData fragment;  // The frame or fragment given by 'frame_num' and
+                      // 'fragment_num'.
+  int has_alpha;      // True if the frame or fragment contains transparency.
+  WebPMuxAnimBlend blend_method;  // Blend operation for the frame.
+
+  uint32_t pad[2];         // padding for later use.
+  void* private_;          // for internal use only.
+};
+
+// Retrieves frame 'frame_number' from 'dmux'.
+// 'iter->fragment' points to the first fragment on return from this function.
+// Individual fragments may be extracted using WebPDemuxSelectFragment().
+// 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->fragment' 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->fragment' to reflect fragment number 'fragment_num'.
+// Returns true if fragment 'fragment_num' is present, false otherwise.
+WEBP_EXTERN(int) WebPDemuxSelectFragment(WebPIterator* iter, int fragment_num);
+
+// Releases any memory associated with 'iter'.
+// Must be called before any subsequent calls to WebPDemuxGetChunk() on the same
+// iter. Also, must be called before destroying the associated WebPDemuxer with
+// WebPDemuxDelete().
+WEBP_EXTERN(void) WebPDemuxReleaseIterator(WebPIterator* iter);
+
+//------------------------------------------------------------------------------
+// Chunk iteration.
+
+struct WebPChunkIterator {
+  // 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_;
+};
+
+// 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", "XMP ", "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);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_WEBP_DEMUX_H_ */
diff --git a/drivers/webp/demux/demux.c b/drivers/webp/demux/demux.c
new file mode 100644
index 0000000000..f66ac6d82b
--- /dev/null
+++ b/drivers/webp/demux/demux.c
@@ -0,0 +1,1002 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//  WebP container demux.
+//
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "../utils/utils.h"
+#include "../webp/decode.h"     // WebPGetFeatures
+#include "../webp/demux.h"
+#include "../webp/format_constants.h"
+
+#define DMUX_MAJ_VERSION 0
+#define DMUX_MIN_VERSION 2
+#define DMUX_REV_VERSION 0
+
+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 has_alpha_;
+  int duration_;
+  WebPMuxAnimDispose dispose_method_;
+  WebPMuxAnimBlend blend_method_;
+  int is_fragment_;  // this is a frame fragment (and not a full frame).
+  int frame_num_;  // the referent frame number for use in assembling fragments.
+  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_;
+  uint32_t bgcolor_;
+  int num_frames_;
+  Frame* frames_;
+  Frame** frames_tail_;
+  Chunk* chunks_;  // non-image chunks
+  Chunk** chunks_tail_;
+};
+
+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 },
+};
+
+//------------------------------------------------------------------------------
+
+int WebPGetDemuxVersion(void) {
+  return (DMUX_MAJ_VERSION << 16) | (DMUX_MIN_VERSION << 8) | DMUX_REV_VERSION;
+}
+
+// -----------------------------------------------------------------------------
+// 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_;
+}
+
+// Read from 'mem' and skip the read bytes.
+static WEBP_INLINE uint8_t ReadByte(MemBuffer* const mem) {
+  const uint8_t byte = mem->buf_[mem->start_];
+  Skip(mem, 1);
+  return byte;
+}
+
+static WEBP_INLINE int ReadLE16s(MemBuffer* const mem) {
+  const uint8_t* const data = mem->buf_ + mem->start_;
+  const int val = GetLE16(data);
+  Skip(mem, 2);
+  return val;
+}
+
+static WEBP_INLINE int ReadLE24s(MemBuffer* const mem) {
+  const uint8_t* const data = mem->buf_ + mem->start_;
+  const int val = GetLE24(data);
+  Skip(mem, 3);
+  return val;
+}
+
+static WEBP_INLINE uint32_t ReadLE32(MemBuffer* const mem) {
+  const uint8_t* const data = mem->buf_ + mem->start_;
+  const uint32_t val = GetLE32(data);
+  Skip(mem, 4);
+  return val;
+}
+
+// -----------------------------------------------------------------------------
+// Secondary chunk parsing
+
+static void AddChunk(WebPDemuxer* const dmux, Chunk* const chunk) {
+  *dmux->chunks_tail_ = chunk;
+  chunk->next_ = NULL;
+  dmux->chunks_tail_ = &chunk->next_;
+}
+
+// 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* const last_frame = *dmux->frames_tail_;
+  if (last_frame != NULL && !last_frame->complete_) return 0;
+
+  *dmux->frames_tail_ = frame;
+  frame->next_ = NULL;
+  dmux->frames_tail_ = &frame->next_;
+  return 1;
+}
+
+// Store image bearing chunks to 'frame'.
+static ParseStatus StoreFrame(int frame_num, uint32_t min_size,
+                              MemBuffer* const mem, Frame* const frame) {
+  int alpha_chunks = 0;
+  int image_chunks = 0;
+  int done = (MemDataSize(mem) < min_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 = ReadLE32(mem);
+    const uint32_t payload_size = ReadLE32(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->has_alpha_ = 1;
+          frame->frame_num_ = frame_num;
+          Skip(mem, payload_available);
+        } else {
+          goto Done;
+        }
+        break;
+      case MKFOURCC('V', 'P', '8', 'L'):
+        if (alpha_chunks > 0) return PARSE_ERROR;  // VP8L has its own alpha
+        // fall through
+      case MKFOURCC('V', 'P', '8', ' '):
+        if (image_chunks == 0) {
+          // Extract the bitstream features, tolerating failures when the data
+          // is incomplete.
+          WebPBitstreamFeatures features;
+          const VP8StatusCode vp8_status =
+              WebPGetFeatures(mem->buf_ + chunk_start_offset, chunk_size,
+                              &features);
+          if (status == PARSE_NEED_MORE_DATA &&
+              vp8_status == VP8_STATUS_NOT_ENOUGH_DATA) {
+            return PARSE_NEED_MORE_DATA;
+          } else if (vp8_status != VP8_STATUS_OK) {
+            // We have enough data, and yet WebPGetFeatures() failed.
+            return PARSE_ERROR;
+          }
+          ++image_chunks;
+          frame->img_components_[0].offset_ = chunk_start_offset;
+          frame->img_components_[0].size_ = chunk_size;
+          frame->width_ = features.width;
+          frame->height_ = features.height;
+          frame->has_alpha_ |= features.has_alpha;
+          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 actual_size,
+                            Frame** frame) {
+  if (SizeIsInvalid(mem, min_size)) return PARSE_ERROR;
+  if (actual_size < min_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 'ANMF' chunk and any image bearing chunks that immediately follow.
+// 'frame_chunk_size' is the previously validated, padded chunk size.
+static ParseStatus ParseAnimationFrame(
+    WebPDemuxer* const dmux, uint32_t frame_chunk_size) {
+  const int is_animation = !!(dmux->feature_flags_ & ANIMATION_FLAG);
+  const uint32_t anmf_payload_size = frame_chunk_size - ANMF_CHUNK_SIZE;
+  int added_frame = 0;
+  int bits;
+  MemBuffer* const mem = &dmux->mem_;
+  Frame* frame;
+  ParseStatus status =
+      NewFrame(mem, ANMF_CHUNK_SIZE, frame_chunk_size, &frame);
+  if (status != PARSE_OK) return status;
+
+  frame->x_offset_       = 2 * ReadLE24s(mem);
+  frame->y_offset_       = 2 * ReadLE24s(mem);
+  frame->width_          = 1 + ReadLE24s(mem);
+  frame->height_         = 1 + ReadLE24s(mem);
+  frame->duration_       = ReadLE24s(mem);
+  bits = ReadByte(mem);
+  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;
+  if (frame->width_ * (uint64_t)frame->height_ >= MAX_IMAGE_AREA) {
+    free(frame);
+    return PARSE_ERROR;
+  }
+
+  // Store a frame only if the animation flag is set there is some data for
+  // this frame is available.
+  status = StoreFrame(dmux->num_frames_ + 1, anmf_payload_size, mem, frame);
+  if (status != PARSE_ERROR && is_animation && 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;
+}
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+// Parse a 'FRGM' chunk and any image bearing chunks that immediately follow.
+// 'fragment_chunk_size' is the previously validated, padded chunk size.
+static ParseStatus ParseFragment(WebPDemuxer* const dmux,
+                                 uint32_t fragment_chunk_size) {
+  const int frame_num = 1;  // All fragments belong to the 1st (and only) frame.
+  const int is_fragmented = !!(dmux->feature_flags_ & FRAGMENTS_FLAG);
+  const uint32_t frgm_payload_size = fragment_chunk_size - FRGM_CHUNK_SIZE;
+  int added_fragment = 0;
+  MemBuffer* const mem = &dmux->mem_;
+  Frame* frame;
+  ParseStatus status =
+      NewFrame(mem, FRGM_CHUNK_SIZE, fragment_chunk_size, &frame);
+  if (status != PARSE_OK) return status;
+
+  frame->is_fragment_ = 1;
+  frame->x_offset_ = 2 * ReadLE24s(mem);
+  frame->y_offset_ = 2 * ReadLE24s(mem);
+
+  // Store a fragment only if the 'fragments' flag is set and there is some
+  // data available.
+  status = StoreFrame(frame_num, frgm_payload_size, mem, frame);
+  if (status != PARSE_ERROR && is_fragmented && frame->frame_num_ > 0) {
+    added_fragment = AddFrame(dmux, frame);
+    if (!added_fragment) {
+      status = PARSE_ERROR;
+    } else {
+      dmux->num_frames_ = 1;
+    }
+  }
+
+  if (!added_fragment) free(frame);
+  return status;
+}
+#endif  // WEBP_EXPERIMENTAL_FEATURES
+
+// 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 ParseStatus 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 PARSE_NEED_MORE_DATA;
+  if (memcmp(GetBuffer(mem), "RIFF", CHUNK_SIZE_BYTES) ||
+      memcmp(GetBuffer(mem) + CHUNK_HEADER_SIZE, "WEBP", CHUNK_SIZE_BYTES)) {
+    return PARSE_ERROR;
+  }
+
+  riff_size = GetLE32(GetBuffer(mem) + TAG_SIZE);
+  if (riff_size < CHUNK_HEADER_SIZE) return PARSE_ERROR;
+  if (riff_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
+
+  // 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 PARSE_OK;
+}
+
+static ParseStatus ParseSingleImage(WebPDemuxer* const dmux) {
+  const size_t min_size = CHUNK_HEADER_SIZE;
+  MemBuffer* const mem = &dmux->mem_;
+  Frame* frame;
+  ParseStatus status;
+  int image_added = 0;
+
+  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;
+
+  // For the single image case we allow parsing of a partial frame, but we need
+  // at least CHUNK_HEADER_SIZE for parsing.
+  status = StoreFrame(1, CHUNK_HEADER_SIZE, &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;
+      frame->has_alpha_ = 0;
+    }
+
+    // Use the frame width/height as the canvas values for non-vp8x files.
+    // Also, set ALPHA_FLAG if this is a lossless image with alpha.
+    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_;
+      dmux->feature_flags_ |= frame->has_alpha_ ? ALPHA_FLAG : 0;
+    }
+    if (!AddFrame(dmux, frame)) {
+      status = PARSE_ERROR;  // last frame was left incomplete
+    } else {
+      image_added = 1;
+      dmux->num_frames_ = 1;
+    }
+  }
+
+  if (!image_added) free(frame);
+  return status;
+}
+
+static ParseStatus ParseVP8XChunks(WebPDemuxer* const dmux) {
+  const int is_animation = !!(dmux->feature_flags_ & ANIMATION_FLAG);
+  MemBuffer* const mem = &dmux->mem_;
+  int anim_chunks = 0;
+  ParseStatus status = PARSE_OK;
+
+  do {
+    int store_chunk = 1;
+    const size_t chunk_start_offset = mem->start_;
+    const uint32_t fourcc = ReadLE32(mem);
+    const uint32_t chunk_size = ReadLE32(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'): {
+        // check that this isn't an animation (all frames should be in an ANMF).
+        if (anim_chunks > 0 || is_animation) return PARSE_ERROR;
+
+        Rewind(mem, CHUNK_HEADER_SIZE);
+        status = ParseSingleImage(dmux);
+        break;
+      }
+      case MKFOURCC('A', 'N', 'I', 'M'): {
+        if (chunk_size_padded < ANIM_CHUNK_SIZE) return PARSE_ERROR;
+
+        if (MemDataSize(mem) < chunk_size_padded) {
+          status = PARSE_NEED_MORE_DATA;
+        } else if (anim_chunks == 0) {
+          ++anim_chunks;
+          dmux->bgcolor_ = ReadLE32(mem);
+          dmux->loop_count_ = ReadLE16s(mem);
+          Skip(mem, chunk_size_padded - ANIM_CHUNK_SIZE);
+        } else {
+          store_chunk = 0;
+          goto Skip;
+        }
+        break;
+      }
+      case MKFOURCC('A', 'N', 'M', 'F'): {
+        if (anim_chunks == 0) return PARSE_ERROR;  // 'ANIM' precedes frames.
+        status = ParseAnimationFrame(dmux, chunk_size_padded);
+        break;
+      }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+      case MKFOURCC('F', 'R', 'G', 'M'): {
+        status = ParseFragment(dmux, chunk_size_padded);
+        break;
+      }
+#endif
+      case MKFOURCC('I', 'C', 'C', 'P'): {
+        store_chunk = !!(dmux->feature_flags_ & ICCP_FLAG);
+        goto Skip;
+      }
+      case MKFOURCC('E', 'X', 'I', 'F'): {
+        store_chunk = !!(dmux->feature_flags_ & EXIF_FLAG);
+        goto Skip;
+      }
+      case MKFOURCC('X', 'M', 'P', ' '): {
+        store_chunk = !!(dmux->feature_flags_ & XMP_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;
+}
+
+static ParseStatus ParseVP8X(WebPDemuxer* const dmux) {
+  MemBuffer* const mem = &dmux->mem_;
+  uint32_t vp8x_size;
+
+  if (MemDataSize(mem) < CHUNK_HEADER_SIZE) return PARSE_NEED_MORE_DATA;
+
+  dmux->is_ext_format_ = 1;
+  Skip(mem, TAG_SIZE);  // VP8X
+  vp8x_size = ReadLE32(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_ = ReadByte(mem);
+  Skip(mem, 3);  // Reserved.
+  dmux->canvas_width_  = 1 + ReadLE24s(mem);
+  dmux->canvas_height_ = 1 + ReadLE24s(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;
+
+  return ParseVP8XChunks(dmux);
+}
+
+// -----------------------------------------------------------------------------
+// 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;
+}
+
+// If 'exact' is true, check that the image resolution matches the canvas.
+// If 'exact' is false, check that the x/y offsets do not exceed the canvas.
+// TODO(jzern): this is insufficient in the fragmented image case if the
+// expectation is that the fragments completely cover the canvas.
+static int CheckFrameBounds(const Frame* const frame, int exact,
+                            int canvas_width, int canvas_height) {
+  if (exact) {
+    if (frame->x_offset_ != 0 || frame->y_offset_ != 0) {
+      return 0;
+    }
+    if (frame->width_ != canvas_width || frame->height_ != canvas_height) {
+      return 0;
+    }
+  } else {
+    if (frame->x_offset_ < 0 || frame->y_offset_ < 0) return 0;
+    if (frame->width_ + frame->x_offset_ > canvas_width) return 0;
+    if (frame->height_ + frame->y_offset_ > canvas_height) return 0;
+  }
+  return 1;
+}
+
+static int IsValidExtendedFormat(const WebPDemuxer* const dmux) {
+  const int is_animation = !!(dmux->feature_flags_ & ANIMATION_FLAG);
+  const int is_fragmented = !!(dmux->feature_flags_ & FRAGMENTS_FLAG);
+  const Frame* f = dmux->frames_;
+
+  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;
+#ifndef WEBP_EXPERIMENTAL_FEATURES
+  if (is_fragmented) return 0;
+#endif
+
+  while (f != NULL) {
+    const int cur_frame_set = f->frame_num_;
+    int frame_count = 0, fragment_count = 0;
+
+    // Check frame properties and if the image is composed of fragments that
+    // each fragment came from a fragment.
+    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 (is_fragmented && !f->is_fragment_) return 0;
+      if (!is_fragmented && f->is_fragment_) return 0;
+      if (!is_animation && f->frame_num_ > 1) 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 {
+        // There shouldn't be a partial frame in a complete file.
+        if (dmux->state_ == WEBP_DEMUX_DONE) return 0;
+
+        // 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;
+      }
+
+      if (f->width_ > 0 && f->height_ > 0 &&
+          !CheckFrameBounds(f, !(is_animation || is_fragmented),
+                            dmux->canvas_width_, dmux->canvas_height_)) {
+        return 0;
+      }
+
+      fragment_count += f->is_fragment_;
+      ++frame_count;
+    }
+    if (!is_fragmented && frame_count > 1) return 0;
+    if (fragment_count > 0 && frame_count != fragment_count) return 0;
+  }
+  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->bgcolor_ = 0xFFFFFFFF;  // White background by default.
+  dmux->canvas_width_ = -1;
+  dmux->canvas_height_ = -1;
+  dmux->frames_tail_ = &dmux->frames_;
+  dmux->chunks_tail_ = &dmux->chunks_;
+  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 (state != NULL) *state = WEBP_DEMUX_PARSE_ERROR;
+
+  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;
+  status = ReadHeader(&mem);
+  if (status != PARSE_OK) {
+    if (state != NULL) {
+      *state = (status == PARSE_NEED_MORE_DATA) ? WEBP_DEMUX_PARSING_HEADER
+                                                : WEBP_DEMUX_PARSE_ERROR;
+    }
+    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);
+
+  status = PARSE_ERROR;
+  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_NEED_MORE_DATA && !partial) status = PARSE_ERROR;
+      if (status != PARSE_ERROR && !parser->valid(dmux)) status = PARSE_ERROR;
+      if (status == PARSE_ERROR) dmux->state_ = WEBP_DEMUX_PARSE_ERROR;
+      break;
+    }
+  }
+  if (state != NULL) *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_;
+    case WEBP_FF_BACKGROUND_COLOR: return dmux->bgcolor_;
+    case WEBP_FF_FRAME_COUNT:      return (uint32_t)dmux->num_frames_;
+  }
+  return 0;
+}
+
+// -----------------------------------------------------------------------------
+// Frame iteration
+
+// Find the first 'frame_num' frame. There may be multiple such frames in a
+// fragmented 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 fragment 'fragment_num' and the total count.
+static const Frame* GetFragment(
+    const Frame* const frame_set, int fragment_num, int* const count) {
+  const int this_frame = frame_set->frame_num_;
+  const Frame* f = frame_set;
+  const Frame* fragment = NULL;
+  int total;
+
+  for (total = 0; f != NULL && f->frame_num_ == this_frame; f = f->next_) {
+    if (++total == fragment_num) fragment = f;
+  }
+  *count = total;
+  return fragment;
+}
+
+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 fragment_num, WebPIterator* const iter) {
+  const uint8_t* const mem_buf = dmux->mem_.buf_;
+  int num_fragments;
+  size_t payload_size = 0;
+  const Frame* const fragment =
+      GetFragment(first_frame, fragment_num, &num_fragments);
+  const uint8_t* const payload =
+      GetFramePayload(mem_buf, fragment, &payload_size);
+  if (payload == NULL) return 0;
+  assert(first_frame != NULL);
+
+  iter->frame_num      = first_frame->frame_num_;
+  iter->num_frames     = dmux->num_frames_;
+  iter->fragment_num   = fragment_num;
+  iter->num_fragments  = num_fragments;
+  iter->x_offset       = fragment->x_offset_;
+  iter->y_offset       = fragment->y_offset_;
+  iter->width          = fragment->width_;
+  iter->height         = fragment->height_;
+  iter->has_alpha      = fragment->has_alpha_;
+  iter->duration       = fragment->duration_;
+  iter->dispose_method = fragment->dispose_method_;
+  iter->blend_method   = fragment->blend_method_;
+  iter->complete       = fragment->complete_;
+  iter->fragment.bytes = payload;
+  iter->fragment.size  = payload_size;
+  // TODO(jzern): adjust offsets for 'FRGM's embedded in 'ANMF'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);
+  if (frame == NULL) return 0;
+
+  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 WebPDemuxSelectFragment(WebPIterator* iter, int fragment_num) {
+  if (iter != NULL && iter->private_ != NULL && fragment_num > 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, fragment_num, 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;
+}
+
diff --git a/drivers/webp/dsp/cpu.c b/drivers/webp/dsp/cpu.c
new file mode 100644
index 0000000000..7a1f417a55
--- /dev/null
+++ b/drivers/webp/dsp/cpu.c
@@ -0,0 +1,80 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// CPU detection
+//
+// Author: Christian Duvivier (cduvivier@google.com)
+
+#include "./dsp.h"
+
+#if defined(__ANDROID__)
+#include <cpu-features.h>
+#endif
+
+//------------------------------------------------------------------------------
+// SSE2 detection.
+//
+
+// apple/darwin gcc-4.0.1 defines __PIC__, but not __pic__ with -fPIC.
+#if (defined(__pic__) || defined(__PIC__)) && defined(__i386__)
+static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) {
+  __asm__ volatile (
+    "mov %%ebx, %%edi\n"
+    "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));
+}
+#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));
+}
+#elif defined(WEBP_MSC_SSE2)
+#define GetCPUInfo __cpuid
+#endif
+
+#if defined(__i386__) || defined(__x86_64__) || defined(WEBP_MSC_SSE2)
+static int x86CPUInfo(CPUFeature feature) {
+  int cpu_info[4];
+  GetCPUInfo(cpu_info, 1);
+  if (feature == kSSE2) {
+    return 0 != (cpu_info[3] & 0x04000000);
+  }
+  if (feature == kSSE3) {
+    return 0 != (cpu_info[2] & 0x00000001);
+  }
+  return 0;
+}
+VP8CPUInfo VP8GetCPUInfo = x86CPUInfo;
+#elif defined(WEBP_ANDROID_NEON)
+static int AndroidCPUInfo(CPUFeature feature) {
+  const AndroidCpuFamily cpu_family = android_getCpuFamily();
+  const uint64_t cpu_features = android_getCpuFeatures();
+  if (feature == kNEON) {
+    return (cpu_family == ANDROID_CPU_FAMILY_ARM &&
+            0 != (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON));
+  }
+  return 0;
+}
+VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo;
+#elif defined(__ARM_NEON__)
+// define a dummy function to enable turning off NEON at runtime by setting
+// VP8DecGetCPUInfo = NULL
+static int armCPUInfo(CPUFeature feature) {
+  (void)feature;
+  return 1;
+}
+VP8CPUInfo VP8GetCPUInfo = armCPUInfo;
+#else
+VP8CPUInfo VP8GetCPUInfo = NULL;
+#endif
+
diff --git a/drivers/webp/dsp/dec.c b/drivers/webp/dsp/dec.c
new file mode 100644
index 0000000000..8b246fad0a
--- /dev/null
+++ b/drivers/webp/dsp/dec.c
@@ -0,0 +1,756 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Speed-critical decoding functions.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./dsp.h"
+#include "../dec/vp8i.h"
+
+//------------------------------------------------------------------------------
+// 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;
+}
+
+//------------------------------------------------------------------------------
+// Transforms (Paragraph 14.4)
+
+#define STORE(x, y, v) \
+  dst[x + y * BPS] = clip_8b(dst[x + y * BPS] + ((v) >> 3))
+
+#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)
+
+static const int kC1 = 20091 + (1 << 16);
+static const int kC2 = 35468;
+#define MUL(a, b) (((a) * (b)) >> 16)
+
+static void TransformOne(const int16_t* in, uint8_t* dst) {
+  int C[4 * 4], *tmp;
+  int i;
+  tmp = C;
+  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]
+    tmp[0] = a + d;   // [-7881, 7875]
+    tmp[1] = b + c;   // [-7878, 7878]
+    tmp[2] = b - c;   // [-7878, 7878]
+    tmp[3] = a - d;   // [-7877, 7879]
+    tmp += 4;
+    in++;
+  }
+  // Each pass is expanding the dynamic range by ~3.85 (upper bound).
+  // The exact value is (2. + (kC1 + kC2) / 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.
+  // In the worst case scenario, the input to clip_8b() can be as large as
+  // [-60713, 60968].
+  tmp = C;
+  for (i = 0; i < 4; ++i) {    // horizontal pass
+    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);
+    STORE(0, 0, a + d);
+    STORE(1, 0, b + c);
+    STORE(2, 0, b - c);
+    STORE(3, 0, a - d);
+    tmp++;
+    dst += BPS;
+  }
+}
+
+// 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 = MUL(in[4], kC2);
+  const int d4 = MUL(in[4], kC1);
+  const int c1 = MUL(in[1], kC2);
+  const int d1 = MUL(in[1], kC1);
+  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 MUL
+#undef STORE2
+
+static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
+  TransformOne(in, dst);
+  if (do_two) {
+    TransformOne(in + 16, dst + 4);
+  }
+}
+
+static void TransformUV(const int16_t* in, uint8_t* dst) {
+  VP8Transform(in + 0 * 16, dst, 1);
+  VP8Transform(in + 2 * 16, dst + 4 * BPS, 1);
+}
+
+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) {
+    for (i = 0; i < 4; ++i) {
+      STORE(i, j, DC);
+    }
+  }
+}
+
+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);
+}
+
+#undef STORE
+
+//------------------------------------------------------------------------------
+// Paragraph 14.3
+
+static void TransformWHT(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;
+  }
+}
+
+void (*VP8TransformWHT)(const int16_t* in, int16_t* out) = TransformWHT;
+
+//------------------------------------------------------------------------------
+// Intra predictions
+
+#define DST(x, y) dst[(x) + (y) * BPS]
+
+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];
+  int y;
+  for (y = 0; y < size; ++y) {
+    const uint8_t* const clip = clip0 + dst[-1];
+    int x;
+    for (x = 0; x < size; ++x) {
+      dst[x] = clip[top[x]];
+    }
+    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); }
+
+//------------------------------------------------------------------------------
+// 16x16
+
+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
+  int j;
+  for (j = 16; j > 0; --j) {
+    memset(dst, dst[-1], 16);
+    dst += BPS;
+  }
+}
+
+static WEBP_INLINE void Put16(int v, uint8_t* dst) {
+  int j;
+  for (j = 0; j < 16; ++j) {
+    memset(dst + j * BPS, v, 16);
+  }
+}
+
+static void DC16(uint8_t *dst) {    // DC
+  int DC = 16;
+  int j;
+  for (j = 0; j < 16; ++j) {
+    DC += dst[-1 + j * BPS] + dst[j - BPS];
+  }
+  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
+  int DC = 8;
+  int i;
+  for (i = 0; i < 16; ++i) {
+    DC += dst[i - BPS];
+  }
+  Put16(DC >> 4, dst);
+}
+
+static void DC16NoTopLeft(uint8_t *dst) {  // DC with no top and left samples
+  Put16(0x80, dst);
+}
+
+//------------------------------------------------------------------------------
+// 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
+  const uint8_t* top = dst - BPS;
+  const uint8_t vals[4] = {
+    AVG3(top[-1], top[0], top[1]),
+    AVG3(top[ 0], top[1], top[2]),
+    AVG3(top[ 1], top[2], top[3]),
+    AVG3(top[ 2], top[3], top[4])
+  };
+  int i;
+  for (i = 0; i < 4; ++i) {
+    memcpy(dst + i * BPS, vals, sizeof(vals));
+  }
+}
+
+static void HE4(uint8_t *dst) {    // horizontal
+  const int A = dst[-1 - BPS];
+  const int B = dst[-1];
+  const int C = dst[-1 + BPS];
+  const int D = dst[-1 + 2 * BPS];
+  const int E = dst[-1 + 3 * BPS];
+  *(uint32_t*)(dst + 0 * BPS) = 0x01010101U * AVG3(A, B, C);
+  *(uint32_t*)(dst + 1 * BPS) = 0x01010101U * AVG3(B, C, D);
+  *(uint32_t*)(dst + 2 * BPS) = 0x01010101U * AVG3(C, D, E);
+  *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(D, E, E);
+}
+
+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];
+  dc >>= 3;
+  for (i = 0; i < 4; ++i) memset(dst + i * BPS, dc, 4);
+}
+
+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];
+  const int L = dst[-1 + 3 * BPS];
+  const int X = dst[-1 - BPS];
+  const int A = dst[0 - BPS];
+  const int B = dst[1 - BPS];
+  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);
+}
+
+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];
+  const int D = dst[3 - BPS];
+  const int E = dst[4 - BPS];
+  const int F = dst[5 - BPS];
+  const int G = dst[6 - BPS];
+  const int H = dst[7 - BPS];
+  DST(0, 0)                                     = AVG3(A, B, C);
+  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);
+}
+
+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];
+  const int X = dst[-1 - BPS];
+  const int A = dst[0 - BPS];
+  const int B = dst[1 - BPS];
+  const int C = dst[2 - BPS];
+  const int D = dst[3 - BPS];
+  DST(0, 0) = DST(1, 2) = AVG2(X, A);
+  DST(1, 0) = DST(2, 2) = AVG2(A, B);
+  DST(2, 0) = DST(3, 2) = AVG2(B, C);
+  DST(3, 0)             = AVG2(C, D);
+
+  DST(0, 3) =             AVG3(K, J, I);
+  DST(0, 2) =             AVG3(J, I, X);
+  DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
+  DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
+  DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
+  DST(3, 1) =             AVG3(B, C, D);
+}
+
+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];
+  const int D = dst[3 - BPS];
+  const int E = dst[4 - BPS];
+  const int F = dst[5 - BPS];
+  const int G = dst[6 - BPS];
+  const int H = dst[7 - BPS];
+  DST(0, 0) =             AVG2(A, B);
+  DST(1, 0) = DST(0, 2) = AVG2(B, C);
+  DST(2, 0) = DST(1, 2) = AVG2(C, D);
+  DST(3, 0) = DST(2, 2) = AVG2(D, E);
+
+  DST(0, 1) =             AVG3(A, B, C);
+  DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
+  DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
+  DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
+              DST(3, 2) = AVG3(E, F, G);
+              DST(3, 3) = AVG3(F, G, H);
+}
+
+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];
+  const int L = dst[-1 + 3 * BPS];
+  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 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];
+  const int L = dst[-1 + 3 * BPS];
+  const int X = dst[-1 - BPS];
+  const int A = dst[0 - BPS];
+  const int B = dst[1 - BPS];
+  const int C = dst[2 - BPS];
+
+  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);
+}
+
+#undef DST
+#undef AVG3
+#undef AVG2
+
+//------------------------------------------------------------------------------
+// Chroma
+
+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
+  int j;
+  for (j = 0; j < 8; ++j) {
+    memset(dst, dst[-1], 8);
+    dst += BPS;
+  }
+}
+
+// helper for chroma-DC predictions
+static WEBP_INLINE void Put8x8uv(uint8_t value, uint8_t* dst) {
+  int j;
+#ifndef WEBP_REFERENCE_IMPLEMENTATION
+  const uint64_t v = (uint64_t)value * 0x0101010101010101ULL;
+  for (j = 0; j < 8; ++j) {
+    *(uint64_t*)(dst + j * BPS) = v;
+  }
+#else
+  for (j = 0; j < 8; ++j) memset(dst + j * BPS, value, 8);
+#endif
+}
+
+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(dc0 >> 4, dst);
+}
+
+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(dc0 >> 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);
+}
+
+//------------------------------------------------------------------------------
+// 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
+};
+
+//------------------------------------------------------------------------------
+// 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];
+}
+
+// 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 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];
+}
+
+static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return (abs0[255 + p1 - p0] > thresh) || (abs0[255 + q1 - q0] > thresh);
+}
+
+static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int thresh) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return (2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) <= thresh;
+}
+
+static WEBP_INLINE int needs_filter2(const uint8_t* p,
+                                     int step, int t, int it) {
+  const int p3 = p[-4*step], p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step];
+  const int q0 = p[0], q1 = p[step], q2 = p[2*step], q3 = p[3*step];
+  if ((2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) > t)
+    return 0;
+  return abs0[255 + p3 - p2] <= it && abs0[255 + p2 - p1] <= it &&
+         abs0[255 + p1 - p0] <= it && abs0[255 + q3 - q2] <= it &&
+         abs0[255 + q2 - q1] <= it && abs0[255 + q1 - q0] <= it;
+}
+
+//------------------------------------------------------------------------------
+// Simple In-loop filtering (Paragraph 15.2)
+
+static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    if (needs_filter(p + i, stride, thresh)) {
+      do_filter2(p + i, stride);
+    }
+  }
+}
+
+static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    if (needs_filter(p + i * stride, 1, thresh)) {
+      do_filter2(p + i * stride, 1);
+    }
+  }
+}
+
+static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    SimpleVFilter16(p, stride, thresh);
+  }
+}
+
+static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    SimpleHFilter16(p, stride, thresh);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Complex In-loop filtering (Paragraph 15.3)
+
+static WEBP_INLINE void FilterLoop26(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_filter6(p, hstride);
+      }
+    }
+    p += vstride;
+  }
+}
+
+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 macroblock edges
+static void VFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop26(p, stride, 1, 16, thresh, ithresh, hev_thresh);
+}
+
+static void HFilter16(uint8_t* p, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop26(p, 1, stride, 16, thresh, ithresh, hev_thresh);
+}
+
+// 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);
+  }
+}
+
+// 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) {
+  FilterLoop26(u, stride, 1, 8, thresh, ithresh, hev_thresh);
+  FilterLoop26(v, stride, 1, 8, thresh, ithresh, hev_thresh);
+}
+
+static void HFilter8(uint8_t* u, uint8_t* v, int stride,
+                     int thresh, int ithresh, int hev_thresh) {
+  FilterLoop26(u, 1, stride, 8, thresh, ithresh, hev_thresh);
+  FilterLoop26(v, 1, stride, 8, 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);
+}
+
+//------------------------------------------------------------------------------
+
+VP8DecIdct2 VP8Transform;
+VP8DecIdct VP8TransformAC3;
+VP8DecIdct VP8TransformUV;
+VP8DecIdct VP8TransformDC;
+VP8DecIdct VP8TransformDCUV;
+
+VP8LumaFilterFunc VP8VFilter16;
+VP8LumaFilterFunc VP8HFilter16;
+VP8ChromaFilterFunc VP8VFilter8;
+VP8ChromaFilterFunc VP8HFilter8;
+VP8LumaFilterFunc VP8VFilter16i;
+VP8LumaFilterFunc VP8HFilter16i;
+VP8ChromaFilterFunc VP8VFilter8i;
+VP8ChromaFilterFunc VP8HFilter8i;
+VP8SimpleFilterFunc VP8SimpleVFilter16;
+VP8SimpleFilterFunc VP8SimpleHFilter16;
+VP8SimpleFilterFunc VP8SimpleVFilter16i;
+VP8SimpleFilterFunc VP8SimpleHFilter16i;
+
+extern void VP8DspInitSSE2(void);
+extern void VP8DspInitNEON(void);
+
+void VP8DspInit(void) {
+  DspInitTables();
+
+  VP8Transform = TransformTwo;
+  VP8TransformUV = TransformUV;
+  VP8TransformDC = TransformDC;
+  VP8TransformDCUV = TransformDCUV;
+  VP8TransformAC3 = TransformAC3;
+
+  VP8VFilter16 = VFilter16;
+  VP8HFilter16 = HFilter16;
+  VP8VFilter8 = VFilter8;
+  VP8HFilter8 = HFilter8;
+  VP8VFilter16i = VFilter16i;
+  VP8HFilter16i = HFilter16i;
+  VP8VFilter8i = VFilter8i;
+  VP8HFilter8i = HFilter8i;
+  VP8SimpleVFilter16 = SimpleVFilter16;
+  VP8SimpleHFilter16 = SimpleHFilter16;
+  VP8SimpleVFilter16i = SimpleVFilter16i;
+  VP8SimpleHFilter16i = SimpleHFilter16i;
+
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
+  if (VP8GetCPUInfo) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      VP8DspInitSSE2();
+    }
+#elif defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      VP8DspInitNEON();
+    }
+#endif
+  }
+}
+
diff --git a/drivers/webp/dsp/dec_neon.c b/drivers/webp/dsp/dec_neon.c
new file mode 100644
index 0000000000..9c3d8cc016
--- /dev/null
+++ b/drivers/webp/dsp/dec_neon.c
@@ -0,0 +1,433 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// ARM NEON version of dsp functions and loop filtering.
+//
+// Authors: Somnath Banerjee (somnath@google.com)
+//          Johann Koenig (johannkoenig@google.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_NEON)
+
+#include "../dec/vp8i.h"
+
+#define QRegs "q0", "q1", "q2", "q3",                                          \
+              "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+
+#define FLIP_SIGN_BIT2(a, b, s)                                                \
+  "veor     " #a "," #a "," #s "               \n"                             \
+  "veor     " #b "," #b "," #s "               \n"                             \
+
+#define FLIP_SIGN_BIT4(a, b, c, d, s)                                          \
+  FLIP_SIGN_BIT2(a, b, s)                                                      \
+  FLIP_SIGN_BIT2(c, d, s)                                                      \
+
+#define NEEDS_FILTER(p1, p0, q0, q1, thresh, mask)                             \
+  "vabd.u8    q15," #p0 "," #q0 "         \n"  /* abs(p0 - q0) */              \
+  "vabd.u8    q14," #p1 "," #q1 "         \n"  /* abs(p1 - q1) */              \
+  "vqadd.u8   q15, q15, q15               \n"  /* abs(p0 - q0) * 2 */          \
+  "vshr.u8    q14, q14, #1                \n"  /* abs(p1 - q1) / 2 */          \
+  "vqadd.u8   q15, q15, q14     \n"  /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */ \
+  "vdup.8     q14, " #thresh "            \n"                                  \
+  "vcge.u8   " #mask ", q14, q15          \n"  /* mask <= thresh */
+
+#define GET_BASE_DELTA(p1, p0, q0, q1, o)                                      \
+  "vqsub.s8   q15," #q0 "," #p0 "         \n"  /* (q0 - p0) */                 \
+  "vqsub.s8  " #o "," #p1 "," #q1 "       \n"  /* (p1 - q1) */                 \
+  "vqadd.s8  " #o "," #o ", q15           \n"  /* (p1 - q1) + 1 * (p0 - q0) */ \
+  "vqadd.s8  " #o "," #o ", q15           \n"  /* (p1 - q1) + 2 * (p0 - q0) */ \
+  "vqadd.s8  " #o "," #o ", q15           \n"  /* (p1 - q1) + 3 * (p0 - q0) */
+
+#define DO_SIMPLE_FILTER(p0, q0, fl)                                           \
+  "vmov.i8    q15, #0x03                  \n"                                  \
+  "vqadd.s8   q15, q15, " #fl "           \n"  /* filter1 = filter + 3 */      \
+  "vshr.s8    q15, q15, #3                \n"  /* filter1 >> 3 */              \
+  "vqadd.s8  " #p0 "," #p0 ", q15         \n"  /* p0 += filter1 */             \
+                                                                               \
+  "vmov.i8    q15, #0x04                  \n"                                  \
+  "vqadd.s8   q15, q15, " #fl "           \n"  /* filter1 = filter + 4 */      \
+  "vshr.s8    q15, q15, #3                \n"  /* filter2 >> 3 */              \
+  "vqsub.s8  " #q0 "," #q0 ", q15         \n"  /* q0 -= filter2 */
+
+// Applies filter on 2 pixels (p0 and q0)
+#define DO_FILTER2(p1, p0, q0, q1, thresh)                                     \
+  NEEDS_FILTER(p1, p0, q0, q1, thresh, q9)     /* filter mask in q9 */         \
+  "vmov.i8    q10, #0x80                  \n"  /* sign bit */                  \
+  FLIP_SIGN_BIT4(p1, p0, q0, q1, q10)          /* convert to signed value */   \
+  GET_BASE_DELTA(p1, p0, q0, q1, q11)          /* get filter level  */         \
+  "vand       q9, q9, q11                 \n"  /* apply filter mask */         \
+  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) {
+  __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    {q12}, [%[p]]                  \n"  // q1
+
+    DO_FILTER2(q1, q2, q3, q12, %[thresh])
+
+    "sub        %[p], %[p], %[stride], lsl #1  \n"  // p -= 2 * stride
+
+    "vst1.u8    {q2}, [%[p]], %[stride]        \n"  // store op0
+    "vst1.u8    {q3}, [%[p]]                   \n"  // store oq0
+    : [p] "+r"(p)
+    : [stride] "r"(stride), [thresh] "r"(thresh)
+    : "memory", QRegs
+  );
+}
+
+static void SimpleHFilter16NEON(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(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, q12, q13, %[thresh])
+
+    "sub        %[p], %[p], #1                 \n"  // p - 1
+
+    "vswp        d5, d24                       \n"
+    STORE8x2(d4, d5, [%[p]], %[stride])
+    STORE8x2(d24, d25, [%[p]], %[stride])
+
+    : [p] "+r"(p)
+    : [stride] "r"(stride), [thresh] "r"(thresh)
+    : "memory", "r4", "r5", "r6", QRegs
+  );
+}
+
+static void SimpleVFilter16iNEON(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    SimpleVFilter16NEON(p, stride, thresh);
+  }
+}
+
+static void SimpleHFilter16iNEON(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    SimpleHFilter16NEON(p, stride, thresh);
+  }
+}
+
+//-----------------------------------------------------------------------------
+// Inverse transforms (Paragraph 14.4)
+
+static void TransformOne(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
+   */
+
+  /* Adapted from libvpx: vp8/common/arm/neon/shortidct4x4llm_neon.asm */
+  __asm__ volatile (
+    "vld1.16         {q1, q2}, [%[in]]           \n"
+    "vld1.16         {d0}, [%[constants]]        \n"
+
+    /* d2: in[0]
+     * d3: in[8]
+     * d4: in[4]
+     * d5: in[12]
+     */
+    "vswp            d3, d4                      \n"
+
+    /* q8 = {in[4], in[12]} * kC1 * 2 >> 16
+     * q9 = {in[4], in[12]} * kC2 >> 16
+     */
+    "vqdmulh.s16     q8, q2, d0[0]               \n"
+    "vqdmulh.s16     q9, q2, d0[1]               \n"
+
+    /* d22 = a = in[0] + in[8]
+     * d23 = b = in[0] - in[8]
+     */
+    "vqadd.s16       d22, d2, d3                 \n"
+    "vqsub.s16       d23, d2, d3                 \n"
+
+    /* The multiplication should be x * kC1 >> 16
+     * However, with vqdmulh we get x * kC1 * 2 >> 16
+     * (multiply, double, return high half)
+     * We avoided this in kC2 by pre-shifting the constant.
+     * q8 = in[4]/[12] * kC1 >> 16
+     */
+    "vshr.s16        q8, q8, #1                  \n"
+
+    /* Add {in[4], in[12]} back after the multiplication. This is handled by
+     * adding 1 << 16 to kC1 in the libwebp C code.
+     */
+    "vqadd.s16       q8, q2, q8                  \n"
+
+    /* d20 = c = in[4]*kC2 - in[12]*kC1
+     * d21 = d = in[4]*kC1 + in[12]*kC2
+     */
+    "vqsub.s16       d20, d18, d17               \n"
+    "vqadd.s16       d21, d19, d16               \n"
+
+    /* d2 = tmp[0] = a + d
+     * d3 = tmp[1] = b + c
+     * d4 = tmp[2] = b - c
+     * d5 = tmp[3] = a - d
+     */
+    "vqadd.s16       d2, d22, d21                \n"
+    "vqadd.s16       d3, d23, d20                \n"
+    "vqsub.s16       d4, d23, d20                \n"
+    "vqsub.s16       d5, d22, d21                \n"
+
+    "vzip.16         q1, q2                      \n"
+    "vzip.16         q1, q2                      \n"
+
+    "vswp            d3, d4                      \n"
+
+    /* q8 = {tmp[4], tmp[12]} * kC1 * 2 >> 16
+     * q9 = {tmp[4], tmp[12]} * kC2 >> 16
+     */
+    "vqdmulh.s16     q8, q2, d0[0]               \n"
+    "vqdmulh.s16     q9, q2, d0[1]               \n"
+
+    /* d22 = a = tmp[0] + tmp[8]
+     * d23 = b = tmp[0] - tmp[8]
+     */
+    "vqadd.s16       d22, d2, d3                 \n"
+    "vqsub.s16       d23, d2, d3                 \n"
+
+    /* See long winded explanations prior */
+    "vshr.s16        q8, q8, #1                  \n"
+    "vqadd.s16       q8, q2, q8                  \n"
+
+    /* d20 = c = in[4]*kC2 - in[12]*kC1
+     * d21 = d = in[4]*kC1 + in[12]*kC2
+     */
+    "vqsub.s16       d20, d18, d17               \n"
+    "vqadd.s16       d21, d19, d16               \n"
+
+    /* d2 = tmp[0] = a + d
+     * d3 = tmp[1] = b + c
+     * d4 = tmp[2] = b - c
+     * d5 = tmp[3] = a - d
+     */
+    "vqadd.s16       d2, d22, d21                \n"
+    "vqadd.s16       d3, d23, d20                \n"
+    "vqsub.s16       d4, d23, d20                \n"
+    "vqsub.s16       d5, d22, d21                \n"
+
+    "vld1.32         d6[0], [%[dst]], %[kBPS]    \n"
+    "vld1.32         d6[1], [%[dst]], %[kBPS]    \n"
+    "vld1.32         d7[0], [%[dst]], %[kBPS]    \n"
+    "vld1.32         d7[1], [%[dst]], %[kBPS]    \n"
+
+    "sub         %[dst], %[dst], %[kBPS], lsl #2 \n"
+
+    /* (val) + 4 >> 3 */
+    "vrshr.s16       d2, d2, #3                  \n"
+    "vrshr.s16       d3, d3, #3                  \n"
+    "vrshr.s16       d4, d4, #3                  \n"
+    "vrshr.s16       d5, d5, #3                  \n"
+
+    "vzip.16         q1, q2                      \n"
+    "vzip.16         q1, q2                      \n"
+
+    /* Must accumulate before saturating */
+    "vmovl.u8        q8, d6                      \n"
+    "vmovl.u8        q9, d7                      \n"
+
+    "vqadd.s16       q1, q1, q8                  \n"
+    "vqadd.s16       q2, q2, q9                  \n"
+
+    "vqmovun.s16     d0, q1                      \n"
+    "vqmovun.s16     d1, q2                      \n"
+
+    "vst1.32         d0[0], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d0[1], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d1[0], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d1[1], [%[dst]]             \n"
+
+    : [in] "+r"(in), [dst] "+r"(dst)  /* modified registers */
+    : [kBPS] "r"(kBPS), [constants] "r"(constants)  /* constants */
+    : "memory", "q0", "q1", "q2", "q8", "q9", "q10", "q11"  /* clobbered */
+  );
+}
+
+static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) {
+  TransformOne(in, dst);
+  if (do_two) {
+    TransformOne(in + 16, dst + 4);
+  }
+}
+
+static void TransformDC(const int16_t* in, uint8_t* dst) {
+  const int DC = (in[0] + 4) >> 3;
+  const int kBPS = BPS;
+  __asm__ volatile (
+    "vdup.16         q1, %[DC]        \n"
+
+    "vld1.32         d0[0], [%[dst]], %[kBPS]    \n"
+    "vld1.32         d1[0], [%[dst]], %[kBPS]    \n"
+    "vld1.32         d0[1], [%[dst]], %[kBPS]    \n"
+    "vld1.32         d1[1], [%[dst]], %[kBPS]    \n"
+
+    "sub         %[dst], %[dst], %[kBPS], lsl #2 \n"
+
+    // add DC and convert to s16.
+    "vaddw.u8        q2, q1, d0                  \n"
+    "vaddw.u8        q3, q1, d1                  \n"
+    // convert back to u8 with saturation
+    "vqmovun.s16     d0,  q2                     \n"
+    "vqmovun.s16     d1,  q3                     \n"
+
+    "vst1.32         d0[0], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d1[0], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d0[1], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d1[1], [%[dst]]             \n"
+    : [in] "+r"(in), [dst] "+r"(dst)  /* modified registers */
+    : [kBPS] "r"(kBPS),   /* constants */
+      [DC] "r"(DC)
+    : "memory", "q0", "q1", "q2", "q3"  /* clobbered */
+  );
+}
+
+static void TransformWHT(const int16_t* in, int16_t* out) {
+  const int kStep = 32;  // The store is only incrementing the pointer as if we
+                         // had stored a single byte.
+  __asm__ volatile (
+    // part 1
+    // load data into q0, q1
+    "vld1.16         {q0, q1}, [%[in]]           \n"
+
+    "vaddl.s16       q2, d0, d3                  \n"  // a0 = in[0] + in[12]
+    "vaddl.s16       q3, d1, d2                  \n"  // a1 = in[4] + in[8]
+    "vsubl.s16       q10, d1, d2                 \n"  // a2 = in[4] - in[8]
+    "vsubl.s16       q11, d0, d3                 \n"  // a3 = in[0] - in[12]
+
+    "vadd.s32        q0, q2, q3                  \n"  // tmp[0] = a0 + a1
+    "vsub.s32        q2, q2, q3                  \n"  // tmp[8] = a0 - a1
+    "vadd.s32        q1, q11, q10                \n"  // tmp[4] = a3 + a2
+    "vsub.s32        q3, q11, q10                \n"  // tmp[12] = a3 - a2
+
+    // Transpose
+    // q0 = tmp[0, 4, 8, 12], q1 = tmp[2, 6, 10, 14]
+    // q2 = tmp[1, 5, 9, 13], q3 = tmp[3, 7, 11, 15]
+    "vswp            d1, d4                      \n"  // vtrn.64 q0, q2
+    "vswp            d3, d6                      \n"  // vtrn.64 q1, q3
+    "vtrn.32         q0, q1                      \n"
+    "vtrn.32         q2, q3                      \n"
+
+    "vmov.s32        q10, #3                     \n"  // dc = 3
+    "vadd.s32        q0, q0, q10                 \n"  // dc = tmp[0] + 3
+    "vadd.s32        q12, q0, q3                 \n"  // a0 = dc + tmp[3]
+    "vadd.s32        q13, q1, q2                 \n"  // a1 = tmp[1] + tmp[2]
+    "vsub.s32        q8, q1, q2                  \n"  // a2 = tmp[1] - tmp[2]
+    "vsub.s32        q9, q0, q3                  \n"  // a3 = dc - tmp[3]
+
+    "vadd.s32        q0, q12, q13                \n"
+    "vshrn.s32       d0, q0, #3                  \n"  // (a0 + a1) >> 3
+    "vadd.s32        q1, q9, q8                  \n"
+    "vshrn.s32       d1, q1, #3                  \n"  // (a3 + a2) >> 3
+    "vsub.s32        q2, q12, q13                \n"
+    "vshrn.s32       d2, q2, #3                  \n"  // (a0 - a1) >> 3
+    "vsub.s32        q3, q9, q8                  \n"
+    "vshrn.s32       d3, q3, #3                  \n"  // (a3 - a2) >> 3
+
+    // set the results to output
+    "vst1.16         d0[0], [%[out]], %[kStep]   \n"
+    "vst1.16         d1[0], [%[out]], %[kStep]   \n"
+    "vst1.16         d2[0], [%[out]], %[kStep]   \n"
+    "vst1.16         d3[0], [%[out]], %[kStep]   \n"
+    "vst1.16         d0[1], [%[out]], %[kStep]   \n"
+    "vst1.16         d1[1], [%[out]], %[kStep]   \n"
+    "vst1.16         d2[1], [%[out]], %[kStep]   \n"
+    "vst1.16         d3[1], [%[out]], %[kStep]   \n"
+    "vst1.16         d0[2], [%[out]], %[kStep]   \n"
+    "vst1.16         d1[2], [%[out]], %[kStep]   \n"
+    "vst1.16         d2[2], [%[out]], %[kStep]   \n"
+    "vst1.16         d3[2], [%[out]], %[kStep]   \n"
+    "vst1.16         d0[3], [%[out]], %[kStep]   \n"
+    "vst1.16         d1[3], [%[out]], %[kStep]   \n"
+    "vst1.16         d2[3], [%[out]], %[kStep]   \n"
+    "vst1.16         d3[3], [%[out]], %[kStep]   \n"
+
+    : [out] "+r"(out)  // modified registers
+    : [in] "r"(in), [kStep] "r"(kStep)  // constants
+    : "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13"  // clobbered
+  );
+}
+
+#endif   // WEBP_USE_NEON
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8DspInitNEON(void);
+
+void VP8DspInitNEON(void) {
+#if defined(WEBP_USE_NEON)
+  VP8Transform = TransformTwo;
+  VP8TransformAC3 = TransformOne;  // no special code here
+  VP8TransformDC = TransformDC;
+  VP8TransformWHT = TransformWHT;
+
+  VP8SimpleVFilter16 = SimpleVFilter16NEON;
+  VP8SimpleHFilter16 = SimpleHFilter16NEON;
+  VP8SimpleVFilter16i = SimpleVFilter16iNEON;
+  VP8SimpleHFilter16i = SimpleHFilter16iNEON;
+#endif   // WEBP_USE_NEON
+}
+
diff --git a/drivers/webp/dsp/dec_sse2.c b/drivers/webp/dsp/dec_sse2.c
new file mode 100644
index 0000000000..150c559f13
--- /dev/null
+++ b/drivers/webp/dsp/dec_sse2.c
@@ -0,0 +1,956 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// SSE2 version of some decoding functions (idct, loop filtering).
+//
+// Author: somnath@google.com (Somnath Banerjee)
+//         cduvivier@google.com (Christian Duvivier)
+
+#include "./dsp.h"
+
+#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"
+
+//------------------------------------------------------------------------------
+// Transforms (Paragraph 14.4)
+
+static void TransformSSE2(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
+  //    K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16
+  //
+  // To be able to use signed 16-bit integers, we use the following trick to
+  // have constants within range:
+  // - Associated constants are obtained by subtracting the 16-bit fixed point
+  //   version of one:
+  //      k = K - (1 << 16)  =>  K = k + (1 << 16)
+  //      K1 = 85267  =>  k1 =  20091
+  //      K2 = 35468  =>  k2 = -30068
+  // - The multiplication of a variable by a constant become the sum of the
+  //   variable and the multiplication of that variable by the associated
+  //   constant:
+  //      (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x
+  const __m128i k1 = _mm_set1_epi16(20091);
+  const __m128i k2 = _mm_set1_epi16(-30068);
+  __m128i T0, T1, T2, T3;
+
+  // Load and concatenate the transform coefficients (we'll do two transforms
+  // in parallel). In the case of only one transform, the second half of the
+  // 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]);
+    // 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]);
+      in0 = _mm_unpacklo_epi64(in0, inB0);
+      in1 = _mm_unpacklo_epi64(in1, inB1);
+      in2 = _mm_unpacklo_epi64(in2, inB2);
+      in3 = _mm_unpacklo_epi64(in3, inB3);
+      // a00 a10 a20 a30   b00 b10 b20 b30
+      // a01 a11 a21 a31   b01 b11 b21 b31
+      // a02 a12 a22 a32   b02 b12 b22 b32
+      // a03 a13 a23 a33   b03 b13 b23 b33
+    }
+  }
+
+  // Vertical pass and subsequent transpose.
+  {
+    // First pass, c and d calculations are longer because of the "trick"
+    // multiplications.
+    const __m128i a = _mm_add_epi16(in0, in2);
+    const __m128i b = _mm_sub_epi16(in0, in2);
+    // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
+    const __m128i c1 = _mm_mulhi_epi16(in1, k2);
+    const __m128i c2 = _mm_mulhi_epi16(in3, k1);
+    const __m128i c3 = _mm_sub_epi16(in1, in3);
+    const __m128i c4 = _mm_sub_epi16(c1, c2);
+    const __m128i c = _mm_add_epi16(c3, c4);
+    // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
+    const __m128i d1 = _mm_mulhi_epi16(in1, k1);
+    const __m128i d2 = _mm_mulhi_epi16(in3, k2);
+    const __m128i d3 = _mm_add_epi16(in1, in3);
+    const __m128i d4 = _mm_add_epi16(d1, d2);
+    const __m128i d = _mm_add_epi16(d3, d4);
+
+    // Second pass.
+    const __m128i tmp0 = _mm_add_epi16(a, d);
+    const __m128i tmp1 = _mm_add_epi16(b, c);
+    const __m128i tmp2 = _mm_sub_epi16(b, c);
+    const __m128i tmp3 = _mm_sub_epi16(a, d);
+
+    // Transpose the two 4x4.
+    // a00 a01 a02 a03   b00 b01 b02 b03
+    // a10 a11 a12 a13   b10 b11 b12 b13
+    // a20 a21 a22 a23   b20 b21 b22 b23
+    // a30 a31 a32 a33   b30 b31 b32 b33
+    const __m128i transpose0_0 = _mm_unpacklo_epi16(tmp0, tmp1);
+    const __m128i transpose0_1 = _mm_unpacklo_epi16(tmp2, tmp3);
+    const __m128i transpose0_2 = _mm_unpackhi_epi16(tmp0, tmp1);
+    const __m128i transpose0_3 = _mm_unpackhi_epi16(tmp2, tmp3);
+    // a00 a10 a01 a11   a02 a12 a03 a13
+    // a20 a30 a21 a31   a22 a32 a23 a33
+    // b00 b10 b01 b11   b02 b12 b03 b13
+    // b20 b30 b21 b31   b22 b32 b23 b33
+    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
+    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
+    // a00 a10 a20 a30 a01 a11 a21 a31
+    // b00 b10 b20 b30 b01 b11 b21 b31
+    // a02 a12 a22 a32 a03 a13 a23 a33
+    // b02 b12 a22 b32 b03 b13 b23 b33
+    T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
+    T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
+    T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
+    T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
+    // a00 a10 a20 a30   b00 b10 b20 b30
+    // a01 a11 a21 a31   b01 b11 b21 b31
+    // a02 a12 a22 a32   b02 b12 b22 b32
+    // a03 a13 a23 a33   b03 b13 b23 b33
+  }
+
+  // Horizontal pass and subsequent transpose.
+  {
+    // First pass, c and d calculations are longer because of the "trick"
+    // multiplications.
+    const __m128i four = _mm_set1_epi16(4);
+    const __m128i dc = _mm_add_epi16(T0, four);
+    const __m128i a =  _mm_add_epi16(dc, T2);
+    const __m128i b =  _mm_sub_epi16(dc, T2);
+    // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
+    const __m128i c1 = _mm_mulhi_epi16(T1, k2);
+    const __m128i c2 = _mm_mulhi_epi16(T3, k1);
+    const __m128i c3 = _mm_sub_epi16(T1, T3);
+    const __m128i c4 = _mm_sub_epi16(c1, c2);
+    const __m128i c = _mm_add_epi16(c3, c4);
+    // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
+    const __m128i d1 = _mm_mulhi_epi16(T1, k1);
+    const __m128i d2 = _mm_mulhi_epi16(T3, k2);
+    const __m128i d3 = _mm_add_epi16(T1, T3);
+    const __m128i d4 = _mm_add_epi16(d1, d2);
+    const __m128i d = _mm_add_epi16(d3, d4);
+
+    // Second pass.
+    const __m128i tmp0 = _mm_add_epi16(a, d);
+    const __m128i tmp1 = _mm_add_epi16(b, c);
+    const __m128i tmp2 = _mm_sub_epi16(b, c);
+    const __m128i tmp3 = _mm_sub_epi16(a, d);
+    const __m128i shifted0 = _mm_srai_epi16(tmp0, 3);
+    const __m128i shifted1 = _mm_srai_epi16(tmp1, 3);
+    const __m128i shifted2 = _mm_srai_epi16(tmp2, 3);
+    const __m128i shifted3 = _mm_srai_epi16(tmp3, 3);
+
+    // Transpose the two 4x4.
+    // a00 a01 a02 a03   b00 b01 b02 b03
+    // a10 a11 a12 a13   b10 b11 b12 b13
+    // a20 a21 a22 a23   b20 b21 b22 b23
+    // a30 a31 a32 a33   b30 b31 b32 b33
+    const __m128i transpose0_0 = _mm_unpacklo_epi16(shifted0, shifted1);
+    const __m128i transpose0_1 = _mm_unpacklo_epi16(shifted2, shifted3);
+    const __m128i transpose0_2 = _mm_unpackhi_epi16(shifted0, shifted1);
+    const __m128i transpose0_3 = _mm_unpackhi_epi16(shifted2, shifted3);
+    // a00 a10 a01 a11   a02 a12 a03 a13
+    // a20 a30 a21 a31   a22 a32 a23 a33
+    // b00 b10 b01 b11   b02 b12 b03 b13
+    // b20 b30 b21 b31   b22 b32 b23 b33
+    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
+    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
+    // a00 a10 a20 a30 a01 a11 a21 a31
+    // b00 b10 b20 b30 b01 b11 b21 b31
+    // a02 a12 a22 a32 a03 a13 a23 a33
+    // b02 b12 a22 b32 b03 b13 b23 b33
+    T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
+    T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
+    T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
+    T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
+    // a00 a10 a20 a30   b00 b10 b20 b30
+    // a01 a11 a21 a31   b01 b11 b21 b31
+    // a02 a12 a22 a32   b02 b12 b22 b32
+    // a03 a13 a23 a33   b03 b13 b23 b33
+  }
+
+  // Add inverse transform to 'dst' and store.
+  {
+    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));
+    } 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));
+    }
+    // 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(s).
+    dst0 = _mm_add_epi16(dst0, T0);
+    dst1 = _mm_add_epi16(dst1, T1);
+    dst2 = _mm_add_epi16(dst2, T2);
+    dst3 = _mm_add_epi16(dst3, T3);
+    // 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.
+    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);
+    } else {
+      // Store four bytes/pixels per line.
+      *(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 TransformAC3SSE2(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)
+
+// Compute abs(p - q) = subs(p - q) OR subs(q - p)
+#define MM_ABS(p, q)  _mm_or_si128(                                            \
+    _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);                                                      \
+}
+
+#define FLIP_SIGN_BIT2(a, b) {                                                 \
+  a = _mm_xor_si128(a, sign_bit);                                              \
+  b = _mm_xor_si128(b, sign_bit);                                              \
+}
+
+#define FLIP_SIGN_BIT4(a, b, c, d) {                                           \
+  FLIP_SIGN_BIT2(a, b);                                                        \
+  FLIP_SIGN_BIT2(c, d);                                                        \
+}
+
+#define GET_NOTHEV(p1, p0, q0, q1, hev_thresh, 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_3 = _mm_subs_epu8(t_1, h);  /* abs(p1 - p0) - hev_tresh */   \
+  const __m128i t_4 = _mm_subs_epu8(t_2, h);  /* abs(q1 - q0) - hev_tresh */   \
+                                                                               \
+  not_hev = _mm_or_si128(t_3, t_4);                                            \
+  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 */                         \
+}
+
+// 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)]
+#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 delta = _mm_packs_epi16(a_lo7, a_hi7);                         \
+  pi = _mm_adds_epi8(pi, delta);                                               \
+  qi = _mm_subs_epi8(qi, delta);                                               \
+}
+
+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
+
+  *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
+
+  t1 = _mm_set1_epi8(thresh);
+  *mask = _mm_subs_epu8(*mask, t1);     // mask <= thresh
+  *mask = _mm_cmpeq_epi8(*mask, _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) {
+  __m128i a, mask;
+  const __m128i sign_bit = _mm_set1_epi8(0x80);
+  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);
+  a = _mm_and_si128(a, mask);     // mask filter values we don't care about
+  DO_SIMPLE_FILTER(*p0, *q0, a);
+
+  // unoffset
+  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) {
+  __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);
+
+  // convert to signed values
+  FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
+
+  t1 = _mm_subs_epi8(*p1, *q1);        // p1 - q1
+  t1 = _mm_andnot_si128(not_hev, t1);  // hev(p1 - q1)
+  t2 = _mm_subs_epi8(*q0, *p0);        // q0 - p0
+  t1 = _mm_adds_epi8(t1, t2);          // hev(p1 - q1) + 1 * (q0 - p0)
+  t1 = _mm_adds_epi8(t1, t2);          // hev(p1 - q1) + 2 * (q0 - 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
+  *p0 = _mm_adds_epi8(*p0, t2);      // p0 += t2
+
+  t2 = _mm_set1_epi8(1);
+  t3 = _mm_adds_epi8(t3, t2);
+  SIGNED_SHIFT_N(t3, 1);             // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 4
+
+  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);
+}
+
+// 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;
+  const __m128i sign_bit = _mm_set1_epi8(0x80);
+
+  // compute hev mask
+  GET_NOTHEV(*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);
+
+  { // 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);
+  }
+  { // 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 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 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 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 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
+
+    UPDATE_2PIXELS(*p2, *q2, a2_lo, a2_hi);
+    UPDATE_2PIXELS(*p1, *q1, a1_lo, a1_hi);
+    UPDATE_2PIXELS(*p0, *q0, a0_lo, a0_hi);
+  }
+
+  // unoffset
+  FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
+  FLIP_SIGN_BIT2(*p2, *q2);
+}
+
+// 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);
+
+  // *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);
+}
+
+static WEBP_INLINE void Load16x4(const uint8_t* r0, const uint8_t* r8,
+                                 int stride,
+                                 __m128i* p1, __m128i* p0,
+                                 __m128i* q0, __m128i* q1) {
+  __m128i t1, t2;
+  // Assume the pixels around the edge (|) are numbered as follows
+  //                00 01 | 02 03
+  //                10 11 | 12 13
+  //                 ...  |  ...
+  //                e0 e1 | e2 e3
+  //                f0 f1 | f2 f3
+  //
+  // r0 is pointing to the 0th row (00)
+  // r8 is pointing to the 8th row (80)
+
+  // Load
+  // p1 = 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00
+  // q0 = 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
+  // p0 = f1 e1 d1 c1 b1 a1 91 81 f0 e0 d0 c0 b0 a0 90 80
+  // q1 = f3 e3 d3 c3 b3 a3 93 83 f2 e2 d2 c2 b2 a2 92 82
+  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);
+}
+
+static WEBP_INLINE void Store4x4(__m128i* x, uint8_t* dst, int stride) {
+  int i;
+  for (i = 0; i < 4; ++i, dst += stride) {
+    *((int32_t*)dst) = _mm_cvtsi128_si32(*x);
+    *x = _mm_srli_si128(*x, 4);
+  }
+}
+
+// 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;
+
+  // 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);
+
+  // 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);
+
+  // 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);
+
+  // 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);
+
+  Store4x4(p0, r0, stride);
+  r0 += 4 * stride;
+  Store4x4(q0, r0, stride);
+
+  Store4x4(p1, r8, stride);
+  r8 += 4 * stride;
+  Store4x4(q1, r8, stride);
+}
+
+//------------------------------------------------------------------------------
+// Simple In-loop filtering (Paragraph 15.2)
+
+static void SimpleVFilter16SSE2(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]);
+  __m128i q0 = _mm_loadu_si128((__m128i*)&p[0]);
+  __m128i q1 = _mm_loadu_si128((__m128i*)&p[stride]);
+
+  DoFilter2(&p1, &p0, &q0, &q1, thresh);
+
+  // Store
+  _mm_storeu_si128((__m128i*)&p[-stride], p0);
+  _mm_storeu_si128((__m128i*)p, q0);
+}
+
+static void SimpleHFilter16SSE2(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);
+  DoFilter2(&p1, &p0, &q0, &q1, thresh);
+  Store16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1);
+}
+
+static void SimpleVFilter16iSSE2(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    SimpleVFilter16SSE2(p, stride, thresh);
+  }
+}
+
+static void SimpleHFilter16iSSE2(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    SimpleHFilter16SSE2(p, stride, thresh);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Complex In-loop filtering (Paragraph 15.3)
+
+#define MAX_DIFF1(p3, p2, p1, p0, m) {                                         \
+  m = MM_ABS(p3, p2);                                                          \
+  m = _mm_max_epu8(m, MM_ABS(p2, p1));                                         \
+  m = _mm_max_epu8(m, MM_ABS(p1, p0));                                         \
+}
+
+#define MAX_DIFF2(p3, p2, p1, p0, m) {                                         \
+  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));                                         \
+}
+
+#define LOAD_H_EDGES4(p, stride, e1, e2, e3, e4) {                             \
+  e1 = _mm_loadu_si128((__m128i*)&(p)[0 * stride]);                            \
+  e2 = _mm_loadu_si128((__m128i*)&(p)[1 * stride]);                            \
+  e3 = _mm_loadu_si128((__m128i*)&(p)[2 * stride]);                            \
+  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_EDGES4(u, v, stride, e1, e2, e3, e4) {                        \
+  LOADUV_H_EDGE(e1, u, v, 0 * stride);                                         \
+  LOADUV_H_EDGE(e2, u, v, 1 * stride);                                         \
+  LOADUV_H_EDGE(e3, u, v, 2 * stride);                                         \
+  LOADUV_H_EDGE(e4, u, v, 3 * stride);                                         \
+}
+
+#define STOREUV(p, u, v, stride) {                                             \
+  _mm_storel_epi64((__m128i*)&u[(stride)], p);                                 \
+  p = _mm_srli_si128(p, 8);                                                    \
+  _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);                                          \
+}
+
+// on macroblock edges
+static void VFilter16SSE2(uint8_t* p, int stride,
+                          int thresh, int ithresh, int hev_thresh) {
+  __m128i t1;
+  __m128i mask;
+  __m128i p2, p1, p0, q0, q1, q2;
+
+  // Load p3, p2, p1, p0
+  LOAD_H_EDGES4(p - 4 * stride, stride, t1, p2, p1, p0);
+  MAX_DIFF1(t1, p2, p1, p0, mask);
+
+  // Load q0, q1, q2, q3
+  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);
+  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);
+}
+
+static void HFilter16SSE2(uint8_t* p, int stride,
+                          int thresh, int ithresh, int hev_thresh) {
+  __m128i mask;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+
+  uint8_t* const b = p - 4;
+  Load16x4(b, b + 8 * stride, stride, &p3, &p2, &p1, &p0);  // p3, p2, p1, p0
+  MAX_DIFF1(p3, p2, p1, p0, mask);
+
+  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);
+  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);
+}
+
+// on three inner edges
+static void VFilter16iSSE2(uint8_t* p, int stride,
+                           int thresh, int ithresh, int hev_thresh) {
+  int k;
+  __m128i mask;
+  __m128i t1, t2, p1, p0, q0, q1;
+
+  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);
+
+    p += 4 * stride;
+
+    // Load q0, q1, q2, q3
+    LOAD_H_EDGES4(p, stride, q0, q1, t1, t2);
+    MAX_DIFF2(t2, t1, q1, q0, mask);
+
+    COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+    DoFilter4(&p1, &p0, &q0, &q1, &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);
+  }
+}
+
+static void HFilter16iSSE2(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;
+
+  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);
+
+    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);
+
+    COMPLEX_FL_MASK(p1, p0, q0, q1, thresh, ithresh, mask);
+    DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
+
+    b -= 2;  // beginning of p1
+    Store16x4(b, b + 8 * stride, stride, &p1, &p0, &q0, &q1);
+
+    p += 4;
+  }
+}
+
+// 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) {
+  __m128i mask;
+  __m128i t1, p2, p1, p0, q0, q1, q2;
+
+  // Load p3, p2, p1, p0
+  LOADUV_H_EDGES4(u - 4 * stride, v - 4 * stride, stride, t1, p2, p1, p0);
+  MAX_DIFF1(t1, p2, p1, p0, mask);
+
+  // Load q0, q1, q2, q3
+  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);
+  DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
+
+  // Store
+  STOREUV(p2, u, v, -3 * stride);
+  STOREUV(p1, u, v, -2 * stride);
+  STOREUV(p0, u, v, -1 * stride);
+  STOREUV(q0, u, v, 0 * stride);
+  STOREUV(q1, u, v, 1 * 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) {
+  __m128i mask;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+
+  uint8_t* const tu = u - 4;
+  uint8_t* const tv = v - 4;
+  Load16x4(tu, tv, stride, &p3, &p2, &p1, &p0);  // p3, p2, p1, p0
+  MAX_DIFF1(p3, p2, p1, p0, mask);
+
+  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);
+  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);
+}
+
+static void VFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
+                          int thresh, int ithresh, int hev_thresh) {
+  __m128i mask;
+  __m128i t1, t2, p1, p0, q0, q1;
+
+  // Load p3, p2, p1, p0
+  LOADUV_H_EDGES4(u, v, stride, t2, t1, p1, p0);
+  MAX_DIFF1(t2, t1, p1, p0, mask);
+
+  u += 4 * stride;
+  v += 4 * stride;
+
+  // Load q0, q1, q2, q3
+  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);
+  DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
+
+  // Store
+  STOREUV(p1, u, v, -2 * stride);
+  STOREUV(p0, u, v, -1 * stride);
+  STOREUV(q0, u, v, 0 * 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) {
+  __m128i mask;
+  __m128i t1, t2, p1, p0, q0, q1;
+  Load16x4(u, v, stride, &t2, &t1, &p1, &p0);   // p3, p2, p1, p0
+  MAX_DIFF1(t2, t1, p1, p0, mask);
+
+  u += 4;  // beginning of q0
+  v += 4;
+  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);
+  DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
+
+  u -= 2;  // beginning of p1
+  v -= 2;
+  Store16x4(u, v, stride, &p1, &p0, &q0, &q1);
+}
+
+#endif   // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8DspInitSSE2(void);
+
+void VP8DspInitSSE2(void) {
+#if defined(WEBP_USE_SSE2)
+  VP8Transform = TransformSSE2;
+#if defined(USE_TRANSFORM_AC3)
+  VP8TransformAC3 = TransformAC3SSE2;
+#endif
+
+  VP8VFilter16 = VFilter16SSE2;
+  VP8HFilter16 = HFilter16SSE2;
+  VP8VFilter8 = VFilter8SSE2;
+  VP8HFilter8 = HFilter8SSE2;
+  VP8VFilter16i = VFilter16iSSE2;
+  VP8HFilter16i = HFilter16iSSE2;
+  VP8VFilter8i = VFilter8iSSE2;
+  VP8HFilter8i = HFilter8iSSE2;
+
+  VP8SimpleVFilter16 = SimpleVFilter16SSE2;
+  VP8SimpleHFilter16 = SimpleHFilter16SSE2;
+  VP8SimpleVFilter16i = SimpleVFilter16iSSE2;
+  VP8SimpleHFilter16i = SimpleHFilter16iSSE2;
+#endif   // WEBP_USE_SSE2
+}
+
diff --git a/drivers/webp/dsp/dsp.h b/drivers/webp/dsp/dsp.h
new file mode 100644
index 0000000000..3be783afe7
--- /dev/null
+++ b/drivers/webp/dsp/dsp.h
@@ -0,0 +1,224 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//   Speed-critical functions.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_DSP_DSP_H_
+#define WEBP_DSP_DSP_H_
+
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// CPU detection
+
+#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)
+#define WEBP_USE_SSE2
+#endif
+
+#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__)
+#define WEBP_ANDROID_NEON  // Android targets that might support NEON
+#endif
+
+#if defined(__ARM_NEON__) || defined(WEBP_ANDROID_NEON)
+#define WEBP_USE_NEON
+#endif
+
+typedef enum {
+  kSSE2,
+  kSSE3,
+  kNEON
+} CPUFeature;
+// returns true if the CPU supports the feature.
+typedef int (*VP8CPUInfo)(CPUFeature feature);
+extern VP8CPUInfo VP8GetCPUInfo;
+
+//------------------------------------------------------------------------------
+// Encoding
+
+// Transforms
+// VP8Idct: Does one of two inverse transforms. If do_two is set, the transforms
+//          will be done for (ref, in, dst) and (ref + 4, in + 16, dst + 4).
+typedef void (*VP8Idct)(const uint8_t* ref, const int16_t* in, uint8_t* dst,
+                        int do_two);
+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 VP8WHT VP8FTransformWHT;
+// Predictions
+// *dst is the destination block. *top and *left can be NULL.
+typedef void (*VP8IntraPreds)(uint8_t *dst, const uint8_t* left,
+                              const uint8_t* top);
+typedef void (*VP8Intra4Preds)(uint8_t *dst, const uint8_t* top);
+extern VP8Intra4Preds VP8EncPredLuma4;
+extern VP8IntraPreds VP8EncPredLuma16;
+extern VP8IntraPreds VP8EncPredChroma8;
+
+typedef int (*VP8Metric)(const uint8_t* pix, const uint8_t* ref);
+extern VP8Metric VP8SSE16x16, VP8SSE16x8, VP8SSE8x8, VP8SSE4x4;
+typedef int (*VP8WMetric)(const uint8_t* pix, const uint8_t* ref,
+                          const uint16_t* const weights);
+extern VP8WMetric VP8TDisto4x4, VP8TDisto16x16;
+
+typedef void (*VP8BlockCopy)(const uint8_t* src, uint8_t* dst);
+extern VP8BlockCopy VP8Copy4x4;
+// Quantization
+struct VP8Matrix;   // forward declaration
+typedef int (*VP8QuantizeBlock)(int16_t in[16], int16_t out[16],
+                                int n, const struct VP8Matrix* const mtx);
+extern VP8QuantizeBlock VP8EncQuantizeBlock;
+
+// specific to 2nd transform:
+typedef int (*VP8QuantizeBlockWHT)(int16_t in[16], int16_t out[16],
+                                   const struct VP8Matrix* const mtx);
+extern VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
+
+// Collect histogram for susceptibility calculation and accumulate in histo[].
+struct VP8Histogram;
+typedef void (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred,
+                          int start_block, int end_block,
+                          struct VP8Histogram* const histo);
+extern const int VP8DspScan[16 + 4 + 4];
+extern VP8CHisto VP8CollectHistogram;
+
+void VP8EncDspInit(void);   // must be called before using any of the above
+
+//------------------------------------------------------------------------------
+// Decoding
+
+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 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 */];
+
+// simple filter (only for luma)
+typedef void (*VP8SimpleFilterFunc)(uint8_t* p, int stride, int thresh);
+extern VP8SimpleFilterFunc VP8SimpleVFilter16;
+extern VP8SimpleFilterFunc VP8SimpleHFilter16;
+extern VP8SimpleFilterFunc VP8SimpleVFilter16i;  // filter 3 inner edges
+extern VP8SimpleFilterFunc VP8SimpleHFilter16i;
+
+// regular filter (on both macroblock edges and inner edges)
+typedef void (*VP8LumaFilterFunc)(uint8_t* luma, int stride,
+                                  int thresh, int ithresh, int hev_t);
+typedef void (*VP8ChromaFilterFunc)(uint8_t* u, uint8_t* v, int stride,
+                                    int thresh, int ithresh, int hev_t);
+// on outer edge
+extern VP8LumaFilterFunc VP8VFilter16;
+extern VP8LumaFilterFunc VP8HFilter16;
+extern VP8ChromaFilterFunc VP8VFilter8;
+extern VP8ChromaFilterFunc VP8HFilter8;
+
+// on inner edge
+extern VP8LumaFilterFunc VP8VFilter16i;   // filtering 3 inner edges altogether
+extern VP8LumaFilterFunc VP8HFilter16i;
+extern VP8ChromaFilterFunc VP8VFilter8i;  // filtering u and v altogether
+extern VP8ChromaFilterFunc VP8HFilter8i;
+
+// must be called before anything using the above
+void VP8DspInit(void);
+
+//------------------------------------------------------------------------------
+// WebP I/O
+
+#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,
+    const uint8_t* cur_u, const uint8_t* cur_v,
+    uint8_t* top_dst, uint8_t* bottom_dst, int len);
+
+#ifdef FANCY_UPSAMPLING
+
+// 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);
+
+// NEON version
+void WebPInitUpsamplersNEON(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);
+
+extern const WebPSampleLinePairFunc 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
+// as 0x00, 0x00, 0x00, 0xff (little endian).
+WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last);
+
+// YUV444->RGB converters
+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 */];
+
+// Main function to be called
+void WebPInitUpsamplers(void);
+
+//------------------------------------------------------------------------------
+// Pre-multiply planes with alpha values
+
+// 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).
+extern void (*WebPApplyAlphaMultiply)(
+    uint8_t* rgba, int alpha_first, int w, int h, int stride);
+
+// Same, buf specifically for RGBA4444 format
+extern void (*WebPApplyAlphaMultiply4444)(
+    uint8_t* rgba4444, int w, int h, int stride);
+
+// To be called first before using the above.
+void WebPInitPremultiply(void);
+
+void WebPInitPremultiplySSE2(void);   // should not be called directly.
+void WebPInitPremultiplyNEON(void);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_DSP_DSP_H_ */
diff --git a/drivers/webp/dsp/enc.c b/drivers/webp/dsp/enc.c
new file mode 100644
index 0000000000..fcc6ec8ea2
--- /dev/null
+++ b/drivers/webp/dsp/enc.c
@@ -0,0 +1,753 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// 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"
+
+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.
+
+const int VP8DspScan[16 + 4 + 4] = {
+  // Luma
+  0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
+  0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
+  0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
+  0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS,
+
+  0 + 0 * BPS,   4 + 0 * BPS, 0 + 4 * BPS,  4 + 4 * BPS,    // U
+  8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
+};
+
+static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
+                             int start_block, int end_block,
+                             VP8Histogram* const histo) {
+  int j;
+  for (j = start_block; j < end_block; ++j) {
+    int k;
+    int16_t out[16];
+
+    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
+
+    // Convert coefficients to bin.
+    for (k = 0; k < 16; ++k) {
+      const int v = abs(out[k]) >> 3;  // TODO(skal): add rounding?
+      const int clipped_value = clip_max(v, MAX_COEFF_THRESH);
+      histo->distribution[clipped_value]++;
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+// run-time tables (~4k)
+
+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 InitTables(void) {
+  if (!tables_ok) {
+    int i;
+    for (i = -255; i <= 255 + 255; ++i) {
+      clip1[255 + i] = clip_8b(i);
+    }
+    tables_ok = 1;
+  }
+}
+
+
+//------------------------------------------------------------------------------
+// Transforms (Paragraph 14.4)
+
+#define STORE(x, y, v) \
+  dst[(x) + (y) * BPS] = clip_8b(ref[(x) + (y) * BPS] + ((v) >> 3))
+
+static const int kC1 = 20091 + (1 << 16);
+static const int kC2 = 35468;
+#define MUL(a, b) (((a) * (b)) >> 16)
+
+static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
+                                      uint8_t* dst) {
+  int C[4 * 4], *tmp;
+  int i;
+  tmp = C;
+  for (i = 0; i < 4; ++i) {    // vertical pass
+    const int a = in[0] + in[8];
+    const int b = in[0] - in[8];
+    const int c = MUL(in[4], kC2) - MUL(in[12], kC1);
+    const int d = MUL(in[4], kC1) + MUL(in[12], kC2);
+    tmp[0] = a + d;
+    tmp[1] = b + c;
+    tmp[2] = b - c;
+    tmp[3] = a - d;
+    tmp += 4;
+    in++;
+  }
+
+  tmp = C;
+  for (i = 0; i < 4; ++i) {    // horizontal pass
+    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);
+    STORE(0, i, a + d);
+    STORE(1, i, b + c);
+    STORE(2, i, b - c);
+    STORE(3, i, a - d);
+    tmp++;
+  }
+}
+
+static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
+                       int do_two) {
+  ITransformOne(ref, in, dst);
+  if (do_two) {
+    ITransformOne(ref + 4, in + 16, dst + 4);
+  }
+}
+
+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];   // 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);         // 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]);  // 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;            // 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 FTransformWHT(const int16_t* in, int16_t* out) {
+  // 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]);  // 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]);  // 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;    // 16b
+    const int b1 = a3 + a2;
+    const int b2 = a3 - a2;
+    const int b3 = a0 - a1;
+    out[ 0 + i] = b0 >> 1;     // 15b
+    out[ 4 + i] = b1 >> 1;
+    out[ 8 + i] = b2 >> 1;
+    out[12 + i] = b3 >> 1;
+  }
+}
+
+#undef MUL
+#undef STORE
+
+//------------------------------------------------------------------------------
+// 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) {
+    memset(dst + j * BPS, value, size);
+  }
+}
+
+static WEBP_INLINE void VerticalPred(uint8_t* dst,
+                                     const uint8_t* top, int size) {
+  int j;
+  if (top) {
+    for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
+  } else {
+    Fill(dst, 127, size);
+  }
+}
+
+static WEBP_INLINE void HorizontalPred(uint8_t* dst,
+                                       const uint8_t* left, int size) {
+  if (left) {
+    int j;
+    for (j = 0; j < size; ++j) {
+      memset(dst + j * BPS, left[j], size);
+    }
+  } else {
+    Fill(dst, 129, size);
+  }
+}
+
+static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
+                                   const uint8_t* top, int size) {
+  int y;
+  if (left) {
+    if (top) {
+      const uint8_t* const clip = clip1 + 255 - left[-1];
+      for (y = 0; y < size; ++y) {
+        const uint8_t* const clip_table = clip + left[y];
+        int x;
+        for (x = 0; x < size; ++x) {
+          dst[x] = clip_table[top[x]];
+        }
+        dst += BPS;
+      }
+    } 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) {
+      VerticalPred(dst, top, size);
+    } else {
+      Fill(dst, 129, size);
+    }
+  }
+}
+
+static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
+                               const uint8_t* top,
+                               int size, int round, int shift) {
+  int DC = 0;
+  int j;
+  if (top) {
+    for (j = 0; j < size; ++j) DC += top[j];
+    if (left) {   // 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
+    for (j = 0; j < size; ++j) DC += left[j];
+    DC += DC;
+    DC = (DC + round) >> shift;
+  } else {   // no top, no left, nothing.
+    DC = 0x80;
+  }
+  Fill(dst, DC, size);
+}
+
+//------------------------------------------------------------------------------
+// Chroma 8x8 prediction (paragraph 12.2)
+
+static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
+                             const uint8_t* top) {
+  // U block
+  DCMode(C8DC8 + dst, left, top, 8, 8, 4);
+  VerticalPred(C8VE8 + dst, top, 8);
+  HorizontalPred(C8HE8 + dst, left, 8);
+  TrueMotion(C8TM8 + dst, left, top, 8);
+  // V block
+  dst += 8;
+  if (top) top += 8;
+  if (left) left += 16;
+  DCMode(C8DC8 + dst, left, top, 8, 8, 4);
+  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) {
+  DCMode(I16DC16 + dst, left, top, 16, 16, 5);
+  VerticalPred(I16VE16 + dst, top, 16);
+  HorizontalPred(I16HE16 + dst, left, 16);
+  TrueMotion(I16TM16 + dst, left, top, 16);
+}
+
+//------------------------------------------------------------------------------
+// luma 4x4 prediction
+
+#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, const uint8_t* top) {    // vertical
+  const uint8_t vals[4] = {
+    AVG3(top[-1], top[0], top[1]),
+    AVG3(top[ 0], top[1], top[2]),
+    AVG3(top[ 1], top[2], top[3]),
+    AVG3(top[ 2], top[3], top[4])
+  };
+  int i;
+  for (i = 0; i < 4; ++i) {
+    memcpy(dst + i * BPS, vals, 4);
+  }
+}
+
+static 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 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 void RD4(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];
+  const int D = top[3];
+  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);
+}
+
+static void LD4(uint8_t* dst, const uint8_t* top) {
+  const int A = top[0];
+  const int B = top[1];
+  const int C = top[2];
+  const int D = top[3];
+  const int E = top[4];
+  const int F = top[5];
+  const int G = top[6];
+  const int H = top[7];
+  DST(0, 0)                                     = AVG3(A, B, C);
+  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);
+}
+
+static void VR4(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 A = top[0];
+  const int B = top[1];
+  const int C = top[2];
+  const int D = top[3];
+  DST(0, 0) = DST(1, 2) = AVG2(X, A);
+  DST(1, 0) = DST(2, 2) = AVG2(A, B);
+  DST(2, 0) = DST(3, 2) = AVG2(B, C);
+  DST(3, 0)             = AVG2(C, D);
+
+  DST(0, 3) =             AVG3(K, J, I);
+  DST(0, 2) =             AVG3(J, I, X);
+  DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
+  DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
+  DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
+  DST(3, 1) =             AVG3(B, C, D);
+}
+
+static void VL4(uint8_t* dst, const uint8_t* top) {
+  const int A = top[0];
+  const int B = top[1];
+  const int C = top[2];
+  const int D = top[3];
+  const int E = top[4];
+  const int F = top[5];
+  const int G = top[6];
+  const int H = top[7];
+  DST(0, 0) =             AVG2(A, B);
+  DST(1, 0) = DST(0, 2) = AVG2(B, C);
+  DST(2, 0) = DST(1, 2) = AVG2(C, D);
+  DST(3, 0) = DST(2, 2) = AVG2(D, E);
+
+  DST(0, 1) =             AVG3(A, B, C);
+  DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
+  DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
+  DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
+              DST(3, 2) = AVG3(E, F, G);
+              DST(3, 3) = AVG3(F, G, H);
+}
+
+static 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 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 void TM4(uint8_t* dst, const uint8_t* top) {
+  int x, y;
+  const uint8_t* const clip = clip1 + 255 - top[-1];
+  for (y = 0; y < 4; ++y) {
+    const uint8_t* const clip_table = clip + top[-2 - y];
+    for (x = 0; x < 4; ++x) {
+      dst[x] = clip_table[top[x]];
+    }
+    dst += BPS;
+  }
+}
+
+#undef DST
+#undef AVG3
+#undef AVG2
+
+// 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);
+}
+
+//------------------------------------------------------------------------------
+// Metric
+
+static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b,
+                              int w, int h) {
+  int count = 0;
+  int y, x;
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      const int diff = (int)a[x] - b[x];
+      count += diff * diff;
+    }
+    a += BPS;
+    b += BPS;
+  }
+  return count;
+}
+
+static int SSE16x16(const uint8_t* a, const uint8_t* b) {
+  return GetSSE(a, b, 16, 16);
+}
+static int SSE16x8(const uint8_t* a, const uint8_t* b) {
+  return GetSSE(a, b, 16, 8);
+}
+static int SSE8x8(const uint8_t* a, const uint8_t* b) {
+  return GetSSE(a, b, 8, 8);
+}
+static int SSE4x4(const uint8_t* a, const uint8_t* b) {
+  return GetSSE(a, b, 4, 4);
+}
+
+//------------------------------------------------------------------------------
+// Texture distortion
+//
+// We try to match the spectral content (weighted) between source and
+// reconstructed samples.
+
+// Hadamard transform
+// Returns the weighted sum of the absolute value of transformed coefficients.
+static int TTransform(const uint8_t* in, const uint16_t* w) {
+  int sum = 0;
+  int tmp[16];
+  int i;
+  // horizontal pass
+  for (i = 0; i < 4; ++i, in += BPS) {
+    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 b0 = a0 + a1;
+    const int b1 = a3 + a2;
+    const int b2 = a3 - a2;
+    const int b3 = a0 - a1;
+
+    sum += w[ 0] * abs(b0);
+    sum += w[ 4] * abs(b1);
+    sum += w[ 8] * abs(b2);
+    sum += w[12] * abs(b3);
+  }
+  return sum;
+}
+
+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) >> 5;
+}
+
+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 += Disto4x4(a + x + y, b + x + y, w);
+    }
+  }
+  return D;
+}
+
+//------------------------------------------------------------------------------
+// Quantization
+//
+
+static const uint8_t kZigzag[16] = {
+  0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
+};
+
+// Simple quantization
+static int QuantizeBlock(int16_t in[16], int16_t out[16],
+                         int n, const VP8Matrix* const mtx) {
+  int last = -1;
+  for (; n < 16; ++n) {
+    const int j = kZigzag[n];
+    const int sign = (in[j] < 0);
+    const int coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
+    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 (out[n] > MAX_LEVEL) out[n] = MAX_LEVEL;
+      if (sign) out[n] = -out[n];
+      in[j] = out[n] * Q;
+      if (out[n]) last = n;
+    } else {
+      out[n] = 0;
+      in[j] = 0;
+    }
+  }
+  return (last >= 0);
+}
+
+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);
+    const int 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 (out[n] > MAX_LEVEL) out[n] = MAX_LEVEL;
+      if (sign) out[n] = -out[n];
+      in[j] = out[n] * Q;
+      if (out[n]) last = n;
+    } else {
+      out[n] = 0;
+      in[j] = 0;
+    }
+  }
+  return (last >= 0);
+}
+
+//------------------------------------------------------------------------------
+// Block copy
+
+static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int size) {
+  int y;
+  for (y = 0; y < size; ++y) {
+    memcpy(dst, src, size);
+    src += BPS;
+    dst += BPS;
+  }
+}
+
+static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); }
+
+//------------------------------------------------------------------------------
+// Initialization
+
+// Speed-critical function pointers. We have to initialize them to the default
+// implementations within VP8EncDspInit().
+VP8CHisto VP8CollectHistogram;
+VP8Idct VP8ITransform;
+VP8Fdct VP8FTransform;
+VP8WHT VP8ITransformWHT;
+VP8WHT VP8FTransformWHT;
+VP8Intra4Preds VP8EncPredLuma4;
+VP8IntraPreds VP8EncPredLuma16;
+VP8IntraPreds VP8EncPredChroma8;
+VP8Metric VP8SSE16x16;
+VP8Metric VP8SSE8x8;
+VP8Metric VP8SSE16x8;
+VP8Metric VP8SSE4x4;
+VP8WMetric VP8TDisto4x4;
+VP8WMetric VP8TDisto16x16;
+VP8QuantizeBlock VP8EncQuantizeBlock;
+VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
+VP8BlockCopy VP8Copy4x4;
+
+extern void VP8EncDspInitSSE2(void);
+extern void VP8EncDspInitNEON(void);
+
+void VP8EncDspInit(void) {
+  InitTables();
+
+  // default C implementations
+  VP8CollectHistogram = CollectHistogram;
+  VP8ITransform = ITransform;
+  VP8FTransform = FTransform;
+  VP8ITransformWHT = ITransformWHT;
+  VP8FTransformWHT = FTransformWHT;
+  VP8EncPredLuma4 = Intra4Preds;
+  VP8EncPredLuma16 = Intra16Preds;
+  VP8EncPredChroma8 = IntraChromaPreds;
+  VP8SSE16x16 = SSE16x16;
+  VP8SSE8x8 = SSE8x8;
+  VP8SSE16x8 = SSE16x8;
+  VP8SSE4x4 = SSE4x4;
+  VP8TDisto4x4 = Disto4x4;
+  VP8TDisto16x16 = Disto16x16;
+  VP8EncQuantizeBlock = QuantizeBlock;
+  VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
+  VP8Copy4x4 = Copy4x4;
+
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
+  if (VP8GetCPUInfo) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      VP8EncDspInitSSE2();
+    }
+#elif defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      VP8EncDspInitNEON();
+    }
+#endif
+  }
+}
+
diff --git a/drivers/webp/dsp/enc_neon.c b/drivers/webp/dsp/enc_neon.c
new file mode 100644
index 0000000000..52cca18682
--- /dev/null
+++ b/drivers/webp/dsp/enc_neon.c
@@ -0,0 +1,632 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// ARM NEON version of speed-critical encoding functions.
+//
+// adapted from libvpx (http://www.webmproject.org/code/)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_NEON)
+
+#include "../enc/vp8enci.h"
+
+//------------------------------------------------------------------------------
+// Transforms (Paragraph 14.4)
+
+// Inverse transform.
+// This code is pretty much the same as TransformOneNEON in the decoder, except
+// for subtraction to *ref. See the comments there for algorithmic explanations.
+static void ITransformOne(const uint8_t* ref,
+                          const int16_t* in, uint8_t* dst) {
+  const int kBPS = BPS;
+  const int16_t kC1C2[] = { 20091, 17734, 0, 0 };  // kC1 / (kC2 >> 1) / 0 / 0
+
+  __asm__ volatile (
+    "vld1.16         {q1, q2}, [%[in]]           \n"
+    "vld1.16         {d0}, [%[kC1C2]]            \n"
+
+    // d2: in[0]
+    // d3: in[8]
+    // d4: in[4]
+    // d5: in[12]
+    "vswp            d3, d4                      \n"
+
+    // q8 = {in[4], in[12]} * kC1 * 2 >> 16
+    // q9 = {in[4], in[12]} * kC2 >> 16
+    "vqdmulh.s16     q8, q2, d0[0]               \n"
+    "vqdmulh.s16     q9, q2, d0[1]               \n"
+
+    // d22 = a = in[0] + in[8]
+    // d23 = b = in[0] - in[8]
+    "vqadd.s16       d22, d2, d3                 \n"
+    "vqsub.s16       d23, d2, d3                 \n"
+
+    //  q8 = in[4]/[12] * kC1 >> 16
+    "vshr.s16        q8, q8, #1                  \n"
+
+    // Add {in[4], in[12]} back after the multiplication.
+    "vqadd.s16       q8, q2, q8                  \n"
+
+    // d20 = c = in[4]*kC2 - in[12]*kC1
+    // d21 = d = in[4]*kC1 + in[12]*kC2
+    "vqsub.s16       d20, d18, d17               \n"
+    "vqadd.s16       d21, d19, d16               \n"
+
+    // d2 = tmp[0] = a + d
+    // d3 = tmp[1] = b + c
+    // d4 = tmp[2] = b - c
+    // d5 = tmp[3] = a - d
+    "vqadd.s16       d2, d22, d21                \n"
+    "vqadd.s16       d3, d23, d20                \n"
+    "vqsub.s16       d4, d23, d20                \n"
+    "vqsub.s16       d5, d22, d21                \n"
+
+    "vzip.16         q1, q2                      \n"
+    "vzip.16         q1, q2                      \n"
+
+    "vswp            d3, d4                      \n"
+
+    // q8 = {tmp[4], tmp[12]} * kC1 * 2 >> 16
+    // q9 = {tmp[4], tmp[12]} * kC2 >> 16
+    "vqdmulh.s16     q8, q2, d0[0]               \n"
+    "vqdmulh.s16     q9, q2, d0[1]               \n"
+
+    // d22 = a = tmp[0] + tmp[8]
+    // d23 = b = tmp[0] - tmp[8]
+    "vqadd.s16       d22, d2, d3                 \n"
+    "vqsub.s16       d23, d2, d3                 \n"
+
+    "vshr.s16        q8, q8, #1                  \n"
+    "vqadd.s16       q8, q2, q8                  \n"
+
+    // d20 = c = in[4]*kC2 - in[12]*kC1
+    // d21 = d = in[4]*kC1 + in[12]*kC2
+    "vqsub.s16       d20, d18, d17               \n"
+    "vqadd.s16       d21, d19, d16               \n"
+
+    // d2 = tmp[0] = a + d
+    // d3 = tmp[1] = b + c
+    // d4 = tmp[2] = b - c
+    // d5 = tmp[3] = a - d
+    "vqadd.s16       d2, d22, d21                \n"
+    "vqadd.s16       d3, d23, d20                \n"
+    "vqsub.s16       d4, d23, d20                \n"
+    "vqsub.s16       d5, d22, d21                \n"
+
+    "vld1.32         d6[0], [%[ref]], %[kBPS]    \n"
+    "vld1.32         d6[1], [%[ref]], %[kBPS]    \n"
+    "vld1.32         d7[0], [%[ref]], %[kBPS]    \n"
+    "vld1.32         d7[1], [%[ref]], %[kBPS]    \n"
+
+    "sub         %[ref], %[ref], %[kBPS], lsl #2 \n"
+
+    // (val) + 4 >> 3
+    "vrshr.s16       d2, d2, #3                  \n"
+    "vrshr.s16       d3, d3, #3                  \n"
+    "vrshr.s16       d4, d4, #3                  \n"
+    "vrshr.s16       d5, d5, #3                  \n"
+
+    "vzip.16         q1, q2                      \n"
+    "vzip.16         q1, q2                      \n"
+
+    // Must accumulate before saturating
+    "vmovl.u8        q8, d6                      \n"
+    "vmovl.u8        q9, d7                      \n"
+
+    "vqadd.s16       q1, q1, q8                  \n"
+    "vqadd.s16       q2, q2, q9                  \n"
+
+    "vqmovun.s16     d0, q1                      \n"
+    "vqmovun.s16     d1, q2                      \n"
+
+    "vst1.32         d0[0], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d0[1], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d1[0], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d1[1], [%[dst]]             \n"
+
+    : [in] "+r"(in), [dst] "+r"(dst)               // modified registers
+    : [kBPS] "r"(kBPS), [kC1C2] "r"(kC1C2), [ref] "r"(ref)  // constants
+    : "memory", "q0", "q1", "q2", "q8", "q9", "q10", "q11"  // clobbered
+  );
+}
+
+static void ITransform(const uint8_t* ref,
+                       const int16_t* in, uint8_t* dst, int do_two) {
+  ITransformOne(ref, in, dst);
+  if (do_two) {
+    ITransformOne(ref + 4, in + 16, dst + 4);
+  }
+}
+
+// Same code as dec_neon.c
+static void ITransformWHT(const int16_t* in, int16_t* out) {
+  const int kStep = 32;  // The store is only incrementing the pointer as if we
+                         // had stored a single byte.
+  __asm__ volatile (
+    // part 1
+    // load data into q0, q1
+    "vld1.16         {q0, q1}, [%[in]]           \n"
+
+    "vaddl.s16       q2, d0, d3                  \n" // a0 = in[0] + in[12]
+    "vaddl.s16       q3, d1, d2                  \n" // a1 = in[4] + in[8]
+    "vsubl.s16       q4, d1, d2                  \n" // a2 = in[4] - in[8]
+    "vsubl.s16       q5, d0, d3                  \n" // a3 = in[0] - in[12]
+
+    "vadd.s32        q0, q2, q3                  \n" // tmp[0] = a0 + a1
+    "vsub.s32        q2, q2, q3                  \n" // tmp[8] = a0 - a1
+    "vadd.s32        q1, q5, q4                  \n" // tmp[4] = a3 + a2
+    "vsub.s32        q3, q5, q4                  \n" // tmp[12] = a3 - a2
+
+    // Transpose
+    // q0 = tmp[0, 4, 8, 12], q1 = tmp[2, 6, 10, 14]
+    // q2 = tmp[1, 5, 9, 13], q3 = tmp[3, 7, 11, 15]
+    "vswp            d1, d4                      \n" // vtrn.64 q0, q2
+    "vswp            d3, d6                      \n" // vtrn.64 q1, q3
+    "vtrn.32         q0, q1                      \n"
+    "vtrn.32         q2, q3                      \n"
+
+    "vmov.s32        q4, #3                      \n" // dc = 3
+    "vadd.s32        q0, q0, q4                  \n" // dc = tmp[0] + 3
+    "vadd.s32        q6, q0, q3                  \n" // a0 = dc + tmp[3]
+    "vadd.s32        q7, q1, q2                  \n" // a1 = tmp[1] + tmp[2]
+    "vsub.s32        q8, q1, q2                  \n" // a2 = tmp[1] - tmp[2]
+    "vsub.s32        q9, q0, q3                  \n" // a3 = dc - tmp[3]
+
+    "vadd.s32        q0, q6, q7                  \n"
+    "vshrn.s32       d0, q0, #3                  \n" // (a0 + a1) >> 3
+    "vadd.s32        q1, q9, q8                  \n"
+    "vshrn.s32       d1, q1, #3                  \n" // (a3 + a2) >> 3
+    "vsub.s32        q2, q6, q7                  \n"
+    "vshrn.s32       d2, q2, #3                  \n" // (a0 - a1) >> 3
+    "vsub.s32        q3, q9, q8                  \n"
+    "vshrn.s32       d3, q3, #3                  \n" // (a3 - a2) >> 3
+
+    // set the results to output
+    "vst1.16         d0[0], [%[out]], %[kStep]      \n"
+    "vst1.16         d1[0], [%[out]], %[kStep]      \n"
+    "vst1.16         d2[0], [%[out]], %[kStep]      \n"
+    "vst1.16         d3[0], [%[out]], %[kStep]      \n"
+    "vst1.16         d0[1], [%[out]], %[kStep]      \n"
+    "vst1.16         d1[1], [%[out]], %[kStep]      \n"
+    "vst1.16         d2[1], [%[out]], %[kStep]      \n"
+    "vst1.16         d3[1], [%[out]], %[kStep]      \n"
+    "vst1.16         d0[2], [%[out]], %[kStep]      \n"
+    "vst1.16         d1[2], [%[out]], %[kStep]      \n"
+    "vst1.16         d2[2], [%[out]], %[kStep]      \n"
+    "vst1.16         d3[2], [%[out]], %[kStep]      \n"
+    "vst1.16         d0[3], [%[out]], %[kStep]      \n"
+    "vst1.16         d1[3], [%[out]], %[kStep]      \n"
+    "vst1.16         d2[3], [%[out]], %[kStep]      \n"
+    "vst1.16         d3[3], [%[out]], %[kStep]      \n"
+
+    : [out] "+r"(out)  // modified registers
+    : [in] "r"(in), [kStep] "r"(kStep)  // constants
+    : "memory", "q0", "q1", "q2", "q3", "q4",
+      "q5", "q6", "q7", "q8", "q9" // clobbered
+  );
+}
+
+// Forward transform.
+
+// adapted from vp8/encoder/arm/neon/shortfdct_neon.asm
+static const int16_t kCoeff16[] = {
+  5352,  5352,  5352, 5352, 2217,  2217,  2217, 2217
+};
+static const int32_t kCoeff32[] = {
+   1812,  1812,  1812,  1812,
+    937,   937,   937,   937,
+  12000, 12000, 12000, 12000,
+  51000, 51000, 51000, 51000
+};
+
+static void FTransform(const uint8_t* src, const uint8_t* ref,
+                       int16_t* out) {
+  const int kBPS = BPS;
+  const uint8_t* src_ptr = src;
+  const uint8_t* ref_ptr = ref;
+  const int16_t* coeff16 = kCoeff16;
+  const int32_t* coeff32 = kCoeff32;
+
+  __asm__ volatile (
+    // load src into q4, q5 in high half
+    "vld1.8 {d8},  [%[src_ptr]], %[kBPS]      \n"
+    "vld1.8 {d10}, [%[src_ptr]], %[kBPS]      \n"
+    "vld1.8 {d9},  [%[src_ptr]], %[kBPS]      \n"
+    "vld1.8 {d11}, [%[src_ptr]]               \n"
+
+    // load ref into q6, q7 in high half
+    "vld1.8 {d12}, [%[ref_ptr]], %[kBPS]      \n"
+    "vld1.8 {d14}, [%[ref_ptr]], %[kBPS]      \n"
+    "vld1.8 {d13}, [%[ref_ptr]], %[kBPS]      \n"
+    "vld1.8 {d15}, [%[ref_ptr]]               \n"
+
+    // Pack the high values in to q4 and q6
+    "vtrn.32     q4, q5                       \n"
+    "vtrn.32     q6, q7                       \n"
+
+    // d[0-3] = src - ref
+    "vsubl.u8    q0, d8, d12                  \n"
+    "vsubl.u8    q1, d9, d13                  \n"
+
+    // load coeff16 into q8(d16=5352, d17=2217)
+    "vld1.16     {q8}, [%[coeff16]]           \n"
+
+    // load coeff32 high half into q9 = 1812, q10 = 937
+    "vld1.32     {q9, q10}, [%[coeff32]]!     \n"
+
+    // load coeff32 low half into q11=12000, q12=51000
+    "vld1.32     {q11,q12}, [%[coeff32]]      \n"
+
+    // part 1
+    // Transpose. Register dN is the same as dN in C
+    "vtrn.32         d0, d2                   \n"
+    "vtrn.32         d1, d3                   \n"
+    "vtrn.16         d0, d1                   \n"
+    "vtrn.16         d2, d3                   \n"
+
+    "vadd.s16        d4, d0, d3               \n" // a0 = d0 + d3
+    "vadd.s16        d5, d1, d2               \n" // a1 = d1 + d2
+    "vsub.s16        d6, d1, d2               \n" // a2 = d1 - d2
+    "vsub.s16        d7, d0, d3               \n" // a3 = d0 - d3
+
+    "vadd.s16        d0, d4, d5               \n" // a0 + a1
+    "vshl.s16        d0, d0, #3               \n" // temp[0+i*4] = (a0+a1) << 3
+    "vsub.s16        d2, d4, d5               \n" // a0 - a1
+    "vshl.s16        d2, d2, #3               \n" // (temp[2+i*4] = (a0-a1) << 3
+
+    "vmlal.s16       q9, d7, d16              \n" // a3*5352 + 1812
+    "vmlal.s16       q10, d7, d17             \n" // a3*2217 + 937
+    "vmlal.s16       q9, d6, d17              \n" // a2*2217 + a3*5352 + 1812
+    "vmlsl.s16       q10, d6, d16             \n" // a3*2217 + 937 - a2*5352
+
+    // temp[1+i*4] = (d2*2217 + d3*5352 + 1812) >> 9
+    // temp[3+i*4] = (d3*2217 + 937 - d2*5352) >> 9
+    "vshrn.s32       d1, q9, #9               \n"
+    "vshrn.s32       d3, q10, #9              \n"
+
+    // part 2
+    // transpose d0=ip[0], d1=ip[4], d2=ip[8], d3=ip[12]
+    "vtrn.32         d0, d2                   \n"
+    "vtrn.32         d1, d3                   \n"
+    "vtrn.16         d0, d1                   \n"
+    "vtrn.16         d2, d3                   \n"
+
+    "vmov.s16        d26, #7                  \n"
+
+    "vadd.s16        d4, d0, d3               \n" // a1 = ip[0] + ip[12]
+    "vadd.s16        d5, d1, d2               \n" // b1 = ip[4] + ip[8]
+    "vsub.s16        d6, d1, d2               \n" // c1 = ip[4] - ip[8]
+    "vadd.s16        d4, d4, d26              \n" // a1 + 7
+    "vsub.s16        d7, d0, d3               \n" // d1 = ip[0] - ip[12]
+
+    "vadd.s16        d0, d4, d5               \n" // op[0] = a1 + b1 + 7
+    "vsub.s16        d2, d4, d5               \n" // op[8] = a1 - b1 + 7
+
+    "vmlal.s16       q11, d7, d16             \n" // d1*5352 + 12000
+    "vmlal.s16       q12, d7, d17             \n" // d1*2217 + 51000
+
+    "vceq.s16        d4, d7, #0               \n"
+
+    "vshr.s16        d0, d0, #4               \n"
+    "vshr.s16        d2, d2, #4               \n"
+
+    "vmlal.s16       q11, d6, d17             \n" // c1*2217 + d1*5352 + 12000
+    "vmlsl.s16       q12, d6, d16             \n" // d1*2217 - c1*5352 + 51000
+
+    "vmvn            d4, d4                   \n" // !(d1 == 0)
+    // op[4] = (c1*2217 + d1*5352 + 12000)>>16
+    "vshrn.s32       d1, q11, #16             \n"
+    // op[4] += (d1!=0)
+    "vsub.s16        d1, d1, d4               \n"
+    // op[12]= (d1*2217 - c1*5352 + 51000)>>16
+    "vshrn.s32       d3, q12, #16             \n"
+
+    // set result to out array
+    "vst1.16         {q0, q1}, [%[out]]   \n"
+    : [src_ptr] "+r"(src_ptr), [ref_ptr] "+r"(ref_ptr),
+      [coeff32] "+r"(coeff32)          // modified registers
+    : [kBPS] "r"(kBPS), [coeff16] "r"(coeff16),
+      [out] "r"(out)                   // constants
+    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+      "q10", "q11", "q12", "q13"       // clobbered
+  );
+}
+
+static void FTransformWHT(const int16_t* in, int16_t* out) {
+  const int kStep = 32;
+  __asm__ volatile (
+    // d0 = in[0 * 16] , d1 = in[1 * 16]
+    // d2 = in[2 * 16] , d3 = in[3 * 16]
+    "vld1.16         d0[0], [%[in]], %[kStep]   \n"
+    "vld1.16         d1[0], [%[in]], %[kStep]   \n"
+    "vld1.16         d2[0], [%[in]], %[kStep]   \n"
+    "vld1.16         d3[0], [%[in]], %[kStep]   \n"
+    "vld1.16         d0[1], [%[in]], %[kStep]   \n"
+    "vld1.16         d1[1], [%[in]], %[kStep]   \n"
+    "vld1.16         d2[1], [%[in]], %[kStep]   \n"
+    "vld1.16         d3[1], [%[in]], %[kStep]   \n"
+    "vld1.16         d0[2], [%[in]], %[kStep]   \n"
+    "vld1.16         d1[2], [%[in]], %[kStep]   \n"
+    "vld1.16         d2[2], [%[in]], %[kStep]   \n"
+    "vld1.16         d3[2], [%[in]], %[kStep]   \n"
+    "vld1.16         d0[3], [%[in]], %[kStep]   \n"
+    "vld1.16         d1[3], [%[in]], %[kStep]   \n"
+    "vld1.16         d2[3], [%[in]], %[kStep]   \n"
+    "vld1.16         d3[3], [%[in]], %[kStep]   \n"
+
+    "vaddl.s16       q2, d0, d2                 \n" // a0=(in[0*16]+in[2*16])
+    "vaddl.s16       q3, d1, d3                 \n" // a1=(in[1*16]+in[3*16])
+    "vsubl.s16       q4, d1, d3                 \n" // a2=(in[1*16]-in[3*16])
+    "vsubl.s16       q5, d0, d2                 \n" // a3=(in[0*16]-in[2*16])
+
+    "vqadd.s32       q6, q2, q3                 \n" // a0 + a1
+    "vqadd.s32       q7, q5, q4                 \n" // a3 + a2
+    "vqsub.s32       q8, q5, q4                 \n" // a3 - a2
+    "vqsub.s32       q9, q2, q3                 \n" // a0 - a1
+
+    // Transpose
+    // q6 = tmp[0, 1,  2,  3] ; q7 = tmp[ 4,  5,  6,  7]
+    // q8 = tmp[8, 9, 10, 11] ; q9 = tmp[12, 13, 14, 15]
+    "vswp            d13, d16                   \n" // vtrn.64 q0, q2
+    "vswp            d15, d18                   \n" // vtrn.64 q1, q3
+    "vtrn.32         q6, q7                     \n"
+    "vtrn.32         q8, q9                     \n"
+
+    "vqadd.s32       q0, q6, q8                 \n" // a0 = tmp[0] + tmp[8]
+    "vqadd.s32       q1, q7, q9                 \n" // a1 = tmp[4] + tmp[12]
+    "vqsub.s32       q2, q7, q9                 \n" // a2 = tmp[4] - tmp[12]
+    "vqsub.s32       q3, q6, q8                 \n" // a3 = tmp[0] - tmp[8]
+
+    "vqadd.s32       q4, q0, q1                 \n" // b0 = a0 + a1
+    "vqadd.s32       q5, q3, q2                 \n" // b1 = a3 + a2
+    "vqsub.s32       q6, q3, q2                 \n" // b2 = a3 - a2
+    "vqsub.s32       q7, q0, q1                 \n" // b3 = a0 - a1
+
+    "vshrn.s32       d18, q4, #1                \n" // b0 >> 1
+    "vshrn.s32       d19, q5, #1                \n" // b1 >> 1
+    "vshrn.s32       d20, q6, #1                \n" // b2 >> 1
+    "vshrn.s32       d21, q7, #1                \n" // b3 >> 1
+
+    "vst1.16         {q9, q10}, [%[out]]        \n"
+
+    : [in] "+r"(in)
+    : [kStep] "r"(kStep), [out] "r"(out)
+    : "memory", "q0", "q1", "q2", "q3", "q4", "q5",
+      "q6", "q7", "q8", "q9", "q10"       // clobbered
+  ) ;
+}
+
+//------------------------------------------------------------------------------
+// Texture distortion
+//
+// We try to match the spectral content (weighted) between source and
+// reconstructed samples.
+
+// Hadamard transform
+// Returns the weighted sum of the absolute value of transformed coefficients.
+// This uses a TTransform helper function in C
+static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
+                    const uint16_t* const w) {
+  const int kBPS = BPS;
+  const uint8_t* A = a;
+  const uint8_t* B = b;
+  const uint16_t* W = w;
+  int sum;
+  __asm__ volatile (
+    "vld1.32         d0[0], [%[a]], %[kBPS]   \n"
+    "vld1.32         d0[1], [%[a]], %[kBPS]   \n"
+    "vld1.32         d2[0], [%[a]], %[kBPS]   \n"
+    "vld1.32         d2[1], [%[a]]            \n"
+
+    "vld1.32         d1[0], [%[b]], %[kBPS]   \n"
+    "vld1.32         d1[1], [%[b]], %[kBPS]   \n"
+    "vld1.32         d3[0], [%[b]], %[kBPS]   \n"
+    "vld1.32         d3[1], [%[b]]            \n"
+
+    // a d0/d2, b d1/d3
+    // d0/d1: 01 01 01 01
+    // d2/d3: 23 23 23 23
+    // But: it goes 01 45 23 67
+    // Notice the middle values are transposed
+    "vtrn.16         q0, q1                   \n"
+
+    // {a0, a1} = {in[0] + in[2], in[1] + in[3]}
+    "vaddl.u8        q2, d0, d2               \n"
+    "vaddl.u8        q10, d1, d3              \n"
+    // {a3, a2} = {in[0] - in[2], in[1] - in[3]}
+    "vsubl.u8        q3, d0, d2               \n"
+    "vsubl.u8        q11, d1, d3              \n"
+
+    // tmp[0] = a0 + a1
+    "vpaddl.s16      q0, q2                   \n"
+    "vpaddl.s16      q8, q10                  \n"
+
+    // tmp[1] = a3 + a2
+    "vpaddl.s16      q1, q3                   \n"
+    "vpaddl.s16      q9, q11                  \n"
+
+    // No pair subtract
+    // q2 = {a0, a3}
+    // q3 = {a1, a2}
+    "vtrn.16         q2, q3                   \n"
+    "vtrn.16         q10, q11                 \n"
+
+    // {tmp[3], tmp[2]} = {a0 - a1, a3 - a2}
+    "vsubl.s16       q12, d4, d6              \n"
+    "vsubl.s16       q13, d5, d7              \n"
+    "vsubl.s16       q14, d20, d22            \n"
+    "vsubl.s16       q15, d21, d23            \n"
+
+    // separate tmp[3] and tmp[2]
+    // q12 = tmp[3]
+    // q13 = tmp[2]
+    "vtrn.32         q12, q13                 \n"
+    "vtrn.32         q14, q15                 \n"
+
+    // Transpose tmp for a
+    "vswp            d1, d26                  \n" // vtrn.64
+    "vswp            d3, d24                  \n" // vtrn.64
+    "vtrn.32         q0, q1                   \n"
+    "vtrn.32         q13, q12                 \n"
+
+    // Transpose tmp for b
+    "vswp            d17, d30                 \n" // vtrn.64
+    "vswp            d19, d28                 \n" // vtrn.64
+    "vtrn.32         q8, q9                   \n"
+    "vtrn.32         q15, q14                 \n"
+
+    // The first Q register is a, the second b.
+    // q0/8 tmp[0-3]
+    // q13/15 tmp[4-7]
+    // q1/9 tmp[8-11]
+    // q12/14 tmp[12-15]
+
+    // These are still in 01 45 23 67 order. We fix it easily in the addition
+    // case but the subtraction propagates them.
+    "vswp            d3, d27                  \n"
+    "vswp            d19, d31                 \n"
+
+    // a0 = tmp[0] + tmp[8]
+    "vadd.s32        q2, q0, q1               \n"
+    "vadd.s32        q3, q8, q9               \n"
+
+    // a1 = tmp[4] + tmp[12]
+    "vadd.s32        q10, q13, q12            \n"
+    "vadd.s32        q11, q15, q14            \n"
+
+    // a2 = tmp[4] - tmp[12]
+    "vsub.s32        q13, q13, q12            \n"
+    "vsub.s32        q15, q15, q14            \n"
+
+    // a3 = tmp[0] - tmp[8]
+    "vsub.s32        q0, q0, q1               \n"
+    "vsub.s32        q8, q8, q9               \n"
+
+    // b0 = a0 + a1
+    "vadd.s32        q1, q2, q10              \n"
+    "vadd.s32        q9, q3, q11              \n"
+
+    // b1 = a3 + a2
+    "vadd.s32        q12, q0, q13             \n"
+    "vadd.s32        q14, q8, q15             \n"
+
+    // b2 = a3 - a2
+    "vsub.s32        q0, q0, q13              \n"
+    "vsub.s32        q8, q8, q15              \n"
+
+    // b3 = a0 - a1
+    "vsub.s32        q2, q2, q10              \n"
+    "vsub.s32        q3, q3, q11              \n"
+
+    "vld1.64         {q10, q11}, [%[w]]       \n"
+
+    // abs(b0)
+    "vabs.s32        q1, q1                   \n"
+    "vabs.s32        q9, q9                   \n"
+    // abs(b1)
+    "vabs.s32        q12, q12                 \n"
+    "vabs.s32        q14, q14                 \n"
+    // abs(b2)
+    "vabs.s32        q0, q0                   \n"
+    "vabs.s32        q8, q8                   \n"
+    // abs(b3)
+    "vabs.s32        q2, q2                   \n"
+    "vabs.s32        q3, q3                   \n"
+
+    // expand w before using.
+    "vmovl.u16       q13, d20                 \n"
+    "vmovl.u16       q15, d21                 \n"
+
+    // w[0] * abs(b0)
+    "vmul.u32        q1, q1, q13              \n"
+    "vmul.u32        q9, q9, q13              \n"
+
+    // w[4] * abs(b1)
+    "vmla.u32        q1, q12, q15             \n"
+    "vmla.u32        q9, q14, q15             \n"
+
+    // expand w before using.
+    "vmovl.u16       q13, d22                 \n"
+    "vmovl.u16       q15, d23                 \n"
+
+    // w[8] * abs(b1)
+    "vmla.u32        q1, q0, q13              \n"
+    "vmla.u32        q9, q8, q13              \n"
+
+    // w[12] * abs(b1)
+    "vmla.u32        q1, q2, q15              \n"
+    "vmla.u32        q9, q3, q15              \n"
+
+    // Sum the arrays
+    "vpaddl.u32      q1, q1                   \n"
+    "vpaddl.u32      q9, q9                   \n"
+    "vadd.u64        d2, d3                   \n"
+    "vadd.u64        d18, d19                 \n"
+
+    // Hadamard transform needs 4 bits of extra precision (2 bits in each
+    // direction) for dynamic raw. Weights w[] are 16bits at max, so the maximum
+    // precision for coeff is 8bit of input + 4bits of Hadamard transform +
+    // 16bits for w[] + 2 bits of abs() summation.
+    //
+    // This uses a maximum of 31 bits (signed). Discarding the top 32 bits is
+    // A-OK.
+
+    // sum2 - sum1
+    "vsub.u32        d0, d2, d18              \n"
+    // abs(sum2 - sum1)
+    "vabs.s32        d0, d0                   \n"
+    // abs(sum2 - sum1) >> 5
+    "vshr.u32        d0, #5                   \n"
+
+    // It would be better to move the value straight into r0 but I'm not
+    // entirely sure how this works with inline assembly.
+    "vmov.32         %[sum], d0[0]            \n"
+
+    : [sum] "=r"(sum), [a] "+r"(A), [b] "+r"(B), [w] "+r"(W)
+    : [kBPS] "r"(kBPS)
+    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+      "q10", "q11", "q12", "q13", "q14", "q15"  // clobbered
+  ) ;
+
+  return sum;
+}
+
+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 += Disto4x4(a + x + y, b + x + y, w);
+    }
+  }
+  return D;
+}
+
+#endif   // WEBP_USE_NEON
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8EncDspInitNEON(void);
+
+void VP8EncDspInitNEON(void) {
+#if defined(WEBP_USE_NEON)
+  VP8ITransform = ITransform;
+  VP8FTransform = FTransform;
+
+  VP8ITransformWHT = ITransformWHT;
+  VP8FTransformWHT = FTransformWHT;
+
+  VP8TDisto4x4 = Disto4x4;
+  VP8TDisto16x16 = Disto16x16;
+#endif   // WEBP_USE_NEON
+}
+
diff --git a/drivers/webp/dsp/enc_sse2.c b/drivers/webp/dsp/enc_sse2.c
new file mode 100644
index 0000000000..540a3cb2db
--- /dev/null
+++ b/drivers/webp/dsp/enc_sse2.c
@@ -0,0 +1,957 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// SSE2 version of speed-critical encoding functions.
+//
+// Author: Christian Duvivier (cduvivier@google.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_SSE2)
+#include <stdlib.h>  // for abs()
+#include <emmintrin.h>
+
+#include "../enc/vp8enci.h"
+
+//------------------------------------------------------------------------------
+// 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;
+  printf("%s\t: ", name);
+  if (size == 8) {
+    for (n = 0; n < 16; ++n) printf("%.2x ", tmp.i8[n]);
+  } else if (size == 16) {
+    for (n = 0; n < 8; ++n) printf("%.4x ", tmp.i16[n]);
+  } else if (size == 32) {
+    for (n = 0; n < 4; ++n) printf("%.8x ", tmp.i32[n]);
+  } else {
+    for (n = 0; n < 2; ++n) printf("%.16lx ", tmp.i64[n]);
+  }
+  printf("\n");
+}
+#endif
+
+//------------------------------------------------------------------------------
+// Compute susceptibility based on DCT-coeff histograms:
+// the higher, the "easier" the macroblock is to compress.
+
+static void CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
+                                 int start_block, int end_block,
+                                 VP8Histogram* const histo) {
+  const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
+  int j;
+  for (j = start_block; j < end_block; ++j) {
+    int16_t out[16];
+    int k;
+
+    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) >> 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) {
+      histo->distribution[out[k]]++;
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+// 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) {
+  // This implementation makes use of 16-bit fixed point versions of two
+  // multiply constants:
+  //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
+  //    K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16
+  //
+  // To be able to use signed 16-bit integers, we use the following trick to
+  // have constants within range:
+  // - Associated constants are obtained by subtracting the 16-bit fixed point
+  //   version of one:
+  //      k = K - (1 << 16)  =>  K = k + (1 << 16)
+  //      K1 = 85267  =>  k1 =  20091
+  //      K2 = 35468  =>  k2 = -30068
+  // - The multiplication of a variable by a constant become the sum of the
+  //   variable and the multiplication of that variable by the associated
+  //   constant:
+  //      (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x
+  const __m128i k1 = _mm_set1_epi16(20091);
+  const __m128i k2 = _mm_set1_epi16(-30068);
+  __m128i T0, T1, T2, T3;
+
+  // Load and concatenate the transform coefficients (we'll do two inverse
+  // transforms in parallel). In the case of only one inverse transform, the
+  // second half of the 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]);
+    // 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]);
+      in0 = _mm_unpacklo_epi64(in0, inB0);
+      in1 = _mm_unpacklo_epi64(in1, inB1);
+      in2 = _mm_unpacklo_epi64(in2, inB2);
+      in3 = _mm_unpacklo_epi64(in3, inB3);
+      // a00 a10 a20 a30   b00 b10 b20 b30
+      // a01 a11 a21 a31   b01 b11 b21 b31
+      // a02 a12 a22 a32   b02 b12 b22 b32
+      // a03 a13 a23 a33   b03 b13 b23 b33
+    }
+  }
+
+  // Vertical pass and subsequent transpose.
+  {
+    // First pass, c and d calculations are longer because of the "trick"
+    // multiplications.
+    const __m128i a = _mm_add_epi16(in0, in2);
+    const __m128i b = _mm_sub_epi16(in0, in2);
+    // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
+    const __m128i c1 = _mm_mulhi_epi16(in1, k2);
+    const __m128i c2 = _mm_mulhi_epi16(in3, k1);
+    const __m128i c3 = _mm_sub_epi16(in1, in3);
+    const __m128i c4 = _mm_sub_epi16(c1, c2);
+    const __m128i c = _mm_add_epi16(c3, c4);
+    // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
+    const __m128i d1 = _mm_mulhi_epi16(in1, k1);
+    const __m128i d2 = _mm_mulhi_epi16(in3, k2);
+    const __m128i d3 = _mm_add_epi16(in1, in3);
+    const __m128i d4 = _mm_add_epi16(d1, d2);
+    const __m128i d = _mm_add_epi16(d3, d4);
+
+    // Second pass.
+    const __m128i tmp0 = _mm_add_epi16(a, d);
+    const __m128i tmp1 = _mm_add_epi16(b, c);
+    const __m128i tmp2 = _mm_sub_epi16(b, c);
+    const __m128i tmp3 = _mm_sub_epi16(a, d);
+
+    // Transpose the two 4x4.
+    // a00 a01 a02 a03   b00 b01 b02 b03
+    // a10 a11 a12 a13   b10 b11 b12 b13
+    // a20 a21 a22 a23   b20 b21 b22 b23
+    // a30 a31 a32 a33   b30 b31 b32 b33
+    const __m128i transpose0_0 = _mm_unpacklo_epi16(tmp0, tmp1);
+    const __m128i transpose0_1 = _mm_unpacklo_epi16(tmp2, tmp3);
+    const __m128i transpose0_2 = _mm_unpackhi_epi16(tmp0, tmp1);
+    const __m128i transpose0_3 = _mm_unpackhi_epi16(tmp2, tmp3);
+    // a00 a10 a01 a11   a02 a12 a03 a13
+    // a20 a30 a21 a31   a22 a32 a23 a33
+    // b00 b10 b01 b11   b02 b12 b03 b13
+    // b20 b30 b21 b31   b22 b32 b23 b33
+    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
+    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
+    // a00 a10 a20 a30 a01 a11 a21 a31
+    // b00 b10 b20 b30 b01 b11 b21 b31
+    // a02 a12 a22 a32 a03 a13 a23 a33
+    // b02 b12 a22 b32 b03 b13 b23 b33
+    T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
+    T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
+    T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
+    T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
+    // a00 a10 a20 a30   b00 b10 b20 b30
+    // a01 a11 a21 a31   b01 b11 b21 b31
+    // a02 a12 a22 a32   b02 b12 b22 b32
+    // a03 a13 a23 a33   b03 b13 b23 b33
+  }
+
+  // Horizontal pass and subsequent transpose.
+  {
+    // First pass, c and d calculations are longer because of the "trick"
+    // multiplications.
+    const __m128i four = _mm_set1_epi16(4);
+    const __m128i dc = _mm_add_epi16(T0, four);
+    const __m128i a =  _mm_add_epi16(dc, T2);
+    const __m128i b =  _mm_sub_epi16(dc, T2);
+    // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
+    const __m128i c1 = _mm_mulhi_epi16(T1, k2);
+    const __m128i c2 = _mm_mulhi_epi16(T3, k1);
+    const __m128i c3 = _mm_sub_epi16(T1, T3);
+    const __m128i c4 = _mm_sub_epi16(c1, c2);
+    const __m128i c = _mm_add_epi16(c3, c4);
+    // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
+    const __m128i d1 = _mm_mulhi_epi16(T1, k1);
+    const __m128i d2 = _mm_mulhi_epi16(T3, k2);
+    const __m128i d3 = _mm_add_epi16(T1, T3);
+    const __m128i d4 = _mm_add_epi16(d1, d2);
+    const __m128i d = _mm_add_epi16(d3, d4);
+
+    // Second pass.
+    const __m128i tmp0 = _mm_add_epi16(a, d);
+    const __m128i tmp1 = _mm_add_epi16(b, c);
+    const __m128i tmp2 = _mm_sub_epi16(b, c);
+    const __m128i tmp3 = _mm_sub_epi16(a, d);
+    const __m128i shifted0 = _mm_srai_epi16(tmp0, 3);
+    const __m128i shifted1 = _mm_srai_epi16(tmp1, 3);
+    const __m128i shifted2 = _mm_srai_epi16(tmp2, 3);
+    const __m128i shifted3 = _mm_srai_epi16(tmp3, 3);
+
+    // Transpose the two 4x4.
+    // a00 a01 a02 a03   b00 b01 b02 b03
+    // a10 a11 a12 a13   b10 b11 b12 b13
+    // a20 a21 a22 a23   b20 b21 b22 b23
+    // a30 a31 a32 a33   b30 b31 b32 b33
+    const __m128i transpose0_0 = _mm_unpacklo_epi16(shifted0, shifted1);
+    const __m128i transpose0_1 = _mm_unpacklo_epi16(shifted2, shifted3);
+    const __m128i transpose0_2 = _mm_unpackhi_epi16(shifted0, shifted1);
+    const __m128i transpose0_3 = _mm_unpackhi_epi16(shifted2, shifted3);
+    // a00 a10 a01 a11   a02 a12 a03 a13
+    // a20 a30 a21 a31   a22 a32 a23 a33
+    // b00 b10 b01 b11   b02 b12 b03 b13
+    // b20 b30 b21 b31   b22 b32 b23 b33
+    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
+    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
+    // a00 a10 a20 a30 a01 a11 a21 a31
+    // b00 b10 b20 b30 b01 b11 b21 b31
+    // a02 a12 a22 a32 a03 a13 a23 a33
+    // b02 b12 a22 b32 b03 b13 b23 b33
+    T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
+    T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
+    T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
+    T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
+    // a00 a10 a20 a30   b00 b10 b20 b30
+    // a01 a11 a21 a31   b01 b11 b21 b31
+    // a02 a12 a22 a32   b02 b12 b22 b32
+    // a03 a13 a23 a33   b03 b13 b23 b33
+  }
+
+  // Add inverse transform to 'ref' and store.
+  {
+    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]);
+    } 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]);
+    }
+    // Convert to 16b.
+    ref0 = _mm_unpacklo_epi8(ref0, zero);
+    ref1 = _mm_unpacklo_epi8(ref1, zero);
+    ref2 = _mm_unpacklo_epi8(ref2, zero);
+    ref3 = _mm_unpacklo_epi8(ref3, zero);
+    // Add the inverse transform(s).
+    ref0 = _mm_add_epi16(ref0, T0);
+    ref1 = _mm_add_epi16(ref1, T1);
+    ref2 = _mm_add_epi16(ref2, T2);
+    ref3 = _mm_add_epi16(ref3, T3);
+    // Unsigned saturate to 8b.
+    ref0 = _mm_packus_epi16(ref0, ref0);
+    ref1 = _mm_packus_epi16(ref1, ref1);
+    ref2 = _mm_packus_epi16(ref2, ref2);
+    ref3 = _mm_packus_epi16(ref3, ref3);
+    // Store the results.
+    if (do_two) {
+      // Store eight bytes/pixels per line.
+      _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0);
+      _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1);
+      _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2);
+      _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3);
+    } else {
+      // Store four bytes/pixels per line.
+      *((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(ref0);
+      *((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(ref1);
+      *((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(ref2);
+      *((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(ref3);
+    }
+  }
+}
+
+static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
+                           int16_t* out) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i seven = _mm_set1_epi16(7);
+  const __m128i k937 = _mm_set1_epi32(937);
+  const __m128i k1812 = _mm_set1_epi32(1812);
+  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 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);
+  __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. -> 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);
+    // 00 01 10 11 02 03 12 13
+    // 20 21 30 31 22 23 32 33
+    const __m128i shuf01_p =
+        _mm_shufflehi_epi16(shuf01, _MM_SHUFFLE(2, 3, 0, 1));
+    const __m128i shuf23_p =
+        _mm_shufflehi_epi16(shuf23, _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);
+    v01 = _mm_unpacklo_epi32(s_lo, s_hi);
+    v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));  // 3 2 3 2 3 2..
+  }
+
+  // Second pass
+  {
+    // 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));
+
+    _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);
+  }
+}
+
+static void FTransformWHTSSE2(const int16_t* in, int16_t* out) {
+  int32_t tmp[16];
+  int i;
+  for (i = 0; i < 4; ++i, in += 64) {
+    const int a0 = (in[0 * 16] + in[2 * 16]);
+    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;
+    tmp[1 + i * 4] = a3 + a2;
+    tmp[2 + i * 4] = a3 - a2;
+    tmp[3 + i * 4] = a0 - a1;
+  }
+  {
+    const __m128i src0 = _mm_loadu_si128((__m128i*)&tmp[0]);
+    const __m128i src1 = _mm_loadu_si128((__m128i*)&tmp[4]);
+    const __m128i src2 = _mm_loadu_si128((__m128i*)&tmp[8]);
+    const __m128i src3 = _mm_loadu_si128((__m128i*)&tmp[12]);
+    const __m128i a0 = _mm_add_epi32(src0, src2);
+    const __m128i a1 = _mm_add_epi32(src1, src3);
+    const __m128i a2 = _mm_sub_epi32(src1, src3);
+    const __m128i a3 = _mm_sub_epi32(src0, src2);
+    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);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Metric
+
+static int SSE_Nx4SSE2(const uint8_t* a, const uint8_t* b,
+                       int num_quads, int do_16) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum1 = zero;
+  __m128i sum2 = zero;
+
+  while (num_quads-- > 0) {
+    // Note: for the !do_16 case, we read 16 pixels instead of 8 but that's ok,
+    // thanks to buffer over-allocation to that effect.
+    const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]);
+    const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]);
+    const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]);
+    const __m128i a3 = _mm_loadu_si128((__m128i*)&a[BPS * 3]);
+    const __m128i b0 = _mm_loadu_si128((__m128i*)&b[BPS * 0]);
+    const __m128i b1 = _mm_loadu_si128((__m128i*)&b[BPS * 1]);
+    const __m128i b2 = _mm_loadu_si128((__m128i*)&b[BPS * 2]);
+    const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]);
+
+    // compute clip0(a-b) and clip0(b-a)
+    const __m128i a0p = _mm_subs_epu8(a0, b0);
+    const __m128i a0m = _mm_subs_epu8(b0, a0);
+    const __m128i a1p = _mm_subs_epu8(a1, b1);
+    const __m128i a1m = _mm_subs_epu8(b1, a1);
+    const __m128i a2p = _mm_subs_epu8(a2, b2);
+    const __m128i a2m = _mm_subs_epu8(b2, a2);
+    const __m128i a3p = _mm_subs_epu8(a3, b3);
+    const __m128i a3m = _mm_subs_epu8(b3, a3);
+
+    // compute |a-b| with 8b arithmetic as clip0(a-b) | clip0(b-a)
+    const __m128i diff0 = _mm_or_si128(a0p, a0m);
+    const __m128i diff1 = _mm_or_si128(a1p, a1m);
+    const __m128i diff2 = _mm_or_si128(a2p, a2m);
+    const __m128i diff3 = _mm_or_si128(a3p, a3m);
+
+    // unpack (only four operations, instead of eight)
+    const __m128i low0 = _mm_unpacklo_epi8(diff0, zero);
+    const __m128i low1 = _mm_unpacklo_epi8(diff1, zero);
+    const __m128i low2 = _mm_unpacklo_epi8(diff2, zero);
+    const __m128i low3 = _mm_unpacklo_epi8(diff3, zero);
+
+    // multiply with self
+    const __m128i low_madd0 = _mm_madd_epi16(low0, low0);
+    const __m128i low_madd1 = _mm_madd_epi16(low1, low1);
+    const __m128i low_madd2 = _mm_madd_epi16(low2, low2);
+    const __m128i low_madd3 = _mm_madd_epi16(low3, low3);
+
+    // collect in a cascading way
+    const __m128i low_sum0 = _mm_add_epi32(low_madd0, low_madd1);
+    const __m128i low_sum1 = _mm_add_epi32(low_madd2, low_madd3);
+    sum1 = _mm_add_epi32(sum1, low_sum0);
+    sum2 = _mm_add_epi32(sum2, low_sum1);
+
+    if (do_16) {  // if necessary, process the higher 8 bytes similarly
+      const __m128i hi0 = _mm_unpackhi_epi8(diff0, zero);
+      const __m128i hi1 = _mm_unpackhi_epi8(diff1, zero);
+      const __m128i hi2 = _mm_unpackhi_epi8(diff2, zero);
+      const __m128i hi3 = _mm_unpackhi_epi8(diff3, zero);
+
+      const __m128i hi_madd0 = _mm_madd_epi16(hi0, hi0);
+      const __m128i hi_madd1 = _mm_madd_epi16(hi1, hi1);
+      const __m128i hi_madd2 = _mm_madd_epi16(hi2, hi2);
+      const __m128i hi_madd3 = _mm_madd_epi16(hi3, hi3);
+      const __m128i hi_sum0 = _mm_add_epi32(hi_madd0, hi_madd1);
+      const __m128i hi_sum1 = _mm_add_epi32(hi_madd2, hi_madd3);
+      sum1 = _mm_add_epi32(sum1, hi_sum0);
+      sum2 = _mm_add_epi32(sum2, hi_sum1);
+    }
+    a += 4 * BPS;
+    b += 4 * BPS;
+  }
+  {
+    int32_t tmp[4];
+    const __m128i sum = _mm_add_epi32(sum1, sum2);
+    _mm_storeu_si128((__m128i*)tmp, sum);
+    return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
+  }
+}
+
+static int SSE16x16SSE2(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4SSE2(a, b, 4, 1);
+}
+
+static int SSE16x8SSE2(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4SSE2(a, b, 2, 1);
+}
+
+static int SSE8x8SSE2(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4SSE2(a, b, 2, 0);
+}
+
+static int SSE4x4SSE2(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((__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]);
+
+  // Combine pair of lines and convert to 16b.
+  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);
+  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);
+
+  // 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);
+  return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
+}
+
+//------------------------------------------------------------------------------
+// Texture distortion
+//
+// We try to match the spectral content (weighted) between source and
+// reconstructed samples.
+
+// 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) {
+  int32_t sum[4];
+  __m128i tmp_0, tmp_1, tmp_2, tmp_3;
+  const __m128i zero = _mm_setzero_si128();
+
+  // 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]);
+
+    // Combine inA and inB (we'll do two transforms in parallel).
+    const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0);
+    const __m128i inAB_1 = _mm_unpacklo_epi8(inA_1, inB_1);
+    const __m128i inAB_2 = _mm_unpacklo_epi8(inA_2, inB_2);
+    const __m128i inAB_3 = _mm_unpacklo_epi8(inA_3, inB_3);
+    // a00 b00 a01 b01 a02 b03 a03 b03   0 0 0 0 0 0 0 0
+    // a10 b10 a11 b11 a12 b12 a13 b13   0 0 0 0 0 0 0 0
+    // a20 b20 a21 b21 a22 b22 a23 b23   0 0 0 0 0 0 0 0
+    // a30 b30 a31 b31 a32 b32 a33 b33   0 0 0 0 0 0 0 0
+
+    // Transpose the two 4x4, discarding the filling zeroes.
+    const __m128i transpose0_0 = _mm_unpacklo_epi8(inAB_0, inAB_2);
+    const __m128i transpose0_1 = _mm_unpacklo_epi8(inAB_1, inAB_3);
+    // a00 a20  b00 b20  a01 a21  b01 b21  a02 a22  b02 b22  a03 a23  b03 b23
+    // a10 a30  b10 b30  a11 a31  b11 b31  a12 a32  b12 b32  a13 a33  b13 b33
+    const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1);
+    const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1);
+    // a00 a10 a20 a30  b00 b10 b20 b30  a01 a11 a21 a31  b01 b11 b21 b31
+    // a02 a12 a22 a32  b02 b12 b22 b32  a03 a13 a23 a33  b03 b13 b23 b33
+
+    // Convert to 16b.
+    tmp_0 = _mm_unpacklo_epi8(transpose1_0, zero);
+    tmp_1 = _mm_unpackhi_epi8(transpose1_0, zero);
+    tmp_2 = _mm_unpacklo_epi8(transpose1_1, zero);
+    tmp_3 = _mm_unpackhi_epi8(transpose1_1, zero);
+    // a00 a10 a20 a30   b00 b10 b20 b30
+    // a01 a11 a21 a31   b01 b11 b21 b31
+    // a02 a12 a22 a32   b02 b12 b22 b32
+    // a03 a13 a23 a33   b03 b13 b23 b33
+  }
+
+  // Horizontal pass and subsequent transpose.
+  {
+    // Calculate a and b (two 4x4 at once).
+    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);
+    // a00 a01 a02 a03   b00 b01 b02 b03
+    // a10 a11 a12 a13   b10 b11 b12 b13
+    // a20 a21 a22 a23   b20 b21 b22 b23
+    // a30 a31 a32 a33   b30 b31 b32 b33
+
+    // Transpose the two 4x4.
+    const __m128i transpose0_0 = _mm_unpacklo_epi16(b0, b1);
+    const __m128i transpose0_1 = _mm_unpacklo_epi16(b2, b3);
+    const __m128i transpose0_2 = _mm_unpackhi_epi16(b0, b1);
+    const __m128i transpose0_3 = _mm_unpackhi_epi16(b2, b3);
+    // a00 a10 a01 a11   a02 a12 a03 a13
+    // a20 a30 a21 a31   a22 a32 a23 a33
+    // b00 b10 b01 b11   b02 b12 b03 b13
+    // b20 b30 b21 b31   b22 b32 b23 b33
+    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
+    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
+    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
+    // a00 a10 a20 a30 a01 a11 a21 a31
+    // b00 b10 b20 b30 b01 b11 b21 b31
+    // a02 a12 a22 a32 a03 a13 a23 a33
+    // b02 b12 a22 b32 b03 b13 b23 b33
+    tmp_0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
+    tmp_1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
+    tmp_2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
+    tmp_3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
+    // a00 a10 a20 a30   b00 b10 b20 b30
+    // a01 a11 a21 a31   b01 b11 b21 b31
+    // a02 a12 a22 a32   b02 b12 b22 b32
+    // a03 a13 a23 a33   b03 b13 b23 b33
+  }
+
+  // Vertical pass and difference of weighted sums.
+  {
+    // 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]);
+
+    // Calculate a and b (two 4x4 at once).
+    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);
+
+    // Separate the transforms of inA and inB.
+    __m128i A_b0 = _mm_unpacklo_epi64(b0, b1);
+    __m128i A_b2 = _mm_unpacklo_epi64(b2, b3);
+    __m128i B_b0 = _mm_unpackhi_epi64(b0, b1);
+    __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);
+    }
+
+    // weighted sums
+    A_b0 = _mm_madd_epi16(A_b0, w_0);
+    A_b2 = _mm_madd_epi16(A_b2, w_8);
+    B_b0 = _mm_madd_epi16(B_b0, w_0);
+    B_b2 = _mm_madd_epi16(B_b2, w_8);
+    A_b0 = _mm_add_epi32(A_b0, A_b2);
+    B_b0 = _mm_add_epi32(B_b0, B_b2);
+
+    // difference of weighted sums
+    A_b0 = _mm_sub_epi32(A_b0, B_b0);
+    _mm_storeu_si128((__m128i*)&sum[0], A_b0);
+  }
+  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) >> 5;
+}
+
+static int Disto16x16SSE2(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);
+    }
+  }
+  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(MAX_LEVEL);
+  const __m128i zero = _mm_setzero_si128();
+  __m128i coeff0, coeff8;
+  __m128i out0, out8;
+  __m128i packed_out;
+
+  // Load all inputs.
+  // TODO(cduvivier): Make variable declarations and allocations aligned so that
+  //                  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]);
+
+  // sign(in) = in >> 15  (0x0000 if positive, 0xffff if negative)
+  const __m128i sign0 = _mm_srai_epi16(in0, 15);
+  const __m128i sign8 = _mm_srai_epi16(in8, 15);
+
+  // coeff = abs(in) = (in ^ sign) - sign
+  coeff0 = _mm_xor_si128(in0, sign0);
+  coeff8 = _mm_xor_si128(in8, sign8);
+  coeff0 = _mm_sub_epi16(coeff0, sign0);
+  coeff8 = _mm_sub_epi16(coeff8, sign8);
+
+  // coeff = abs(in) + sharpen
+  coeff0 = _mm_add_epi16(coeff0, sharpen0);
+  coeff8 = _mm_add_epi16(coeff8, sharpen8);
+
+  // 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);
+    __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)
+    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_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)
+  out0 = _mm_xor_si128(out0, sign0);
+  out8 = _mm_xor_si128(out8, sign8);
+  out0 = _mm_sub_epi16(out0, sign0);
+  out8 = _mm_sub_epi16(out8, sign8);
+
+  // in = out * Q
+  in0 = _mm_mullo_epi16(out0, q0);
+  in8 = _mm_mullo_epi16(out8, q8);
+
+  _mm_storeu_si128((__m128i*)&in[0], in0);
+  _mm_storeu_si128((__m128i*)&in[8], in8);
+
+  // zigzag the output before storing it.
+  //
+  // The zigzag pattern can almost be reproduced with a small sequence of
+  // shuffles. After it, we only need to swap the 7th (ending up in third
+  // position instead of twelfth) and 8th values.
+  {
+    __m128i outZ0, outZ8;
+    outZ0 = _mm_shufflehi_epi16(out0,  _MM_SHUFFLE(2, 1, 3, 0));
+    outZ0 = _mm_shuffle_epi32  (outZ0, _MM_SHUFFLE(3, 1, 2, 0));
+    outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2));
+    outZ8 = _mm_shufflelo_epi16(out8,  _MM_SHUFFLE(3, 0, 2, 1));
+    outZ8 = _mm_shuffle_epi32  (outZ8, _MM_SHUFFLE(3, 1, 2, 0));
+    outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0));
+    _mm_storeu_si128((__m128i*)&out[0], outZ0);
+    _mm_storeu_si128((__m128i*)&out[8], outZ8);
+    packed_out = _mm_packs_epi16(outZ0, outZ8);
+  }
+  {
+    const int16_t outZ_12 = out[12];
+    const int16_t outZ_3 = out[3];
+    out[3] = outZ_12;
+    out[12] = outZ_3;
+  }
+
+  // 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]);
+  }
+}
+
+static int QuantizeBlockWHTSSE2(int16_t in[16], int16_t out[16],
+                                const VP8Matrix* const mtx) {
+  return QuantizeBlockSSE2(in, out, 0, mtx);
+}
+
+#endif   // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8EncDspInitSSE2(void);
+
+void VP8EncDspInitSSE2(void) {
+#if defined(WEBP_USE_SSE2)
+  VP8CollectHistogram = CollectHistogramSSE2;
+  VP8EncQuantizeBlock = QuantizeBlockSSE2;
+  VP8EncQuantizeBlockWHT = QuantizeBlockWHTSSE2;
+  VP8ITransform = ITransformSSE2;
+  VP8FTransform = FTransformSSE2;
+  VP8FTransformWHT = FTransformWHTSSE2;
+  VP8SSE16x16 = SSE16x16SSE2;
+  VP8SSE16x8 = SSE16x8SSE2;
+  VP8SSE8x8 = SSE8x8SSE2;
+  VP8SSE4x4 = SSE4x4SSE2;
+  VP8TDisto4x4 = Disto4x4SSE2;
+  VP8TDisto16x16 = Disto16x16SSE2;
+#endif   // WEBP_USE_SSE2
+}
+
diff --git a/drivers/webp/dsp/lossless.c b/drivers/webp/dsp/lossless.c
new file mode 100644
index 0000000000..bab76d22de
--- /dev/null
+++ b/drivers/webp/dsp/lossless.c
@@ -0,0 +1,1532 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Image transforms and color space conversion methods for lossless decoder.
+//
+// Authors: Vikas Arora (vikaas.arora@gmail.com)
+//          Jyrki Alakuijala (jyrki@google.com)
+//          Urvang Joshi (urvang@google.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_SSE2)
+#include <emmintrin.h>
+#endif
+
+#include <math.h>
+#include <stdlib.h>
+#include "./lossless.h"
+#include "../dec/vp8li.h"
+#include "./yuv.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
+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
+};
+
+const float kSLog2Table[LOG_LOOKUP_IDX_MAX] = {
+  0.00000000f,    0.00000000f,  2.00000000f,   4.75488750f,
+  8.00000000f,   11.60964047f,  15.50977500f,  19.65148445f,
+  24.00000000f,  28.52932501f,  33.21928095f,  38.05374781f,
+  43.01955001f,  48.10571634f,  53.30296891f,  58.60335893f,
+  64.00000000f,  69.48686830f,  75.05865003f,  80.71062276f,
+  86.43856190f,  92.23866588f,  98.10749561f,  104.04192499f,
+  110.03910002f, 116.09640474f, 122.21143267f, 128.38196256f,
+  134.60593782f, 140.88144886f, 147.20671787f, 153.58008562f,
+  160.00000000f, 166.46500594f, 172.97373660f, 179.52490559f,
+  186.11730005f, 192.74977453f, 199.42124551f, 206.13068654f,
+  212.87712380f, 219.65963219f, 226.47733176f, 233.32938445f,
+  240.21499122f, 247.13338933f, 254.08384998f, 261.06567603f,
+  268.07820003f, 275.12078236f, 282.19280949f, 289.29369244f,
+  296.42286534f, 303.57978409f, 310.76392512f, 317.97478424f,
+  325.21187564f, 332.47473081f, 339.76289772f, 347.07593991f,
+  354.41343574f, 361.77497759f, 369.16017124f, 376.56863518f,
+  384.00000000f, 391.45390785f, 398.93001188f, 406.42797576f,
+  413.94747321f, 421.48818752f, 429.04981119f, 436.63204548f,
+  444.23460010f, 451.85719280f, 459.49954906f, 467.16140179f,
+  474.84249102f, 482.54256363f, 490.26137307f, 497.99867911f,
+  505.75424759f, 513.52785023f, 521.31926438f, 529.12827280f,
+  536.95466351f, 544.79822957f, 552.65876890f, 560.53608414f,
+  568.42998244f, 576.34027536f, 584.26677867f, 592.20931226f,
+  600.16769996f, 608.14176943f, 616.13135206f, 624.13628279f,
+  632.15640007f, 640.19154569f, 648.24156472f, 656.30630539f,
+  664.38561898f, 672.47935976f, 680.58738488f, 688.70955430f,
+  696.84573069f, 704.99577935f, 713.15956818f, 721.33696754f,
+  729.52785023f, 737.73209140f, 745.94956849f, 754.18016116f,
+  762.42375127f, 770.68022275f, 778.94946161f, 787.23135586f,
+  795.52579543f, 803.83267219f, 812.15187982f, 820.48331383f,
+  828.82687147f, 837.18245171f, 845.54995518f, 853.92928416f,
+  862.32034249f, 870.72303558f, 879.13727036f, 887.56295522f,
+  896.00000000f, 904.44831595f, 912.90781569f, 921.37841320f,
+  929.86002376f, 938.35256392f, 946.85595152f, 955.37010560f,
+  963.89494641f, 972.43039537f, 980.97637504f, 989.53280911f,
+  998.09962237f, 1006.67674069f, 1015.26409097f, 1023.86160116f,
+  1032.46920021f, 1041.08681805f, 1049.71438560f, 1058.35183469f,
+  1066.99909811f, 1075.65610955f, 1084.32280357f, 1092.99911564f,
+  1101.68498204f, 1110.38033993f, 1119.08512727f, 1127.79928282f,
+  1136.52274614f, 1145.25545758f, 1153.99735821f, 1162.74838989f,
+  1171.50849518f, 1180.27761738f, 1189.05570047f, 1197.84268914f,
+  1206.63852876f, 1215.44316535f, 1224.25654560f, 1233.07861684f,
+  1241.90932703f, 1250.74862473f, 1259.59645914f, 1268.45278005f,
+  1277.31753781f, 1286.19068338f, 1295.07216828f, 1303.96194457f,
+  1312.85996488f, 1321.76618236f, 1330.68055071f, 1339.60302413f,
+  1348.53355734f, 1357.47210556f, 1366.41862452f, 1375.37307041f,
+  1384.33539991f, 1393.30557020f, 1402.28353887f, 1411.26926400f,
+  1420.26270412f, 1429.26381818f, 1438.27256558f, 1447.28890615f,
+  1456.31280014f, 1465.34420819f, 1474.38309138f, 1483.42941118f,
+  1492.48312945f, 1501.54420843f, 1510.61261078f, 1519.68829949f,
+  1528.77123795f, 1537.86138993f, 1546.95871952f, 1556.06319119f,
+  1565.17476976f, 1574.29342040f, 1583.41910860f, 1592.55180020f,
+  1601.69146137f, 1610.83805860f, 1619.99155871f, 1629.15192882f,
+  1638.31913637f, 1647.49314911f, 1656.67393509f, 1665.86146266f,
+  1675.05570047f, 1684.25661744f, 1693.46418280f, 1702.67836605f,
+  1711.89913698f, 1721.12646563f, 1730.36032233f, 1739.60067768f,
+  1748.84750254f, 1758.10076802f, 1767.36044551f, 1776.62650662f,
+  1785.89892323f, 1795.17766747f, 1804.46271172f, 1813.75402857f,
+  1823.05159087f, 1832.35537170f, 1841.66534438f, 1850.98148244f,
+  1860.30375965f, 1869.63214999f, 1878.96662767f, 1888.30716711f,
+  1897.65374295f, 1907.00633003f, 1916.36490342f, 1925.72943838f,
+  1935.09991037f, 1944.47629506f, 1953.85856831f, 1963.24670620f,
+  1972.64068498f, 1982.04048108f, 1991.44607117f, 2000.85743204f,
+  2010.27454072f, 2019.69737440f, 2029.12591044f, 2038.56012640f
+};
+
+const VP8LPrefixCode kPrefixEncodeCode[PREFIX_LOOKUP_IDX_MAX] = {
+  { 0, 0}, { 0, 0}, { 1, 0}, { 2, 0}, { 3, 0}, { 4, 1}, { 4, 1}, { 5, 1},
+  { 5, 1}, { 6, 2}, { 6, 2}, { 6, 2}, { 6, 2}, { 7, 2}, { 7, 2}, { 7, 2},
+  { 7, 2}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3},
+  { 8, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3},
+  { 9, 3}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4},
+  {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4},
+  {10, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4},
+  {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4},
+  {11, 4}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5},
+  {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5},
+  {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5},
+  {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5},
+  {12, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5},
+  {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5},
+  {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5},
+  {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5},
+  {13, 5}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6},
+  {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6},
+  {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6},
+  {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6},
+  {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6},
+  {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6},
+  {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6},
+  {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6},
+  {14, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6},
+  {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6},
+  {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6},
+  {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6},
+  {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6},
+  {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6},
+  {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6},
+  {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6},
+  {15, 6}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7},
+  {16, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+  {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7},
+};
+
+const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX] = {
+   0,  0,  0,  0,  0,  0,  1,  0,  1,  0,  1,  2,  3,  0,  1,  2,  3,
+   0,  1,  2,  3,  4,  5,  6,  7,  0,  1,  2,  3,  4,  5,  6,  7,
+   0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+   0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+   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,
+   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,
+   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,  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,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+  32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+  48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+  64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79,
+  80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95,
+  96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111,
+  112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126,
+  127,
+   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,
+  64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79,
+  80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95,
+  96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111,
+  112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126
+};
+
+float VP8LFastSLog2Slow(int v) {
+  assert(v >= LOG_LOOKUP_IDX_MAX);
+  if (v < APPROX_LOG_MAX) {
+    int log_cnt = 0;
+    const float v_f = (float)v;
+    while (v >= LOG_LOOKUP_IDX_MAX) {
+      ++log_cnt;
+      v = v >> 1;
+    }
+    return v_f * (kLog2Table[v] + log_cnt);
+  } else {
+    return (float)(LOG_2_RECIPROCAL * v * log((double)v));
+  }
+}
+
+float VP8LFastLog2Slow(int v) {
+  assert(v >= LOG_LOOKUP_IDX_MAX);
+  if (v < APPROX_LOG_MAX) {
+    int log_cnt = 0;
+    while (v >= LOG_LOOKUP_IDX_MAX) {
+      ++log_cnt;
+      v = v >> 1;
+    }
+    return kLog2Table[v] + log_cnt;
+  } else {
+    return (float)(LOG_2_RECIPROCAL * log((double)v));
+  }
+}
+
+//------------------------------------------------------------------------------
+// Image transforms.
+
+// In-place sum of each component with mod 256.
+static WEBP_INLINE void AddPixelsEq(uint32_t* a, uint32_t b) {
+  const uint32_t alpha_and_green = (*a & 0xff00ff00u) + (b & 0xff00ff00u);
+  const uint32_t red_and_blue = (*a & 0x00ff00ffu) + (b & 0x00ff00ffu);
+  *a = (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu);
+}
+
+static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) {
+  return (((a0 ^ a1) & 0xfefefefeL) >> 1) + (a0 & a1);
+}
+
+static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) {
+  return Average2(Average2(a0, a2), a1);
+}
+
+static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1,
+                                     uint32_t a2, uint32_t a3) {
+  return Average2(Average2(a0, a1), Average2(a2, a3));
+}
+
+static WEBP_INLINE uint32_t Clip255(uint32_t a) {
+  if (a < 256) {
+    return a;
+  }
+  // return 0, when a is a negative integer.
+  // return 255, when a is positive.
+  return ~a >> 24;
+}
+
+static WEBP_INLINE int AddSubtractComponentFull(int a, int b, int c) {
+  return Clip255(a + b - c);
+}
+
+static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1,
+                                                   uint32_t c2) {
+  const int a = AddSubtractComponentFull(c0 >> 24, c1 >> 24, c2 >> 24);
+  const int r = AddSubtractComponentFull((c0 >> 16) & 0xff,
+                                         (c1 >> 16) & 0xff,
+                                         (c2 >> 16) & 0xff);
+  const int g = AddSubtractComponentFull((c0 >> 8) & 0xff,
+                                         (c1 >> 8) & 0xff,
+                                         (c2 >> 8) & 0xff);
+  const int b = AddSubtractComponentFull(c0 & 0xff, c1 & 0xff, c2 & 0xff);
+  return (a << 24) | (r << 16) | (g << 8) | b;
+}
+
+static WEBP_INLINE int AddSubtractComponentHalf(int a, int b) {
+  return Clip255(a + (a - b) / 2);
+}
+
+static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
+                                                   uint32_t c2) {
+  const uint32_t ave = Average2(c0, c1);
+  const int a = AddSubtractComponentHalf(ave >> 24, c2 >> 24);
+  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;
+}
+
+static WEBP_INLINE int Sub3(int a, int b, int c) {
+  const int pb = b - c;
+  const int pa = a - c;
+  return abs(pb) - abs(pa);
+}
+
+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;
+}
+
+//------------------------------------------------------------------------------
+// Predictors
+
+static uint32_t Predictor0(uint32_t left, const uint32_t* const top) {
+  (void)top;
+  (void)left;
+  return ARGB_BLACK;
+}
+static uint32_t Predictor1(uint32_t left, const uint32_t* const top) {
+  (void)top;
+  return left;
+}
+static uint32_t Predictor2(uint32_t left, const uint32_t* const top) {
+  (void)left;
+  return top[0];
+}
+static uint32_t Predictor3(uint32_t left, const uint32_t* const top) {
+  (void)left;
+  return top[1];
+}
+static uint32_t Predictor4(uint32_t left, const uint32_t* const top) {
+  (void)left;
+  return top[-1];
+}
+static uint32_t Predictor5(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average3(left, top[0], top[1]);
+  return pred;
+}
+static uint32_t Predictor6(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average2(left, top[-1]);
+  return pred;
+}
+static uint32_t Predictor7(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average2(left, top[0]);
+  return pred;
+}
+static uint32_t Predictor8(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average2(top[-1], top[0]);
+  (void)left;
+  return pred;
+}
+static uint32_t Predictor9(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average2(top[0], top[1]);
+  (void)left;
+  return pred;
+}
+static uint32_t Predictor10(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = Average4(left, top[-1], top[0], top[1]);
+  return pred;
+}
+static uint32_t Predictor11(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = VP8LSelect(top[0], left, top[-1]);
+  return pred;
+}
+static uint32_t Predictor12(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = VP8LClampedAddSubtractFull(left, top[0], top[-1]);
+  return pred;
+}
+static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
+  const uint32_t pred = VP8LClampedAddSubtractHalf(left, top[0], top[-1]);
+  return pred;
+}
+
+// TODO(vikasa): Export the predictor array, to allow SSE2 variants.
+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 combined Shanon's entropy for distribution {X} and {X+Y}
+static float CombinedShannonEntropy(const int* const X,
+                                    const int* const Y, int n) {
+  int i;
+  double retval = 0.;
+  int sumX = 0, sumXY = 0;
+  for (i = 0; i < n; ++i) {
+    const int x = X[i];
+    const int xy = X[i] + Y[i];
+    if (x != 0) {
+      sumX += x;
+      retval -= VP8LFastSLog2(x);
+    }
+    if (xy != 0) {
+      sumXY += xy;
+      retval -= VP8LFastSLog2(xy);
+    }
+  }
+  retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
+  return (float)retval;
+}
+
+static float PredictionCostSpatialHistogram(int accumulated[4][256],
+                                            int tile[4][256]) {
+  int i;
+  double retval = 0;
+  for (i = 0; i < 4; ++i) {
+    const double kExpValue = 0.94;
+    retval += PredictionCostSpatial(tile[i], 1, kExpValue);
+    retval += CombinedShannonEntropy(tile[i], accumulated[i], 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,
+                                      int y_start, int y_end, uint32_t* data) {
+  const int width = transform->xsize_;
+  if (y_start == 0) {  // First Row follows the L (mode=1) mode.
+    int x;
+    const uint32_t pred0 = Predictor0(data[-1], NULL);
+    AddPixelsEq(data, pred0);
+    for (x = 1; x < width; ++x) {
+      const uint32_t pred1 = Predictor1(data[x - 1], NULL);
+      AddPixelsEq(data + x, pred1);
+    }
+    data += width;
+    ++y_start;
+  }
+
+  {
+    int y = y_start;
+    const int mask = (1 << transform->bits_) - 1;
+    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;
+
+      // 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];
+        }
+        pred = pred_func(data[x - 1], data + x - width);
+        AddPixelsEq(data + x, pred);
+      }
+      data += width;
+      ++y;
+      if ((y & mask) == 0) {   // Use the same mask, since tiles are squares.
+        pred_mode_base += tiles_per_row;
+      }
+    }
+  }
+}
+
+static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs) {
+  int i = 0;
+  for (; 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(uint32_t* data, const uint32_t* data_end) {
+  while (data < data_end) {
+    const uint32_t argb = *data;
+    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;
+  }
+}
+
+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) {
+  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) {
+    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 uint8_t TransformColorRed(uint8_t green_to_red,
+                                             uint32_t argb) {
+  const uint32_t green = argb >> 8;
+  uint32_t new_red = argb >> 16;
+  new_red -= ColorTransformDelta(green_to_red, green);
+  return (new_red & 0xff);
+}
+
+static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue,
+                                              uint8_t red_to_blue,
+                                              uint32_t argb) {
+  const uint32_t green = argb >> 8;
+  const uint32_t red = argb >> 16;
+  uint8_t new_blue = argb;
+  new_blue -= ColorTransformDelta(green_to_blue, green);
+  new_blue -= ColorTransformDelta(red_to_blue, red);
+  return (new_blue & 0xff);
+}
+
+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.
+  // Favor small absolute values for PredictionCostSpatial
+  static const double kExpValue = 2.4;
+  return CombinedShannonEntropy(counts, accumulated, 256) +
+         PredictionCostSpatial(counts, 3, kExpValue);
+}
+
+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;
+
+    for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
+      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;
+        }
+        ++histo[TransformColorRed(green_to_red, argb[ix])];  // red.
+      }
+    }
+    cur_diff = PredictionCostCrossColor(&accumulated_red_histo[0], &histo[0]);
+    if ((uint8_t)green_to_red == prevX.green_to_red_) {
+      cur_diff -= 3;  // favor keeping the areas locally similar
+    }
+    if ((uint8_t)green_to_red == prevY.green_to_red_) {
+      cur_diff -= 3;  // favor keeping the areas locally similar
+    }
+    if (green_to_red == 0) {
+      cur_diff -= 3;
+    }
+    if (cur_diff < best_diff) {
+      best_diff = cur_diff;
+      best_tx.green_to_red_ = green_to_red;
+    }
+  }
+  best_diff = MAX_DIFF_COST;
+  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 };
+      for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
+        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;
+          }
+          ++histo[TransformColorBlue(green_to_blue, red_to_blue, argb[ix])];
+        }
+      }
+      cur_diff =
+          PredictionCostCrossColor(&accumulated_blue_histo[0], &histo[0]);
+      if ((uint8_t)green_to_blue == prevX.green_to_blue_) {
+        cur_diff -= 3;  // favor keeping the areas locally similar
+      }
+      if ((uint8_t)green_to_blue == prevY.green_to_blue_) {
+        cur_diff -= 3;  // favor keeping the areas locally similar
+      }
+      if ((uint8_t)red_to_blue == prevX.red_to_blue_) {
+        cur_diff -= 3;  // favor keeping the areas locally similar
+      }
+      if ((uint8_t)red_to_blue == prevY.red_to_blue_) {
+        cur_diff -= 3;  // favor keeping the areas locally similar
+      }
+      if (green_to_blue == 0) {
+        cur_diff -= 3;
+      }
+      if (red_to_blue == 0) {
+        cur_diff -= 3;
+      }
+      if (cur_diff < best_diff) {
+        best_diff = cur_diff;
+        best_tx.green_to_blue_ = green_to_blue;
+        best_tx.red_to_blue_ = red_to_blue;
+      }
+    }
+  }
+  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];
+        }
+      }
+    }
+  }
+}
+
+// Color space inverse transform.
+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 tiles_per_row = VP8LSubSampleSize(width, transform->bits_);
+  int y = y_start;
+  const uint32_t* pred_row =
+      transform->data_ + (y >> transform->bits_) * tiles_per_row;
+
+  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);
+    }
+    data += width;
+    ++y;
+    if ((y & mask) == 0) pred_row += tiles_per_row;;
+  }
+}
+
+// Separate out pixels packed together using pixel-bundling.
+// We define two methods for ARGB data (uint32_t) and alpha-only data (uint8_t).
+#define COLOR_INDEX_INVERSE(FUNC_NAME, TYPE, GET_INDEX, GET_VALUE)             \
+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 {                                                                     \
+    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 WEBP_INLINE uint32_t GetARGBIndex(uint32_t idx) {
+  return (idx >> 8) & 0xff;
+}
+
+static WEBP_INLINE uint8_t GetAlphaIndex(uint8_t idx) {
+  return idx;
+}
+
+static WEBP_INLINE uint32_t GetARGBValue(uint32_t val) {
+  return val;
+}
+
+static WEBP_INLINE uint8_t GetAlphaValue(uint32_t val) {
+  return (val >> 8) & 0xff;
+}
+
+static COLOR_INDEX_INVERSE(ColorIndexInverseTransform, uint32_t, GetARGBIndex,
+                           GetARGBValue)
+COLOR_INDEX_INVERSE(VP8LColorIndexInverseTransformAlpha, uint8_t, GetAlphaIndex,
+                    GetAlphaValue)
+
+#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:
+      VP8LAddGreenToBlueAndRed(out, 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.
+        memcpy(out - width, out + (row_end - row_start - 1) * width,
+               width * sizeof(*out));
+      }
+      break;
+    case CROSS_COLOR_TRANSFORM:
+      ColorSpaceInverseTransform(transform, row_start, row_end, out);
+      break;
+    case COLOR_INDEXING_TRANSFORM:
+      if (in == out && transform->bits_ > 0) {
+        // Move packed pixels to the end of unpacked region, so that unpacking
+        // can occur seamlessly.
+        // 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) * width;
+        const int in_stride = (row_end - row_start) *
+            VP8LSubSampleSize(transform->xsize_, transform->bits_);
+        uint32_t* const src = out + out_stride - in_stride;
+        memmove(src, out, in_stride * sizeof(*src));
+        ColorIndexInverseTransform(transform, row_start, row_end, src, out);
+      } else {
+        ColorIndexInverseTransform(transform, row_start, row_end, in, out);
+      }
+      break;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Color space conversion.
+
+static int is_big_endian(void) {
+  static const union {
+    uint16_t w;
+    uint8_t b[2];
+  } tmp = { 1 };
+  return (tmp.b[0] != 1);
+}
+
+static void ConvertBGRAToRGB(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) & 0xff;
+    *dst++ = (argb >>  8) & 0xff;
+    *dst++ = (argb >>  0) & 0xff;
+  }
+}
+
+static void ConvertBGRAToRGBA(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) & 0xff;
+    *dst++ = (argb >>  8) & 0xff;
+    *dst++ = (argb >>  0) & 0xff;
+    *dst++ = (argb >> 24) & 0xff;
+  }
+}
+
+static void ConvertBGRAToRGBA4444(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++;
+    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) {
+  const uint32_t* const src_end = src + num_pixels;
+  while (src < src_end) {
+    const uint32_t argb = *src++;
+    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) {
+  const uint32_t* const src_end = src + num_pixels;
+  while (src < src_end) {
+    const uint32_t argb = *src++;
+    *dst++ = (argb >>  0) & 0xff;
+    *dst++ = (argb >>  8) & 0xff;
+    *dst++ = (argb >> 16) & 0xff;
+  }
+}
+
+static void CopyOrSwap(const uint32_t* src, int num_pixels, uint8_t* dst,
+                       int swap_on_big_endian) {
+  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__)
+#if !defined(WEBP_REFERENCE_IMPLEMENTATION)
+#if defined(__i386__) || defined(__x86_64__)
+      __asm__ volatile("bswap %0" : "=r"(argb) : "0"(argb));
+      *(uint32_t*)dst = argb;
+#elif defined(_MSC_VER)
+      argb = _byteswap_ulong(argb);
+      *(uint32_t*)dst = argb;
+#else
+      dst[0] = (argb >> 24) & 0xff;
+      dst[1] = (argb >> 16) & 0xff;
+      dst[2] = (argb >>  8) & 0xff;
+      dst[3] = (argb >>  0) & 0xff;
+#endif
+#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  // __BIG_ENDIAN__
+      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));
+  }
+}
+
+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);
+      break;
+    case MODE_RGBA:
+      ConvertBGRAToRGBA(in_data, num_pixels, rgba);
+      break;
+    case MODE_rgbA:
+      ConvertBGRAToRGBA(in_data, num_pixels, rgba);
+      WebPApplyAlphaMultiply(rgba, 0, num_pixels, 1, 0);
+      break;
+    case MODE_BGR:
+      ConvertBGRAToBGR(in_data, num_pixels, rgba);
+      break;
+    case MODE_BGRA:
+      CopyOrSwap(in_data, num_pixels, rgba, 1);
+      break;
+    case MODE_bgrA:
+      CopyOrSwap(in_data, num_pixels, rgba, 1);
+      WebPApplyAlphaMultiply(rgba, 0, num_pixels, 1, 0);
+      break;
+    case MODE_ARGB:
+      CopyOrSwap(in_data, num_pixels, rgba, 0);
+      break;
+    case MODE_Argb:
+      CopyOrSwap(in_data, num_pixels, rgba, 0);
+      WebPApplyAlphaMultiply(rgba, 1, num_pixels, 1, 0);
+      break;
+    case MODE_RGBA_4444:
+      ConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
+      break;
+    case MODE_rgbA_4444:
+      ConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
+      WebPApplyAlphaMultiply4444(rgba, num_pixels, 1, 0);
+      break;
+    case MODE_RGB_565:
+      ConvertBGRAToRGB565(in_data, num_pixels, rgba);
+      break;
+    default:
+      assert(0);          // Code flow should not reach here.
+  }
+}
+
+// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
+void VP8LBundleColorMap(const uint8_t* const row, int width,
+                        int xbits, uint32_t* const dst) {
+  int x;
+  if (xbits > 0) {
+    const int bit_depth = 1 << (3 - xbits);
+    const int mask = (1 << xbits) - 1;
+    uint32_t code = 0xff000000;
+    for (x = 0; x < width; ++x) {
+      const int xsub = x & mask;
+      if (xsub == 0) {
+        code = 0xff000000;
+      }
+      code |= row[x] << (8 + bit_depth * xsub);
+      dst[x >> xbits] = code;
+    }
+  } else {
+    for (x = 0; x < width; ++x) dst[x] = 0xff000000 | (row[x] << 8);
+  }
+}
+
+//------------------------------------------------------------------------------
+
+// TODO(vikasa): Move the SSE2 functions to lossless_dsp.c (new file), once
+// color-space conversion methods (ConvertFromBGRA) are also updated for SSE2.
+#if defined(WEBP_USE_SSE2)
+static WEBP_INLINE uint32_t ClampedAddSubtractFullSSE2(uint32_t c0, uint32_t c1,
+                                                       uint32_t c2) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
+  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
+  const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
+  const __m128i V1 = _mm_add_epi16(C0, C1);
+  const __m128i V2 = _mm_sub_epi16(V1, C2);
+  const __m128i b = _mm_packus_epi16(V2, V2);
+  const uint32_t output = _mm_cvtsi128_si32(b);
+  return output;
+}
+
+static WEBP_INLINE uint32_t ClampedAddSubtractHalfSSE2(uint32_t c0, uint32_t c1,
+                                                       uint32_t c2) {
+  const uint32_t ave = Average2(c0, c1);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ave), zero);
+  const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
+  const __m128i A1 = _mm_sub_epi16(A0, B0);
+  const __m128i BgtA = _mm_cmpgt_epi16(B0, A0);
+  const __m128i A2 = _mm_sub_epi16(A1, BgtA);
+  const __m128i A3 = _mm_srai_epi16(A2, 1);
+  const __m128i A4 = _mm_add_epi16(A0, A3);
+  const __m128i A5 = _mm_packus_epi16(A4, A4);
+  const uint32_t output = _mm_cvtsi128_si32(A5);
+  return output;
+}
+
+static WEBP_INLINE uint32_t SelectSSE2(uint32_t a, uint32_t b, uint32_t c) {
+  int pa_minus_pb;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i A0 = _mm_cvtsi32_si128(a);
+  const __m128i B0 = _mm_cvtsi32_si128(b);
+  const __m128i C0 = _mm_cvtsi32_si128(c);
+  const __m128i AC0 = _mm_subs_epu8(A0, C0);
+  const __m128i CA0 = _mm_subs_epu8(C0, A0);
+  const __m128i BC0 = _mm_subs_epu8(B0, C0);
+  const __m128i CB0 = _mm_subs_epu8(C0, B0);
+  const __m128i AC = _mm_or_si128(AC0, CA0);
+  const __m128i BC = _mm_or_si128(BC0, CB0);
+  const __m128i pa = _mm_unpacklo_epi8(AC, zero);  // |a - c|
+  const __m128i pb = _mm_unpacklo_epi8(BC, zero);  // |b - c|
+  const __m128i diff = _mm_sub_epi16(pb, pa);
+  {
+    int16_t out[8];
+    _mm_storeu_si128((__m128i*)out, diff);
+    pa_minus_pb = out[0] + out[1] + out[2] + out[3];
+  }
+  return (pa_minus_pb <= 0) ? a : b;
+}
+
+static void SubtractGreenFromBlueAndRedSSE2(uint32_t* argb_data, int num_pixs) {
+  int i = 0;
+  const __m128i mask = _mm_set1_epi32(0x0000ff00);
+  for (; i + 4 < num_pixs; i += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
+    const __m128i in_00g0 = _mm_and_si128(in, mask);     // 00g0|00g0|...
+    const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8);  // 0g00|0g00|...
+    const __m128i in_000g = _mm_srli_epi32(in_00g0, 8);  // 000g|000g|...
+    const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g);
+    const __m128i out = _mm_sub_epi8(in, in_0g0g);
+    _mm_storeu_si128((__m128i*)&argb_data[i], out);
+  }
+  // fallthrough and finish off with plain-C
+  for (; 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;
+  }
+}
+
+static void AddGreenToBlueAndRedSSE2(uint32_t* data, const uint32_t* data_end) {
+  const __m128i mask = _mm_set1_epi32(0x0000ff00);
+  for (; data + 4 < data_end; data += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)data);
+    const __m128i in_00g0 = _mm_and_si128(in, mask);     // 00g0|00g0|...
+    const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8);  // 0g00|0g00|...
+    const __m128i in_000g = _mm_srli_epi32(in_00g0, 8);  // 000g|000g|...
+    const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g);
+    const __m128i out = _mm_add_epi8(in, in_0g0g);
+    _mm_storeu_si128((__m128i*)data, out);
+  }
+  // fallthrough and finish off with plain-C
+  while (data < data_end) {
+    const uint32_t argb = *data;
+    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;
+  }
+}
+
+extern void VP8LDspInitSSE2(void);
+
+void VP8LDspInitSSE2(void) {
+  VP8LClampedAddSubtractFull = ClampedAddSubtractFullSSE2;
+  VP8LClampedAddSubtractHalf = ClampedAddSubtractHalfSSE2;
+  VP8LSelect = SelectSSE2;
+  VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRedSSE2;
+  VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRedSSE2;
+}
+#endif
+//------------------------------------------------------------------------------
+
+VP8LPredClampedAddSubFunc VP8LClampedAddSubtractFull;
+VP8LPredClampedAddSubFunc VP8LClampedAddSubtractHalf;
+VP8LPredSelectFunc VP8LSelect;
+VP8LSubtractGreenFromBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
+VP8LAddGreenToBlueAndRedFunc VP8LAddGreenToBlueAndRed;
+
+void VP8LDspInit(void) {
+  VP8LClampedAddSubtractFull = ClampedAddSubtractFull;
+  VP8LClampedAddSubtractHalf = ClampedAddSubtractHalf;
+  VP8LSelect = Select;
+  VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
+  VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
+
+  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
+  if (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      VP8LDspInitSSE2();
+    }
+#endif
+  }
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/dsp/lossless.h b/drivers/webp/dsp/lossless.h
new file mode 100644
index 0000000000..0f1d44200b
--- /dev/null
+++ b/drivers/webp/dsp/lossless.h
@@ -0,0 +1,220 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Image transforms and color space conversion methods for lossless decoder.
+//
+// Authors: Vikas Arora (vikaas.arora@gmail.com)
+//          Jyrki Alakuijala (jyrki@google.com)
+
+#ifndef WEBP_DSP_LOSSLESS_H_
+#define WEBP_DSP_LOSSLESS_H_
+
+#include "../webp/types.h"
+#include "../webp/decode.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+//
+
+typedef uint32_t (*VP8LPredClampedAddSubFunc)(uint32_t c0, uint32_t c1,
+                                              uint32_t c2);
+typedef uint32_t (*VP8LPredSelectFunc)(uint32_t c0, uint32_t c1, uint32_t c2);
+typedef void (*VP8LSubtractGreenFromBlueAndRedFunc)(uint32_t* argb_data,
+                                                    int num_pixs);
+typedef void (*VP8LAddGreenToBlueAndRedFunc)(uint32_t* data_start,
+                                             const uint32_t* data_end);
+
+extern VP8LPredClampedAddSubFunc VP8LClampedAddSubtractFull;
+extern VP8LPredClampedAddSubFunc VP8LClampedAddSubtractHalf;
+extern VP8LPredSelectFunc VP8LSelect;
+extern VP8LSubtractGreenFromBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
+extern VP8LAddGreenToBlueAndRedFunc VP8LAddGreenToBlueAndRed;
+
+// Must be called before calling any of the above methods.
+void VP8LDspInit(void);
+
+//------------------------------------------------------------------------------
+// Image transforms.
+
+struct VP8LTransform;  // Defined in dec/vp8li.h.
+
+// Performs inverse transform of data given transform information, start and end
+// rows. Transform will be applied to rows [row_start, row_end[.
+// The *in and *out pointers refer to source and destination data respectively
+// corresponding to the intermediate row (row_start).
+void VP8LInverseTransform(const struct VP8LTransform* const transform,
+                          int row_start, int row_end,
+                          const uint32_t* const in, uint32_t* const out);
+
+// 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);
+
+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.
+
+// 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);
+
+//------------------------------------------------------------------------------
+// Misc methods.
+
+// Computes sampled size of 'size' when sampling using 'sampling bits'.
+static WEBP_INLINE uint32_t VP8LSubSampleSize(uint32_t size,
+                                              uint32_t sampling_bits) {
+  return (size + (1 << sampling_bits) - 1) >> sampling_bits;
+}
+
+// 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];
+float VP8LFastLog2Slow(int v);
+float VP8LFastSLog2Slow(int v);
+static WEBP_INLINE float VP8LFastLog2(int 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 (v < LOG_LOOKUP_IDX_MAX) ? kSLog2Table[v] : VP8LFastSLog2Slow(v);
+}
+
+// -----------------------------------------------------------------------------
+// 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 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
+// Returns (int)floor(log2(n)). n must be > 0.
+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
+
+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) {
+  const uint32_t alpha_and_green =
+      0x00ff00ffu + (a & 0xff00ff00u) - (b & 0xff00ff00u);
+  const uint32_t red_and_blue =
+      0xff00ff00u + (a & 0x00ff00ffu) - (b & 0x00ff00ffu);
+  return (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu);
+}
+
+void VP8LBundleColorMap(const uint8_t* const row, int width,
+                        int xbits, uint32_t* const dst);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // WEBP_DSP_LOSSLESS_H_
diff --git a/drivers/webp/dsp/upsampling.c b/drivers/webp/dsp/upsampling.c
new file mode 100644
index 0000000000..978e3ce250
--- /dev/null
+++ b/drivers/webp/dsp/upsampling.c
@@ -0,0 +1,366 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// YUV to RGB upsampling functions.
+//
+// Author: somnath@google.com (Somnath Banerjee)
+
+#include "./dsp.h"
+#include "./yuv.h"
+
+#include <assert.h>
+
+//------------------------------------------------------------------------------
+// Fancy upsampler
+
+#ifdef FANCY_UPSAMPLING
+
+// Fancy upsampling functions to convert YUV to RGB
+WebPUpsampleLinePairFunc WebPUpsamplers[MODE_LAST];
+
+// Given samples laid out in a square as:
+//  [a b]
+//  [c d]
+// we interpolate u/v as:
+//  ([9*a + 3*b + 3*c +   d    3*a + 9*b + 3*c +   d] + [8 8]) / 16
+//  ([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 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 x;                                                                       \
+  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 */            \
+  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 != NULL) {                                                      \
+    const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2;                \
+    FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst);                    \
+  }                                                                            \
+  for (x = 1; x <= last_pixel_pair; ++x) {                                     \
+    const uint32_t t_uv = LOAD_UV(top_u[x], top_v[x]);  /* top sample */       \
+    const uint32_t uv   = LOAD_UV(cur_u[x], cur_v[x]);  /* sample */           \
+    /* precompute invariant values associated with first and second diagonals*/\
+    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;                    \
+    {                                                                          \
+      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),                          \
+           top_dst + (2 * x - 1) * XSTEP);                                     \
+      FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16),                          \
+           top_dst + (2 * x - 0) * XSTEP);                                     \
+    }                                                                          \
+    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),                       \
+           bottom_dst + (2 * x - 1) * XSTEP);                                  \
+      FUNC(bottom_y[2 * x + 0], uv1 & 0xff, (uv1 >> 16),                       \
+           bottom_dst + (2 * x + 0) * XSTEP);                                  \
+    }                                                                          \
+    tl_uv = t_uv;                                                              \
+    l_uv = uv;                                                                 \
+  }                                                                            \
+  if (!(len & 1)) {                                                            \
+    {                                                                          \
+      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 != 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);                                    \
+    }                                                                          \
+  }                                                                            \
+}
+
+// All variants implemented.
+UPSAMPLE_FUNC(UpsampleRgbLinePair,  VP8YuvToRgb,  3)
+UPSAMPLE_FUNC(UpsampleBgrLinePair,  VP8YuvToBgr,  3)
+UPSAMPLE_FUNC(UpsampleRgbaLinePair, VP8YuvToRgba, 4)
+UPSAMPLE_FUNC(UpsampleBgraLinePair, VP8YuvToBgra, 4)
+UPSAMPLE_FUNC(UpsampleArgbLinePair, VP8YuvToArgb, 4)
+UPSAMPLE_FUNC(UpsampleRgba4444LinePair, VP8YuvToRgba4444, 2)
+UPSAMPLE_FUNC(UpsampleRgb565LinePair,  VP8YuvToRgb565,  2)
+
+#undef LOAD_UV
+#undef UPSAMPLE_FUNC
+
+#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)                                      \
+static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y,              \
+                      const uint8_t* top_u, const uint8_t* top_v,              \
+                      const uint8_t* bot_u, const uint8_t* bot_v,              \
+                      uint8_t* top_dst, uint8_t* bot_dst, int len) {           \
+  const int half_len = len >> 1;                                               \
+  int x;                                                                       \
+  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);         \
+    }                                                                          \
+    if (len & 1) FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x);  \
+  }                                                                            \
+  if (bot_dst != NULL) {                                                       \
+    for (x = 0; x < half_len; ++x) {                                           \
+      FUNC(bot_y[2 * x + 0], bot_u[x], bot_v[x], bot_dst + 8 * x + 0);         \
+      FUNC(bot_y[2 * x + 1], bot_u[x], bot_v[x], bot_dst + 8 * x + 4);         \
+    }                                                                          \
+    if (len & 1) FUNC(bot_y[2 * x + 0], bot_u[x], bot_v[x], bot_dst + 8 * x);  \
+  }                                                                            \
+}
+
+DUAL_SAMPLE_FUNC(DualLineSamplerBGRA, VP8YuvToBgra)
+DUAL_SAMPLE_FUNC(DualLineSamplerARGB, VP8YuvToArgb)
+#undef DUAL_SAMPLE_FUNC
+
+#endif  // !FANCY_UPSAMPLING
+
+WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last) {
+  WebPInitUpsamplers();
+  VP8YUVInit();
+#ifdef FANCY_UPSAMPLING
+  return WebPUpsamplers[alpha_is_last ? MODE_BGRA : MODE_ARGB];
+#else
+  return (alpha_is_last ? DualLineSamplerBGRA : DualLineSamplerARGB);
+#endif
+}
+
+//------------------------------------------------------------------------------
+// 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) {                                 \
+  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)
+
+#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
+
+#define MULTIPLIER(a)  ((a) * 0x1111)    // 0x1111 ~= (1 << 16) / 15
+
+static WEBP_INLINE uint8_t dither_hi(uint8_t x) {
+  return (x & 0xf0) | (x >> 4);
+}
+
+static WEBP_INLINE uint8_t dither_lo(uint8_t x) {
+  return (x & 0x0f) | (x << 4);
+}
+
+static WEBP_INLINE uint8_t multiply(uint8_t x, uint32_t m) {
+  return (x * m) >> 16;
+}
+
+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;
+    }
+    rgba4444 += stride;
+  }
+}
+#undef MULTIPLIER
+
+void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int)
+    = ApplyAlphaMultiply;
+void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int)
+    = ApplyAlphaMultiply4444;
+
+//------------------------------------------------------------------------------
+// Main call
+
+void WebPInitUpsamplers(void) {
+#ifdef FANCY_UPSAMPLING
+  WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair;
+  WebPUpsamplers[MODE_RGBA]      = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_BGR]       = UpsampleBgrLinePair;
+  WebPUpsamplers[MODE_BGRA]      = UpsampleBgraLinePair;
+  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
+#if defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      WebPInitUpsamplersNEON();
+    }
+#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 (VP8GetCPUInfo != NULL) {
+#if defined(WEBP_USE_SSE2)
+    if (VP8GetCPUInfo(kSSE2)) {
+      WebPInitPremultiplySSE2();
+    }
+#endif
+#if defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      WebPInitPremultiplyNEON();
+    }
+#endif
+  }
+#endif  // FANCY_UPSAMPLING
+}
+
diff --git a/drivers/webp/dsp/upsampling_neon.c b/drivers/webp/dsp/upsampling_neon.c
new file mode 100644
index 0000000000..791222f81e
--- /dev/null
+++ b/drivers/webp/dsp/upsampling_neon.c
@@ -0,0 +1,265 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// NEON version of YUV to RGB upsampling functions.
+//
+// Author: mans@mansr.com (Mans Rullgard)
+// Based on SSE code by: somnath@google.com (Somnath Banerjee)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_NEON)
+
+#include <assert.h>
+#include <arm_neon.h>
+#include <string.h>
+#include "./yuv.h"
+
+#ifdef FANCY_UPSAMPLING
+
+//-----------------------------------------------------------------------------
+// U/V upsampling
+
+// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels.
+#define UPSAMPLE_16PIXELS(r1, r2, out) {                                \
+  uint8x8_t a = vld1_u8(r1);                                            \
+  uint8x8_t b = vld1_u8(r1 + 1);                                        \
+  uint8x8_t c = vld1_u8(r2);                                            \
+  uint8x8_t d = vld1_u8(r2 + 1);                                        \
+                                                                        \
+  uint16x8_t al = vshll_n_u8(a, 1);                                     \
+  uint16x8_t bl = vshll_n_u8(b, 1);                                     \
+  uint16x8_t cl = vshll_n_u8(c, 1);                                     \
+  uint16x8_t dl = vshll_n_u8(d, 1);                                     \
+                                                                        \
+  uint8x8_t diag1, diag2;                                               \
+  uint16x8_t sl;                                                        \
+                                                                        \
+  /* a + b + c + d */                                                   \
+  sl = vaddl_u8(a,  b);                                                 \
+  sl = vaddw_u8(sl, c);                                                 \
+  sl = vaddw_u8(sl, d);                                                 \
+                                                                        \
+  al = vaddq_u16(sl, al); /* 3a +  b +  c +  d */                       \
+  bl = vaddq_u16(sl, bl); /*  a + 3b +  c +  d */                       \
+                                                                        \
+  al = vaddq_u16(al, dl); /* 3a +  b +  c + 3d */                       \
+  bl = vaddq_u16(bl, cl); /*  a + 3b + 3c +  d */                       \
+                                                                        \
+  diag2 = vshrn_n_u16(al, 3);                                           \
+  diag1 = vshrn_n_u16(bl, 3);                                           \
+                                                                        \
+  a = vrhadd_u8(a, diag1);                                              \
+  b = vrhadd_u8(b, diag2);                                              \
+  c = vrhadd_u8(c, diag2);                                              \
+  d = vrhadd_u8(d, diag1);                                              \
+                                                                        \
+  {                                                                     \
+    const uint8x8x2_t a_b = {{ a, b }};                                 \
+    const uint8x8x2_t c_d = {{ c, d }};                                 \
+    vst2_u8(out,      a_b);                                             \
+    vst2_u8(out + 32, c_d);                                             \
+  }                                                                     \
+}
+
+// Turn the macro into a function for reducing code-size when non-critical
+static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,
+                             uint8_t *out) {
+  UPSAMPLE_16PIXELS(r1, r2, out);
+}
+
+#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) {                  \
+  uint8_t r1[9], r2[9];                                                 \
+  memcpy(r1, (tb), (num_pixels));                                       \
+  memcpy(r2, (bb), (num_pixels));                                       \
+  /* replicate last byte */                                             \
+  memset(r1 + (num_pixels), r1[(num_pixels) - 1], 9 - (num_pixels));    \
+  memset(r2 + (num_pixels), r2[(num_pixels) - 1], 9 - (num_pixels));    \
+  Upsample16Pixels(r1, r2, out);                                        \
+}
+
+//-----------------------------------------------------------------------------
+// YUV->RGB conversion
+
+static const int16_t kCoeffs[4] = { kYScale, kVToR, kUToG, kVToG };
+
+#define v255 vmov_n_u8(255)
+
+#define STORE_Rgb(out, r, g, b) do {                                    \
+  const uint8x8x3_t r_g_b = {{ r, g, b }};                              \
+  vst3_u8(out, r_g_b);                                                  \
+} while (0)
+
+#define STORE_Bgr(out, r, g, b) do {                                    \
+  const uint8x8x3_t b_g_r = {{ b, g, r }};                              \
+  vst3_u8(out, b_g_r);                                                  \
+} while (0)
+
+#define STORE_Rgba(out, r, g, b) do {                                   \
+  const uint8x8x4_t r_g_b_v255 = {{ r, g, b, v255 }};                   \
+  vst4_u8(out, r_g_b_v255);                                             \
+} while (0)
+
+#define STORE_Bgra(out, r, g, b) do {                                   \
+  const uint8x8x4_t b_g_r_v255 = {{ b, g, r, v255 }};                   \
+  vst4_u8(out, b_g_r_v255);                                             \
+} while (0)
+
+#define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) {            \
+  int i;                                                                \
+  for (i = 0; i < N; i += 8) {                                          \
+    const int off = ((cur_x) + i) * XSTEP;                              \
+    uint8x8_t y  = vld1_u8((src_y) + (cur_x)  + i);                     \
+    uint8x8_t u  = vld1_u8((src_uv) + i);                               \
+    uint8x8_t v  = vld1_u8((src_uv) + i + 16);                          \
+    const int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16));       \
+    const int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128));      \
+    const int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128));      \
+    int32x4_t yl = vmull_lane_s16(vget_low_s16(yy),  cf16, 0);          \
+    int32x4_t yh = vmull_lane_s16(vget_high_s16(yy), cf16, 0);          \
+    const int32x4_t rl = vmlal_lane_s16(yl, vget_low_s16(vv),  cf16, 1);\
+    const int32x4_t rh = vmlal_lane_s16(yh, vget_high_s16(vv), cf16, 1);\
+    int32x4_t gl = vmlsl_lane_s16(yl, vget_low_s16(uu),  cf16, 2);      \
+    int32x4_t gh = vmlsl_lane_s16(yh, vget_high_s16(uu), cf16, 2);      \
+    const int32x4_t bl = vmovl_s16(vget_low_s16(uu));                   \
+    const int32x4_t bh = vmovl_s16(vget_high_s16(uu));                  \
+    gl = vmlsl_lane_s16(gl, vget_low_s16(vv),  cf16, 3);                \
+    gh = vmlsl_lane_s16(gh, vget_high_s16(vv), cf16, 3);                \
+    yl = vmlaq_lane_s32(yl, bl, cf32, 0);                               \
+    yh = vmlaq_lane_s32(yh, bh, cf32, 0);                               \
+    /* vrshrn_n_s32() already incorporates the rounding constant */     \
+    y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, YUV_FIX2),            \
+                                 vrshrn_n_s32(rh, YUV_FIX2)));          \
+    u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, YUV_FIX2),            \
+                                 vrshrn_n_s32(gh, YUV_FIX2)));          \
+    v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(yl, YUV_FIX2),            \
+                                 vrshrn_n_s32(yh, YUV_FIX2)));          \
+    STORE_ ## FMT(out + off, y, u, v);                                  \
+  }                                                                     \
+}
+
+#define CONVERT1(FUNC, XSTEP, N, src_y, src_uv, rgb, cur_x) {           \
+  int i;                                                                \
+  for (i = 0; i < N; i++) {                                             \
+    const int off = ((cur_x) + i) * XSTEP;                              \
+    const int y = src_y[(cur_x) + i];                                   \
+    const int u = (src_uv)[i];                                          \
+    const int v = (src_uv)[i + 16];                                     \
+    FUNC(y, u, v, rgb + off);                                           \
+  }                                                                     \
+}
+
+#define CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, uv,                  \
+                      top_dst, bottom_dst, cur_x, len) {                \
+  CONVERT8(FMT, XSTEP, len, top_y, uv, top_dst, cur_x)                  \
+  if (bottom_y != NULL) {                                               \
+    CONVERT8(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x)   \
+  }                                                                     \
+}
+
+#define CONVERT2RGB_1(FUNC, XSTEP, top_y, bottom_y, uv,                 \
+                      top_dst, bottom_dst, cur_x, len) {                \
+  CONVERT1(FUNC, XSTEP, len, top_y, uv, top_dst, cur_x);                \
+  if (bottom_y != NULL) {                                               \
+    CONVERT1(FUNC, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x); \
+  }                                                                     \
+}
+
+#define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, 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 block;                                                            \
+  /* 16 byte aligned array to cache reconstructed u and v */            \
+  uint8_t uv_buf[2 * 32 + 15];                                          \
+  uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);     \
+  const int uv_len = (len + 1) >> 1;                                    \
+  /* 9 pixels must be read-able for each block */                       \
+  const int num_blocks = (uv_len - 1) >> 3;                             \
+  const int leftover = uv_len - num_blocks * 8;                         \
+  const int last_pos = 1 + 16 * 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;                  \
+                                                                        \
+  const int16x4_t cf16 = vld1_s16(kCoeffs);                             \
+  const int32x2_t cf32 = vmov_n_s32(kUToB);                             \
+  const uint8x8_t u16  = vmov_n_u8(16);                                 \
+  const uint8x8_t u128 = vmov_n_u8(128);                                \
+                                                                        \
+  /* Treat the first pixel in regular way */                            \
+  assert(top_y != NULL);                                                \
+  {                                                                     \
+    const int u0 = (top_u[0] + u_diag) >> 1;                            \
+    const int v0 = (top_v[0] + v_diag) >> 1;                            \
+    VP8YuvTo ## FMT(top_y[0], u0, v0, top_dst);                         \
+  }                                                                     \
+  if (bottom_y != NULL) {                                               \
+    const int u0 = (cur_u[0] + u_diag) >> 1;                            \
+    const int v0 = (cur_v[0] + v_diag) >> 1;                            \
+    VP8YuvTo ## FMT(bottom_y[0], u0, v0, bottom_dst);                   \
+  }                                                                     \
+                                                                        \
+  for (block = 0; block < num_blocks; ++block) {                        \
+    UPSAMPLE_16PIXELS(top_u, cur_u, r_uv);                              \
+    UPSAMPLE_16PIXELS(top_v, cur_v, r_uv + 16);                         \
+    CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, r_uv,                    \
+                  top_dst, bottom_dst, 16 * block + 1, 16);             \
+    top_u += 8;                                                         \
+    cur_u += 8;                                                         \
+    top_v += 8;                                                         \
+    cur_v += 8;                                                         \
+  }                                                                     \
+                                                                        \
+  UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv);                    \
+  UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 16);               \
+  CONVERT2RGB_1(VP8YuvTo ## FMT, XSTEP, top_y, bottom_y, r_uv,          \
+                top_dst, bottom_dst, last_pos, len - last_pos);         \
+}
+
+// NEON variants of the fancy upsampler.
+NEON_UPSAMPLE_FUNC(UpsampleRgbLinePairNEON,  Rgb,  3)
+NEON_UPSAMPLE_FUNC(UpsampleBgrLinePairNEON,  Bgr,  3)
+NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePairNEON, Rgba, 4)
+NEON_UPSAMPLE_FUNC(UpsampleBgraLinePairNEON, Bgra, 4)
+
+#endif  // FANCY_UPSAMPLING
+
+#endif   // WEBP_USE_NEON
+
+//------------------------------------------------------------------------------
+
+#ifdef FANCY_UPSAMPLING
+
+extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
+
+void WebPInitUpsamplersNEON(void) {
+#if defined(WEBP_USE_NEON)
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePairNEON;
+  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairNEON;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePairNEON;
+  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairNEON;
+#endif   // WEBP_USE_NEON
+}
+
+void WebPInitPremultiplyNEON(void) {
+#if defined(WEBP_USE_NEON)
+  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairNEON;
+  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairNEON;
+#endif   // WEBP_USE_NEON
+}
+
+#else
+
+// this empty function is to avoid an empty .o
+void WebPInitPremultiplyNEON(void) {}
+
+#endif  // FANCY_UPSAMPLING
+
diff --git a/drivers/webp/dsp/upsampling_sse2.c b/drivers/webp/dsp/upsampling_sse2.c
new file mode 100644
index 0000000000..0db0798c6d
--- /dev/null
+++ b/drivers/webp/dsp/upsampling_sse2.c
@@ -0,0 +1,218 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// SSE2 version of YUV to RGB upsampling functions.
+//
+// Author: somnath@google.com (Somnath Banerjee)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_SSE2)
+
+#include <assert.h>
+#include <emmintrin.h>
+#include <string.h>
+#include "./yuv.h"
+
+#ifdef FANCY_UPSAMPLING
+
+// We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows
+// u = (9*a + 3*b + 3*c + d + 8) / 16
+//   = (a + (a + 3*b + 3*c + d) / 8 + 1) / 2
+//   = (a + m + 1) / 2
+// where m = (a + 3*b + 3*c + d) / 8
+//         = ((a + b + c + d) / 2 + b + c) / 4
+//
+// Let's say  k = (a + b + c + d) / 4.
+// We can compute k as
+// k = (s + t + 1) / 2 - ((a^d) | (b^c) | (s^t)) & 1
+// where s = (a + d + 1) / 2 and t = (b + c + 1) / 2
+//
+// Then m can be written as
+// m = (k + t + 1) / 2 - (((b^c) & (s^t)) | (k^t)) & 1
+
+// Computes out = (k + in + 1) / 2 - ((ij & (s^t)) | (k^in)) & 1
+#define GET_M(ij, in, out) do {                                                \
+  const __m128i tmp0 = _mm_avg_epu8(k, (in));     /* (k + in + 1) / 2 */       \
+  const __m128i tmp1 = _mm_and_si128((ij), st);   /* (ij) & (s^t) */           \
+  const __m128i tmp2 = _mm_xor_si128(k, (in));    /* (k^in) */                 \
+  const __m128i tmp3 = _mm_or_si128(tmp1, tmp2);  /* ((ij) & (s^t)) | (k^in) */\
+  const __m128i tmp4 = _mm_and_si128(tmp3, one);  /* & 1 -> lsb_correction */  \
+  (out) = _mm_sub_epi8(tmp0, tmp4);    /* (k + in + 1) / 2 - lsb_correction */ \
+} while (0)
+
+// pack and store two alternating pixel rows
+#define PACK_AND_STORE(a, b, da, db, out) do {                                 \
+  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 s = _mm_avg_epu8(a, d);        /* s = (a + d + 1) / 2 */       \
+  const __m128i t = _mm_avg_epu8(b, c);        /* t = (b + c + 1) / 2 */       \
+  const __m128i st = _mm_xor_si128(s, t);      /* st = s^t */                  \
+                                                                               \
+  const __m128i ad = _mm_xor_si128(a, d);      /* ad = a^d */                  \
+  const __m128i bc = _mm_xor_si128(b, c);      /* bc = b^c */                  \
+                                                                               \
+  const __m128i t1 = _mm_or_si128(ad, bc);     /* (a^d) | (b^c) */             \
+  const __m128i t2 = _mm_or_si128(t1, st);     /* (a^d) | (b^c) | (s^t) */     \
+  const __m128i t3 = _mm_and_si128(t2, one);   /* (a^d) | (b^c) | (s^t) & 1 */ \
+  const __m128i t4 = _mm_avg_epu8(s, t);                                       \
+  const __m128i k = _mm_sub_epi8(t4, t3);      /* k = (a + b + c + d) / 4 */   \
+  __m128i diag1, diag2;                                                        \
+                                                                               \
+  GET_M(bc, t, diag1);                  /* diag1 = (a + 3b + 3c + d) / 8 */    \
+  GET_M(ad, s, diag2);                  /* diag2 = (3a + b + c + 3d) / 8 */    \
+                                                                               \
+  /* pack the alternate pixels */                                              \
+  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
+static void Upsample32Pixels(const uint8_t r1[], const uint8_t r2[],
+                             uint8_t* const out) {
+  UPSAMPLE_32PIXELS(r1, r2, out);
+}
+
+#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) {                         \
+  uint8_t r1[17], r2[17];                                                      \
+  memcpy(r1, (tb), (num_pixels));                                              \
+  memcpy(r2, (bb), (num_pixels));                                              \
+  /* replicate last byte */                                                    \
+  memset(r1 + (num_pixels), r1[(num_pixels) - 1], 17 - (num_pixels));          \
+  memset(r2 + (num_pixels), r2[(num_pixels) - 1], 17 - (num_pixels));          \
+  /* using the shared function instead of the macro saves ~3k code size */     \
+  Upsample32Pixels(r1, r2, out);                                               \
+}
+
+#define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y,                              \
+                    top_dst, bottom_dst, cur_x, num_pixels) {                  \
+  int n;                                                                       \
+  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 != NULL) {                                                      \
+    for (n = 0; n < (num_pixels); ++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 uv_pos, pos;                                                             \
+  /* 16byte-aligned array to cache reconstructed u and v */                    \
+  uint8_t uv_buf[4 * 32 + 15];                                                 \
+  uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);             \
+  uint8_t* const r_v = r_u + 32;                                               \
+                                                                               \
+  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);                               \
+    }                                                                          \
+  }                                                                            \
+  /* 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);    \
+  }                                                                            \
+  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);                                               \
+  }                                                                            \
+}
+
+// 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)
+
+#undef GET_M
+#undef PACK_AND_STORE
+#undef UPSAMPLE_32PIXELS
+#undef UPSAMPLE_LAST_BLOCK
+#undef CONVERT2RGB
+#undef CONVERT2RGB_32
+#undef SSE2_UPSAMPLE_FUNC
+
+#endif  // FANCY_UPSAMPLING
+
+#endif   // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+
+#ifdef FANCY_UPSAMPLING
+
+extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
+
+void WebPInitUpsamplersSSE2(void) {
+#if defined(WEBP_USE_SSE2)
+  VP8YUVInitSSE2();
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePairSSE2;
+  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePairSSE2;
+  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2;
+#endif   // WEBP_USE_SSE2
+}
+
+void WebPInitPremultiplySSE2(void) {
+#if defined(WEBP_USE_SSE2)
+  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2;
+  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2;
+#endif   // WEBP_USE_SSE2
+}
+
+#else
+
+// this empty function is to avoid an empty .o
+void WebPInitPremultiplySSE2(void) {}
+
+#endif  // FANCY_UPSAMPLING
+
diff --git a/drivers/webp/dsp/yuv.c b/drivers/webp/dsp/yuv.c
new file mode 100644
index 0000000000..4f9cafc104
--- /dev/null
+++ b/drivers/webp/dsp/yuv.c
@@ -0,0 +1,207 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// YUV->RGB conversion function
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./yuv.h"
+
+
+#if defined(WEBP_YUV_USE_TABLE)
+
+static int done = 0;
+
+static WEBP_INLINE uint8_t clip(int v, int max_value) {
+  return v < 0 ? 0 : v > max_value ? max_value : v;
+}
+
+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];
+
+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;
+    VP8kVToG[i] = -45773 * (i - 128);
+    VP8kUToB[i] = (113618 * (i - 128) + YUV_HALF) >> YUV_FIX;
+  }
+  for (i = YUV_RANGE_MIN; i < YUV_RANGE_MAX; ++i) {
+    const int k = ((i - 16) * 76283 + YUV_HALF) >> YUV_FIX;
+    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;
+}
+
+#else
+
+void VP8YUVInit(void) {}
+
+#endif  // WEBP_YUV_USE_TABLE
+
+//-----------------------------------------------------------------------------
+// SSE2 extras
+
+#if defined(WEBP_USE_SSE2)
+
+#ifdef FANCY_UPSAMPLING
+
+#include <emmintrin.h>
+#include <string.h>   // for memcpy
+
+typedef union {   // handy struct for converting SSE2 registers
+  int32_t i32[4];
+  uint8_t u8[16];
+  __m128i m;
+} VP8kCstSSE2;
+
+static int done_sse2 = 0;
+static VP8kCstSSE2 VP8kUtoRGBA[256], VP8kVtoRGBA[256], VP8kYtoRGBA[256];
+
+void VP8YUVInitSSE2(void) {
+  if (!done_sse2) {
+    int i;
+    for (i = 0; i < 256; ++i) {
+      VP8kYtoRGBA[i].i32[0] =
+        VP8kYtoRGBA[i].i32[1] =
+        VP8kYtoRGBA[i].i32[2] = (i - 16) * kYScale + YUV_HALF2;
+      VP8kYtoRGBA[i].i32[3] = 0xff << YUV_FIX2;
+
+      VP8kUtoRGBA[i].i32[0] = 0;
+      VP8kUtoRGBA[i].i32[1] = -kUToG * (i - 128);
+      VP8kUtoRGBA[i].i32[2] =  kUToB * (i - 128);
+      VP8kUtoRGBA[i].i32[3] = 0;
+
+      VP8kVtoRGBA[i].i32[0] =  kVToR * (i - 128);
+      VP8kVtoRGBA[i].i32[1] = -kVToG * (i - 128);
+      VP8kVtoRGBA[i].i32[2] = 0;
+      VP8kVtoRGBA[i].i32[3] = 0;
+    }
+    done_sse2 = 1;
+  }
+}
+
+static WEBP_INLINE __m128i VP8GetRGBA32b(int y, int u, int v) {
+  const __m128i u_part = _mm_loadu_si128(&VP8kUtoRGBA[u].m);
+  const __m128i v_part = _mm_loadu_si128(&VP8kVtoRGBA[v].m);
+  const __m128i y_part = _mm_loadu_si128(&VP8kYtoRGBA[y].m);
+  const __m128i uv_part = _mm_add_epi32(u_part, v_part);
+  const __m128i rgba1 = _mm_add_epi32(y_part, uv_part);
+  const __m128i rgba2 = _mm_srai_epi32(rgba1, YUV_FIX2);
+  return rgba2;
+}
+
+static WEBP_INLINE void VP8YuvToRgbSSE2(uint8_t y, uint8_t u, uint8_t v,
+                                        uint8_t* const rgb) {
+  const __m128i tmp0 = VP8GetRGBA32b(y, u, v);
+  const __m128i tmp1 = _mm_packs_epi32(tmp0, tmp0);
+  const __m128i tmp2 = _mm_packus_epi16(tmp1, tmp1);
+  // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
+  _mm_storel_epi64((__m128i*)rgb, tmp2);
+}
+
+static WEBP_INLINE void VP8YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
+                                        uint8_t* const bgr) {
+  const __m128i tmp0 = VP8GetRGBA32b(y, u, v);
+  const __m128i tmp1 = _mm_shuffle_epi32(tmp0, _MM_SHUFFLE(3, 0, 1, 2));
+  const __m128i tmp2 = _mm_packs_epi32(tmp1, tmp1);
+  const __m128i tmp3 = _mm_packus_epi16(tmp2, tmp2);
+  // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
+  _mm_storel_epi64((__m128i*)bgr, tmp3);
+}
+
+void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                    uint8_t* dst) {
+  int n;
+  for (n = 0; n < 32; n += 4) {
+    const __m128i tmp0_1 = VP8GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
+    const __m128i tmp0_2 = VP8GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
+    const __m128i tmp0_3 = VP8GetRGBA32b(y[n + 2], u[n + 2], v[n + 2]);
+    const __m128i tmp0_4 = VP8GetRGBA32b(y[n + 3], u[n + 3], v[n + 3]);
+    const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
+    const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
+    const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
+    _mm_storeu_si128((__m128i*)dst, tmp2);
+    dst += 4 * 4;
+  }
+}
+
+void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                    uint8_t* dst) {
+  int n;
+  for (n = 0; n < 32; n += 2) {
+    const __m128i tmp0_1 = VP8GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
+    const __m128i tmp0_2 = VP8GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
+    const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
+    const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
+    const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
+    const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
+    _mm_storel_epi64((__m128i*)dst, tmp3);
+    dst += 4 * 2;
+  }
+}
+
+void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                   uint8_t* dst) {
+  int n;
+  uint8_t tmp0[2 * 3 + 5 + 15];
+  uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15);  // align
+  for (n = 0; n < 30; ++n) {   // we directly stomp the *dst memory
+    VP8YuvToRgbSSE2(y[n], u[n], v[n], dst + n * 3);
+  }
+  // Last two pixels are special: we write in a tmp buffer before sending
+  // to dst.
+  VP8YuvToRgbSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
+  VP8YuvToRgbSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
+  memcpy(dst + n * 3, tmp, 2 * 3);
+}
+
+void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                   uint8_t* dst) {
+  int n;
+  uint8_t tmp0[2 * 3 + 5 + 15];
+  uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15);  // align
+  for (n = 0; n < 30; ++n) {
+    VP8YuvToBgrSSE2(y[n], u[n], v[n], dst + n * 3);
+  }
+  VP8YuvToBgrSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
+  VP8YuvToBgrSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
+  memcpy(dst + n * 3, tmp, 2 * 3);
+}
+
+#else
+
+void VP8YUVInitSSE2(void) {}
+
+#endif  // FANCY_UPSAMPLING
+
+#endif  // WEBP_USE_SSE2
+
diff --git a/drivers/webp/dsp/yuv.h b/drivers/webp/dsp/yuv.h
new file mode 100644
index 0000000000..dd778f9cbe
--- /dev/null
+++ b/drivers/webp/dsp/yuv.h
@@ -0,0 +1,317 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// 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
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+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(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];
+  rgb[0] = VP8kClip[y + r_off - YUV_RANGE_MIN];
+  rgb[1] = VP8kClip[y + g_off - YUV_RANGE_MIN];
+  rgb[2] = VP8kClip[y + b_off - YUV_RANGE_MIN];
+}
+
+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];
+  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 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];
+  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 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];
+  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,
+                                     uint8_t* const bgra) {
+  VP8YuvToBgr(y, u, v, bgra);
+  bgra[3] = 0xff;
+}
+
+static WEBP_INLINE void VP8YuvToRgba(uint8_t y, uint8_t u, uint8_t v,
+                                     uint8_t* const rgba) {
+  VP8YuvToRgb(y, u, v, rgba);
+  rgba[3] = 0xff;
+}
+
+// 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)
+
+#if defined(FANCY_UPSAMPLING)
+// 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);
+#endif  // FANCY_UPSAMPLING
+
+// Must be called to initialize tables before using the functions.
+void VP8YUVInitSSE2(void);
+
+#endif    // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// RGB -> YUV conversion
+
+// 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;
+}
+
+#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 + rounding + (16 << YUV_FIX)) >> YUV_FIX;  // no need to clip
+}
+
+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, 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
+}
+
+static WEBP_INLINE int VP8_RGB_TO_U(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 VP8_RGB_TO_V(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
+
+#endif  /* WEBP_DSP_YUV_H_ */
diff --git a/drivers/webp/enc/alpha.c b/drivers/webp/enc/alpha.c
new file mode 100644
index 0000000000..21d4b5cbde
--- /dev/null
+++ b/drivers/webp/enc/alpha.c
@@ -0,0 +1,410 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Alpha-plane compression.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include "./vp8enci.h"
+#include "../utils/filters.h"
+#include "../utils/quant_levels.h"
+#include "../webp/format_constants.h"
+
+// -----------------------------------------------------------------------------
+// Encodes the given alpha data via specified compression method 'method'.
+// The pre-processing (quantization) is performed if 'quality' is less than 100.
+// For such cases, the encoding is lossy. The valid range is [0, 100] for
+// 'quality' and [0, 1] for 'method':
+//   'method = 0' - No compression;
+//   'method = 1' - Use lossless coder on the alpha plane only
+// 'filter' values [0, 4] correspond to prediction modes none, horizontal,
+// vertical & gradient filters. The prediction mode 4 will try all the
+// prediction modes 0 to 3 and pick the best one.
+// 'effort_level': specifies how much effort must be spent to try and reduce
+//  the compressed output size. In range 0 (quick) to 6 (slow).
+//
+// 'output' corresponds to the buffer containing compressed alpha data.
+//          This buffer is allocated by this method and caller should call
+//          free(*output) when done.
+// 'output_size' corresponds to size of this compressed alpha buffer.
+//
+// Returns 1 on successfully encoding the alpha and
+//         0 if either:
+//           invalid quality or method, or
+//           memory allocation for the compressed data fails.
+
+#include "../enc/vp8li.h"
+
+static int EncodeLossless(const uint8_t* const data, int width, int height,
+                          int effort_level,  // in [0..6] range
+                          VP8BitWriter* const bw,
+                          WebPAuxStats* const stats) {
+  int ok = 0;
+  WebPConfig config;
+  WebPPicture picture;
+  VP8LBitWriter tmp_bw;
+
+  WebPPictureInit(&picture);
+  picture.width = width;
+  picture.height = height;
+  picture.use_argb = 1;
+  picture.stats = stats;
+  if (!WebPPictureAlloc(&picture)) return 0;
+
+  // Transfer the alpha values to the green channel.
+  {
+    int i, j;
+    uint32_t* dst = picture.argb;
+    const uint8_t* src = data;
+    for (j = 0; j < picture.height; ++j) {
+      for (i = 0; i < picture.width; ++i) {
+        dst[i] = src[i] << 8;  // we leave A/R/B channels zero'd.
+      }
+      src += width;
+      dst += picture.argb_stride;
+    }
+  }
+
+  WebPConfigInit(&config);
+  config.lossless = 1;
+  config.method = effort_level;  // impact is very small
+  // Set a low default quality for encoding alpha. Ensure that Alpha quality at
+  // lower methods (3 and below) is less than the threshold for triggering
+  // costly 'BackwardReferencesTraceBackwards'.
+  config.quality = 8.f * effort_level;
+  assert(config.quality >= 0 && config.quality <= 100.f);
+
+  ok = VP8LBitWriterInit(&tmp_bw, (width * height) >> 3);
+  ok = ok && (VP8LEncodeStream(&config, &picture, &tmp_bw) == VP8_ENC_OK);
+  WebPPictureFree(&picture);
+  if (ok) {
+    const uint8_t* const buffer = VP8LBitWriterFinish(&tmp_bw);
+    const size_t buffer_size = VP8LBitWriterNumBytes(&tmp_bw);
+    VP8BitWriterAppend(bw, buffer, buffer_size);
+  }
+  VP8LBitWriterDestroy(&tmp_bw);
+  return ok && !bw->error_;
+}
+
+// -----------------------------------------------------------------------------
+
+// Small struct to hold the result of a filter mode compression attempt.
+typedef struct {
+  size_t score;
+  VP8BitWriter bw;
+  WebPAuxStats stats;
+} FilterTrial;
+
+// This function always returns an initialized 'bw' object, even upon error.
+static int EncodeAlphaInternal(const uint8_t* const data, int width, int height,
+                               int method, int filter, int reduce_levels,
+                               int effort_level,  // in [0..6] range
+                               uint8_t* const tmp_alpha,
+                               FilterTrial* result) {
+  int ok = 0;
+  const uint8_t* alpha_src;
+  WebPFilterFunc filter_func;
+  uint8_t header;
+  size_t expected_size;
+  const size_t data_size = width * height;
+
+  assert((uint64_t)data_size == (uint64_t)width * height);  // as per spec
+  assert(filter >= 0 && filter < WEBP_FILTER_LAST);
+  assert(method >= ALPHA_NO_COMPRESSION);
+  assert(method <= ALPHA_LOSSLESS_COMPRESSION);
+  assert(sizeof(header) == ALPHA_HEADER_LEN);
+  // TODO(skal): have a common function and #define's to validate alpha params.
+
+  expected_size =
+      (method == ALPHA_NO_COMPRESSION) ? (ALPHA_HEADER_LEN + data_size)
+                                       : (data_size >> 5);
+  header = method | (filter << 2);
+  if (reduce_levels) header |= ALPHA_PREPROCESSED_LEVELS << 4;
+
+  VP8BitWriterInit(&result->bw, expected_size);
+  VP8BitWriterAppend(&result->bw, &header, ALPHA_HEADER_LEN);
+
+  filter_func = WebPFilters[filter];
+  if (filter_func != NULL) {
+    filter_func(data, width, height, width, tmp_alpha);
+    alpha_src = tmp_alpha;
+  }  else {
+    alpha_src = data;
+  }
+
+  if (method == ALPHA_NO_COMPRESSION) {
+    ok = VP8BitWriterAppend(&result->bw, alpha_src, width * height);
+    ok = ok && !result->bw.error_;
+  } else {
+    ok = EncodeLossless(alpha_src, width, height, effort_level,
+                        &result->bw, &result->stats);
+    VP8BitWriterFinish(&result->bw);
+  }
+  result->score = VP8BitWriterSize(&result->bw);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+
+// TODO(skal): move to dsp/ ?
+static void CopyPlane(const uint8_t* src, int src_stride,
+                      uint8_t* dst, int dst_stride, int width, int height) {
+  while (height-- > 0) {
+    memcpy(dst, src, width);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static int GetNumColors(const uint8_t* data, int width, int height,
+                        int stride) {
+  int j;
+  int colors = 0;
+  uint8_t color[256] = { 0 };
+
+  for (j = 0; j < height; ++j) {
+    int i;
+    const uint8_t* const p = data + j * stride;
+    for (i = 0; i < width; ++i) {
+      color[p[i]] = 1;
+    }
+  }
+  for (j = 0; j < 256; ++j) {
+    if (color[j] > 0) ++colors;
+  }
+  return colors;
+}
+
+#define FILTER_TRY_NONE (1 << WEBP_FILTER_NONE)
+#define FILTER_TRY_ALL ((1 << WEBP_FILTER_LAST) - 1)
+
+// Given the input 'filter' option, return an OR'd bit-set of filters to try.
+static uint32_t GetFilterMap(const uint8_t* alpha, int width, int height,
+                             int filter, int effort_level) {
+  uint32_t bit_map = 0U;
+  if (filter == WEBP_FILTER_FAST) {
+    // Quick estimate of the best candidate.
+    int try_filter_none = (effort_level > 3);
+    const int kMinColorsForFilterNone = 16;
+    const int kMaxColorsForFilterNone = 192;
+    const int num_colors = GetNumColors(alpha, width, height, width);
+    // For low number of colors, NONE yields better compression.
+    filter = (num_colors <= kMinColorsForFilterNone) ? WEBP_FILTER_NONE :
+             EstimateBestFilter(alpha, width, height, width);
+    bit_map |= 1 << filter;
+    // For large number of colors, try FILTER_NONE in addition to the best
+    // filter as well.
+    if (try_filter_none || num_colors > kMaxColorsForFilterNone) {
+      bit_map |= FILTER_TRY_NONE;
+    }
+  } else if (filter == WEBP_FILTER_NONE) {
+    bit_map = FILTER_TRY_NONE;
+  } else {  // WEBP_FILTER_BEST -> try all
+    bit_map = FILTER_TRY_ALL;
+  }
+  return bit_map;
+}
+
+static void InitFilterTrial(FilterTrial* const score) {
+  score->score = (size_t)~0U;
+  VP8BitWriterInit(&score->bw, 0);
+}
+
+static int ApplyFiltersAndEncode(const uint8_t* alpha, int width, int height,
+                                 size_t data_size, int method, int filter,
+                                 int reduce_levels, int effort_level,
+                                 uint8_t** const output,
+                                 size_t* const output_size,
+                                 WebPAuxStats* const stats) {
+  int ok = 1;
+  FilterTrial best;
+  uint32_t try_map =
+      GetFilterMap(alpha, width, height, filter, effort_level);
+  InitFilterTrial(&best);
+  if (try_map != FILTER_TRY_NONE) {
+    uint8_t* filtered_alpha =  (uint8_t*)malloc(data_size);
+    if (filtered_alpha == NULL) return 0;
+
+    for (filter = WEBP_FILTER_NONE; ok && try_map; ++filter, try_map >>= 1) {
+      if (try_map & 1) {
+        FilterTrial trial;
+        ok = EncodeAlphaInternal(alpha, width, height, method, filter,
+                                 reduce_levels, effort_level, filtered_alpha,
+                                 &trial);
+        if (ok && trial.score < best.score) {
+          VP8BitWriterWipeOut(&best.bw);
+          best = trial;
+        } else {
+          VP8BitWriterWipeOut(&trial.bw);
+        }
+      }
+    }
+    free(filtered_alpha);
+  } else {
+    ok = EncodeAlphaInternal(alpha, width, height, method, WEBP_FILTER_NONE,
+                             reduce_levels, effort_level, NULL, &best);
+  }
+  if (ok) {
+    if (stats != NULL) *stats = best.stats;
+    *output_size = VP8BitWriterSize(&best.bw);
+    *output = VP8BitWriterBuf(&best.bw);
+  } else {
+    VP8BitWriterWipeOut(&best.bw);
+  }
+  return ok;
+}
+
+static int EncodeAlpha(VP8Encoder* const enc,
+                       int quality, int method, int filter,
+                       int effort_level,
+                       uint8_t** const output, size_t* const output_size) {
+  const WebPPicture* const pic = enc->pic_;
+  const int width = pic->width;
+  const int height = pic->height;
+
+  uint8_t* quant_alpha = NULL;
+  const size_t data_size = width * height;
+  uint64_t sse = 0;
+  int ok = 1;
+  const int reduce_levels = (quality < 100);
+
+  // quick sanity checks
+  assert((uint64_t)data_size == (uint64_t)width * height);  // as per spec
+  assert(enc != NULL && pic != NULL && pic->a != NULL);
+  assert(output != NULL && output_size != NULL);
+  assert(width > 0 && height > 0);
+  assert(pic->a_stride >= width);
+  assert(filter >= WEBP_FILTER_NONE && filter <= WEBP_FILTER_FAST);
+
+  if (quality < 0 || quality > 100) {
+    return 0;
+  }
+
+  if (method < ALPHA_NO_COMPRESSION || method > ALPHA_LOSSLESS_COMPRESSION) {
+    return 0;
+  }
+
+  if (method == ALPHA_NO_COMPRESSION) {
+    // Don't filter, as filtering will make no impact on compressed size.
+    filter = WEBP_FILTER_NONE;
+  }
+
+  quant_alpha = (uint8_t*)malloc(data_size);
+  if (quant_alpha == NULL) {
+    return 0;
+  }
+
+  // Extract alpha data (width x height) from raw_data (stride x height).
+  CopyPlane(pic->a, pic->a_stride, quant_alpha, width, width, height);
+
+  if (reduce_levels) {  // No Quantization required for 'quality = 100'.
+    // 16 alpha levels gives quite a low MSE w.r.t original alpha plane hence
+    // mapped to moderate quality 70. Hence Quality:[0, 70] -> Levels:[2, 16]
+    // and Quality:]70, 100] -> Levels:]16, 256].
+    const int alpha_levels = (quality <= 70) ? (2 + quality / 5)
+                                             : (16 + (quality - 70) * 8);
+    ok = QuantizeLevels(quant_alpha, width, height, alpha_levels, &sse);
+  }
+
+  if (ok) {
+    ok = ApplyFiltersAndEncode(quant_alpha, width, height, data_size, method,
+                               filter, reduce_levels, effort_level, output,
+                               output_size, pic->stats);
+    if (pic->stats != NULL) {  // need stats?
+      pic->stats->coded_size += (int)(*output_size);
+      enc->sse_[3] = sse;
+    }
+  }
+
+  free(quant_alpha);
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+// Main calls
+
+static int CompressAlphaJob(VP8Encoder* const enc, void* dummy) {
+  const WebPConfig* config = enc->config_;
+  uint8_t* alpha_data = NULL;
+  size_t alpha_size = 0;
+  const int effort_level = config->method;  // maps to [0..6]
+  const WEBP_FILTER_TYPE filter =
+      (config->alpha_filtering == 0) ? WEBP_FILTER_NONE :
+      (config->alpha_filtering == 1) ? WEBP_FILTER_FAST :
+                                       WEBP_FILTER_BEST;
+  if (!EncodeAlpha(enc, config->alpha_quality, config->alpha_compression,
+                   filter, effort_level, &alpha_data, &alpha_size)) {
+    return 0;
+  }
+  if (alpha_size != (uint32_t)alpha_size) {  // Sanity check.
+    free(alpha_data);
+    return 0;
+  }
+  enc->alpha_data_size_ = (uint32_t)alpha_size;
+  enc->alpha_data_ = alpha_data;
+  (void)dummy;
+  return 1;
+}
+
+void VP8EncInitAlpha(VP8Encoder* const enc) {
+  enc->has_alpha_ = WebPPictureHasTransparency(enc->pic_);
+  enc->alpha_data_ = NULL;
+  enc->alpha_data_size_ = 0;
+  if (enc->thread_level_ > 0) {
+    WebPWorker* const worker = &enc->alpha_worker_;
+    WebPWorkerInit(worker);
+    worker->data1 = enc;
+    worker->data2 = NULL;
+    worker->hook = (WebPWorkerHook)CompressAlphaJob;
+  }
+}
+
+int VP8EncStartAlpha(VP8Encoder* const enc) {
+  if (enc->has_alpha_) {
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      if (!WebPWorkerReset(worker)) {    // Makes sure worker is good to go.
+        return 0;
+      }
+      WebPWorkerLaunch(worker);
+      return 1;
+    } else {
+      return CompressAlphaJob(enc, NULL);   // just do the job right away
+    }
+  }
+  return 1;
+}
+
+int VP8EncFinishAlpha(VP8Encoder* const enc) {
+  if (enc->has_alpha_) {
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      if (!WebPWorkerSync(worker)) return 0;  // error
+    }
+  }
+  return WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
+}
+
+int VP8EncDeleteAlpha(VP8Encoder* const enc) {
+  int ok = 1;
+  if (enc->thread_level_ > 0) {
+    WebPWorker* const worker = &enc->alpha_worker_;
+    ok = WebPWorkerSync(worker);  // finish anything left in flight
+    WebPWorkerEnd(worker);  // still need to end the worker, even if !ok
+  }
+  free(enc->alpha_data_);
+  enc->alpha_data_ = NULL;
+  enc->alpha_data_size_ = 0;
+  enc->has_alpha_ = 0;
+  return ok;
+}
+
diff --git a/drivers/webp/enc/analysis.c b/drivers/webp/enc/analysis.c
new file mode 100644
index 0000000000..7d4cfdc190
--- /dev/null
+++ b/drivers/webp/enc/analysis.c
@@ -0,0 +1,497 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Macroblock analysis
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+#include <string.h>
+#include <assert.h>
+
+#include "./vp8enci.h"
+#include "./cost.h"
+#include "../utils/utils.h"
+
+#define MAX_ITERS_K_MEANS  6
+
+//------------------------------------------------------------------------------
+// Smooth the segment map by replacing isolated block by the majority of its
+// neighbours.
+
+static void SmoothSegmentMap(VP8Encoder* const enc) {
+  int n, x, y;
+  const int w = enc->mb_w_;
+  const int h = enc->mb_h_;
+  const int majority_cnt_3_x_3_grid = 5;
+  uint8_t* const tmp = (uint8_t*)WebPSafeMalloc((uint64_t)w * h, sizeof(*tmp));
+  assert((uint64_t)(w * h) == (uint64_t)w * h);   // no overflow, as per spec
+
+  if (tmp == NULL) return;
+  for (y = 1; y < h - 1; ++y) {
+    for (x = 1; x < w - 1; ++x) {
+      int cnt[NUM_MB_SEGMENTS] = { 0 };
+      const VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
+      int majority_seg = mb->segment_;
+      // Check the 8 neighbouring segment values.
+      cnt[mb[-w - 1].segment_]++;  // top-left
+      cnt[mb[-w + 0].segment_]++;  // top
+      cnt[mb[-w + 1].segment_]++;  // top-right
+      cnt[mb[   - 1].segment_]++;  // left
+      cnt[mb[   + 1].segment_]++;  // right
+      cnt[mb[ w - 1].segment_]++;  // bottom-left
+      cnt[mb[ w + 0].segment_]++;  // bottom
+      cnt[mb[ w + 1].segment_]++;  // bottom-right
+      for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
+        if (cnt[n] >= majority_cnt_3_x_3_grid) {
+          majority_seg = n;
+          break;
+        }
+      }
+      tmp[x + y * w] = majority_seg;
+    }
+  }
+  for (y = 1; y < h - 1; ++y) {
+    for (x = 1; x < w - 1; ++x) {
+      VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
+      mb->segment_ = tmp[x + y * w];
+    }
+  }
+  free(tmp);
+}
+
+//------------------------------------------------------------------------------
+// set segment susceptibility alpha_ / beta_
+
+static WEBP_INLINE int clip(int v, int m, int M) {
+  return (v < m) ? m : (v > M) ? M : v;
+}
+
+static void SetSegmentAlphas(VP8Encoder* const enc,
+                             const int centers[NUM_MB_SEGMENTS],
+                             int mid) {
+  const int nb = enc->segment_hdr_.num_segments_;
+  int min = centers[0], max = centers[0];
+  int n;
+
+  if (nb > 1) {
+    for (n = 0; n < nb; ++n) {
+      if (min > centers[n]) min = centers[n];
+      if (max < centers[n]) max = centers[n];
+    }
+  }
+  if (max == min) max = min + 1;
+  assert(mid <= max && mid >= min);
+  for (n = 0; n < nb; ++n) {
+    const int alpha = 255 * (centers[n] - mid) / (max - min);
+    const int beta = 255 * (centers[n] - min) / (max - min);
+    enc->dqm_[n].alpha_ = clip(alpha, -127, 127);
+    enc->dqm_[n].beta_ = clip(beta, 0, 255);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Compute susceptibility based on DCT-coeff histograms:
+// the higher, the "easier" the macroblock is to compress.
+
+#define MAX_ALPHA 255                // 8b of precision for susceptibilities.
+#define ALPHA_SCALE (2 * MAX_ALPHA)  // scaling factor for alpha.
+#define DEFAULT_ALPHA (-1)
+#define IS_BETTER_ALPHA(alpha, best_alpha) ((alpha) > (best_alpha))
+
+static int FinalAlphaValue(int alpha) {
+  alpha = MAX_ALPHA - alpha;
+  return clip(alpha, 0, MAX_ALPHA);
+}
+
+static int GetAlpha(const VP8Histogram* const histo) {
+  int max_value = 0, last_non_zero = 1;
+  int k;
+  int alpha;
+  for (k = 0; k <= MAX_COEFF_THRESH; ++k) {
+    const int value = histo->distribution[k];
+    if (value > 0) {
+      if (value > max_value) max_value = value;
+      last_non_zero = k;
+    }
+  }
+  // 'alpha' will later be clipped to [0..MAX_ALPHA] range, clamping outer
+  // values which happen to be mostly noise. This leaves the maximum precision
+  // for handling the useful small values which contribute most.
+  alpha = (max_value > 1) ? ALPHA_SCALE * last_non_zero / max_value : 0;
+  return alpha;
+}
+
+static void MergeHistograms(const VP8Histogram* const in,
+                            VP8Histogram* const out) {
+  int i;
+  for (i = 0; i <= MAX_COEFF_THRESH; ++i) {
+    out->distribution[i] += in->distribution[i];
+  }
+}
+
+//------------------------------------------------------------------------------
+// Simplified k-Means, to assign Nb segments based on alpha-histogram
+
+static void AssignSegments(VP8Encoder* const enc,
+                           const int alphas[MAX_ALPHA + 1]) {
+  const int nb = enc->segment_hdr_.num_segments_;
+  int centers[NUM_MB_SEGMENTS];
+  int weighted_average = 0;
+  int map[MAX_ALPHA + 1];
+  int a, n, k;
+  int min_a = 0, max_a = MAX_ALPHA, range_a;
+  // 'int' type is ok for histo, and won't overflow
+  int accum[NUM_MB_SEGMENTS], dist_accum[NUM_MB_SEGMENTS];
+
+  assert(nb >= 1);
+
+  // bracket the input
+  for (n = 0; n <= MAX_ALPHA && alphas[n] == 0; ++n) {}
+  min_a = n;
+  for (n = MAX_ALPHA; n > min_a && alphas[n] == 0; --n) {}
+  max_a = n;
+  range_a = max_a - min_a;
+
+  // Spread initial centers evenly
+  for (k = 0, n = 1; k < nb; ++k, n += 2) {
+    assert(n < 2 * nb);
+    centers[k] = min_a + (n * range_a) / (2 * nb);
+  }
+
+  for (k = 0; k < MAX_ITERS_K_MEANS; ++k) {     // few iters are enough
+    int total_weight;
+    int displaced;
+    // Reset stats
+    for (n = 0; n < nb; ++n) {
+      accum[n] = 0;
+      dist_accum[n] = 0;
+    }
+    // Assign nearest center for each 'a'
+    n = 0;    // track the nearest center for current 'a'
+    for (a = min_a; a <= max_a; ++a) {
+      if (alphas[a]) {
+        while (n + 1 < nb && abs(a - centers[n + 1]) < abs(a - centers[n])) {
+          n++;
+        }
+        map[a] = n;
+        // accumulate contribution into best centroid
+        dist_accum[n] += a * alphas[a];
+        accum[n] += alphas[a];
+      }
+    }
+    // All point are classified. Move the centroids to the
+    // center of their respective cloud.
+    displaced = 0;
+    weighted_average = 0;
+    total_weight = 0;
+    for (n = 0; n < nb; ++n) {
+      if (accum[n]) {
+        const int new_center = (dist_accum[n] + accum[n] / 2) / accum[n];
+        displaced += abs(centers[n] - new_center);
+        centers[n] = new_center;
+        weighted_average += new_center * accum[n];
+        total_weight += accum[n];
+      }
+    }
+    weighted_average = (weighted_average + total_weight / 2) / total_weight;
+    if (displaced < 5) break;   // no need to keep on looping...
+  }
+
+  // Map each original value to the closest centroid
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    VP8MBInfo* const mb = &enc->mb_info_[n];
+    const int alpha = mb->alpha_;
+    mb->segment_ = map[alpha];
+    mb->alpha_ = centers[map[alpha]];  // for the record.
+  }
+
+  if (nb > 1) {
+    const int smooth = (enc->config_->preprocessing & 1);
+    if (smooth) SmoothSegmentMap(enc);
+  }
+
+  SetSegmentAlphas(enc, centers, weighted_average);  // pick some alphas.
+}
+
+//------------------------------------------------------------------------------
+// Macroblock analysis: collect histogram for each mode, deduce the maximal
+// susceptibility and set best modes for this macroblock.
+// Segment assignment is done later.
+
+// Number of modes to inspect for alpha_ evaluation. For high-quality settings
+// (method >= FAST_ANALYSIS_METHOD) we don't need to test all the possible modes
+// during the analysis phase.
+#define FAST_ANALYSIS_METHOD 4  // method above which we do partial analysis
+#define MAX_INTRA16_MODE 2
+#define MAX_INTRA4_MODE  2
+#define MAX_UV_MODE      2
+
+static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
+  const int max_mode =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA16_MODE
+                                                  : NUM_PRED_MODES;
+  int mode;
+  int best_alpha = DEFAULT_ALPHA;
+  int best_mode = 0;
+
+  VP8MakeLuma16Preds(it);
+  for (mode = 0; mode < max_mode; ++mode) {
+    VP8Histogram histo = { { 0 } };
+    int alpha;
+
+    VP8CollectHistogram(it->yuv_in_ + Y_OFF,
+                        it->yuv_p_ + VP8I16ModeOffsets[mode],
+                        0, 16, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
+      best_alpha = alpha;
+      best_mode = mode;
+    }
+  }
+  VP8SetIntra16Mode(it, best_mode);
+  return best_alpha;
+}
+
+static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
+                                   int best_alpha) {
+  uint8_t modes[16];
+  const int max_mode =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA4_MODE
+                                                  : NUM_BMODES;
+  int i4_alpha;
+  VP8Histogram total_histo = { { 0 } };
+  int cur_histo = 0;
+
+  VP8IteratorStartI4(it);
+  do {
+    int mode;
+    int best_mode_alpha = DEFAULT_ALPHA;
+    VP8Histogram histos[2];
+    const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
+
+    VP8MakeIntra4Preds(it);
+    for (mode = 0; mode < max_mode; ++mode) {
+      int alpha;
+
+      memset(&histos[cur_histo], 0, sizeof(histos[cur_histo]));
+      VP8CollectHistogram(src, it->yuv_p_ + VP8I4ModeOffsets[mode],
+                          0, 1, &histos[cur_histo]);
+      alpha = GetAlpha(&histos[cur_histo]);
+      if (IS_BETTER_ALPHA(alpha, best_mode_alpha)) {
+        best_mode_alpha = alpha;
+        modes[it->i4_] = mode;
+        cur_histo ^= 1;   // keep track of best histo so far.
+      }
+    }
+    // accumulate best histogram
+    MergeHistograms(&histos[cur_histo ^ 1], &total_histo);
+    // Note: we reuse the original samples for predictors
+  } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF));
+
+  i4_alpha = GetAlpha(&total_histo);
+  if (IS_BETTER_ALPHA(i4_alpha, best_alpha)) {
+    VP8SetIntra4Mode(it, modes);
+    best_alpha = i4_alpha;
+  }
+  return best_alpha;
+}
+
+static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
+  int best_alpha = DEFAULT_ALPHA;
+  int best_mode = 0;
+  const int max_mode =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_UV_MODE
+                                                  : NUM_PRED_MODES;
+  int mode;
+  VP8MakeChroma8Preds(it);
+  for (mode = 0; mode < max_mode; ++mode) {
+    VP8Histogram histo = { { 0 } };
+    int alpha;
+    VP8CollectHistogram(it->yuv_in_ + U_OFF,
+                        it->yuv_p_ + VP8UVModeOffsets[mode],
+                        16, 16 + 4 + 4, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
+      best_alpha = alpha;
+      best_mode = mode;
+    }
+  }
+  VP8SetIntraUVMode(it, best_mode);
+  return best_alpha;
+}
+
+static void MBAnalyze(VP8EncIterator* const it,
+                      int alphas[MAX_ALPHA + 1],
+                      int* const alpha, int* const uv_alpha) {
+  const VP8Encoder* const enc = it->enc_;
+  int best_alpha, best_uv_alpha;
+
+  VP8SetIntra16Mode(it, 0);  // default: Intra16, DC_PRED
+  VP8SetSkip(it, 0);         // not skipped
+  VP8SetSegment(it, 0);      // default segment, spec-wise.
+
+  best_alpha = MBAnalyzeBestIntra16Mode(it);
+  if (enc->method_ >= 5) {
+    // We go and make a fast decision for intra4/intra16.
+    // It's usually not a good and definitive pick, but helps seeding the stats
+    // about level bit-cost.
+    // TODO(skal): improve criterion.
+    best_alpha = MBAnalyzeBestIntra4Mode(it, best_alpha);
+  }
+  best_uv_alpha = MBAnalyzeBestUVMode(it);
+
+  // Final susceptibility mix
+  best_alpha = (3 * best_alpha + best_uv_alpha + 2) >> 2;
+  best_alpha = FinalAlphaValue(best_alpha);
+  alphas[best_alpha]++;
+  it->mb_->alpha_ = best_alpha;   // for later remapping.
+
+  // Accumulate for later complexity analysis.
+  *alpha += best_alpha;   // mixed susceptibility (not just luma)
+  *uv_alpha += best_uv_alpha;
+}
+
+static void DefaultMBInfo(VP8MBInfo* const mb) {
+  mb->type_ = 1;     // I16x16
+  mb->uv_mode_ = 0;
+  mb->skip_ = 0;     // not skipped
+  mb->segment_ = 0;  // default segment
+  mb->alpha_ = 0;
+}
+
+//------------------------------------------------------------------------------
+// Main analysis loop:
+// Collect all susceptibilities for each macroblock and record their
+// distribution in alphas[]. Segments is assigned a-posteriori, based on
+// this histogram.
+// We also pick an intra16 prediction mode, which shouldn't be considered
+// final except for fast-encode settings. We can also pick some intra4 modes
+// and decide intra4/intra16, but that's usually almost always a bad choice at
+// this stage.
+
+static void ResetAllMBInfo(VP8Encoder* const enc) {
+  int n;
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    DefaultMBInfo(&enc->mb_info_[n]);
+  }
+  // Default susceptibilities.
+  enc->dqm_[0].alpha_ = 0;
+  enc->dqm_[0].beta_ = 0;
+  // Note: we can't compute this alpha_ / uv_alpha_ -> set to default value.
+  enc->alpha_ = 0;
+  enc->uv_alpha_ = 0;
+  WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
+}
+
+// struct used to collect job result
+typedef struct {
+  WebPWorker worker;
+  int alphas[MAX_ALPHA + 1];
+  int alpha, uv_alpha;
+  VP8EncIterator it;
+  int delta_progress;
+} SegmentJob;
+
+// main work call
+static int DoSegmentsJob(SegmentJob* const job, VP8EncIterator* const it) {
+  int ok = 1;
+  if (!VP8IteratorIsDone(it)) {
+    uint8_t tmp[32 + ALIGN_CST];
+    uint8_t* const scratch = (uint8_t*)DO_ALIGN(tmp);
+    do {
+      // Let's pretend we have perfect lossless reconstruction.
+      VP8IteratorImport(it, scratch);
+      MBAnalyze(it, job->alphas, &job->alpha, &job->uv_alpha);
+      ok = VP8IteratorProgress(it, job->delta_progress);
+    } while (ok && VP8IteratorNext(it));
+  }
+  return ok;
+}
+
+static void MergeJobs(const SegmentJob* const src, SegmentJob* const dst) {
+  int i;
+  for (i = 0; i <= MAX_ALPHA; ++i) dst->alphas[i] += src->alphas[i];
+  dst->alpha += src->alpha;
+  dst->uv_alpha += src->uv_alpha;
+}
+
+// initialize the job struct with some TODOs
+static void InitSegmentJob(VP8Encoder* const enc, SegmentJob* const job,
+                           int start_row, int end_row) {
+  WebPWorkerInit(&job->worker);
+  job->worker.data1 = job;
+  job->worker.data2 = &job->it;
+  job->worker.hook = (WebPWorkerHook)DoSegmentsJob;
+  VP8IteratorInit(enc, &job->it);
+  VP8IteratorSetRow(&job->it, start_row);
+  VP8IteratorSetCountDown(&job->it, (end_row - start_row) * enc->mb_w_);
+  memset(job->alphas, 0, sizeof(job->alphas));
+  job->alpha = 0;
+  job->uv_alpha = 0;
+  // only one of both jobs can record the progress, since we don't
+  // expect the user's hook to be multi-thread safe
+  job->delta_progress = (start_row == 0) ? 20 : 0;
+}
+
+// main entry point
+int VP8EncAnalyze(VP8Encoder* const enc) {
+  int ok = 1;
+  const int do_segments =
+      enc->config_->emulate_jpeg_size ||   // We need the complexity evaluation.
+      (enc->segment_hdr_.num_segments_ > 1) ||
+      (enc->method_ == 0);  // for method 0, we need preds_[] to be filled.
+  if (do_segments) {
+    const int last_row = enc->mb_h_;
+    // We give a little more than a half work to the main thread.
+    const int split_row = (9 * last_row + 15) >> 4;
+    const int total_mb = last_row * enc->mb_w_;
+#ifdef WEBP_USE_THREAD
+    const int kMinSplitRow = 2;  // minimal rows needed for mt to be worth it
+    const int do_mt = (enc->thread_level_ > 0) && (split_row >= kMinSplitRow);
+#else
+    const int do_mt = 0;
+#endif
+    SegmentJob main_job;
+    if (do_mt) {
+      SegmentJob side_job;
+      // Note the use of '&' instead of '&&' because we must call the functions
+      // no matter what.
+      InitSegmentJob(enc, &main_job, 0, split_row);
+      InitSegmentJob(enc, &side_job, split_row, last_row);
+      // we don't need to call Reset() on main_job.worker, since we're calling
+      // WebPWorkerExecute() on it
+      ok &= WebPWorkerReset(&side_job.worker);
+      // launch the two jobs in parallel
+      if (ok) {
+        WebPWorkerLaunch(&side_job.worker);
+        WebPWorkerExecute(&main_job.worker);
+        ok &= WebPWorkerSync(&side_job.worker);
+        ok &= WebPWorkerSync(&main_job.worker);
+      }
+      WebPWorkerEnd(&side_job.worker);
+      if (ok) MergeJobs(&side_job, &main_job);  // merge results together
+    } else {
+      // Even for single-thread case, we use the generic Worker tools.
+      InitSegmentJob(enc, &main_job, 0, last_row);
+      WebPWorkerExecute(&main_job.worker);
+      ok &= WebPWorkerSync(&main_job.worker);
+    }
+    WebPWorkerEnd(&main_job.worker);
+    if (ok) {
+      enc->alpha_ = main_job.alpha / total_mb;
+      enc->uv_alpha_ = main_job.uv_alpha / total_mb;
+      AssignSegments(enc, main_job.alphas);
+    }
+  } else {   // Use only one default segment.
+    ResetAllMBInfo(enc);
+  }
+  return ok;
+}
+
diff --git a/drivers/webp/enc/backward_references.c b/drivers/webp/enc/backward_references.c
new file mode 100644
index 0000000000..77b4be7432
--- /dev/null
+++ b/drivers/webp/enc/backward_references.c
@@ -0,0 +1,894 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+//
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./backward_references.h"
+#include "./histogram.h"
+#include "../dsp/lossless.h"
+#include "../utils/color_cache.h"
+#include "../utils/utils.h"
+
+#define VALUES_IN_BYTE 256
+
+#define HASH_BITS 18
+#define HASH_SIZE (1 << HASH_BITS)
+#define HASH_MULTIPLIER (0xc6a4a7935bd1e995ULL)
+
+// 1M window (4M bytes) minus 120 special codes for short distances.
+#define WINDOW_SIZE ((1 << 20) - 120)
+
+// Bounds for the match length.
+#define MIN_LENGTH 2
+#define MAX_LENGTH 4096
+
+typedef struct {
+  // Stores the most recently added position with the given hash value.
+  int32_t hash_to_first_index_[HASH_SIZE];
+  // chain_[pos] stores the previous position with the same hash value
+  // for every pixel in the image.
+  int32_t* chain_;
+} HashChain;
+
+// -----------------------------------------------------------------------------
+
+static const uint8_t plane_to_code_lut[128] = {
+ 96,   73,  55,  39,  23,  13,   5,  1,  255, 255, 255, 255, 255, 255, 255, 255,
+ 101,  78,  58,  42,  26,  16,   8,  2,    0,   3,  9,   17,  27,  43,  59,  79,
+ 102,  86,  62,  46,  32,  20,  10,  6,    4,   7,  11,  21,  33,  47,  63,  87,
+ 105,  90,  70,  52,  37,  28,  18,  14,  12,  15,  19,  29,  38,  53,  71,  91,
+ 110,  99,  82,  66,  48,  35,  30,  24,  22,  25,  31,  36,  49,  67,  83, 100,
+ 115, 108,  94,  76,  64,  50,  44,  40,  34,  41,  45,  51,  65,  77,  95, 109,
+ 118, 113, 103,  92,  80,  68,  60,  56,  54,  57,  61,  69,  81,  93, 104, 114,
+ 119, 116, 111, 106,  97,  88,  84,  74,  72,  75,  85,  89,  98, 107, 112, 117
+};
+
+static int DistanceToPlaneCode(int xsize, int dist) {
+  const int yoffset = dist / xsize;
+  const int xoffset = dist - yoffset * xsize;
+  if (xoffset <= 8 && yoffset < 8) {
+    return plane_to_code_lut[yoffset * 16 + 8 - xoffset] + 1;
+  } else if (xoffset > xsize - 8 && yoffset < 7) {
+    return plane_to_code_lut[(yoffset + 1) * 16 + 8 + (xsize - xoffset)] + 1;
+  }
+  return dist + 120;
+}
+
+static WEBP_INLINE int FindMatchLength(const uint32_t* const array1,
+                                       const uint32_t* const array2,
+                                       const int max_limit) {
+  int match_len = 0;
+  while (match_len < max_limit && array1[match_len] == array2[match_len]) {
+    ++match_len;
+  }
+  return match_len;
+}
+
+// -----------------------------------------------------------------------------
+//  VP8LBackwardRefs
+
+void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs) {
+  if (refs != NULL) {
+    refs->refs = NULL;
+    refs->size = 0;
+    refs->max_size = 0;
+  }
+}
+
+void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs) {
+  if (refs != NULL) {
+    free(refs->refs);
+    VP8LInitBackwardRefs(refs);
+  }
+}
+
+int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size) {
+  assert(refs != NULL);
+  refs->size = 0;
+  refs->max_size = 0;
+  refs->refs = (PixOrCopy*)WebPSafeMalloc((uint64_t)max_size,
+                                          sizeof(*refs->refs));
+  if (refs->refs == NULL) return 0;
+  refs->max_size = max_size;
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+// Hash chains
+
+static WEBP_INLINE uint64_t GetPixPairHash64(const uint32_t* const argb) {
+  uint64_t key = ((uint64_t)(argb[1]) << 32) | argb[0];
+  key = (key * HASH_MULTIPLIER) >> (64 - HASH_BITS);
+  return key;
+}
+
+static int HashChainInit(HashChain* const p, int size) {
+  int i;
+  p->chain_ = (int*)WebPSafeMalloc((uint64_t)size, sizeof(*p->chain_));
+  if (p->chain_ == NULL) {
+    return 0;
+  }
+  for (i = 0; i < size; ++i) {
+    p->chain_[i] = -1;
+  }
+  for (i = 0; i < HASH_SIZE; ++i) {
+    p->hash_to_first_index_[i] = -1;
+  }
+  return 1;
+}
+
+static void HashChainDelete(HashChain* const p) {
+  if (p != NULL) {
+    free(p->chain_);
+    free(p);
+  }
+}
+
+// Insertion of two pixels at a time.
+static void HashChainInsert(HashChain* const p,
+                            const uint32_t* const argb, int pos) {
+  const uint64_t hash_code = GetPixPairHash64(argb);
+  p->chain_[pos] = p->hash_to_first_index_[hash_code];
+  p->hash_to_first_index_[hash_code] = pos;
+}
+
+static void GetParamsForHashChainFindCopy(int quality, int xsize,
+                                          int cache_bits, int* window_size,
+                                          int* iter_pos, int* iter_limit) {
+  const int iter_mult = (quality < 27) ? 1 : 1 + ((quality - 27) >> 4);
+  const int iter_neg = -iter_mult * (quality >> 1);
+  // Limit the backward-ref window size for lower qualities.
+  const int max_window_size = (quality > 50) ? WINDOW_SIZE
+                            : (quality > 25) ? (xsize << 8)
+                            : (xsize << 4);
+  assert(xsize > 0);
+  *window_size = (max_window_size > WINDOW_SIZE) ? WINDOW_SIZE
+               : max_window_size;
+  *iter_pos = 8 + (quality >> 3);
+  // For lower entropy images, the rigorous search loop in HashChainFindCopy
+  // can be relaxed.
+  *iter_limit = (cache_bits > 0) ? iter_neg : iter_neg / 2;
+}
+
+static int HashChainFindCopy(const HashChain* const p,
+                             int base_position, int xsize_signed,
+                             const uint32_t* const argb, int max_len,
+                             int window_size, int iter_pos, int iter_limit,
+                             int* const distance_ptr,
+                             int* const length_ptr) {
+  const uint32_t* const argb_start = argb + base_position;
+  uint64_t best_val = 0;
+  uint32_t best_length = 1;
+  uint32_t best_distance = 0;
+  const uint32_t xsize = (uint32_t)xsize_signed;
+  const int min_pos =
+      (base_position > window_size) ? base_position - window_size : 0;
+  int pos;
+  assert(xsize > 0);
+  if (max_len > MAX_LENGTH) {
+    max_len = MAX_LENGTH;
+  }
+  for (pos = p->hash_to_first_index_[GetPixPairHash64(argb_start)];
+       pos >= min_pos;
+       pos = p->chain_[pos]) {
+    uint64_t val;
+    uint32_t curr_length;
+    uint32_t distance;
+    const uint64_t* const ptr1 =
+        (const uint64_t*)(argb + pos + best_length - 1);
+    const uint64_t* const ptr2 =
+        (const uint64_t*)(argb_start + best_length - 1);
+
+    if (iter_pos < 0) {
+      if (iter_pos < iter_limit || best_val >= 0xff0000) {
+        break;
+      }
+    }
+    --iter_pos;
+
+    // Before 'expensive' linear match, check if the two arrays match at the
+    // current best length index and also for the succeeding elements.
+    if (*ptr1 != *ptr2) continue;
+
+    curr_length = FindMatchLength(argb + pos, argb_start, max_len);
+    if (curr_length < best_length) continue;
+
+    distance = (uint32_t)(base_position - pos);
+    val = curr_length << 16;
+    // Favoring 2d locality here gives savings for certain images.
+    if (distance < 9 * xsize) {
+      const uint32_t y = distance / xsize;
+      uint32_t x = distance % xsize;
+      if (x > (xsize >> 1)) {
+        x = xsize - x;
+      }
+      if (x <= 7) {
+        val += 9 * 9 + 9 * 9;
+        val -= y * y + x * x;
+      }
+    }
+    if (best_val < val) {
+      best_val = val;
+      best_length = curr_length;
+      best_distance = distance;
+      if (curr_length >= (uint32_t)max_len) {
+        break;
+      }
+      if ((best_distance == 1 || distance == xsize) &&
+          best_length >= 128) {
+        break;
+      }
+    }
+  }
+  *distance_ptr = (int)best_distance;
+  *length_ptr = best_length;
+  return (best_length >= MIN_LENGTH);
+}
+
+static WEBP_INLINE void PushBackCopy(VP8LBackwardRefs* const refs, int length) {
+  int size = refs->size;
+  while (length >= MAX_LENGTH) {
+    refs->refs[size++] = PixOrCopyCreateCopy(1, MAX_LENGTH);
+    length -= MAX_LENGTH;
+  }
+  if (length > 0) {
+    refs->refs[size++] = PixOrCopyCreateCopy(1, length);
+  }
+  refs->size = size;
+}
+
+static void BackwardReferencesRle(int xsize, int ysize,
+                                  const uint32_t* const argb,
+                                  VP8LBackwardRefs* const refs) {
+  const int pix_count = xsize * ysize;
+  int match_len = 0;
+  int i;
+  refs->size = 0;
+  PushBackCopy(refs, match_len);    // i=0 case
+  refs->refs[refs->size++] = PixOrCopyCreateLiteral(argb[0]);
+  for (i = 1; i < pix_count; ++i) {
+    if (argb[i] == argb[i - 1]) {
+      ++match_len;
+    } else {
+      PushBackCopy(refs, match_len);
+      match_len = 0;
+      refs->refs[refs->size++] = PixOrCopyCreateLiteral(argb[i]);
+    }
+  }
+  PushBackCopy(refs, match_len);
+}
+
+static int BackwardReferencesHashChain(int xsize, int ysize,
+                                       const uint32_t* const argb,
+                                       int cache_bits, int quality,
+                                       VP8LBackwardRefs* const refs) {
+  int i;
+  int ok = 0;
+  int cc_init = 0;
+  const int use_color_cache = (cache_bits > 0);
+  const int pix_count = xsize * ysize;
+  HashChain* const hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+  VP8LColorCache hashers;
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
+
+  if (hash_chain == NULL) return 0;
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+
+  if (!HashChainInit(hash_chain, pix_count)) goto Error;
+
+  refs->size = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
+                                &window_size, &iter_pos, &iter_limit);
+  for (i = 0; i < pix_count; ) {
+    // Alternative#1: Code the pixels starting at 'i' using backward reference.
+    int offset = 0;
+    int len = 0;
+    if (i < pix_count - 1) {  // FindCopy(i,..) reads pixels at [i] and [i + 1].
+      int max_len = pix_count - i;
+      HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
+                        window_size, iter_pos, iter_limit,
+                        &offset, &len);
+    }
+    if (len >= MIN_LENGTH) {
+      // Alternative#2: Insert the pixel at 'i' as literal, and code the
+      // pixels starting at 'i + 1' using backward reference.
+      int offset2 = 0;
+      int len2 = 0;
+      int k;
+      HashChainInsert(hash_chain, &argb[i], i);
+      if (i < pix_count - 2) {  // FindCopy(i+1,..) reads [i + 1] and [i + 2].
+        int max_len = pix_count - (i + 1);
+        HashChainFindCopy(hash_chain, i + 1, xsize, argb, max_len,
+                          window_size, iter_pos, iter_limit,
+                          &offset2, &len2);
+        if (len2 > len + 1) {
+          const uint32_t pixel = argb[i];
+          // Alternative#2 is a better match. So push pixel at 'i' as literal.
+          if (use_color_cache && VP8LColorCacheContains(&hashers, pixel)) {
+            const int ix = VP8LColorCacheGetIndex(&hashers, pixel);
+            refs->refs[refs->size] = PixOrCopyCreateCacheIdx(ix);
+          } else {
+            if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
+            refs->refs[refs->size] = PixOrCopyCreateLiteral(pixel);
+          }
+          ++refs->size;
+          i++;  // Backward reference to be done for next pixel.
+          len = len2;
+          offset = offset2;
+        }
+      }
+      if (len >= MAX_LENGTH) {
+        len = MAX_LENGTH - 1;
+      }
+      refs->refs[refs->size++] = PixOrCopyCreateCopy(offset, len);
+      if (use_color_cache) {
+        for (k = 0; k < len; ++k) {
+          VP8LColorCacheInsert(&hashers, argb[i + k]);
+        }
+      }
+      // Add to the hash_chain (but cannot add the last pixel).
+      {
+        const int last = (len < pix_count - 1 - i) ? len : pix_count - 1 - i;
+        for (k = 1; k < last; ++k) {
+          HashChainInsert(hash_chain, &argb[i + k], i + k);
+        }
+      }
+      i += len;
+    } else {
+      const uint32_t pixel = argb[i];
+      if (use_color_cache && VP8LColorCacheContains(&hashers, pixel)) {
+        // push pixel as a PixOrCopyCreateCacheIdx pixel
+        const int ix = VP8LColorCacheGetIndex(&hashers, pixel);
+        refs->refs[refs->size] = PixOrCopyCreateCacheIdx(ix);
+      } else {
+        if (use_color_cache) VP8LColorCacheInsert(&hashers, pixel);
+        refs->refs[refs->size] = PixOrCopyCreateLiteral(pixel);
+      }
+      ++refs->size;
+      if (i + 1 < pix_count) {
+        HashChainInsert(hash_chain, &argb[i], i);
+      }
+      ++i;
+    }
+  }
+  ok = 1;
+Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  HashChainDelete(hash_chain);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+
+typedef struct {
+  double alpha_[VALUES_IN_BYTE];
+  double red_[VALUES_IN_BYTE];
+  double literal_[PIX_OR_COPY_CODES_MAX];
+  double blue_[VALUES_IN_BYTE];
+  double distance_[NUM_DISTANCE_CODES];
+} CostModel;
+
+static int BackwardReferencesTraceBackwards(
+    int xsize, int ysize, int recursive_cost_model,
+    const uint32_t* const argb, int quality, int cache_bits,
+    VP8LBackwardRefs* const refs);
+
+static void ConvertPopulationCountTableToBitEstimates(
+    int num_symbols, const int population_counts[], double output[]) {
+  int sum = 0;
+  int nonzeros = 0;
+  int i;
+  for (i = 0; i < num_symbols; ++i) {
+    sum += population_counts[i];
+    if (population_counts[i] > 0) {
+      ++nonzeros;
+    }
+  }
+  if (nonzeros <= 1) {
+    memset(output, 0, num_symbols * sizeof(*output));
+  } else {
+    const double logsum = VP8LFastLog2(sum);
+    for (i = 0; i < num_symbols; ++i) {
+      output[i] = logsum - VP8LFastLog2(population_counts[i]);
+    }
+  }
+}
+
+static int CostModelBuild(CostModel* const m, int xsize, int ysize,
+                          int recursion_level, const uint32_t* const argb,
+                          int quality, int cache_bits) {
+  int ok = 0;
+  VP8LHistogram histo;
+  VP8LBackwardRefs refs;
+
+  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize)) goto Error;
+
+  if (recursion_level > 0) {
+    if (!BackwardReferencesTraceBackwards(xsize, ysize, recursion_level - 1,
+                                          argb, quality, cache_bits, &refs)) {
+      goto Error;
+    }
+  } else {
+    if (!BackwardReferencesHashChain(xsize, ysize, argb, cache_bits, quality,
+                                     &refs)) {
+      goto Error;
+    }
+  }
+  VP8LHistogramCreate(&histo, &refs, cache_bits);
+  ConvertPopulationCountTableToBitEstimates(
+      VP8LHistogramNumCodes(&histo), histo.literal_, m->literal_);
+  ConvertPopulationCountTableToBitEstimates(
+      VALUES_IN_BYTE, histo.red_, m->red_);
+  ConvertPopulationCountTableToBitEstimates(
+      VALUES_IN_BYTE, histo.blue_, m->blue_);
+  ConvertPopulationCountTableToBitEstimates(
+      VALUES_IN_BYTE, histo.alpha_, m->alpha_);
+  ConvertPopulationCountTableToBitEstimates(
+      NUM_DISTANCE_CODES, histo.distance_, m->distance_);
+  ok = 1;
+
+ Error:
+  VP8LClearBackwardRefs(&refs);
+  return ok;
+}
+
+static WEBP_INLINE double GetLiteralCost(const CostModel* const m, uint32_t v) {
+  return m->alpha_[v >> 24] +
+         m->red_[(v >> 16) & 0xff] +
+         m->literal_[(v >> 8) & 0xff] +
+         m->blue_[v & 0xff];
+}
+
+static WEBP_INLINE double GetCacheCost(const CostModel* const m, uint32_t idx) {
+  const int literal_idx = VALUES_IN_BYTE + NUM_LENGTH_CODES + idx;
+  return m->literal_[literal_idx];
+}
+
+static WEBP_INLINE double GetLengthCost(const CostModel* const m,
+                                        uint32_t length) {
+  int code, extra_bits;
+  VP8LPrefixEncodeBits(length, &code, &extra_bits);
+  return m->literal_[VALUES_IN_BYTE + code] + extra_bits;
+}
+
+static WEBP_INLINE double GetDistanceCost(const CostModel* const m,
+                                          uint32_t distance) {
+  int code, extra_bits;
+  VP8LPrefixEncodeBits(distance, &code, &extra_bits);
+  return m->distance_[code] + extra_bits;
+}
+
+static int BackwardReferencesHashChainDistanceOnly(
+    int xsize, int ysize, int recursive_cost_model, const uint32_t* const argb,
+    int quality, int cache_bits, uint32_t* const dist_array) {
+  int i;
+  int ok = 0;
+  int cc_init = 0;
+  const int pix_count = xsize * ysize;
+  const int use_color_cache = (cache_bits > 0);
+  float* const cost =
+      (float*)WebPSafeMalloc((uint64_t)pix_count, sizeof(*cost));
+  CostModel* cost_model = (CostModel*)malloc(sizeof(*cost_model));
+  HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+  VP8LColorCache hashers;
+  const double mul0 = (recursive_cost_model != 0) ? 1.0 : 0.68;
+  const double mul1 = (recursive_cost_model != 0) ? 1.0 : 0.82;
+  const int min_distance_code = 2;  // TODO(vikasa): tune as function of quality
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
+
+  if (cost == NULL || cost_model == NULL || hash_chain == NULL) goto Error;
+
+  if (!HashChainInit(hash_chain, pix_count)) goto Error;
+
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+
+  if (!CostModelBuild(cost_model, xsize, ysize, recursive_cost_model, argb,
+                      quality, cache_bits)) {
+    goto Error;
+  }
+
+  for (i = 0; i < pix_count; ++i) cost[i] = 1e38f;
+
+  // We loop one pixel at a time, but store all currently best points to
+  // non-processed locations from this point.
+  dist_array[0] = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
+                                &window_size, &iter_pos, &iter_limit);
+  for (i = 0; i < pix_count; ++i) {
+    double prev_cost = 0.0;
+    int shortmax;
+    if (i > 0) {
+      prev_cost = cost[i - 1];
+    }
+    for (shortmax = 0; shortmax < 2; ++shortmax) {
+      int offset = 0;
+      int len = 0;
+      if (i < pix_count - 1) {  // FindCopy reads pixels at [i] and [i + 1].
+        int max_len = shortmax ? 2 : pix_count - i;
+        HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
+                          window_size, iter_pos, iter_limit,
+                          &offset, &len);
+      }
+      if (len >= MIN_LENGTH) {
+        const int code = DistanceToPlaneCode(xsize, offset);
+        const double distance_cost =
+            prev_cost + GetDistanceCost(cost_model, code);
+        int k;
+        for (k = 1; k < len; ++k) {
+          const double cost_val = distance_cost + GetLengthCost(cost_model, k);
+          if (cost[i + k] > cost_val) {
+            cost[i + k] = (float)cost_val;
+            dist_array[i + k] = k + 1;
+          }
+        }
+        // This if is for speedup only. It roughly doubles the speed, and
+        // makes compression worse by .1 %.
+        if (len >= 128 && code <= min_distance_code) {
+          // Long copy for short distances, let's skip the middle
+          // lookups for better copies.
+          // 1) insert the hashes.
+          if (use_color_cache) {
+            for (k = 0; k < len; ++k) {
+              VP8LColorCacheInsert(&hashers, argb[i + k]);
+            }
+          }
+          // 2) Add to the hash_chain (but cannot add the last pixel)
+          {
+            const int last = (len + i < pix_count - 1) ? len + i
+                                                       : pix_count - 1;
+            for (k = i; k < last; ++k) {
+              HashChainInsert(hash_chain, &argb[k], k);
+            }
+          }
+          // 3) jump.
+          i += len - 1;  // for loop does ++i, thus -1 here.
+          goto next_symbol;
+        }
+      }
+    }
+    if (i < pix_count - 1) {
+      HashChainInsert(hash_chain, &argb[i], i);
+    }
+    {
+      // inserting a literal pixel
+      double cost_val = prev_cost;
+      if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
+        const int ix = VP8LColorCacheGetIndex(&hashers, argb[i]);
+        cost_val += GetCacheCost(cost_model, ix) * mul0;
+      } else {
+        if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
+        cost_val += GetLiteralCost(cost_model, argb[i]) * mul1;
+      }
+      if (cost[i] > cost_val) {
+        cost[i] = (float)cost_val;
+        dist_array[i] = 1;  // only one is inserted.
+      }
+    }
+ next_symbol: ;
+  }
+  // Last pixel still to do, it can only be a single step if not reached
+  // through cheaper means already.
+  ok = 1;
+Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  HashChainDelete(hash_chain);
+  free(cost_model);
+  free(cost);
+  return ok;
+}
+
+// We pack the path at the end of *dist_array and return
+// a pointer to this part of the array. Example:
+// dist_array = [1x2xx3x2] => packed [1x2x1232], chosen_path = [1232]
+static void TraceBackwards(uint32_t* const dist_array,
+                           int dist_array_size,
+                           uint32_t** const chosen_path,
+                           int* const chosen_path_size) {
+  uint32_t* path = dist_array + dist_array_size;
+  uint32_t* cur = dist_array + dist_array_size - 1;
+  while (cur >= dist_array) {
+    const int k = *cur;
+    --path;
+    *path = k;
+    cur -= k;
+  }
+  *chosen_path = path;
+  *chosen_path_size = (int)(dist_array + dist_array_size - path);
+}
+
+static int BackwardReferencesHashChainFollowChosenPath(
+    int xsize, int ysize, const uint32_t* const argb,
+    int quality, int cache_bits,
+    const uint32_t* const chosen_path, int chosen_path_size,
+    VP8LBackwardRefs* const refs) {
+  const int pix_count = xsize * ysize;
+  const int use_color_cache = (cache_bits > 0);
+  int size = 0;
+  int i = 0;
+  int k;
+  int ix;
+  int ok = 0;
+  int cc_init = 0;
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
+  HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
+  VP8LColorCache hashers;
+
+  if (hash_chain == NULL || !HashChainInit(hash_chain, pix_count)) {
+    goto Error;
+  }
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) goto Error;
+  }
+
+  refs->size = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, cache_bits,
+                                &window_size, &iter_pos, &iter_limit);
+  for (ix = 0; ix < chosen_path_size; ++ix, ++size) {
+    int offset = 0;
+    int len = 0;
+    int max_len = chosen_path[ix];
+    if (max_len != 1) {
+      HashChainFindCopy(hash_chain, i, xsize, argb, max_len,
+                        window_size, iter_pos, iter_limit,
+                        &offset, &len);
+      assert(len == max_len);
+      refs->refs[size] = PixOrCopyCreateCopy(offset, len);
+      if (use_color_cache) {
+        for (k = 0; k < len; ++k) {
+          VP8LColorCacheInsert(&hashers, argb[i + k]);
+        }
+      }
+      {
+        const int last = (len < pix_count - 1 - i) ? len : pix_count - 1 - i;
+        for (k = 0; k < last; ++k) {
+          HashChainInsert(hash_chain, &argb[i + k], i + k);
+        }
+      }
+      i += len;
+    } else {
+      if (use_color_cache && VP8LColorCacheContains(&hashers, argb[i])) {
+        // push pixel as a color cache index
+        const int idx = VP8LColorCacheGetIndex(&hashers, argb[i]);
+        refs->refs[size] = PixOrCopyCreateCacheIdx(idx);
+      } else {
+        if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
+        refs->refs[size] = PixOrCopyCreateLiteral(argb[i]);
+      }
+      if (i + 1 < pix_count) {
+        HashChainInsert(hash_chain, &argb[i], i);
+      }
+      ++i;
+    }
+  }
+  assert(size <= refs->max_size);
+  refs->size = size;
+  ok = 1;
+Error:
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  HashChainDelete(hash_chain);
+  return ok;
+}
+
+// Returns 1 on success.
+static int BackwardReferencesTraceBackwards(int xsize, int ysize,
+                                            int recursive_cost_model,
+                                            const uint32_t* const argb,
+                                            int quality, int cache_bits,
+                                            VP8LBackwardRefs* const refs) {
+  int ok = 0;
+  const int dist_array_size = xsize * ysize;
+  uint32_t* chosen_path = NULL;
+  int chosen_path_size = 0;
+  uint32_t* dist_array =
+      (uint32_t*)WebPSafeMalloc((uint64_t)dist_array_size, sizeof(*dist_array));
+
+  if (dist_array == NULL) goto Error;
+
+  if (!BackwardReferencesHashChainDistanceOnly(
+      xsize, ysize, recursive_cost_model, argb, quality, cache_bits,
+      dist_array)) {
+    goto Error;
+  }
+  TraceBackwards(dist_array, dist_array_size, &chosen_path, &chosen_path_size);
+  if (!BackwardReferencesHashChainFollowChosenPath(
+      xsize, ysize, argb, quality, cache_bits, chosen_path, chosen_path_size,
+      refs)) {
+    goto Error;
+  }
+  ok = 1;
+ Error:
+  free(dist_array);
+  return ok;
+}
+
+static void BackwardReferences2DLocality(int xsize,
+                                         VP8LBackwardRefs* const refs) {
+  int i;
+  for (i = 0; i < refs->size; ++i) {
+    if (PixOrCopyIsCopy(&refs->refs[i])) {
+      const int dist = refs->refs[i].argb_or_distance;
+      const int transformed_dist = DistanceToPlaneCode(xsize, dist);
+      refs->refs[i].argb_or_distance = transformed_dist;
+    }
+  }
+}
+
+int VP8LGetBackwardReferences(int width, int height,
+                              const uint32_t* const argb,
+                              int quality, int cache_bits, int use_2d_locality,
+                              VP8LBackwardRefs* const best) {
+  int ok = 0;
+  int lz77_is_useful;
+  VP8LBackwardRefs refs_rle, refs_lz77;
+  const int num_pix = width * height;
+
+  VP8LBackwardRefsAlloc(&refs_rle, num_pix);
+  VP8LBackwardRefsAlloc(&refs_lz77, num_pix);
+  VP8LInitBackwardRefs(best);
+  if (refs_rle.refs == NULL || refs_lz77.refs == NULL) {
+ Error1:
+    VP8LClearBackwardRefs(&refs_rle);
+    VP8LClearBackwardRefs(&refs_lz77);
+    goto End;
+  }
+
+  if (!BackwardReferencesHashChain(width, height, argb, cache_bits, quality,
+                                   &refs_lz77)) {
+    goto End;
+  }
+  // Backward Reference using RLE only.
+  BackwardReferencesRle(width, height, argb, &refs_rle);
+
+  {
+    double bit_cost_lz77, bit_cost_rle;
+    VP8LHistogram* const histo = (VP8LHistogram*)malloc(sizeof(*histo));
+    if (histo == NULL) goto Error1;
+    // Evaluate lz77 coding
+    VP8LHistogramCreate(histo, &refs_lz77, cache_bits);
+    bit_cost_lz77 = VP8LHistogramEstimateBits(histo);
+    // Evaluate RLE coding
+    VP8LHistogramCreate(histo, &refs_rle, cache_bits);
+    bit_cost_rle = VP8LHistogramEstimateBits(histo);
+    // Decide if LZ77 is useful.
+    lz77_is_useful = (bit_cost_lz77 < bit_cost_rle);
+    free(histo);
+  }
+
+  // Choose appropriate backward reference.
+  if (lz77_is_useful) {
+    // TraceBackwards is costly. Don't execute it at lower quality.
+    const int try_lz77_trace_backwards = (quality >= 25);
+    *best = refs_lz77;   // default guess: lz77 is better
+    VP8LClearBackwardRefs(&refs_rle);
+    if (try_lz77_trace_backwards) {
+      // Set recursion level for large images using a color cache.
+      const int recursion_level =
+          (num_pix < 320 * 200) && (cache_bits > 0) ? 1 : 0;
+      VP8LBackwardRefs refs_trace;
+      if (!VP8LBackwardRefsAlloc(&refs_trace, num_pix)) {
+        goto End;
+      }
+      if (BackwardReferencesTraceBackwards(width, height, recursion_level, argb,
+                                           quality, cache_bits, &refs_trace)) {
+        VP8LClearBackwardRefs(&refs_lz77);
+        *best = refs_trace;
+      }
+    }
+  } else {
+    VP8LClearBackwardRefs(&refs_lz77);
+    *best = refs_rle;
+  }
+
+  if (use_2d_locality) BackwardReferences2DLocality(width, best);
+
+  ok = 1;
+
+ End:
+  if (!ok) {
+    VP8LClearBackwardRefs(best);
+  }
+  return ok;
+}
+
+// Returns 1 on success.
+static int ComputeCacheHistogram(const uint32_t* const argb,
+                                 int xsize, int ysize,
+                                 const VP8LBackwardRefs* const refs,
+                                 int cache_bits,
+                                 VP8LHistogram* const histo) {
+  int pixel_index = 0;
+  int i;
+  uint32_t k;
+  VP8LColorCache hashers;
+  const int use_color_cache = (cache_bits > 0);
+  int cc_init = 0;
+
+  if (use_color_cache) {
+    cc_init = VP8LColorCacheInit(&hashers, cache_bits);
+    if (!cc_init) return 0;
+  }
+
+  for (i = 0; i < refs->size; ++i) {
+    const PixOrCopy* const v = &refs->refs[i];
+    if (PixOrCopyIsLiteral(v)) {
+      if (use_color_cache &&
+          VP8LColorCacheContains(&hashers, argb[pixel_index])) {
+        // push pixel as a cache index
+        const int ix = VP8LColorCacheGetIndex(&hashers, argb[pixel_index]);
+        const PixOrCopy token = PixOrCopyCreateCacheIdx(ix);
+        VP8LHistogramAddSinglePixOrCopy(histo, &token);
+      } else {
+        VP8LHistogramAddSinglePixOrCopy(histo, v);
+      }
+    } else {
+      VP8LHistogramAddSinglePixOrCopy(histo, v);
+    }
+    if (use_color_cache) {
+      for (k = 0; k < PixOrCopyLength(v); ++k) {
+        VP8LColorCacheInsert(&hashers, argb[pixel_index + k]);
+      }
+    }
+    pixel_index += PixOrCopyLength(v);
+  }
+  assert(pixel_index == xsize * ysize);
+  (void)xsize;  // xsize is not used in non-debug compilations otherwise.
+  (void)ysize;  // ysize is not used in non-debug compilations otherwise.
+  if (cc_init) VP8LColorCacheClear(&hashers);
+  return 1;
+}
+
+// Returns how many bits are to be used for a color cache.
+int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
+                                      int xsize, int ysize,
+                                      int* const best_cache_bits) {
+  int ok = 0;
+  int cache_bits;
+  double lowest_entropy = 1e99;
+  VP8LBackwardRefs refs;
+  static const double kSmallPenaltyForLargeCache = 4.0;
+  static const int quality = 30;
+  if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize) ||
+      !BackwardReferencesHashChain(xsize, ysize, argb, 0, quality, &refs)) {
+    goto Error;
+  }
+  for (cache_bits = 0; cache_bits <= MAX_COLOR_CACHE_BITS; ++cache_bits) {
+    double cur_entropy;
+    VP8LHistogram histo;
+    VP8LHistogramInit(&histo, cache_bits);
+    ComputeCacheHistogram(argb, xsize, ysize, &refs, cache_bits, &histo);
+    cur_entropy = VP8LHistogramEstimateBits(&histo) +
+        kSmallPenaltyForLargeCache * cache_bits;
+    if (cache_bits == 0 || cur_entropy < lowest_entropy) {
+      *best_cache_bits = cache_bits;
+      lowest_entropy = cur_entropy;
+    }
+  }
+  ok = 1;
+ Error:
+  VP8LClearBackwardRefs(&refs);
+  return ok;
+}
diff --git a/drivers/webp/enc/backward_references.h b/drivers/webp/enc/backward_references.h
new file mode 100644
index 0000000000..e1c75f04f9
--- /dev/null
+++ b/drivers/webp/enc/backward_references.h
@@ -0,0 +1,152 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+//
+
+#ifndef WEBP_ENC_BACKWARD_REFERENCES_H_
+#define WEBP_ENC_BACKWARD_REFERENCES_H_
+
+#include <assert.h>
+#include <stdlib.h>
+#include "../webp/types.h"
+#include "../webp/format_constants.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The spec allows 11, we use 9 bits to reduce memory consumption in encoding.
+// Having 9 instead of 11 only removes about 0.25 % of compression density.
+#define MAX_COLOR_CACHE_BITS 9
+
+// Max ever number of codes we'll use:
+#define PIX_OR_COPY_CODES_MAX \
+    (NUM_LITERAL_CODES + NUM_LENGTH_CODES + (1 << MAX_COLOR_CACHE_BITS))
+
+// -----------------------------------------------------------------------------
+// PixOrCopy
+
+enum Mode {
+  kLiteral,
+  kCacheIdx,
+  kCopy,
+  kNone
+};
+
+typedef struct {
+  // mode as uint8_t to make the memory layout to be exactly 8 bytes.
+  uint8_t mode;
+  uint16_t len;
+  uint32_t argb_or_distance;
+} PixOrCopy;
+
+static WEBP_INLINE PixOrCopy PixOrCopyCreateCopy(uint32_t distance,
+                                                 uint16_t len) {
+  PixOrCopy retval;
+  retval.mode = kCopy;
+  retval.argb_or_distance = distance;
+  retval.len = len;
+  return retval;
+}
+
+static WEBP_INLINE PixOrCopy PixOrCopyCreateCacheIdx(int idx) {
+  PixOrCopy retval;
+  assert(idx >= 0);
+  assert(idx < (1 << MAX_COLOR_CACHE_BITS));
+  retval.mode = kCacheIdx;
+  retval.argb_or_distance = idx;
+  retval.len = 1;
+  return retval;
+}
+
+static WEBP_INLINE PixOrCopy PixOrCopyCreateLiteral(uint32_t argb) {
+  PixOrCopy retval;
+  retval.mode = kLiteral;
+  retval.argb_or_distance = argb;
+  retval.len = 1;
+  return retval;
+}
+
+static WEBP_INLINE int PixOrCopyIsLiteral(const PixOrCopy* const p) {
+  return (p->mode == kLiteral);
+}
+
+static WEBP_INLINE int PixOrCopyIsCacheIdx(const PixOrCopy* const p) {
+  return (p->mode == kCacheIdx);
+}
+
+static WEBP_INLINE int PixOrCopyIsCopy(const PixOrCopy* const p) {
+  return (p->mode == kCopy);
+}
+
+static WEBP_INLINE uint32_t PixOrCopyLiteral(const PixOrCopy* const p,
+                                             int component) {
+  assert(p->mode == kLiteral);
+  return (p->argb_or_distance >> (component * 8)) & 0xff;
+}
+
+static WEBP_INLINE uint32_t PixOrCopyLength(const PixOrCopy* const p) {
+  return p->len;
+}
+
+static WEBP_INLINE uint32_t PixOrCopyArgb(const PixOrCopy* const p) {
+  assert(p->mode == kLiteral);
+  return p->argb_or_distance;
+}
+
+static WEBP_INLINE uint32_t PixOrCopyCacheIdx(const PixOrCopy* const p) {
+  assert(p->mode == kCacheIdx);
+  assert(p->argb_or_distance < (1U << MAX_COLOR_CACHE_BITS));
+  return p->argb_or_distance;
+}
+
+static WEBP_INLINE uint32_t PixOrCopyDistance(const PixOrCopy* const p) {
+  assert(p->mode == kCopy);
+  return p->argb_or_distance;
+}
+
+// -----------------------------------------------------------------------------
+// VP8LBackwardRefs
+
+typedef struct {
+  PixOrCopy* refs;
+  int size;      // currently used
+  int max_size;  // maximum capacity
+} VP8LBackwardRefs;
+
+// Initialize the object. Must be called first. 'refs' can be NULL.
+void VP8LInitBackwardRefs(VP8LBackwardRefs* const refs);
+
+// Release memory and re-initialize the object. 'refs' can be NULL.
+void VP8LClearBackwardRefs(VP8LBackwardRefs* const refs);
+
+// Allocate 'max_size' references. Returns false in case of memory error.
+int VP8LBackwardRefsAlloc(VP8LBackwardRefs* const refs, int max_size);
+
+// -----------------------------------------------------------------------------
+// Main entry points
+
+// Evaluates best possible backward references for specified quality.
+// Further optimize for 2D locality if use_2d_locality flag is set.
+int VP8LGetBackwardReferences(int width, int height,
+                              const uint32_t* const argb,
+                              int quality, int cache_bits, int use_2d_locality,
+                              VP8LBackwardRefs* const best);
+
+// Produce an estimate for a good color cache size for the image.
+int VP8LCalculateEstimateForCacheSize(const uint32_t* const argb,
+                                      int xsize, int ysize,
+                                      int* const best_cache_bits);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // WEBP_ENC_BACKWARD_REFERENCES_H_
diff --git a/drivers/webp/enc/config.c b/drivers/webp/enc/config.c
new file mode 100644
index 0000000000..af7f0b09e8
--- /dev/null
+++ b/drivers/webp/enc/config.c
@@ -0,0 +1,140 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Coding tools configuration
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "../webp/encode.h"
+
+//------------------------------------------------------------------------------
+// WebPConfig
+//------------------------------------------------------------------------------
+
+int WebPConfigInitInternal(WebPConfig* config,
+                           WebPPreset preset, float quality, int version) {
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_ENCODER_ABI_VERSION)) {
+    return 0;   // caller/system version mismatch!
+  }
+  if (config == NULL) return 0;
+
+  config->quality = quality;
+  config->target_size = 0;
+  config->target_PSNR = 0.;
+  config->method = 4;
+  config->sns_strength = 50;
+  config->filter_strength = 60;   // mid-filtering
+  config->filter_sharpness = 0;
+  config->filter_type = 1;        // default: strong (so U/V is filtered too)
+  config->partitions = 0;
+  config->segments = 4;
+  config->pass = 1;
+  config->show_compressed = 0;
+  config->preprocessing = 0;
+  config->autofilter = 0;
+  config->partition_limit = 0;
+  config->alpha_compression = 1;
+  config->alpha_filtering = 1;
+  config->alpha_quality = 100;
+  config->lossless = 0;
+  config->image_hint = WEBP_HINT_DEFAULT;
+  config->emulate_jpeg_size = 0;
+  config->thread_level = 0;
+  config->low_memory = 0;
+
+  // TODO(skal): tune.
+  switch (preset) {
+    case WEBP_PRESET_PICTURE:
+      config->sns_strength = 80;
+      config->filter_sharpness = 4;
+      config->filter_strength = 35;
+      config->preprocessing &= ~2;   // no dithering
+      break;
+    case WEBP_PRESET_PHOTO:
+      config->sns_strength = 80;
+      config->filter_sharpness = 3;
+      config->filter_strength = 30;
+      config->preprocessing |= 2;
+      break;
+    case WEBP_PRESET_DRAWING:
+      config->sns_strength = 25;
+      config->filter_sharpness = 6;
+      config->filter_strength = 10;
+      break;
+    case WEBP_PRESET_ICON:
+      config->sns_strength = 0;
+      config->filter_strength = 0;   // disable filtering to retain sharpness
+      config->preprocessing &= ~2;   // no dithering
+      break;
+    case WEBP_PRESET_TEXT:
+      config->sns_strength = 0;
+      config->filter_strength = 0;   // disable filtering to retain sharpness
+      config->preprocessing &= ~2;   // no dithering
+      config->segments = 2;
+      break;
+    case WEBP_PRESET_DEFAULT:
+    default:
+      break;
+  }
+  return WebPValidateConfig(config);
+}
+
+int WebPValidateConfig(const WebPConfig* config) {
+  if (config == NULL) return 0;
+  if (config->quality < 0 || config->quality > 100)
+    return 0;
+  if (config->target_size < 0)
+    return 0;
+  if (config->target_PSNR < 0)
+    return 0;
+  if (config->method < 0 || config->method > 6)
+    return 0;
+  if (config->segments < 1 || config->segments > 4)
+    return 0;
+  if (config->sns_strength < 0 || config->sns_strength > 100)
+    return 0;
+  if (config->filter_strength < 0 || config->filter_strength > 100)
+    return 0;
+  if (config->filter_sharpness < 0 || config->filter_sharpness > 7)
+    return 0;
+  if (config->filter_type < 0 || config->filter_type > 1)
+    return 0;
+  if (config->autofilter < 0 || config->autofilter > 1)
+    return 0;
+  if (config->pass < 1 || config->pass > 10)
+    return 0;
+  if (config->show_compressed < 0 || config->show_compressed > 1)
+    return 0;
+  if (config->preprocessing < 0 || config->preprocessing > 3)
+    return 0;
+  if (config->partitions < 0 || config->partitions > 3)
+    return 0;
+  if (config->partition_limit < 0 || config->partition_limit > 100)
+    return 0;
+  if (config->alpha_compression < 0)
+    return 0;
+  if (config->alpha_filtering < 0)
+    return 0;
+  if (config->alpha_quality < 0 || config->alpha_quality > 100)
+    return 0;
+  if (config->lossless < 0 || config->lossless > 1)
+    return 0;
+  if (config->image_hint >= WEBP_HINT_LAST)
+    return 0;
+  if (config->emulate_jpeg_size < 0 || config->emulate_jpeg_size > 1)
+    return 0;
+  if (config->thread_level < 0 || config->thread_level > 1)
+    return 0;
+  if (config->low_memory < 0 || config->low_memory > 1)
+    return 0;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/enc/cost.c b/drivers/webp/enc/cost.c
new file mode 100644
index 0000000000..09699f8044
--- /dev/null
+++ b/drivers/webp/enc/cost.c
@@ -0,0 +1,489 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Cost tables for level and modes
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./cost.h"
+
+//------------------------------------------------------------------------------
+// Boolean-cost cost table
+
+const uint16_t VP8EntropyCost[256] = {
+  1792, 1792, 1792, 1536, 1536, 1408, 1366, 1280, 1280, 1216,
+  1178, 1152, 1110, 1076, 1061, 1024, 1024,  992,  968,  951,
+   939,  911,  896,  878,  871,  854,  838,  820,  811,  794,
+   786,  768,  768,  752,  740,  732,  720,  709,  704,  690,
+   683,  672,  666,  655,  647,  640,  631,  622,  615,  607,
+   598,  592,  586,  576,  572,  564,  559,  555,  547,  541,
+   534,  528,  522,  512,  512,  504,  500,  494,  488,  483,
+   477,  473,  467,  461,  458,  452,  448,  443,  438,  434,
+   427,  424,  419,  415,  410,  406,  403,  399,  394,  390,
+   384,  384,  377,  374,  370,  366,  362,  359,  355,  351,
+   347,  342,  342,  336,  333,  330,  326,  323,  320,  316,
+   312,  308,  305,  302,  299,  296,  293,  288,  287,  283,
+   280,  277,  274,  272,  268,  266,  262,  256,  256,  256,
+   251,  248,  245,  242,  240,  237,  234,  232,  228,  226,
+   223,  221,  218,  216,  214,  211,  208,  205,  203,  201,
+   198,  196,  192,  191,  188,  187,  183,  181,  179,  176,
+   175,  171,  171,  168,  165,  163,  160,  159,  156,  154,
+   152,  150,  148,  146,  144,  142,  139,  138,  135,  133,
+   131,  128,  128,  125,  123,  121,  119,  117,  115,  113,
+   111,  110,  107,  105,  103,  102,  100,   98,   96,   94,
+    92,   91,   89,   86,   86,   83,   82,   80,   77,   76,
+    74,   73,   71,   69,   67,   66,   64,   63,   61,   59,
+    57,   55,   54,   52,   51,   49,   47,   46,   44,   43,
+    41,   40,   38,   36,   35,   33,   32,   30,   29,   27,
+    25,   24,   22,   21,   19,   18,   16,   15,   13,   12,
+    10,    9,    7,    6,    4,    3
+};
+
+//------------------------------------------------------------------------------
+// Level cost tables
+
+// For each given level, the following table gives the pattern of contexts to
+// use for coding it (in [][0]) as well as the bit value to use for each
+// context (in [][1]).
+const uint16_t VP8LevelCodes[MAX_VARIABLE_LEVEL][2] = {
+                  {0x001, 0x000}, {0x007, 0x001}, {0x00f, 0x005},
+  {0x00f, 0x00d}, {0x033, 0x003}, {0x033, 0x003}, {0x033, 0x023},
+  {0x033, 0x023}, {0x033, 0x023}, {0x033, 0x023}, {0x0d3, 0x013},
+  {0x0d3, 0x013}, {0x0d3, 0x013}, {0x0d3, 0x013}, {0x0d3, 0x013},
+  {0x0d3, 0x013}, {0x0d3, 0x013}, {0x0d3, 0x013}, {0x0d3, 0x093},
+  {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093},
+  {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093},
+  {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093},
+  {0x0d3, 0x093}, {0x0d3, 0x093}, {0x0d3, 0x093}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053},
+  {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x053}, {0x153, 0x153}
+};
+
+// fixed costs for coding levels, deduce from the coding tree.
+// This is only the part that doesn't depend on the probability state.
+const uint16_t VP8LevelFixedCosts[MAX_LEVEL + 1] = {
+     0,  256,  256,  256,  256,  432,  618,  630,
+   731,  640,  640,  828,  901,  948, 1021, 1101,
+  1174, 1221, 1294, 1042, 1085, 1115, 1158, 1202,
+  1245, 1275, 1318, 1337, 1380, 1410, 1453, 1497,
+  1540, 1570, 1613, 1280, 1295, 1317, 1332, 1358,
+  1373, 1395, 1410, 1454, 1469, 1491, 1506, 1532,
+  1547, 1569, 1584, 1601, 1616, 1638, 1653, 1679,
+  1694, 1716, 1731, 1775, 1790, 1812, 1827, 1853,
+  1868, 1890, 1905, 1727, 1733, 1742, 1748, 1759,
+  1765, 1774, 1780, 1800, 1806, 1815, 1821, 1832,
+  1838, 1847, 1853, 1878, 1884, 1893, 1899, 1910,
+  1916, 1925, 1931, 1951, 1957, 1966, 1972, 1983,
+  1989, 1998, 2004, 2027, 2033, 2042, 2048, 2059,
+  2065, 2074, 2080, 2100, 2106, 2115, 2121, 2132,
+  2138, 2147, 2153, 2178, 2184, 2193, 2199, 2210,
+  2216, 2225, 2231, 2251, 2257, 2266, 2272, 2283,
+  2289, 2298, 2304, 2168, 2174, 2183, 2189, 2200,
+  2206, 2215, 2221, 2241, 2247, 2256, 2262, 2273,
+  2279, 2288, 2294, 2319, 2325, 2334, 2340, 2351,
+  2357, 2366, 2372, 2392, 2398, 2407, 2413, 2424,
+  2430, 2439, 2445, 2468, 2474, 2483, 2489, 2500,
+  2506, 2515, 2521, 2541, 2547, 2556, 2562, 2573,
+  2579, 2588, 2594, 2619, 2625, 2634, 2640, 2651,
+  2657, 2666, 2672, 2692, 2698, 2707, 2713, 2724,
+  2730, 2739, 2745, 2540, 2546, 2555, 2561, 2572,
+  2578, 2587, 2593, 2613, 2619, 2628, 2634, 2645,
+  2651, 2660, 2666, 2691, 2697, 2706, 2712, 2723,
+  2729, 2738, 2744, 2764, 2770, 2779, 2785, 2796,
+  2802, 2811, 2817, 2840, 2846, 2855, 2861, 2872,
+  2878, 2887, 2893, 2913, 2919, 2928, 2934, 2945,
+  2951, 2960, 2966, 2991, 2997, 3006, 3012, 3023,
+  3029, 3038, 3044, 3064, 3070, 3079, 3085, 3096,
+  3102, 3111, 3117, 2981, 2987, 2996, 3002, 3013,
+  3019, 3028, 3034, 3054, 3060, 3069, 3075, 3086,
+  3092, 3101, 3107, 3132, 3138, 3147, 3153, 3164,
+  3170, 3179, 3185, 3205, 3211, 3220, 3226, 3237,
+  3243, 3252, 3258, 3281, 3287, 3296, 3302, 3313,
+  3319, 3328, 3334, 3354, 3360, 3369, 3375, 3386,
+  3392, 3401, 3407, 3432, 3438, 3447, 3453, 3464,
+  3470, 3479, 3485, 3505, 3511, 3520, 3526, 3537,
+  3543, 3552, 3558, 2816, 2822, 2831, 2837, 2848,
+  2854, 2863, 2869, 2889, 2895, 2904, 2910, 2921,
+  2927, 2936, 2942, 2967, 2973, 2982, 2988, 2999,
+  3005, 3014, 3020, 3040, 3046, 3055, 3061, 3072,
+  3078, 3087, 3093, 3116, 3122, 3131, 3137, 3148,
+  3154, 3163, 3169, 3189, 3195, 3204, 3210, 3221,
+  3227, 3236, 3242, 3267, 3273, 3282, 3288, 3299,
+  3305, 3314, 3320, 3340, 3346, 3355, 3361, 3372,
+  3378, 3387, 3393, 3257, 3263, 3272, 3278, 3289,
+  3295, 3304, 3310, 3330, 3336, 3345, 3351, 3362,
+  3368, 3377, 3383, 3408, 3414, 3423, 3429, 3440,
+  3446, 3455, 3461, 3481, 3487, 3496, 3502, 3513,
+  3519, 3528, 3534, 3557, 3563, 3572, 3578, 3589,
+  3595, 3604, 3610, 3630, 3636, 3645, 3651, 3662,
+  3668, 3677, 3683, 3708, 3714, 3723, 3729, 3740,
+  3746, 3755, 3761, 3781, 3787, 3796, 3802, 3813,
+  3819, 3828, 3834, 3629, 3635, 3644, 3650, 3661,
+  3667, 3676, 3682, 3702, 3708, 3717, 3723, 3734,
+  3740, 3749, 3755, 3780, 3786, 3795, 3801, 3812,
+  3818, 3827, 3833, 3853, 3859, 3868, 3874, 3885,
+  3891, 3900, 3906, 3929, 3935, 3944, 3950, 3961,
+  3967, 3976, 3982, 4002, 4008, 4017, 4023, 4034,
+  4040, 4049, 4055, 4080, 4086, 4095, 4101, 4112,
+  4118, 4127, 4133, 4153, 4159, 4168, 4174, 4185,
+  4191, 4200, 4206, 4070, 4076, 4085, 4091, 4102,
+  4108, 4117, 4123, 4143, 4149, 4158, 4164, 4175,
+  4181, 4190, 4196, 4221, 4227, 4236, 4242, 4253,
+  4259, 4268, 4274, 4294, 4300, 4309, 4315, 4326,
+  4332, 4341, 4347, 4370, 4376, 4385, 4391, 4402,
+  4408, 4417, 4423, 4443, 4449, 4458, 4464, 4475,
+  4481, 4490, 4496, 4521, 4527, 4536, 4542, 4553,
+  4559, 4568, 4574, 4594, 4600, 4609, 4615, 4626,
+  4632, 4641, 4647, 3515, 3521, 3530, 3536, 3547,
+  3553, 3562, 3568, 3588, 3594, 3603, 3609, 3620,
+  3626, 3635, 3641, 3666, 3672, 3681, 3687, 3698,
+  3704, 3713, 3719, 3739, 3745, 3754, 3760, 3771,
+  3777, 3786, 3792, 3815, 3821, 3830, 3836, 3847,
+  3853, 3862, 3868, 3888, 3894, 3903, 3909, 3920,
+  3926, 3935, 3941, 3966, 3972, 3981, 3987, 3998,
+  4004, 4013, 4019, 4039, 4045, 4054, 4060, 4071,
+  4077, 4086, 4092, 3956, 3962, 3971, 3977, 3988,
+  3994, 4003, 4009, 4029, 4035, 4044, 4050, 4061,
+  4067, 4076, 4082, 4107, 4113, 4122, 4128, 4139,
+  4145, 4154, 4160, 4180, 4186, 4195, 4201, 4212,
+  4218, 4227, 4233, 4256, 4262, 4271, 4277, 4288,
+  4294, 4303, 4309, 4329, 4335, 4344, 4350, 4361,
+  4367, 4376, 4382, 4407, 4413, 4422, 4428, 4439,
+  4445, 4454, 4460, 4480, 4486, 4495, 4501, 4512,
+  4518, 4527, 4533, 4328, 4334, 4343, 4349, 4360,
+  4366, 4375, 4381, 4401, 4407, 4416, 4422, 4433,
+  4439, 4448, 4454, 4479, 4485, 4494, 4500, 4511,
+  4517, 4526, 4532, 4552, 4558, 4567, 4573, 4584,
+  4590, 4599, 4605, 4628, 4634, 4643, 4649, 4660,
+  4666, 4675, 4681, 4701, 4707, 4716, 4722, 4733,
+  4739, 4748, 4754, 4779, 4785, 4794, 4800, 4811,
+  4817, 4826, 4832, 4852, 4858, 4867, 4873, 4884,
+  4890, 4899, 4905, 4769, 4775, 4784, 4790, 4801,
+  4807, 4816, 4822, 4842, 4848, 4857, 4863, 4874,
+  4880, 4889, 4895, 4920, 4926, 4935, 4941, 4952,
+  4958, 4967, 4973, 4993, 4999, 5008, 5014, 5025,
+  5031, 5040, 5046, 5069, 5075, 5084, 5090, 5101,
+  5107, 5116, 5122, 5142, 5148, 5157, 5163, 5174,
+  5180, 5189, 5195, 5220, 5226, 5235, 5241, 5252,
+  5258, 5267, 5273, 5293, 5299, 5308, 5314, 5325,
+  5331, 5340, 5346, 4604, 4610, 4619, 4625, 4636,
+  4642, 4651, 4657, 4677, 4683, 4692, 4698, 4709,
+  4715, 4724, 4730, 4755, 4761, 4770, 4776, 4787,
+  4793, 4802, 4808, 4828, 4834, 4843, 4849, 4860,
+  4866, 4875, 4881, 4904, 4910, 4919, 4925, 4936,
+  4942, 4951, 4957, 4977, 4983, 4992, 4998, 5009,
+  5015, 5024, 5030, 5055, 5061, 5070, 5076, 5087,
+  5093, 5102, 5108, 5128, 5134, 5143, 5149, 5160,
+  5166, 5175, 5181, 5045, 5051, 5060, 5066, 5077,
+  5083, 5092, 5098, 5118, 5124, 5133, 5139, 5150,
+  5156, 5165, 5171, 5196, 5202, 5211, 5217, 5228,
+  5234, 5243, 5249, 5269, 5275, 5284, 5290, 5301,
+  5307, 5316, 5322, 5345, 5351, 5360, 5366, 5377,
+  5383, 5392, 5398, 5418, 5424, 5433, 5439, 5450,
+  5456, 5465, 5471, 5496, 5502, 5511, 5517, 5528,
+  5534, 5543, 5549, 5569, 5575, 5584, 5590, 5601,
+  5607, 5616, 5622, 5417, 5423, 5432, 5438, 5449,
+  5455, 5464, 5470, 5490, 5496, 5505, 5511, 5522,
+  5528, 5537, 5543, 5568, 5574, 5583, 5589, 5600,
+  5606, 5615, 5621, 5641, 5647, 5656, 5662, 5673,
+  5679, 5688, 5694, 5717, 5723, 5732, 5738, 5749,
+  5755, 5764, 5770, 5790, 5796, 5805, 5811, 5822,
+  5828, 5837, 5843, 5868, 5874, 5883, 5889, 5900,
+  5906, 5915, 5921, 5941, 5947, 5956, 5962, 5973,
+  5979, 5988, 5994, 5858, 5864, 5873, 5879, 5890,
+  5896, 5905, 5911, 5931, 5937, 5946, 5952, 5963,
+  5969, 5978, 5984, 6009, 6015, 6024, 6030, 6041,
+  6047, 6056, 6062, 6082, 6088, 6097, 6103, 6114,
+  6120, 6129, 6135, 6158, 6164, 6173, 6179, 6190,
+  6196, 6205, 6211, 6231, 6237, 6246, 6252, 6263,
+  6269, 6278, 6284, 6309, 6315, 6324, 6330, 6341,
+  6347, 6356, 6362, 6382, 6388, 6397, 6403, 6414,
+  6420, 6429, 6435, 3515, 3521, 3530, 3536, 3547,
+  3553, 3562, 3568, 3588, 3594, 3603, 3609, 3620,
+  3626, 3635, 3641, 3666, 3672, 3681, 3687, 3698,
+  3704, 3713, 3719, 3739, 3745, 3754, 3760, 3771,
+  3777, 3786, 3792, 3815, 3821, 3830, 3836, 3847,
+  3853, 3862, 3868, 3888, 3894, 3903, 3909, 3920,
+  3926, 3935, 3941, 3966, 3972, 3981, 3987, 3998,
+  4004, 4013, 4019, 4039, 4045, 4054, 4060, 4071,
+  4077, 4086, 4092, 3956, 3962, 3971, 3977, 3988,
+  3994, 4003, 4009, 4029, 4035, 4044, 4050, 4061,
+  4067, 4076, 4082, 4107, 4113, 4122, 4128, 4139,
+  4145, 4154, 4160, 4180, 4186, 4195, 4201, 4212,
+  4218, 4227, 4233, 4256, 4262, 4271, 4277, 4288,
+  4294, 4303, 4309, 4329, 4335, 4344, 4350, 4361,
+  4367, 4376, 4382, 4407, 4413, 4422, 4428, 4439,
+  4445, 4454, 4460, 4480, 4486, 4495, 4501, 4512,
+  4518, 4527, 4533, 4328, 4334, 4343, 4349, 4360,
+  4366, 4375, 4381, 4401, 4407, 4416, 4422, 4433,
+  4439, 4448, 4454, 4479, 4485, 4494, 4500, 4511,
+  4517, 4526, 4532, 4552, 4558, 4567, 4573, 4584,
+  4590, 4599, 4605, 4628, 4634, 4643, 4649, 4660,
+  4666, 4675, 4681, 4701, 4707, 4716, 4722, 4733,
+  4739, 4748, 4754, 4779, 4785, 4794, 4800, 4811,
+  4817, 4826, 4832, 4852, 4858, 4867, 4873, 4884,
+  4890, 4899, 4905, 4769, 4775, 4784, 4790, 4801,
+  4807, 4816, 4822, 4842, 4848, 4857, 4863, 4874,
+  4880, 4889, 4895, 4920, 4926, 4935, 4941, 4952,
+  4958, 4967, 4973, 4993, 4999, 5008, 5014, 5025,
+  5031, 5040, 5046, 5069, 5075, 5084, 5090, 5101,
+  5107, 5116, 5122, 5142, 5148, 5157, 5163, 5174,
+  5180, 5189, 5195, 5220, 5226, 5235, 5241, 5252,
+  5258, 5267, 5273, 5293, 5299, 5308, 5314, 5325,
+  5331, 5340, 5346, 4604, 4610, 4619, 4625, 4636,
+  4642, 4651, 4657, 4677, 4683, 4692, 4698, 4709,
+  4715, 4724, 4730, 4755, 4761, 4770, 4776, 4787,
+  4793, 4802, 4808, 4828, 4834, 4843, 4849, 4860,
+  4866, 4875, 4881, 4904, 4910, 4919, 4925, 4936,
+  4942, 4951, 4957, 4977, 4983, 4992, 4998, 5009,
+  5015, 5024, 5030, 5055, 5061, 5070, 5076, 5087,
+  5093, 5102, 5108, 5128, 5134, 5143, 5149, 5160,
+  5166, 5175, 5181, 5045, 5051, 5060, 5066, 5077,
+  5083, 5092, 5098, 5118, 5124, 5133, 5139, 5150,
+  5156, 5165, 5171, 5196, 5202, 5211, 5217, 5228,
+  5234, 5243, 5249, 5269, 5275, 5284, 5290, 5301,
+  5307, 5316, 5322, 5345, 5351, 5360, 5366, 5377,
+  5383, 5392, 5398, 5418, 5424, 5433, 5439, 5450,
+  5456, 5465, 5471, 5496, 5502, 5511, 5517, 5528,
+  5534, 5543, 5549, 5569, 5575, 5584, 5590, 5601,
+  5607, 5616, 5622, 5417, 5423, 5432, 5438, 5449,
+  5455, 5464, 5470, 5490, 5496, 5505, 5511, 5522,
+  5528, 5537, 5543, 5568, 5574, 5583, 5589, 5600,
+  5606, 5615, 5621, 5641, 5647, 5656, 5662, 5673,
+  5679, 5688, 5694, 5717, 5723, 5732, 5738, 5749,
+  5755, 5764, 5770, 5790, 5796, 5805, 5811, 5822,
+  5828, 5837, 5843, 5868, 5874, 5883, 5889, 5900,
+  5906, 5915, 5921, 5941, 5947, 5956, 5962, 5973,
+  5979, 5988, 5994, 5858, 5864, 5873, 5879, 5890,
+  5896, 5905, 5911, 5931, 5937, 5946, 5952, 5963,
+  5969, 5978, 5984, 6009, 6015, 6024, 6030, 6041,
+  6047, 6056, 6062, 6082, 6088, 6097, 6103, 6114,
+  6120, 6129, 6135, 6158, 6164, 6173, 6179, 6190,
+  6196, 6205, 6211, 6231, 6237, 6246, 6252, 6263,
+  6269, 6278, 6284, 6309, 6315, 6324, 6330, 6341,
+  6347, 6356, 6362, 6382, 6388, 6397, 6403, 6414,
+  6420, 6429, 6435, 5303, 5309, 5318, 5324, 5335,
+  5341, 5350, 5356, 5376, 5382, 5391, 5397, 5408,
+  5414, 5423, 5429, 5454, 5460, 5469, 5475, 5486,
+  5492, 5501, 5507, 5527, 5533, 5542, 5548, 5559,
+  5565, 5574, 5580, 5603, 5609, 5618, 5624, 5635,
+  5641, 5650, 5656, 5676, 5682, 5691, 5697, 5708,
+  5714, 5723, 5729, 5754, 5760, 5769, 5775, 5786,
+  5792, 5801, 5807, 5827, 5833, 5842, 5848, 5859,
+  5865, 5874, 5880, 5744, 5750, 5759, 5765, 5776,
+  5782, 5791, 5797, 5817, 5823, 5832, 5838, 5849,
+  5855, 5864, 5870, 5895, 5901, 5910, 5916, 5927,
+  5933, 5942, 5948, 5968, 5974, 5983, 5989, 6000,
+  6006, 6015, 6021, 6044, 6050, 6059, 6065, 6076,
+  6082, 6091, 6097, 6117, 6123, 6132, 6138, 6149,
+  6155, 6164, 6170, 6195, 6201, 6210, 6216, 6227,
+  6233, 6242, 6248, 6268, 6274, 6283, 6289, 6300,
+  6306, 6315, 6321, 6116, 6122, 6131, 6137, 6148,
+  6154, 6163, 6169, 6189, 6195, 6204, 6210, 6221,
+  6227, 6236, 6242, 6267, 6273, 6282, 6288, 6299,
+  6305, 6314, 6320, 6340, 6346, 6355, 6361, 6372,
+  6378, 6387, 6393, 6416, 6422, 6431, 6437, 6448,
+  6454, 6463, 6469, 6489, 6495, 6504, 6510, 6521,
+  6527, 6536, 6542, 6567, 6573, 6582, 6588, 6599,
+  6605, 6614, 6620, 6640, 6646, 6655, 6661, 6672,
+  6678, 6687, 6693, 6557, 6563, 6572, 6578, 6589,
+  6595, 6604, 6610, 6630, 6636, 6645, 6651, 6662,
+  6668, 6677, 6683, 6708, 6714, 6723, 6729, 6740,
+  6746, 6755, 6761, 6781, 6787, 6796, 6802, 6813,
+  6819, 6828, 6834, 6857, 6863, 6872, 6878, 6889,
+  6895, 6904, 6910, 6930, 6936, 6945, 6951, 6962,
+  6968, 6977, 6983, 7008, 7014, 7023, 7029, 7040,
+  7046, 7055, 7061, 7081, 7087, 7096, 7102, 7113,
+  7119, 7128, 7134, 6392, 6398, 6407, 6413, 6424,
+  6430, 6439, 6445, 6465, 6471, 6480, 6486, 6497,
+  6503, 6512, 6518, 6543, 6549, 6558, 6564, 6575,
+  6581, 6590, 6596, 6616, 6622, 6631, 6637, 6648,
+  6654, 6663, 6669, 6692, 6698, 6707, 6713, 6724,
+  6730, 6739, 6745, 6765, 6771, 6780, 6786, 6797,
+  6803, 6812, 6818, 6843, 6849, 6858, 6864, 6875,
+  6881, 6890, 6896, 6916, 6922, 6931, 6937, 6948,
+  6954, 6963, 6969, 6833, 6839, 6848, 6854, 6865,
+  6871, 6880, 6886, 6906, 6912, 6921, 6927, 6938,
+  6944, 6953, 6959, 6984, 6990, 6999, 7005, 7016,
+  7022, 7031, 7037, 7057, 7063, 7072, 7078, 7089,
+  7095, 7104, 7110, 7133, 7139, 7148, 7154, 7165,
+  7171, 7180, 7186, 7206, 7212, 7221, 7227, 7238,
+  7244, 7253, 7259, 7284, 7290, 7299, 7305, 7316,
+  7322, 7331, 7337, 7357, 7363, 7372, 7378, 7389,
+  7395, 7404, 7410, 7205, 7211, 7220, 7226, 7237,
+  7243, 7252, 7258, 7278, 7284, 7293, 7299, 7310,
+  7316, 7325, 7331, 7356, 7362, 7371, 7377, 7388,
+  7394, 7403, 7409, 7429, 7435, 7444, 7450, 7461,
+  7467, 7476, 7482, 7505, 7511, 7520, 7526, 7537,
+  7543, 7552, 7558, 7578, 7584, 7593, 7599, 7610,
+  7616, 7625, 7631, 7656, 7662, 7671, 7677, 7688,
+  7694, 7703, 7709, 7729, 7735, 7744, 7750, 7761
+};
+
+static int VariableLevelCost(int level, const uint8_t probas[NUM_PROBAS]) {
+  int pattern = VP8LevelCodes[level - 1][0];
+  int bits = VP8LevelCodes[level - 1][1];
+  int cost = 0;
+  int i;
+  for (i = 2; pattern; ++i) {
+    if (pattern & 1) {
+      cost += VP8BitCost(bits & 1, probas[i]);
+    }
+    bits >>= 1;
+    pattern >>= 1;
+  }
+  return cost;
+}
+
+//------------------------------------------------------------------------------
+// Pre-calc level costs once for all
+
+void VP8CalculateLevelCosts(VP8Proba* const proba) {
+  int ctype, band, ctx;
+
+  if (!proba->dirty_) return;  // nothing to do.
+
+  for (ctype = 0; ctype < NUM_TYPES; ++ctype) {
+    for (band = 0; band < NUM_BANDS; ++band) {
+      for (ctx = 0; ctx < NUM_CTX; ++ctx) {
+        const uint8_t* const p = proba->coeffs_[ctype][band][ctx];
+        uint16_t* const table = proba->level_cost_[ctype][band][ctx];
+        const int cost_base = VP8BitCost(1, p[1]);
+        int v;
+        table[0] = VP8BitCost(0, p[1]);
+        for (v = 1; v <= MAX_VARIABLE_LEVEL; ++v) {
+          table[v] = cost_base + VariableLevelCost(v, p);
+        }
+        // Starting at level 67 and up, the variable part of the cost is
+        // actually constant.
+      }
+    }
+  }
+  proba->dirty_ = 0;
+}
+
+//------------------------------------------------------------------------------
+// Mode cost tables.
+
+// These are the fixed probabilities (in the coding trees) turned into bit-cost
+// by calling VP8BitCost().
+const uint16_t VP8FixedCostsUV[4] = { 302, 984, 439, 642 };
+// note: these values include the fixed VP8BitCost(1, 145) mode selection cost.
+const uint16_t VP8FixedCostsI16[4] = { 663, 919, 872, 919 };
+const uint16_t VP8FixedCostsI4[NUM_BMODES][NUM_BMODES][NUM_BMODES] = {
+  { {   40, 1151, 1723, 1874, 2103, 2019, 1628, 1777, 2226, 2137 },
+    {  192,  469, 1296, 1308, 1849, 1794, 1781, 1703, 1713, 1522 },
+    {  142,  910,  762, 1684, 1849, 1576, 1460, 1305, 1801, 1657 },
+    {  559,  641, 1370,  421, 1182, 1569, 1612, 1725,  863, 1007 },
+    {  299, 1059, 1256, 1108,  636, 1068, 1581, 1883,  869, 1142 },
+    {  277, 1111,  707, 1362, 1089,  672, 1603, 1541, 1545, 1291 },
+    {  214,  781, 1609, 1303, 1632, 2229,  726, 1560, 1713,  918 },
+    {  152, 1037, 1046, 1759, 1983, 2174, 1358,  742, 1740, 1390 },
+    {  512, 1046, 1420,  753,  752, 1297, 1486, 1613,  460, 1207 },
+    {  424,  827, 1362,  719, 1462, 1202, 1199, 1476, 1199,  538 } },
+  { {  240,  402, 1134, 1491, 1659, 1505, 1517, 1555, 1979, 2099 },
+    {  467,  242,  960, 1232, 1714, 1620, 1834, 1570, 1676, 1391 },
+    {  500,  455,  463, 1507, 1699, 1282, 1564,  982, 2114, 2114 },
+    {  672,  643, 1372,  331, 1589, 1667, 1453, 1938,  996,  876 },
+    {  458,  783, 1037,  911,  738,  968, 1165, 1518,  859, 1033 },
+    {  504,  815,  504, 1139, 1219,  719, 1506, 1085, 1268, 1268 },
+    {  333,  630, 1445, 1239, 1883, 3672,  799, 1548, 1865,  598 },
+    {  399,  644,  746, 1342, 1856, 1350, 1493,  613, 1855, 1015 },
+    {  622,  749, 1205,  608, 1066, 1408, 1290, 1406,  546,  971 },
+    {  500,  753, 1041,  668, 1230, 1617, 1297, 1425, 1383,  523 } },
+  { {  394,  553,  523, 1502, 1536,  981, 1608, 1142, 1666, 2181 },
+    {  655,  430,  375, 1411, 1861, 1220, 1677, 1135, 1978, 1553 },
+    {  690,  640,  245, 1954, 2070, 1194, 1528,  982, 1972, 2232 },
+    {  559,  834,  741,  867, 1131,  980, 1225,  852, 1092,  784 },
+    {  690,  875,  516,  959,  673,  894, 1056, 1190, 1528, 1126 },
+    {  740,  951,  384, 1277, 1177,  492, 1579, 1155, 1846, 1513 },
+    {  323,  775, 1062, 1776, 3062, 1274,  813, 1188, 1372,  655 },
+    {  488,  971,  484, 1767, 1515, 1775, 1115,  503, 1539, 1461 },
+    {  740, 1006,  998,  709,  851, 1230, 1337,  788,  741,  721 },
+    {  522, 1073,  573, 1045, 1346,  887, 1046, 1146, 1203,  697 } },
+  { {  105,  864, 1442, 1009, 1934, 1840, 1519, 1920, 1673, 1579 },
+    {  534,  305, 1193,  683, 1388, 2164, 1802, 1894, 1264, 1170 },
+    {  305,  518,  877, 1108, 1426, 3215, 1425, 1064, 1320, 1242 },
+    {  683,  732, 1927,  257, 1493, 2048, 1858, 1552, 1055,  947 },
+    {  394,  814, 1024,  660,  959, 1556, 1282, 1289,  893, 1047 },
+    {  528,  615,  996,  940, 1201,  635, 1094, 2515,  803, 1358 },
+    {  347,  614, 1609, 1187, 3133, 1345, 1007, 1339, 1017,  667 },
+    {  218,  740,  878, 1605, 3650, 3650, 1345,  758, 1357, 1617 },
+    {  672,  750, 1541,  558, 1257, 1599, 1870, 2135,  402, 1087 },
+    {  592,  684, 1161,  430, 1092, 1497, 1475, 1489, 1095,  822 } },
+  { {  228, 1056, 1059, 1368,  752,  982, 1512, 1518,  987, 1782 },
+    {  494,  514,  818,  942,  965,  892, 1610, 1356, 1048, 1363 },
+    {  512,  648,  591, 1042,  761,  991, 1196, 1454, 1309, 1463 },
+    {  683,  749, 1043,  676,  841, 1396, 1133, 1138,  654,  939 },
+    {  622, 1101, 1126,  994,  361, 1077, 1203, 1318,  877, 1219 },
+    {  631, 1068,  857, 1650,  651,  477, 1650, 1419,  828, 1170 },
+    {  555,  727, 1068, 1335, 3127, 1339,  820, 1331, 1077,  429 },
+    {  504,  879,  624, 1398,  889,  889, 1392,  808,  891, 1406 },
+    {  683, 1602, 1289,  977,  578,  983, 1280, 1708,  406, 1122 },
+    {  399,  865, 1433, 1070, 1072,  764,  968, 1477, 1223,  678 } },
+  { {  333,  760,  935, 1638, 1010,  529, 1646, 1410, 1472, 2219 },
+    {  512,  494,  750, 1160, 1215,  610, 1870, 1868, 1628, 1169 },
+    {  572,  646,  492, 1934, 1208,  603, 1580, 1099, 1398, 1995 },
+    {  786,  789,  942,  581, 1018,  951, 1599, 1207,  731,  768 },
+    {  690, 1015,  672, 1078,  582,  504, 1693, 1438, 1108, 2897 },
+    {  768, 1267,  571, 2005, 1243,  244, 2881, 1380, 1786, 1453 },
+    {  452,  899, 1293,  903, 1311, 3100,  465, 1311, 1319,  813 },
+    {  394,  927,  942, 1103, 1358, 1104,  946,  593, 1363, 1109 },
+    {  559, 1005, 1007, 1016,  658, 1173, 1021, 1164,  623, 1028 },
+    {  564,  796,  632, 1005, 1014,  863, 2316, 1268,  938,  764 } },
+  { {  266,  606, 1098, 1228, 1497, 1243,  948, 1030, 1734, 1461 },
+    {  366,  585,  901, 1060, 1407, 1247,  876, 1134, 1620, 1054 },
+    {  452,  565,  542, 1729, 1479, 1479, 1016,  886, 2938, 1150 },
+    {  555, 1088, 1533,  950, 1354,  895,  834, 1019, 1021,  496 },
+    {  704,  815, 1193,  971,  973,  640, 1217, 2214,  832,  578 },
+    {  672, 1245,  579,  871,  875,  774,  872, 1273, 1027,  949 },
+    {  296, 1134, 2050, 1784, 1636, 3425,  442, 1550, 2076,  722 },
+    {  342,  982, 1259, 1846, 1848, 1848,  622,  568, 1847, 1052 },
+    {  555, 1064, 1304,  828,  746, 1343, 1075, 1329, 1078,  494 },
+    {  288, 1167, 1285, 1174, 1639, 1639,  833, 2254, 1304,  509 } },
+  { {  342,  719,  767, 1866, 1757, 1270, 1246,  550, 1746, 2151 },
+    {  483,  653,  694, 1509, 1459, 1410, 1218,  507, 1914, 1266 },
+    {  488,  757,  447, 2979, 1813, 1268, 1654,  539, 1849, 2109 },
+    {  522, 1097, 1085,  851, 1365, 1111,  851,  901,  961,  605 },
+    {  709,  716,  841,  728,  736,  945,  941,  862, 2845, 1057 },
+    {  512, 1323,  500, 1336, 1083,  681, 1342,  717, 1604, 1350 },
+    {  452, 1155, 1372, 1900, 1501, 3290,  311,  944, 1919,  922 },
+    {  403, 1520,  977, 2132, 1733, 3522, 1076,  276, 3335, 1547 },
+    {  559, 1374, 1101,  615,  673, 2462,  974,  795,  984,  984 },
+    {  547, 1122, 1062,  812, 1410,  951, 1140,  622, 1268,  651 } },
+  { {  165,  982, 1235,  938, 1334, 1366, 1659, 1578,  964, 1612 },
+    {  592,  422,  925,  847, 1139, 1112, 1387, 2036,  861, 1041 },
+    {  403,  837,  732,  770,  941, 1658, 1250,  809, 1407, 1407 },
+    {  896,  874, 1071,  381, 1568, 1722, 1437, 2192,  480, 1035 },
+    {  640, 1098, 1012, 1032,  684, 1382, 1581, 2106,  416,  865 },
+    {  559, 1005,  819,  914,  710,  770, 1418,  920,  838, 1435 },
+    {  415, 1258, 1245,  870, 1278, 3067,  770, 1021, 1287,  522 },
+    {  406,  990,  601, 1009, 1265, 1265, 1267,  759, 1017, 1277 },
+    {  968, 1182, 1329,  788, 1032, 1292, 1705, 1714,  203, 1403 },
+    {  732,  877, 1279,  471,  901, 1161, 1545, 1294,  755,  755 } },
+  { {  111,  931, 1378, 1185, 1933, 1648, 1148, 1714, 1873, 1307 },
+    {  406,  414, 1030, 1023, 1910, 1404, 1313, 1647, 1509,  793 },
+    {  342,  640,  575, 1088, 1241, 1349, 1161, 1350, 1756, 1502 },
+    {  559,  766, 1185,  357, 1682, 1428, 1329, 1897, 1219,  802 },
+    {  473,  909, 1164,  771,  719, 2508, 1427, 1432,  722,  782 },
+    {  342,  892,  785, 1145, 1150,  794, 1296, 1550,  973, 1057 },
+    {  208, 1036, 1326, 1343, 1606, 3395,  815, 1455, 1618,  712 },
+    {  228,  928,  890, 1046, 3499, 1711,  994,  829, 1720, 1318 },
+    {  768,  724, 1058,  636,  991, 1075, 1319, 1324,  616,  825 },
+    {  305, 1167, 1358,  899, 1587, 1587,  987, 1988, 1332,  501 } }
+};
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/enc/cost.h b/drivers/webp/enc/cost.h
new file mode 100644
index 0000000000..3cbad1ae4c
--- /dev/null
+++ b/drivers/webp/enc/cost.h
@@ -0,0 +1,51 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Cost tables for level and modes.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_ENC_COST_H_
+#define WEBP_ENC_COST_H_
+
+#include "./vp8enci.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// approximate cost per level:
+extern const uint16_t VP8LevelFixedCosts[MAX_LEVEL + 1];
+extern const uint16_t VP8EntropyCost[256];        // 8bit fixed-point log(p)
+
+// Cost of coding one event with probability 'proba'.
+static WEBP_INLINE int VP8BitCost(int bit, uint8_t proba) {
+  return !bit ? VP8EntropyCost[proba] : VP8EntropyCost[255 - proba];
+}
+
+// Level cost calculations
+extern const uint16_t VP8LevelCodes[MAX_VARIABLE_LEVEL][2];
+void VP8CalculateLevelCosts(VP8Proba* const proba);
+static WEBP_INLINE int VP8LevelCost(const uint16_t* const table, int level) {
+  return VP8LevelFixedCosts[level]
+       + table[(level > MAX_VARIABLE_LEVEL) ? MAX_VARIABLE_LEVEL : level];
+}
+
+// Mode costs
+extern const uint16_t VP8FixedCostsUV[4];
+extern const uint16_t VP8FixedCostsI16[4];
+extern const uint16_t VP8FixedCostsI4[NUM_BMODES][NUM_BMODES][NUM_BMODES];
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_ENC_COST_H_ */
diff --git a/drivers/webp/enc/filter.c b/drivers/webp/enc/filter.c
new file mode 100644
index 0000000000..dd27804b55
--- /dev/null
+++ b/drivers/webp/enc/filter.c
@@ -0,0 +1,471 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Selecting filter level
+//
+// Author: somnath@google.com (Somnath Banerjee)
+
+#include <assert.h>
+#include "./vp8enci.h"
+
+// This table gives, for a given sharpness, the filtering strength to be
+// used (at least) in order to filter a given edge step delta.
+// This is constructed by brute force inspection: for all delta, we iterate
+// over all possible filtering strength / thresh until needs_filter() returns
+// true.
+#define MAX_DELTA_SIZE 64
+static const uint8_t kLevelsFromDelta[8][MAX_DELTA_SIZE] = {
+  { 0,   1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+    16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+    32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+    48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63 },
+  { 0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 14, 15, 17, 18,
+    20, 21, 23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42,
+    44, 45, 47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 14, 16, 17, 19,
+    20, 22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43,
+    44, 46, 47, 49, 50, 52, 53, 55, 56, 58, 59, 61, 62, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 15, 16, 18, 19,
+    21, 22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43,
+    45, 46, 48, 49, 51, 52, 54, 55, 57, 58, 60, 61, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 14, 15, 17, 18, 20,
+    21, 23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42, 44,
+    45, 47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 13, 15, 16, 17, 19, 20,
+    22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43, 44,
+    46, 47, 49, 50, 52, 53, 55, 56, 58, 59, 61, 62, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 13, 15, 16, 18, 19, 21,
+    22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43, 45,
+    46, 48, 49, 51, 52, 54, 55, 57, 58, 60, 61, 63, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 14, 15, 17, 18, 20, 21,
+    23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42, 44, 45,
+    47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 }
+};
+
+int VP8FilterStrengthFromDelta(int sharpness, int delta) {
+  const int pos = (delta < MAX_DELTA_SIZE) ? delta : MAX_DELTA_SIZE - 1;
+  assert(sharpness >= 0 && sharpness <= 7);
+  return kLevelsFromDelta[sharpness][pos];
+}
+
+// -----------------------------------------------------------------------------
+// NOTE: clip1, tables and InitTables are repeated entries of dsp.c
+static uint8_t abs0[255 + 255 + 1];     // abs(i)
+static uint8_t abs1[255 + 255 + 1];     // abs(i)>>1
+static int8_t sclip1[1020 + 1020 + 1];  // clips [-1020, 1020] to [-128, 127]
+static int8_t sclip2[112 + 112 + 1];    // clips [-112, 112] to [-16, 15]
+static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
+
+static int tables_ok = 0;
+
+static void InitTables(void) {
+  if (!tables_ok) {
+    int i;
+    for (i = -255; i <= 255; ++i) {
+      abs0[255 + i] = (i < 0) ? -i : i;
+      abs1[255 + i] = abs0[255 + i] >> 1;
+    }
+    for (i = -1020; i <= 1020; ++i) {
+      sclip1[1020 + i] = (i < -128) ? -128 : (i > 127) ? 127 : i;
+    }
+    for (i = -112; i <= 112; ++i) {
+      sclip2[112 + i] = (i < -16) ? -16 : (i > 15) ? 15 : i;
+    }
+    for (i = -255; i <= 255 + 255; ++i) {
+      clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i;
+    }
+    tables_ok = 1;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Edge filtering functions
+
+// 4 pixels in, 2 pixels out
+static WEBP_INLINE void do_filter2(uint8_t* p, int step) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  const int a = 3 * (q0 - p0) + sclip1[1020 + p1 - q1];
+  const int a1 = sclip2[112 + ((a + 4) >> 3)];
+  const int a2 = sclip2[112 + ((a + 3) >> 3)];
+  p[-step] = clip1[255 + p0 + a2];
+  p[    0] = clip1[255 + q0 - a1];
+}
+
+// 4 pixels in, 4 pixels out
+static WEBP_INLINE void do_filter4(uint8_t* p, int step) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  const int a = 3 * (q0 - p0);
+  const int a1 = sclip2[112 + ((a + 4) >> 3)];
+  const int a2 = sclip2[112 + ((a + 3) >> 3)];
+  const int a3 = (a1 + 1) >> 1;
+  p[-2*step] = clip1[255 + p1 + a3];
+  p[-  step] = clip1[255 + p0 + a2];
+  p[      0] = clip1[255 + q0 - a1];
+  p[   step] = clip1[255 + q1 - a3];
+}
+
+// high edge-variance
+static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return (abs0[255 + p1 - p0] > thresh) || (abs0[255 + q1 - q0] > thresh);
+}
+
+static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int thresh) {
+  const int p1 = p[-2*step], p0 = p[-step], q0 = p[0], q1 = p[step];
+  return (2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) <= thresh;
+}
+
+static WEBP_INLINE int needs_filter2(const uint8_t* p,
+                                     int step, int t, int it) {
+  const int p3 = p[-4*step], p2 = p[-3*step], p1 = p[-2*step], p0 = p[-step];
+  const int q0 = p[0], q1 = p[step], q2 = p[2*step], q3 = p[3*step];
+  if ((2 * abs0[255 + p0 - q0] + abs1[255 + p1 - q1]) > t)
+    return 0;
+  return abs0[255 + p3 - p2] <= it && abs0[255 + p2 - p1] <= it &&
+         abs0[255 + p1 - p0] <= it && abs0[255 + q3 - q2] <= it &&
+         abs0[255 + q2 - q1] <= it && abs0[255 + q1 - q0] <= it;
+}
+
+//------------------------------------------------------------------------------
+// Simple In-loop filtering (Paragraph 15.2)
+
+static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    if (needs_filter(p + i, stride, thresh)) {
+      do_filter2(p + i, stride);
+    }
+  }
+}
+
+static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    if (needs_filter(p + i * stride, 1, thresh)) {
+      do_filter2(p + i * stride, 1);
+    }
+  }
+}
+
+static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    SimpleVFilter16(p, stride, thresh);
+  }
+}
+
+static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    SimpleHFilter16(p, stride, thresh);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Complex In-loop filtering (Paragraph 15.3)
+
+static WEBP_INLINE void FilterLoop24(uint8_t* p,
+                                     int hstride, int vstride, int size,
+                                     int thresh, int ithresh, int hev_thresh) {
+  while (size-- > 0) {
+    if (needs_filter2(p, hstride, thresh, ithresh)) {
+      if (hev(p, hstride, hev_thresh)) {
+        do_filter2(p, hstride);
+      } else {
+        do_filter4(p, hstride);
+      }
+    }
+    p += vstride;
+  }
+}
+
+// on three inner edges
+static void VFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4 * stride;
+    FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh);
+  }
+}
+
+static void HFilter16i(uint8_t* p, int stride,
+                       int thresh, int ithresh, int hev_thresh) {
+  int k;
+  for (k = 3; k > 0; --k) {
+    p += 4;
+    FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh);
+  }
+}
+
+static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
+  FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh);
+}
+
+static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
+                      int thresh, int ithresh, int hev_thresh) {
+  FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
+  FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh);
+}
+
+//------------------------------------------------------------------------------
+
+void (*VP8EncVFilter16i)(uint8_t*, int, int, int, int) = VFilter16i;
+void (*VP8EncHFilter16i)(uint8_t*, int, int, int, int) = HFilter16i;
+void (*VP8EncVFilter8i)(uint8_t*, uint8_t*, int, int, int, int) = VFilter8i;
+void (*VP8EncHFilter8i)(uint8_t*, uint8_t*, int, int, int, int) = HFilter8i;
+
+void (*VP8EncSimpleVFilter16i)(uint8_t*, int, int) = SimpleVFilter16i;
+void (*VP8EncSimpleHFilter16i)(uint8_t*, int, int) = SimpleHFilter16i;
+
+//------------------------------------------------------------------------------
+// Paragraph 15.4: compute the inner-edge filtering strength
+
+static int GetILevel(int sharpness, int level) {
+  if (sharpness > 0) {
+    if (sharpness > 4) {
+      level >>= 2;
+    } else {
+      level >>= 1;
+    }
+    if (level > 9 - sharpness) {
+      level = 9 - sharpness;
+    }
+  }
+  if (level < 1) level = 1;
+  return level;
+}
+
+static void DoFilter(const VP8EncIterator* const it, int level) {
+  const VP8Encoder* const enc = it->enc_;
+  const int ilevel = GetILevel(enc->config_->filter_sharpness, level);
+  const int limit = 2 * level + ilevel;
+
+  uint8_t* const y_dst = it->yuv_out2_ + Y_OFF;
+  uint8_t* const u_dst = it->yuv_out2_ + U_OFF;
+  uint8_t* const v_dst = it->yuv_out2_ + V_OFF;
+
+  // copy current block to yuv_out2_
+  memcpy(y_dst, it->yuv_out_, YUV_SIZE * sizeof(uint8_t));
+
+  if (enc->filter_hdr_.simple_ == 1) {   // simple
+    VP8EncSimpleHFilter16i(y_dst, BPS, limit);
+    VP8EncSimpleVFilter16i(y_dst, BPS, limit);
+  } else {    // complex
+    const int hev_thresh = (level >= 40) ? 2 : (level >= 15) ? 1 : 0;
+    VP8EncHFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8EncHFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+    VP8EncVFilter16i(y_dst, BPS, limit, ilevel, hev_thresh);
+    VP8EncVFilter8i(u_dst, v_dst, BPS, limit, ilevel, hev_thresh);
+  }
+}
+
+//------------------------------------------------------------------------------
+// SSIM metric
+
+enum { KERNEL = 3 };
+static const double kMinValue = 1.e-10;  // minimal threshold
+
+void VP8SSIMAddStats(const DistoStats* const src, DistoStats* const dst) {
+  dst->w   += src->w;
+  dst->xm  += src->xm;
+  dst->ym  += src->ym;
+  dst->xxm += src->xxm;
+  dst->xym += src->xym;
+  dst->yym += src->yym;
+}
+
+static void VP8SSIMAccumulate(const uint8_t* src1, int stride1,
+                              const uint8_t* src2, int stride2,
+                              int xo, int yo, int W, int H,
+                              DistoStats* const stats) {
+  const int ymin = (yo - KERNEL < 0) ? 0 : yo - KERNEL;
+  const int ymax = (yo + KERNEL > H - 1) ? H - 1 : yo + KERNEL;
+  const int xmin = (xo - KERNEL < 0) ? 0 : xo - KERNEL;
+  const int xmax = (xo + KERNEL > W - 1) ? W - 1 : xo + KERNEL;
+  int x, y;
+  src1 += ymin * stride1;
+  src2 += ymin * stride2;
+  for (y = ymin; y <= ymax; ++y, src1 += stride1, src2 += stride2) {
+    for (x = xmin; x <= xmax; ++x) {
+      const int s1 = src1[x];
+      const int s2 = src2[x];
+      stats->w   += 1;
+      stats->xm  += s1;
+      stats->ym  += s2;
+      stats->xxm += s1 * s1;
+      stats->xym += s1 * s2;
+      stats->yym += s2 * s2;
+    }
+  }
+}
+
+double VP8SSIMGet(const DistoStats* const stats) {
+  const double xmxm = stats->xm * stats->xm;
+  const double ymym = stats->ym * stats->ym;
+  const double xmym = stats->xm * stats->ym;
+  const double w2 = stats->w * stats->w;
+  double sxx = stats->xxm * stats->w - xmxm;
+  double syy = stats->yym * stats->w - ymym;
+  double sxy = stats->xym * stats->w - xmym;
+  double C1, C2;
+  double fnum;
+  double fden;
+  // small errors are possible, due to rounding. Clamp to zero.
+  if (sxx < 0.) sxx = 0.;
+  if (syy < 0.) syy = 0.;
+  C1 = 6.5025 * w2;
+  C2 = 58.5225 * w2;
+  fnum = (2 * xmym + C1) * (2 * sxy + C2);
+  fden = (xmxm + ymym + C1) * (sxx + syy + C2);
+  return (fden != 0.) ? fnum / fden : kMinValue;
+}
+
+double VP8SSIMGetSquaredError(const DistoStats* const s) {
+  if (s->w > 0.) {
+    const double iw2 = 1. / (s->w * s->w);
+    const double sxx = s->xxm * s->w - s->xm * s->xm;
+    const double syy = s->yym * s->w - s->ym * s->ym;
+    const double sxy = s->xym * s->w - s->xm * s->ym;
+    const double SSE = iw2 * (sxx + syy - 2. * sxy);
+    if (SSE > kMinValue) return SSE;
+  }
+  return kMinValue;
+}
+
+void VP8SSIMAccumulatePlane(const uint8_t* src1, int stride1,
+                            const uint8_t* src2, int stride2,
+                            int W, int H, DistoStats* const stats) {
+  int x, y;
+  for (y = 0; y < H; ++y) {
+    for (x = 0; x < W; ++x) {
+      VP8SSIMAccumulate(src1, stride1, src2, stride2, x, y, W, H, stats);
+    }
+  }
+}
+
+static double GetMBSSIM(const uint8_t* yuv1, const uint8_t* yuv2) {
+  int x, y;
+  DistoStats s = { .0, .0, .0, .0, .0, .0 };
+
+  // compute SSIM in a 10 x 10 window
+  for (x = 3; x < 13; x++) {
+    for (y = 3; y < 13; y++) {
+      VP8SSIMAccumulate(yuv1 + Y_OFF, BPS, yuv2 + Y_OFF, BPS, x, y, 16, 16, &s);
+    }
+  }
+  for (x = 1; x < 7; x++) {
+    for (y = 1; y < 7; y++) {
+      VP8SSIMAccumulate(yuv1 + U_OFF, BPS, yuv2 + U_OFF, BPS, x, y, 8, 8, &s);
+      VP8SSIMAccumulate(yuv1 + V_OFF, BPS, yuv2 + V_OFF, BPS, x, y, 8, 8, &s);
+    }
+  }
+  return VP8SSIMGet(&s);
+}
+
+//------------------------------------------------------------------------------
+// Exposed APIs: Encoder should call the following 3 functions to adjust
+// loop filter strength
+
+void VP8InitFilter(VP8EncIterator* const it) {
+  if (it->lf_stats_ != NULL) {
+    int s, i;
+    InitTables();
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      for (i = 0; i < MAX_LF_LEVELS; i++) {
+        (*it->lf_stats_)[s][i] = 0;
+      }
+    }
+  }
+}
+
+void VP8StoreFilterStats(VP8EncIterator* const it) {
+  int d;
+  VP8Encoder* const enc = it->enc_;
+  const int s = it->mb_->segment_;
+  const int level0 = enc->dqm_[s].fstrength_;  // TODO: ref_lf_delta[]
+
+  // explore +/-quant range of values around level0
+  const int delta_min = -enc->dqm_[s].quant_;
+  const int delta_max = enc->dqm_[s].quant_;
+  const int step_size = (delta_max - delta_min >= 4) ? 4 : 1;
+
+  if (it->lf_stats_ == NULL) return;
+
+  // NOTE: Currently we are applying filter only across the sublock edges
+  // There are two reasons for that.
+  // 1. Applying filter on macro block edges will change the pixels in
+  // the left and top macro blocks. That will be hard to restore
+  // 2. Macro Blocks on the bottom and right are not yet compressed. So we
+  // cannot apply filter on the right and bottom macro block edges.
+  if (it->mb_->type_ == 1 && it->mb_->skip_) return;
+
+  // Always try filter level  zero
+  (*it->lf_stats_)[s][0] += GetMBSSIM(it->yuv_in_, it->yuv_out_);
+
+  for (d = delta_min; d <= delta_max; d += step_size) {
+    const int level = level0 + d;
+    if (level <= 0 || level >= MAX_LF_LEVELS) {
+      continue;
+    }
+    DoFilter(it, level);
+    (*it->lf_stats_)[s][level] += GetMBSSIM(it->yuv_in_, it->yuv_out2_);
+  }
+}
+
+void VP8AdjustFilterStrength(VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  if (it->lf_stats_ != NULL) {
+    int s;
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      int i, best_level = 0;
+      // Improvement over filter level 0 should be at least 1e-5 (relatively)
+      double best_v = 1.00001 * (*it->lf_stats_)[s][0];
+      for (i = 1; i < MAX_LF_LEVELS; i++) {
+        const double v = (*it->lf_stats_)[s][i];
+        if (v > best_v) {
+          best_v = v;
+          best_level = i;
+        }
+      }
+      enc->dqm_[s].fstrength_ = best_level;
+    }
+  } else if (enc->config_->filter_strength > 0) {
+    int max_level = 0;
+    int s;
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      VP8SegmentInfo* const dqm = &enc->dqm_[s];
+      // this '>> 3' accounts for some inverse WHT scaling
+      const int delta = (dqm->max_edge_ * dqm->y2_.q_[1]) >> 3;
+      const int level =
+          VP8FilterStrengthFromDelta(enc->filter_hdr_.sharpness_, delta);
+      if (level > dqm->fstrength_) {
+        dqm->fstrength_ = level;
+      }
+      if (max_level < dqm->fstrength_) {
+        max_level = dqm->fstrength_;
+      }
+    }
+    enc->filter_hdr_.level_ = max_level;
+  }
+}
+
+// -----------------------------------------------------------------------------
+
diff --git a/drivers/webp/enc/frame.c b/drivers/webp/enc/frame.c
new file mode 100644
index 0000000000..2582244c6c
--- /dev/null
+++ b/drivers/webp/enc/frame.c
@@ -0,0 +1,1068 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//   frame coding and analysis
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+
+#include "./vp8enci.h"
+#include "./cost.h"
+#include "../webp/format_constants.h"  // RIFF constants
+
+#define SEGMENT_VISU 0
+#define DEBUG_SEARCH 0    // useful to track search convergence
+
+// On-the-fly info about the current set of residuals. Handy to avoid
+// passing zillions of params.
+typedef struct {
+  int first;
+  int last;
+  const int16_t* coeffs;
+
+  int coeff_type;
+  ProbaArray* prob;
+  StatsArray* stats;
+  CostArray*  cost;
+} VP8Residual;
+
+//------------------------------------------------------------------------------
+// multi-pass convergence
+
+#define HEADER_SIZE_ESTIMATE (RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE +  \
+                              VP8_FRAME_HEADER_SIZE)
+#define DQ_LIMIT 0.4  // convergence is considered reached if dq < DQ_LIMIT
+// we allow 2k of extra head-room in PARTITION0 limit.
+#define PARTITION0_SIZE_LIMIT ((VP8_MAX_PARTITION0_SIZE - 2048ULL) << 11)
+
+typedef struct {  // struct for organizing convergence in either size or PSNR
+  int is_first;
+  float dq;
+  float q, last_q;
+  double value, last_value;   // PSNR or size
+  double target;
+  int do_size_search;
+} PassStats;
+
+static int InitPassStats(const VP8Encoder* const enc, PassStats* const s) {
+  const uint64_t target_size = (uint64_t)enc->config_->target_size;
+  const int do_size_search = (target_size != 0);
+  const float target_PSNR = enc->config_->target_PSNR;
+
+  s->is_first = 1;
+  s->dq = 10.f;
+  s->q = s->last_q = enc->config_->quality;
+  s->target = do_size_search ? (double)target_size
+            : (target_PSNR > 0.) ? target_PSNR
+            : 40.;   // default, just in case
+  s->value = s->last_value = 0.;
+  s->do_size_search = do_size_search;
+  return do_size_search;
+}
+
+static float Clamp(float v, float min, float max) {
+  return (v < min) ? min : (v > max) ? max : v;
+}
+
+static float ComputeNextQ(PassStats* const s) {
+  float dq;
+  if (s->is_first) {
+    dq = (s->value > s->target) ? -s->dq : s->dq;
+    s->is_first = 0;
+  } else if (s->value != s->last_value) {
+    const double slope = (s->target - s->value) / (s->last_value - s->value);
+    dq = (float)(slope * (s->last_q - s->q));
+  } else {
+    dq = 0.;  // we're done?!
+  }
+  // Limit variable to avoid large swings.
+  s->dq = Clamp(dq, -30.f, 30.f);
+  s->last_q = s->q;
+  s->last_value = s->value;
+  s->q = Clamp(s->q + s->dq, 0.f, 100.f);
+  return s->q;
+}
+
+//------------------------------------------------------------------------------
+// Tables for level coding
+
+const uint8_t VP8EncBands[16 + 1] = {
+  0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
+  0  // sentinel
+};
+
+const uint8_t VP8Cat3[] = { 173, 148, 140 };
+const uint8_t VP8Cat4[] = { 176, 155, 140, 135 };
+const uint8_t VP8Cat5[] = { 180, 157, 141, 134, 130 };
+const uint8_t VP8Cat6[] =
+    { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129 };
+
+//------------------------------------------------------------------------------
+// Reset the statistics about: number of skips, token proba, level cost,...
+
+static void ResetStats(VP8Encoder* const enc) {
+  VP8Proba* const proba = &enc->proba_;
+  VP8CalculateLevelCosts(proba);
+  proba->nb_skip_ = 0;
+}
+
+//------------------------------------------------------------------------------
+// Skip decision probability
+
+#define SKIP_PROBA_THRESHOLD 250  // value below which using skip_proba is OK.
+
+static int CalcSkipProba(uint64_t nb, uint64_t total) {
+  return (int)(total ? (total - nb) * 255 / total : 255);
+}
+
+// Returns the bit-cost for coding the skip probability.
+static int FinalizeSkipProba(VP8Encoder* const enc) {
+  VP8Proba* const proba = &enc->proba_;
+  const int nb_mbs = enc->mb_w_ * enc->mb_h_;
+  const int nb_events = proba->nb_skip_;
+  int size;
+  proba->skip_proba_ = CalcSkipProba(nb_events, nb_mbs);
+  proba->use_skip_proba_ = (proba->skip_proba_ < SKIP_PROBA_THRESHOLD);
+  size = 256;   // 'use_skip_proba' bit
+  if (proba->use_skip_proba_) {
+    size +=  nb_events * VP8BitCost(1, proba->skip_proba_)
+         + (nb_mbs - nb_events) * VP8BitCost(0, proba->skip_proba_);
+    size += 8 * 256;   // cost of signaling the skip_proba_ itself.
+  }
+  return size;
+}
+
+//------------------------------------------------------------------------------
+// Recording of token probabilities.
+
+static void ResetTokenStats(VP8Encoder* const enc) {
+  VP8Proba* const proba = &enc->proba_;
+  memset(proba->stats_, 0, sizeof(proba->stats_));
+}
+
+// Record proba context used
+static int Record(int bit, proba_t* const stats) {
+  proba_t p = *stats;
+  if (p >= 0xffff0000u) {               // an overflow is inbound.
+    p = ((p + 1u) >> 1) & 0x7fff7fffu;  // -> divide the stats by 2.
+  }
+  // record bit count (lower 16 bits) and increment total count (upper 16 bits).
+  p += 0x00010000u + bit;
+  *stats = p;
+  return bit;
+}
+
+// We keep the table free variant around for reference, in case.
+#define USE_LEVEL_CODE_TABLE
+
+// Simulate block coding, but only record statistics.
+// Note: no need to record the fixed probas.
+static int RecordCoeffs(int ctx, const VP8Residual* const res) {
+  int n = res->first;
+  // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  proba_t* s = res->stats[n][ctx];
+  if (res->last  < 0) {
+    Record(0, s + 0);
+    return 0;
+  }
+  while (n <= res->last) {
+    int v;
+    Record(1, s + 0);  // order of record doesn't matter
+    while ((v = res->coeffs[n++]) == 0) {
+      Record(0, s + 1);
+      s = res->stats[VP8EncBands[n]][0];
+    }
+    Record(1, s + 1);
+    if (!Record(2u < (unsigned int)(v + 1), s + 2)) {  // v = -1 or 1
+      s = res->stats[VP8EncBands[n]][1];
+    } else {
+      v = abs(v);
+#if !defined(USE_LEVEL_CODE_TABLE)
+      if (!Record(v > 4, s + 3)) {
+        if (Record(v != 2, s + 4))
+          Record(v == 4, s + 5);
+      } else if (!Record(v > 10, s + 6)) {
+        Record(v > 6, s + 7);
+      } else if (!Record((v >= 3 + (8 << 2)), s + 8)) {
+        Record((v >= 3 + (8 << 1)), s + 9);
+      } else {
+        Record((v >= 3 + (8 << 3)), s + 10);
+      }
+#else
+      if (v > MAX_VARIABLE_LEVEL)
+        v = MAX_VARIABLE_LEVEL;
+
+      {
+        const int bits = VP8LevelCodes[v - 1][1];
+        int pattern = VP8LevelCodes[v - 1][0];
+        int i;
+        for (i = 0; (pattern >>= 1) != 0; ++i) {
+          const int mask = 2 << i;
+          if (pattern & 1) Record(!!(bits & mask), s + 3 + i);
+        }
+      }
+#endif
+      s = res->stats[VP8EncBands[n]][2];
+    }
+  }
+  if (n < 16) Record(0, s + 0);
+  return 1;
+}
+
+// Collect statistics and deduce probabilities for next coding pass.
+// Return the total bit-cost for coding the probability updates.
+static int CalcTokenProba(int nb, int total) {
+  assert(nb <= total);
+  return nb ? (255 - nb * 255 / total) : 255;
+}
+
+// Cost of coding 'nb' 1's and 'total-nb' 0's using 'proba' probability.
+static int BranchCost(int nb, int total, int proba) {
+  return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba);
+}
+
+static int FinalizeTokenProbas(VP8Proba* const proba) {
+  int has_changed = 0;
+  int size = 0;
+  int t, b, c, p;
+  for (t = 0; t < NUM_TYPES; ++t) {
+    for (b = 0; b < NUM_BANDS; ++b) {
+      for (c = 0; c < NUM_CTX; ++c) {
+        for (p = 0; p < NUM_PROBAS; ++p) {
+          const proba_t stats = proba->stats_[t][b][c][p];
+          const int nb = (stats >> 0) & 0xffff;
+          const int total = (stats >> 16) & 0xffff;
+          const int update_proba = VP8CoeffsUpdateProba[t][b][c][p];
+          const int old_p = VP8CoeffsProba0[t][b][c][p];
+          const int new_p = CalcTokenProba(nb, total);
+          const int old_cost = BranchCost(nb, total, old_p)
+                             + VP8BitCost(0, update_proba);
+          const int new_cost = BranchCost(nb, total, new_p)
+                             + VP8BitCost(1, update_proba)
+                             + 8 * 256;
+          const int use_new_p = (old_cost > new_cost);
+          size += VP8BitCost(use_new_p, update_proba);
+          if (use_new_p) {  // only use proba that seem meaningful enough.
+            proba->coeffs_[t][b][c][p] = new_p;
+            has_changed |= (new_p != old_p);
+            size += 8 * 256;
+          } else {
+            proba->coeffs_[t][b][c][p] = old_p;
+          }
+        }
+      }
+    }
+  }
+  proba->dirty_ = has_changed;
+  return size;
+}
+
+//------------------------------------------------------------------------------
+// Finalize Segment probability based on the coding tree
+
+static int GetProba(int a, int b) {
+  const int total = a + b;
+  return (total == 0) ? 255     // that's the default probability.
+                      : (255 * a + total / 2) / total;  // rounded proba
+}
+
+static void SetSegmentProbas(VP8Encoder* const enc) {
+  int p[NUM_MB_SEGMENTS] = { 0 };
+  int n;
+
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    const VP8MBInfo* const mb = &enc->mb_info_[n];
+    p[mb->segment_]++;
+  }
+  if (enc->pic_->stats != NULL) {
+    for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
+      enc->pic_->stats->segment_size[n] = p[n];
+    }
+  }
+  if (enc->segment_hdr_.num_segments_ > 1) {
+    uint8_t* const probas = enc->proba_.segments_;
+    probas[0] = GetProba(p[0] + p[1], p[2] + p[3]);
+    probas[1] = GetProba(p[0], p[1]);
+    probas[2] = GetProba(p[2], p[3]);
+
+    enc->segment_hdr_.update_map_ =
+        (probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255);
+    enc->segment_hdr_.size_ =
+        p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) +
+        p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) +
+        p[2] * (VP8BitCost(1, probas[0]) + VP8BitCost(0, probas[2])) +
+        p[3] * (VP8BitCost(1, probas[0]) + VP8BitCost(1, probas[2]));
+  } else {
+    enc->segment_hdr_.update_map_ = 0;
+    enc->segment_hdr_.size_ = 0;
+  }
+}
+
+//------------------------------------------------------------------------------
+// helper functions for residuals struct VP8Residual.
+
+static void InitResidual(int first, int coeff_type,
+                         VP8Encoder* const enc, VP8Residual* const res) {
+  res->coeff_type = coeff_type;
+  res->prob  = enc->proba_.coeffs_[coeff_type];
+  res->stats = enc->proba_.stats_[coeff_type];
+  res->cost  = enc->proba_.level_cost_[coeff_type];
+  res->first = first;
+}
+
+static void SetResidualCoeffs(const int16_t* const coeffs,
+                              VP8Residual* const res) {
+  int n;
+  res->last = -1;
+  for (n = 15; n >= res->first; --n) {
+    if (coeffs[n]) {
+      res->last = n;
+      break;
+    }
+  }
+  res->coeffs = coeffs;
+}
+
+//------------------------------------------------------------------------------
+// Mode costs
+
+static int GetResidualCost(int ctx0, const VP8Residual* const res) {
+  int n = res->first;
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  int p0 = res->prob[n][ctx0][0];
+  const uint16_t* t = res->cost[n][ctx0];
+  int cost;
+
+  if (res->last < 0) {
+    return VP8BitCost(0, p0);
+  }
+  cost = VP8BitCost(1, p0);
+  for (; n < res->last; ++n) {
+    const int v = abs(res->coeffs[n]);
+    const int b = VP8EncBands[n + 1];
+    const int ctx = (v >= 2) ? 2 : v;
+    cost += VP8LevelCost(t, v);
+    t = res->cost[b][ctx];
+    // the masking trick is faster than "if (v) cost += ..." with clang
+    cost += (v ? ~0U : 0) & VP8BitCost(1, res->prob[b][ctx][0]);
+  }
+  // Last coefficient is always non-zero
+  {
+    const int v = abs(res->coeffs[n]);
+    assert(v != 0);
+    cost += VP8LevelCost(t, v);
+    if (n < 15) {
+      const int b = VP8EncBands[n + 1];
+      const int ctx = (v == 1) ? 1 : 2;
+      const int last_p0 = res->prob[b][ctx][0];
+      cost += VP8BitCost(0, last_p0);
+    }
+  }
+  return cost;
+}
+
+int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]) {
+  const int x = (it->i4_ & 3), y = (it->i4_ >> 2);
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int R = 0;
+  int ctx;
+
+  InitResidual(0, 3, enc, &res);
+  ctx = it->top_nz_[x] + it->left_nz_[y];
+  SetResidualCoeffs(levels, &res);
+  R += GetResidualCost(ctx, &res);
+  return R;
+}
+
+int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd) {
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int x, y;
+  int R = 0;
+
+  VP8IteratorNzToBytes(it);   // re-import the non-zero context
+
+  // DC
+  InitResidual(0, 1, enc, &res);
+  SetResidualCoeffs(rd->y_dc_levels, &res);
+  R += GetResidualCost(it->top_nz_[8] + it->left_nz_[8], &res);
+
+  // AC
+  InitResidual(1, 0, enc, &res);
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      R += GetResidualCost(ctx, &res);
+      it->top_nz_[x] = it->left_nz_[y] = (res.last >= 0);
+    }
+  }
+  return R;
+}
+
+int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+  int ch, x, y;
+  int R = 0;
+
+  VP8IteratorNzToBytes(it);  // re-import the non-zero context
+
+  InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        R += GetResidualCost(ctx, &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = (res.last >= 0);
+      }
+    }
+  }
+  return R;
+}
+
+//------------------------------------------------------------------------------
+// Coefficient coding
+
+static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
+  int n = res->first;
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  const uint8_t* p = res->prob[n][ctx];
+  if (!VP8PutBit(bw, res->last >= 0, p[0])) {
+    return 0;
+  }
+
+  while (n < 16) {
+    const int c = res->coeffs[n++];
+    const int sign = c < 0;
+    int v = sign ? -c : c;
+    if (!VP8PutBit(bw, v != 0, p[1])) {
+      p = res->prob[VP8EncBands[n]][0];
+      continue;
+    }
+    if (!VP8PutBit(bw, v > 1, p[2])) {
+      p = res->prob[VP8EncBands[n]][1];
+    } else {
+      if (!VP8PutBit(bw, v > 4, p[3])) {
+        if (VP8PutBit(bw, v != 2, p[4]))
+          VP8PutBit(bw, v == 4, p[5]);
+      } else if (!VP8PutBit(bw, v > 10, p[6])) {
+        if (!VP8PutBit(bw, v > 6, p[7])) {
+          VP8PutBit(bw, v == 6, 159);
+        } else {
+          VP8PutBit(bw, v >= 9, 165);
+          VP8PutBit(bw, !(v & 1), 145);
+        }
+      } else {
+        int mask;
+        const uint8_t* tab;
+        if (v < 3 + (8 << 1)) {          // VP8Cat3  (3b)
+          VP8PutBit(bw, 0, p[8]);
+          VP8PutBit(bw, 0, p[9]);
+          v -= 3 + (8 << 0);
+          mask = 1 << 2;
+          tab = VP8Cat3;
+        } else if (v < 3 + (8 << 2)) {   // VP8Cat4  (4b)
+          VP8PutBit(bw, 0, p[8]);
+          VP8PutBit(bw, 1, p[9]);
+          v -= 3 + (8 << 1);
+          mask = 1 << 3;
+          tab = VP8Cat4;
+        } else if (v < 3 + (8 << 3)) {   // VP8Cat5  (5b)
+          VP8PutBit(bw, 1, p[8]);
+          VP8PutBit(bw, 0, p[10]);
+          v -= 3 + (8 << 2);
+          mask = 1 << 4;
+          tab = VP8Cat5;
+        } else {                         // VP8Cat6 (11b)
+          VP8PutBit(bw, 1, p[8]);
+          VP8PutBit(bw, 1, p[10]);
+          v -= 3 + (8 << 3);
+          mask = 1 << 10;
+          tab = VP8Cat6;
+        }
+        while (mask) {
+          VP8PutBit(bw, !!(v & mask), *tab++);
+          mask >>= 1;
+        }
+      }
+      p = res->prob[VP8EncBands[n]][2];
+    }
+    VP8PutBitUniform(bw, sign);
+    if (n == 16 || !VP8PutBit(bw, n <= res->last, p[0])) {
+      return 1;   // EOB
+    }
+  }
+  return 1;
+}
+
+static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it,
+                          const VP8ModeScore* const rd) {
+  int x, y, ch;
+  VP8Residual res;
+  uint64_t pos1, pos2, pos3;
+  const int i16 = (it->mb_->type_ == 1);
+  const int segment = it->mb_->segment_;
+  VP8Encoder* const enc = it->enc_;
+
+  VP8IteratorNzToBytes(it);
+
+  pos1 = VP8BitWriterPos(bw);
+  if (i16) {
+    InitResidual(0, 1, enc, &res);
+    SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+      PutCoeffs(bw, it->top_nz_[8] + it->left_nz_[8], &res);
+    InitResidual(1, 0, enc, &res);
+  } else {
+    InitResidual(0, 3, enc, &res);
+  }
+
+  // luma-AC
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] = PutCoeffs(bw, ctx, &res);
+    }
+  }
+  pos2 = VP8BitWriterPos(bw);
+
+  // U/V
+  InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
+            PutCoeffs(bw, ctx, &res);
+      }
+    }
+  }
+  pos3 = VP8BitWriterPos(bw);
+  it->luma_bits_ = pos2 - pos1;
+  it->uv_bits_ = pos3 - pos2;
+  it->bit_count_[segment][i16] += it->luma_bits_;
+  it->bit_count_[segment][2] += it->uv_bits_;
+  VP8IteratorBytesToNz(it);
+}
+
+// Same as CodeResiduals, but doesn't actually write anything.
+// Instead, it just records the event distribution.
+static void RecordResiduals(VP8EncIterator* const it,
+                            const VP8ModeScore* const rd) {
+  int x, y, ch;
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+
+  VP8IteratorNzToBytes(it);
+
+  if (it->mb_->type_ == 1) {   // i16x16
+    InitResidual(0, 1, enc, &res);
+    SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+      RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res);
+    InitResidual(1, 0, enc, &res);
+  } else {
+    InitResidual(0, 3, enc, &res);
+  }
+
+  // luma-AC
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] = RecordCoeffs(ctx, &res);
+    }
+  }
+
+  // U/V
+  InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
+            RecordCoeffs(ctx, &res);
+      }
+    }
+  }
+
+  VP8IteratorBytesToNz(it);
+}
+
+//------------------------------------------------------------------------------
+// Token buffer
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+static void RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
+                         VP8TBuffer* const tokens) {
+  int x, y, ch;
+  VP8Residual res;
+  VP8Encoder* const enc = it->enc_;
+
+  VP8IteratorNzToBytes(it);
+  if (it->mb_->type_ == 1) {   // i16x16
+    const int ctx = it->top_nz_[8] + it->left_nz_[8];
+    InitResidual(0, 1, enc, &res);
+    SetResidualCoeffs(rd->y_dc_levels, &res);
+    it->top_nz_[8] = it->left_nz_[8] =
+        VP8RecordCoeffTokens(ctx, 1,
+                             res.first, res.last, res.coeffs, tokens);
+    RecordCoeffs(ctx, &res);
+    InitResidual(1, 0, enc, &res);
+  } else {
+    InitResidual(0, 3, enc, &res);
+  }
+
+  // luma-AC
+  for (y = 0; y < 4; ++y) {
+    for (x = 0; x < 4; ++x) {
+      const int ctx = it->top_nz_[x] + it->left_nz_[y];
+      SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
+      it->top_nz_[x] = it->left_nz_[y] =
+          VP8RecordCoeffTokens(ctx, res.coeff_type,
+                               res.first, res.last, res.coeffs, tokens);
+      RecordCoeffs(ctx, &res);
+    }
+  }
+
+  // U/V
+  InitResidual(0, 2, enc, &res);
+  for (ch = 0; ch <= 2; ch += 2) {
+    for (y = 0; y < 2; ++y) {
+      for (x = 0; x < 2; ++x) {
+        const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+        SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
+        it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
+            VP8RecordCoeffTokens(ctx, 2,
+                                 res.first, res.last, res.coeffs, tokens);
+        RecordCoeffs(ctx, &res);
+      }
+    }
+  }
+  VP8IteratorBytesToNz(it);
+}
+
+#endif    // !DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+// ExtraInfo map / Debug function
+
+#if SEGMENT_VISU
+static void SetBlock(uint8_t* p, int value, int size) {
+  int y;
+  for (y = 0; y < size; ++y) {
+    memset(p, value, size);
+    p += BPS;
+  }
+}
+#endif
+
+static void ResetSSE(VP8Encoder* const enc) {
+  enc->sse_[0] = 0;
+  enc->sse_[1] = 0;
+  enc->sse_[2] = 0;
+  // Note: enc->sse_[3] is managed by alpha.c
+  enc->sse_count_ = 0;
+}
+
+static void StoreSSE(const VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  const uint8_t* const in = it->yuv_in_;
+  const uint8_t* const out = it->yuv_out_;
+  // Note: not totally accurate at boundary. And doesn't include in-loop filter.
+  enc->sse_[0] += VP8SSE16x16(in + Y_OFF, out + Y_OFF);
+  enc->sse_[1] += VP8SSE8x8(in + U_OFF, out + U_OFF);
+  enc->sse_[2] += VP8SSE8x8(in + V_OFF, out + V_OFF);
+  enc->sse_count_ += 16 * 16;
+}
+
+static void StoreSideInfo(const VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  const VP8MBInfo* const mb = it->mb_;
+  WebPPicture* const pic = enc->pic_;
+
+  if (pic->stats != NULL) {
+    StoreSSE(it);
+    enc->block_count_[0] += (mb->type_ == 0);
+    enc->block_count_[1] += (mb->type_ == 1);
+    enc->block_count_[2] += (mb->skip_ != 0);
+  }
+
+  if (pic->extra_info != NULL) {
+    uint8_t* const info = &pic->extra_info[it->x_ + it->y_ * enc->mb_w_];
+    switch (pic->extra_info_type) {
+      case 1: *info = mb->type_; break;
+      case 2: *info = mb->segment_; break;
+      case 3: *info = enc->dqm_[mb->segment_].quant_; break;
+      case 4: *info = (mb->type_ == 1) ? it->preds_[0] : 0xff; break;
+      case 5: *info = mb->uv_mode_; break;
+      case 6: {
+        const int b = (int)((it->luma_bits_ + it->uv_bits_ + 7) >> 3);
+        *info = (b > 255) ? 255 : b; break;
+      }
+      case 7: *info = mb->alpha_; break;
+      default: *info = 0; break;
+    };
+  }
+#if SEGMENT_VISU  // visualize segments and prediction modes
+  SetBlock(it->yuv_out_ + Y_OFF, mb->segment_ * 64, 16);
+  SetBlock(it->yuv_out_ + U_OFF, it->preds_[0] * 64, 8);
+  SetBlock(it->yuv_out_ + V_OFF, mb->uv_mode_ * 64, 8);
+#endif
+}
+
+static double GetPSNR(uint64_t mse, uint64_t size) {
+  return (mse > 0 && size > 0) ? 10. * log10(255. * 255. * size / mse) : 99;
+}
+
+//------------------------------------------------------------------------------
+//  StatLoop(): only collect statistics (number of skips, token usage, ...).
+//  This is used for deciding optimal probabilities. It also modifies the
+//  quantizer value if some target (size, PSNR) was specified.
+
+static void SetLoopParams(VP8Encoder* const enc, float q) {
+  // Make sure the quality parameter is inside valid bounds
+  q = Clamp(q, 0.f, 100.f);
+
+  VP8SetSegmentParams(enc, q);      // setup segment quantizations and filters
+  SetSegmentProbas(enc);            // compute segment probabilities
+
+  ResetStats(enc);
+  ResetSSE(enc);
+}
+
+static uint64_t OneStatPass(VP8Encoder* const enc, VP8RDLevel rd_opt,
+                            int nb_mbs, int percent_delta,
+                            PassStats* const s) {
+  VP8EncIterator it;
+  uint64_t size = 0;
+  uint64_t size_p0 = 0;
+  uint64_t distortion = 0;
+  const uint64_t pixel_count = nb_mbs * 384;
+
+  VP8IteratorInit(enc, &it);
+  SetLoopParams(enc, s->q);
+  do {
+    VP8ModeScore info;
+    VP8IteratorImport(&it, NULL);
+    if (VP8Decimate(&it, &info, rd_opt)) {
+      // Just record the number of skips and act like skip_proba is not used.
+      enc->proba_.nb_skip_++;
+    }
+    RecordResiduals(&it, &info);
+    size += info.R + info.H;
+    size_p0 += info.H;
+    distortion += info.D;
+    if (percent_delta && !VP8IteratorProgress(&it, percent_delta))
+      return 0;
+    VP8IteratorSaveBoundary(&it);
+  } while (VP8IteratorNext(&it) && --nb_mbs > 0);
+
+  size_p0 += enc->segment_hdr_.size_;
+  if (s->do_size_search) {
+    size += FinalizeSkipProba(enc);
+    size += FinalizeTokenProbas(&enc->proba_);
+    size = ((size + size_p0 + 1024) >> 11) + HEADER_SIZE_ESTIMATE;
+    s->value = (double)size;
+  } else {
+    s->value = GetPSNR(distortion, pixel_count);
+  }
+  return size_p0;
+}
+
+static int StatLoop(VP8Encoder* const enc) {
+  const int method = enc->method_;
+  const int do_search = enc->do_search_;
+  const int fast_probe = ((method == 0 || method == 3) && !do_search);
+  int num_pass_left = enc->config_->pass;
+  const int task_percent = 20;
+  const int percent_per_pass =
+      (task_percent + num_pass_left / 2) / num_pass_left;
+  const int final_percent = enc->percent_ + task_percent;
+  const VP8RDLevel rd_opt =
+      (method >= 3 || do_search) ? RD_OPT_BASIC : RD_OPT_NONE;
+  int nb_mbs = enc->mb_w_ * enc->mb_h_;
+  PassStats stats;
+
+  InitPassStats(enc, &stats);
+  ResetTokenStats(enc);
+
+  // Fast mode: quick analysis pass over few mbs. Better than nothing.
+  if (fast_probe) {
+    if (method == 3) {  // we need more stats for method 3 to be reliable.
+      nb_mbs = (nb_mbs > 200) ? nb_mbs >> 1 : 100;
+    } else {
+      nb_mbs = (nb_mbs > 200) ? nb_mbs >> 2 : 50;
+    }
+  }
+
+  while (num_pass_left-- > 0) {
+    const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
+                             (num_pass_left == 0) ||
+                             (enc->max_i4_header_bits_ == 0);
+    const uint64_t size_p0 =
+        OneStatPass(enc, rd_opt, nb_mbs, percent_per_pass, &stats);
+    if (size_p0 == 0) return 0;
+#if (DEBUG_SEARCH > 0)
+    printf("#%d value:%.1lf -> %.1lf   q:%.2f -> %.2f\n",
+           num_pass_left, stats.last_value, stats.value, stats.last_q, stats.q);
+#endif
+    if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
+      ++num_pass_left;
+      enc->max_i4_header_bits_ >>= 1;  // strengthen header bit limitation...
+      continue;                        // ...and start over
+    }
+    if (is_last_pass) {
+      break;
+    }
+    // If no target size: just do several pass without changing 'q'
+    if (do_search) {
+      ComputeNextQ(&stats);
+      if (fabs(stats.dq) <= DQ_LIMIT) break;
+    }
+  }
+  if (!do_search || !stats.do_size_search) {
+    // Need to finalize probas now, since it wasn't done during the search.
+    FinalizeSkipProba(enc);
+    FinalizeTokenProbas(&enc->proba_);
+  }
+  VP8CalculateLevelCosts(&enc->proba_);  // finalize costs
+  return WebPReportProgress(enc->pic_, final_percent, &enc->percent_);
+}
+
+//------------------------------------------------------------------------------
+// Main loops
+//
+
+static const int kAverageBytesPerMB[8] = { 50, 24, 16, 9, 7, 5, 3, 2 };
+
+static int PreLoopInitialize(VP8Encoder* const enc) {
+  int p;
+  int ok = 1;
+  const int average_bytes_per_MB = kAverageBytesPerMB[enc->base_quant_ >> 4];
+  const int bytes_per_parts =
+      enc->mb_w_ * enc->mb_h_ * average_bytes_per_MB / enc->num_parts_;
+  // Initialize the bit-writers
+  for (p = 0; ok && p < enc->num_parts_; ++p) {
+    ok = VP8BitWriterInit(enc->parts_ + p, bytes_per_parts);
+  }
+  if (!ok) VP8EncFreeBitWriters(enc);  // malloc error occurred
+  return ok;
+}
+
+static int PostLoopFinalize(VP8EncIterator* const it, int ok) {
+  VP8Encoder* const enc = it->enc_;
+  if (ok) {      // Finalize the partitions, check for extra errors.
+    int p;
+    for (p = 0; p < enc->num_parts_; ++p) {
+      VP8BitWriterFinish(enc->parts_ + p);
+      ok &= !enc->parts_[p].error_;
+    }
+  }
+
+  if (ok) {      // All good. Finish up.
+    if (enc->pic_->stats != NULL) {  // finalize byte counters...
+      int i, s;
+      for (i = 0; i <= 2; ++i) {
+        for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+          enc->residual_bytes_[i][s] = (int)((it->bit_count_[s][i] + 7) >> 3);
+        }
+      }
+    }
+    VP8AdjustFilterStrength(it);     // ...and store filter stats.
+  } else {
+    // Something bad happened -> need to do some memory cleanup.
+    VP8EncFreeBitWriters(enc);
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+//  VP8EncLoop(): does the final bitstream coding.
+
+static void ResetAfterSkip(VP8EncIterator* const it) {
+  if (it->mb_->type_ == 1) {
+    *it->nz_ = 0;  // reset all predictors
+    it->left_nz_[8] = 0;
+  } else {
+    *it->nz_ &= (1 << 24);  // preserve the dc_nz bit
+  }
+}
+
+int VP8EncLoop(VP8Encoder* const enc) {
+  VP8EncIterator it;
+  int ok = PreLoopInitialize(enc);
+  if (!ok) return 0;
+
+  StatLoop(enc);  // stats-collection loop
+
+  VP8IteratorInit(enc, &it);
+  VP8InitFilter(&it);
+  do {
+    VP8ModeScore info;
+    const int dont_use_skip = !enc->proba_.use_skip_proba_;
+    const VP8RDLevel rd_opt = enc->rd_opt_level_;
+
+    VP8IteratorImport(&it, NULL);
+    // Warning! order is important: first call VP8Decimate() and
+    // *then* decide how to code the skip decision if there's one.
+    if (!VP8Decimate(&it, &info, rd_opt) || dont_use_skip) {
+      CodeResiduals(it.bw_, &it, &info);
+    } else {   // reset predictors after a skip
+      ResetAfterSkip(&it);
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (enc->use_layer_) {
+      VP8EncCodeLayerBlock(&it);
+    }
+#endif
+    StoreSideInfo(&it);
+    VP8StoreFilterStats(&it);
+    VP8IteratorExport(&it);
+    ok = VP8IteratorProgress(&it, 20);
+    VP8IteratorSaveBoundary(&it);
+  } while (ok && VP8IteratorNext(&it));
+
+  return PostLoopFinalize(&it, ok);
+}
+
+//------------------------------------------------------------------------------
+// Single pass using Token Buffer.
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+#define MIN_COUNT 96  // minimum number of macroblocks before updating stats
+
+int VP8EncTokenLoop(VP8Encoder* const enc) {
+  // Roughly refresh the proba eight times per pass
+  int max_count = (enc->mb_w_ * enc->mb_h_) >> 3;
+  int num_pass_left = enc->config_->pass;
+  const int do_search = enc->do_search_;
+  VP8EncIterator it;
+  VP8Proba* const proba = &enc->proba_;
+  const VP8RDLevel rd_opt = enc->rd_opt_level_;
+  const uint64_t pixel_count = enc->mb_w_ * enc->mb_h_ * 384;
+  PassStats stats;
+  int ok;
+
+  InitPassStats(enc, &stats);
+  ok = PreLoopInitialize(enc);
+  if (!ok) return 0;
+
+  if (max_count < MIN_COUNT) max_count = MIN_COUNT;
+
+  assert(enc->num_parts_ == 1);
+  assert(enc->use_tokens_);
+  assert(proba->use_skip_proba_ == 0);
+  assert(rd_opt >= RD_OPT_BASIC);   // otherwise, token-buffer won't be useful
+  assert(num_pass_left > 0);
+
+  while (ok && num_pass_left-- > 0) {
+    const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
+                             (num_pass_left == 0) ||
+                             (enc->max_i4_header_bits_ == 0);
+    uint64_t size_p0 = 0;
+    uint64_t distortion = 0;
+    int cnt = max_count;
+    VP8IteratorInit(enc, &it);
+    SetLoopParams(enc, stats.q);
+    if (is_last_pass) {
+      ResetTokenStats(enc);
+      VP8InitFilter(&it);  // don't collect stats until last pass (too costly)
+    }
+    VP8TBufferClear(&enc->tokens_);
+    do {
+      VP8ModeScore info;
+      VP8IteratorImport(&it, NULL);
+      if (--cnt < 0) {
+        FinalizeTokenProbas(proba);
+        VP8CalculateLevelCosts(proba);  // refresh cost tables for rd-opt
+        cnt = max_count;
+      }
+      VP8Decimate(&it, &info, rd_opt);
+      RecordTokens(&it, &info, &enc->tokens_);
+      size_p0 += info.H;
+      distortion += info.D;
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+      if (enc->use_layer_) {
+        VP8EncCodeLayerBlock(&it);
+      }
+#endif
+      if (is_last_pass) {
+        StoreSideInfo(&it);
+        VP8StoreFilterStats(&it);
+        VP8IteratorExport(&it);
+        ok = VP8IteratorProgress(&it, 20);
+      }
+      VP8IteratorSaveBoundary(&it);
+    } while (ok && VP8IteratorNext(&it));
+    if (!ok) break;
+
+    size_p0 += enc->segment_hdr_.size_;
+    if (stats.do_size_search) {
+      uint64_t size = FinalizeTokenProbas(&enc->proba_);
+      size += VP8EstimateTokenSize(&enc->tokens_,
+                                   (const uint8_t*)proba->coeffs_);
+      size = (size + size_p0 + 1024) >> 11;  // -> size in bytes
+      size += HEADER_SIZE_ESTIMATE;
+      stats.value = (double)size;
+    } else {  // compute and store PSNR
+      stats.value = GetPSNR(distortion, pixel_count);
+    }
+
+#if (DEBUG_SEARCH > 0)
+    printf("#%2d metric:%.1lf -> %.1lf   last_q=%.2lf q=%.2lf dq=%.2lf\n",
+           num_pass_left, stats.last_value, stats.value,
+           stats.last_q, stats.q, stats.dq);
+#endif
+    if (size_p0 > PARTITION0_SIZE_LIMIT) {
+      ++num_pass_left;
+      enc->max_i4_header_bits_ >>= 1;  // strengthen header bit limitation...
+      continue;                        // ...and start over
+    }
+    if (is_last_pass) {
+      break;   // done
+    }
+    if (do_search) {
+      ComputeNextQ(&stats);  // Adjust q
+    }
+  }
+  if (ok) {
+    if (!stats.do_size_search) {
+      FinalizeTokenProbas(&enc->proba_);
+    }
+    ok = VP8EmitTokens(&enc->tokens_, enc->parts_ + 0,
+                       (const uint8_t*)proba->coeffs_, 1);
+  }
+  ok = ok && WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
+  return PostLoopFinalize(&it, ok);
+}
+
+#else
+
+int VP8EncTokenLoop(VP8Encoder* const enc) {
+  (void)enc;
+  return 0;   // we shouldn't be here.
+}
+
+#endif    // DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/enc/histogram.c b/drivers/webp/enc/histogram.c
new file mode 100644
index 0000000000..abd253bd7c
--- /dev/null
+++ b/drivers/webp/enc/histogram.c
@@ -0,0 +1,512 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+//
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <math.h>
+#include <stdio.h>
+
+#include "./backward_references.h"
+#include "./histogram.h"
+#include "../dsp/lossless.h"
+#include "../utils/utils.h"
+
+static void HistogramClear(VP8LHistogram* const p) {
+  memset(p->literal_, 0, sizeof(p->literal_));
+  memset(p->red_, 0, sizeof(p->red_));
+  memset(p->blue_, 0, sizeof(p->blue_));
+  memset(p->alpha_, 0, sizeof(p->alpha_));
+  memset(p->distance_, 0, sizeof(p->distance_));
+  p->bit_cost_ = 0;
+}
+
+void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
+                            VP8LHistogram* const histo) {
+  int i;
+  for (i = 0; i < refs->size; ++i) {
+    VP8LHistogramAddSinglePixOrCopy(histo, &refs->refs[i]);
+  }
+}
+
+void VP8LHistogramCreate(VP8LHistogram* const p,
+                         const VP8LBackwardRefs* const refs,
+                         int palette_code_bits) {
+  if (palette_code_bits >= 0) {
+    p->palette_code_bits_ = palette_code_bits;
+  }
+  HistogramClear(p);
+  VP8LHistogramStoreRefs(refs, p);
+}
+
+void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits) {
+  p->palette_code_bits_ = palette_code_bits;
+  HistogramClear(p);
+}
+
+VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
+  int i;
+  VP8LHistogramSet* set;
+  VP8LHistogram* bulk;
+  const uint64_t total_size = sizeof(*set)
+                            + (uint64_t)size * sizeof(*set->histograms)
+                            + (uint64_t)size * sizeof(**set->histograms);
+  uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
+  if (memory == NULL) return NULL;
+
+  set = (VP8LHistogramSet*)memory;
+  memory += sizeof(*set);
+  set->histograms = (VP8LHistogram**)memory;
+  memory += size * sizeof(*set->histograms);
+  bulk = (VP8LHistogram*)memory;
+  set->max_size = size;
+  set->size = size;
+  for (i = 0; i < size; ++i) {
+    set->histograms[i] = bulk + i;
+    VP8LHistogramInit(set->histograms[i], cache_bits);
+  }
+  return set;
+}
+
+// -----------------------------------------------------------------------------
+
+void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
+                                     const PixOrCopy* const v) {
+  if (PixOrCopyIsLiteral(v)) {
+    ++histo->alpha_[PixOrCopyLiteral(v, 3)];
+    ++histo->red_[PixOrCopyLiteral(v, 2)];
+    ++histo->literal_[PixOrCopyLiteral(v, 1)];
+    ++histo->blue_[PixOrCopyLiteral(v, 0)];
+  } else if (PixOrCopyIsCacheIdx(v)) {
+    int literal_ix = 256 + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
+    ++histo->literal_[literal_ix];
+  } else {
+    int code, extra_bits;
+    VP8LPrefixEncodeBits(PixOrCopyLength(v), &code, &extra_bits);
+    ++histo->literal_[256 + code];
+    VP8LPrefixEncodeBits(PixOrCopyDistance(v), &code, &extra_bits);
+    ++histo->distance_[code];
+  }
+}
+
+static double BitsEntropy(const int* const array, int n) {
+  double retval = 0.;
+  int sum = 0;
+  int nonzeros = 0;
+  int max_val = 0;
+  int i;
+  double mix;
+  for (i = 0; i < n; ++i) {
+    if (array[i] != 0) {
+      sum += array[i];
+      ++nonzeros;
+      retval -= VP8LFastSLog2(array[i]);
+      if (max_val < array[i]) {
+        max_val = array[i];
+      }
+    }
+  }
+  retval += VP8LFastSLog2(sum);
+
+  if (nonzeros < 5) {
+    if (nonzeros <= 1) {
+      return 0;
+    }
+    // Two symbols, they will be 0 and 1 in a Huffman code.
+    // Let's mix in a bit of entropy to favor good clustering when
+    // distributions of these are combined.
+    if (nonzeros == 2) {
+      return 0.99 * sum + 0.01 * retval;
+    }
+    // No matter what the entropy says, we cannot be better than min_limit
+    // with Huffman coding. I am mixing a bit of entropy into the
+    // min_limit since it produces much better (~0.5 %) compression results
+    // perhaps because of better entropy clustering.
+    if (nonzeros == 3) {
+      mix = 0.95;
+    } else {
+      mix = 0.7;  // nonzeros == 4.
+    }
+  } else {
+    mix = 0.627;
+  }
+
+  {
+    double min_limit = 2 * sum - max_val;
+    min_limit = mix * min_limit + (1.0 - mix) * retval;
+    return (retval < min_limit) ? min_limit : retval;
+  }
+}
+
+// Returns the cost encode the rle-encoded entropy code.
+// The constants in this function are experimental.
+static double HuffmanCost(const int* const population, int length) {
+  // Small bias because Huffman code length is typically not stored in
+  // full length.
+  static const int kHuffmanCodeOfHuffmanCodeSize = CODE_LENGTH_CODES * 3;
+  static const double kSmallBias = 9.1;
+  double retval = kHuffmanCodeOfHuffmanCodeSize - kSmallBias;
+  int streak = 0;
+  int i = 0;
+  for (; i < length - 1; ++i) {
+    ++streak;
+    if (population[i] == population[i + 1]) {
+      continue;
+    }
+ last_streak_hack:
+    // population[i] points now to the symbol in the streak of same values.
+    if (streak > 3) {
+      if (population[i] == 0) {
+        retval += 1.5625 + 0.234375 * streak;
+      } else {
+        retval += 2.578125 + 0.703125 * streak;
+      }
+    } else {
+      if (population[i] == 0) {
+        retval += 1.796875 * streak;
+      } else {
+        retval += 3.28125 * streak;
+      }
+    }
+    streak = 0;
+  }
+  if (i == length - 1) {
+    ++streak;
+    goto last_streak_hack;
+  }
+  return retval;
+}
+
+static double PopulationCost(const int* const population, int length) {
+  return BitsEntropy(population, length) + HuffmanCost(population, length);
+}
+
+static double ExtraCost(const int* const population, int length) {
+  int i;
+  double cost = 0.;
+  for (i = 2; i < length - 2; ++i) cost += (i >> 1) * population[i + 2];
+  return cost;
+}
+
+// Estimates the Entropy + Huffman + other block overhead size cost.
+double VP8LHistogramEstimateBits(const VP8LHistogram* const p) {
+  return PopulationCost(p->literal_, VP8LHistogramNumCodes(p))
+       + PopulationCost(p->red_, 256)
+       + PopulationCost(p->blue_, 256)
+       + PopulationCost(p->alpha_, 256)
+       + PopulationCost(p->distance_, NUM_DISTANCE_CODES)
+       + ExtraCost(p->literal_ + 256, NUM_LENGTH_CODES)
+       + ExtraCost(p->distance_, NUM_DISTANCE_CODES);
+}
+
+double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p) {
+  return BitsEntropy(p->literal_, VP8LHistogramNumCodes(p))
+       + BitsEntropy(p->red_, 256)
+       + BitsEntropy(p->blue_, 256)
+       + BitsEntropy(p->alpha_, 256)
+       + BitsEntropy(p->distance_, NUM_DISTANCE_CODES)
+       + ExtraCost(p->literal_ + 256, NUM_LENGTH_CODES)
+       + ExtraCost(p->distance_, NUM_DISTANCE_CODES);
+}
+
+// -----------------------------------------------------------------------------
+// Various histogram combine/cost-eval functions
+
+// Adds 'in' histogram to 'out'
+static void HistogramAdd(const VP8LHistogram* const in,
+                         VP8LHistogram* const out) {
+  int i;
+  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
+    out->literal_[i] += in->literal_[i];
+  }
+  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
+    out->distance_[i] += in->distance_[i];
+  }
+  for (i = 0; i < 256; ++i) {
+    out->red_[i] += in->red_[i];
+    out->blue_[i] += in->blue_[i];
+    out->alpha_[i] += in->alpha_[i];
+  }
+}
+
+// Performs out = a + b, computing the cost C(a+b) - C(a) - C(b) while comparing
+// to the threshold value 'cost_threshold'. The score returned is
+//  Score = C(a+b) - C(a) - C(b), where C(a) + C(b) is known and fixed.
+// Since the previous score passed is 'cost_threshold', we only need to compare
+// the partial cost against 'cost_threshold + C(a) + C(b)' to possibly bail-out
+// early.
+static double HistogramAddEval(const VP8LHistogram* const a,
+                               const VP8LHistogram* const b,
+                               VP8LHistogram* const out,
+                               double cost_threshold) {
+  double cost = 0;
+  const double sum_cost = a->bit_cost_ + b->bit_cost_;
+  int i;
+
+  cost_threshold += sum_cost;
+
+  // palette_code_bits_ is part of the cost evaluation for literal_.
+  // TODO(skal): remove/simplify this palette_code_bits_?
+  out->palette_code_bits_ =
+      (a->palette_code_bits_ > b->palette_code_bits_) ? a->palette_code_bits_ :
+                                                        b->palette_code_bits_;
+  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
+    out->literal_[i] = a->literal_[i] + b->literal_[i];
+  }
+  cost += PopulationCost(out->literal_, VP8LHistogramNumCodes(out));
+  cost += ExtraCost(out->literal_ + 256, NUM_LENGTH_CODES);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) out->red_[i] = a->red_[i] + b->red_[i];
+  cost += PopulationCost(out->red_, 256);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) out->blue_[i] = a->blue_[i] + b->blue_[i];
+  cost += PopulationCost(out->blue_, 256);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
+    out->distance_[i] = a->distance_[i] + b->distance_[i];
+  }
+  cost += PopulationCost(out->distance_, NUM_DISTANCE_CODES);
+  cost += ExtraCost(out->distance_, NUM_DISTANCE_CODES);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) out->alpha_[i] = a->alpha_[i] + b->alpha_[i];
+  cost += PopulationCost(out->alpha_, 256);
+
+  out->bit_cost_ = cost;
+  return cost - sum_cost;
+}
+
+// Same as HistogramAddEval(), except that the resulting histogram
+// is not stored. Only the cost C(a+b) - C(a) is evaluated. We omit
+// the term C(b) which is constant over all the evaluations.
+static double HistogramAddThresh(const VP8LHistogram* const a,
+                                 const VP8LHistogram* const b,
+                                 double cost_threshold) {
+  int tmp[PIX_OR_COPY_CODES_MAX];  // <= max storage we'll need
+  int i;
+  double cost = -a->bit_cost_;
+
+  for (i = 0; i < PIX_OR_COPY_CODES_MAX; ++i) {
+    tmp[i] = a->literal_[i] + b->literal_[i];
+  }
+  // note that the tests are ordered so that the usually largest
+  // cost shares come first.
+  cost += PopulationCost(tmp, VP8LHistogramNumCodes(a));
+  cost += ExtraCost(tmp + 256, NUM_LENGTH_CODES);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) tmp[i] = a->red_[i] + b->red_[i];
+  cost += PopulationCost(tmp, 256);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) tmp[i] = a->blue_[i] + b->blue_[i];
+  cost += PopulationCost(tmp, 256);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
+    tmp[i] = a->distance_[i] + b->distance_[i];
+  }
+  cost += PopulationCost(tmp, NUM_DISTANCE_CODES);
+  cost += ExtraCost(tmp, NUM_DISTANCE_CODES);
+  if (cost > cost_threshold) return cost;
+
+  for (i = 0; i < 256; ++i) tmp[i] = a->alpha_[i] + b->alpha_[i];
+  cost += PopulationCost(tmp, 256);
+
+  return cost;
+}
+
+// -----------------------------------------------------------------------------
+
+static void HistogramBuildImage(int xsize, int histo_bits,
+                                const VP8LBackwardRefs* const backward_refs,
+                                VP8LHistogramSet* const image) {
+  int i;
+  int x = 0, y = 0;
+  const int histo_xsize = VP8LSubSampleSize(xsize, histo_bits);
+  VP8LHistogram** const histograms = image->histograms;
+  assert(histo_bits > 0);
+  for (i = 0; i < backward_refs->size; ++i) {
+    const PixOrCopy* const v = &backward_refs->refs[i];
+    const int ix = (y >> histo_bits) * histo_xsize + (x >> histo_bits);
+    VP8LHistogramAddSinglePixOrCopy(histograms[ix], v);
+    x += PixOrCopyLength(v);
+    while (x >= xsize) {
+      x -= xsize;
+      ++y;
+    }
+  }
+}
+
+static uint32_t MyRand(uint32_t *seed) {
+  *seed *= 16807U;
+  if (*seed == 0) {
+    *seed = 1;
+  }
+  return *seed;
+}
+
+static int HistogramCombine(const VP8LHistogramSet* const in,
+                            VP8LHistogramSet* const out, int iter_mult,
+                            int num_pairs, int num_tries_no_success) {
+  int ok = 0;
+  int i, iter;
+  uint32_t seed = 0;
+  int tries_with_no_success = 0;
+  int out_size = in->size;
+  const int outer_iters = in->size * iter_mult;
+  const int min_cluster_size = 2;
+  VP8LHistogram* const histos = (VP8LHistogram*)malloc(2 * sizeof(*histos));
+  VP8LHistogram* cur_combo = histos + 0;    // trial merged histogram
+  VP8LHistogram* best_combo = histos + 1;   // best merged histogram so far
+  if (histos == NULL) goto End;
+
+  // Copy histograms from in[] to out[].
+  assert(in->size <= out->size);
+  for (i = 0; i < in->size; ++i) {
+    in->histograms[i]->bit_cost_ = VP8LHistogramEstimateBits(in->histograms[i]);
+    *out->histograms[i] = *in->histograms[i];
+  }
+
+  // Collapse similar histograms in 'out'.
+  for (iter = 0; iter < outer_iters && out_size >= min_cluster_size; ++iter) {
+    double best_cost_diff = 0.;
+    int best_idx1 = -1, best_idx2 = 1;
+    int j;
+    const int num_tries = (num_pairs < out_size) ? num_pairs : out_size;
+    seed += iter;
+    for (j = 0; j < num_tries; ++j) {
+      double curr_cost_diff;
+      // Choose two histograms at random and try to combine them.
+      const uint32_t idx1 = MyRand(&seed) % out_size;
+      const uint32_t tmp = (j & 7) + 1;
+      const uint32_t diff = (tmp < 3) ? tmp : MyRand(&seed) % (out_size - 1);
+      const uint32_t idx2 = (idx1 + diff + 1) % out_size;
+      if (idx1 == idx2) {
+        continue;
+      }
+      // Calculate cost reduction on combining.
+      curr_cost_diff = HistogramAddEval(out->histograms[idx1],
+                                        out->histograms[idx2],
+                                        cur_combo, best_cost_diff);
+      if (curr_cost_diff < best_cost_diff) {    // found a better pair?
+        {     // swap cur/best combo histograms
+          VP8LHistogram* const tmp_histo = cur_combo;
+          cur_combo = best_combo;
+          best_combo = tmp_histo;
+        }
+        best_cost_diff = curr_cost_diff;
+        best_idx1 = idx1;
+        best_idx2 = idx2;
+      }
+    }
+
+    if (best_idx1 >= 0) {
+      *out->histograms[best_idx1] = *best_combo;
+      // swap best_idx2 slot with last one (which is now unused)
+      --out_size;
+      if (best_idx2 != out_size) {
+        out->histograms[best_idx2] = out->histograms[out_size];
+        out->histograms[out_size] = NULL;   // just for sanity check.
+      }
+      tries_with_no_success = 0;
+    }
+    if (++tries_with_no_success >= num_tries_no_success) {
+      break;
+    }
+  }
+  out->size = out_size;
+  ok = 1;
+
+ End:
+  free(histos);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+// Histogram refinement
+
+// What is the bit cost of moving square_histogram from cur_symbol to candidate.
+static double HistogramDistance(const VP8LHistogram* const square_histogram,
+                                const VP8LHistogram* const candidate,
+                                double cost_threshold) {
+  return HistogramAddThresh(candidate, square_histogram, cost_threshold);
+}
+
+// Find the best 'out' histogram for each of the 'in' histograms.
+// Note: we assume that out[]->bit_cost_ is already up-to-date.
+static void HistogramRemap(const VP8LHistogramSet* const in,
+                           const VP8LHistogramSet* const out,
+                           uint16_t* const symbols) {
+  int i;
+  for (i = 0; i < in->size; ++i) {
+    int best_out = 0;
+    double best_bits =
+        HistogramDistance(in->histograms[i], out->histograms[0], 1.e38);
+    int k;
+    for (k = 1; k < out->size; ++k) {
+      const double cur_bits =
+          HistogramDistance(in->histograms[i], out->histograms[k], best_bits);
+      if (cur_bits < best_bits) {
+        best_bits = cur_bits;
+        best_out = k;
+      }
+    }
+    symbols[i] = best_out;
+  }
+
+  // Recompute each out based on raw and symbols.
+  for (i = 0; i < out->size; ++i) {
+    HistogramClear(out->histograms[i]);
+  }
+  for (i = 0; i < in->size; ++i) {
+    HistogramAdd(in->histograms[i], out->histograms[symbols[i]]);
+  }
+}
+
+int VP8LGetHistoImageSymbols(int xsize, int ysize,
+                             const VP8LBackwardRefs* const refs,
+                             int quality, int histo_bits, int cache_bits,
+                             VP8LHistogramSet* const image_in,
+                             uint16_t* const histogram_symbols) {
+  int ok = 0;
+  const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
+  const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
+  const int histo_image_raw_size = histo_xsize * histo_ysize;
+
+  // Heuristic params for HistogramCombine().
+  const int num_tries_no_success = 8 + (quality >> 1);
+  const int iter_mult = (quality < 27) ? 1 : 1 + ((quality - 27) >> 4);
+  const int num_pairs = (quality < 25) ? 10 : (5 * quality) >> 3;
+
+  VP8LHistogramSet* const image_out =
+      VP8LAllocateHistogramSet(histo_image_raw_size, cache_bits);
+  if (image_out == NULL) return 0;
+
+  // Build histogram image.
+  HistogramBuildImage(xsize, histo_bits, refs, image_out);
+  // Collapse similar histograms.
+  if (!HistogramCombine(image_out, image_in, iter_mult, num_pairs,
+                        num_tries_no_success)) {
+    goto Error;
+  }
+  // Find the optimal map from original histograms to the final ones.
+  HistogramRemap(image_out, image_in, histogram_symbols);
+  ok = 1;
+
+Error:
+  free(image_out);
+  return ok;
+}
diff --git a/drivers/webp/enc/histogram.h b/drivers/webp/enc/histogram.h
new file mode 100644
index 0000000000..4d346a857b
--- /dev/null
+++ b/drivers/webp/enc/histogram.h
@@ -0,0 +1,101 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+//
+// Models the histograms of literal and distance codes.
+
+#ifndef WEBP_ENC_HISTOGRAM_H_
+#define WEBP_ENC_HISTOGRAM_H_
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "./backward_references.h"
+#include "../webp/format_constants.h"
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// A simple container for histograms of data.
+typedef struct {
+  // literal_ contains green literal, palette-code and
+  // copy-length-prefix histogram
+  int literal_[PIX_OR_COPY_CODES_MAX];
+  int red_[256];
+  int blue_[256];
+  int alpha_[256];
+  // Backward reference prefix-code histogram.
+  int distance_[NUM_DISTANCE_CODES];
+  int palette_code_bits_;
+  double bit_cost_;   // cached value of VP8LHistogramEstimateBits(this)
+} VP8LHistogram;
+
+// Collection of histograms with fixed capacity, allocated as one
+// big memory chunk. Can be destroyed by simply calling 'free()'.
+typedef struct {
+  int size;         // number of slots currently in use
+  int max_size;     // maximum capacity
+  VP8LHistogram** histograms;
+} VP8LHistogramSet;
+
+// Create the histogram.
+//
+// The input data is the PixOrCopy data, which models the literals, stop
+// codes and backward references (both distances and lengths).  Also: if
+// palette_code_bits is >= 0, initialize the histogram with this value.
+void VP8LHistogramCreate(VP8LHistogram* const p,
+                         const VP8LBackwardRefs* const refs,
+                         int palette_code_bits);
+
+// Set the palette_code_bits and reset the stats.
+void VP8LHistogramInit(VP8LHistogram* const p, int palette_code_bits);
+
+// Collect all the references into a histogram (without reset)
+void VP8LHistogramStoreRefs(const VP8LBackwardRefs* const refs,
+                            VP8LHistogram* const histo);
+
+// Allocate an array of pointer to histograms, allocated and initialized
+// using 'cache_bits'. Return NULL in case of memory error.
+VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits);
+
+// Accumulate a token 'v' into a histogram.
+void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
+                                     const PixOrCopy* const v);
+
+// Estimate how many bits the combined entropy of literals and distance
+// approximately maps to.
+double VP8LHistogramEstimateBits(const VP8LHistogram* const p);
+
+// This function estimates the cost in bits excluding the bits needed to
+// represent the entropy code itself.
+double VP8LHistogramEstimateBitsBulk(const VP8LHistogram* const p);
+
+static WEBP_INLINE int VP8LHistogramNumCodes(const VP8LHistogram* const p) {
+  return 256 + NUM_LENGTH_CODES +
+      ((p->palette_code_bits_ > 0) ? (1 << p->palette_code_bits_) : 0);
+}
+
+// Builds the histogram image.
+int VP8LGetHistoImageSymbols(int xsize, int ysize,
+                             const VP8LBackwardRefs* const refs,
+                             int quality, int histogram_bits, int cache_bits,
+                             VP8LHistogramSet* const image_in,
+                             uint16_t* const histogram_symbols);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // WEBP_ENC_HISTOGRAM_H_
diff --git a/drivers/webp/enc/iterator.c b/drivers/webp/enc/iterator.c
new file mode 100644
index 0000000000..e42ad001ac
--- /dev/null
+++ b/drivers/webp/enc/iterator.c
@@ -0,0 +1,456 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// VP8Iterator: block iterator
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <string.h>
+
+#include "./vp8enci.h"
+
+//------------------------------------------------------------------------------
+// VP8Iterator
+//------------------------------------------------------------------------------
+
+static void InitLeft(VP8EncIterator* const it) {
+  it->y_left_[-1] = it->u_left_[-1] = it->v_left_[-1] =
+      (it->y_ > 0) ? 129 : 127;
+  memset(it->y_left_, 129, 16);
+  memset(it->u_left_, 129, 8);
+  memset(it->v_left_, 129, 8);
+  it->left_nz_[8] = 0;
+}
+
+static void InitTop(VP8EncIterator* const it) {
+  const VP8Encoder* const enc = it->enc_;
+  const size_t top_size = enc->mb_w_ * 16;
+  memset(enc->y_top_, 127, 2 * top_size);
+  memset(enc->nz_, 0, enc->mb_w_ * sizeof(*enc->nz_));
+}
+
+void VP8IteratorSetRow(VP8EncIterator* const it, int y) {
+  VP8Encoder* const enc = it->enc_;
+  it->x_ = 0;
+  it->y_ = y;
+  it->bw_ = &enc->parts_[y & (enc->num_parts_ - 1)];
+  it->preds_ = enc->preds_ + y * 4 * enc->preds_w_;
+  it->nz_ = enc->nz_;
+  it->mb_ = enc->mb_info_ + y * enc->mb_w_;
+  it->y_top_ = enc->y_top_;
+  it->uv_top_ = enc->uv_top_;
+  InitLeft(it);
+}
+
+void VP8IteratorReset(VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  VP8IteratorSetRow(it, 0);
+  VP8IteratorSetCountDown(it, enc->mb_w_ * enc->mb_h_);  // default
+  InitTop(it);
+  InitLeft(it);
+  memset(it->bit_count_, 0, sizeof(it->bit_count_));
+  it->do_trellis_ = 0;
+}
+
+void VP8IteratorSetCountDown(VP8EncIterator* const it, int count_down) {
+  it->count_down_ = it->count_down0_ = count_down;
+}
+
+int VP8IteratorIsDone(const VP8EncIterator* const it) {
+  return (it->count_down_ <= 0);
+}
+
+void VP8IteratorInit(VP8Encoder* const enc, VP8EncIterator* const it) {
+  it->enc_ = enc;
+  it->y_stride_  = enc->pic_->y_stride;
+  it->uv_stride_ = enc->pic_->uv_stride;
+  it->yuv_in_   = (uint8_t*)DO_ALIGN(it->yuv_mem_);
+  it->yuv_out_  = it->yuv_in_ + YUV_SIZE;
+  it->yuv_out2_ = it->yuv_out_ + YUV_SIZE;
+  it->yuv_p_    = it->yuv_out2_ + YUV_SIZE;
+  it->lf_stats_ = enc->lf_stats_;
+  it->percent0_ = enc->percent_;
+  it->y_left_ = (uint8_t*)DO_ALIGN(it->yuv_left_mem_ + 1);
+  it->u_left_ = it->y_left_ + 16 + 16;
+  it->v_left_ = it->u_left_ + 16;
+  VP8IteratorReset(it);
+}
+
+int VP8IteratorProgress(const VP8EncIterator* const it, int delta) {
+  VP8Encoder* const enc = it->enc_;
+  if (delta && enc->pic_->progress_hook != NULL) {
+    const int done = it->count_down0_ - it->count_down_;
+    const int percent = (it->count_down0_ <= 0)
+                      ? it->percent0_
+                      : it->percent0_ + delta * done / it->count_down0_;
+    return WebPReportProgress(enc->pic_, percent, &enc->percent_);
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Import the source samples into the cache. Takes care of replicating
+// boundary pixels if necessary.
+
+static WEBP_INLINE int MinSize(int a, int b) { return (a < b) ? a : b; }
+
+static void ImportBlock(const uint8_t* src, int src_stride,
+                        uint8_t* dst, int w, int h, int size) {
+  int i;
+  for (i = 0; i < h; ++i) {
+    memcpy(dst, src, w);
+    if (w < size) {
+      memset(dst + w, dst[w - 1], size - w);
+    }
+    dst += BPS;
+    src += src_stride;
+  }
+  for (i = h; i < size; ++i) {
+    memcpy(dst, dst - BPS, size);
+    dst += BPS;
+  }
+}
+
+static void ImportLine(const uint8_t* src, int src_stride,
+                       uint8_t* dst, int len, int total_len) {
+  int i;
+  for (i = 0; i < len; ++i, src += src_stride) dst[i] = *src;
+  for (; i < total_len; ++i) dst[i] = dst[len - 1];
+}
+
+void VP8IteratorImport(VP8EncIterator* const it, uint8_t* tmp_32) {
+  const VP8Encoder* const enc = it->enc_;
+  const int x = it->x_, y = it->y_;
+  const WebPPicture* const pic = enc->pic_;
+  const uint8_t* const ysrc = pic->y + (y * pic->y_stride  + x) * 16;
+  const uint8_t* const usrc = pic->u + (y * pic->uv_stride + x) * 8;
+  const uint8_t* const vsrc = pic->v + (y * pic->uv_stride + x) * 8;
+  const int w = MinSize(pic->width - x * 16, 16);
+  const int h = MinSize(pic->height - y * 16, 16);
+  const int uv_w = (w + 1) >> 1;
+  const int uv_h = (h + 1) >> 1;
+
+  ImportBlock(ysrc, pic->y_stride,  it->yuv_in_ + Y_OFF, w, h, 16);
+  ImportBlock(usrc, pic->uv_stride, it->yuv_in_ + U_OFF, uv_w, uv_h, 8);
+  ImportBlock(vsrc, pic->uv_stride, it->yuv_in_ + V_OFF, uv_w, uv_h, 8);
+
+  if (tmp_32 == NULL) return;
+
+  // Import source (uncompressed) samples into boundary.
+  if (x == 0) {
+    InitLeft(it);
+  } else {
+    if (y == 0) {
+      it->y_left_[-1] = it->u_left_[-1] = it->v_left_[-1] = 127;
+    } else {
+      it->y_left_[-1] = ysrc[- 1 - pic->y_stride];
+      it->u_left_[-1] = usrc[- 1 - pic->uv_stride];
+      it->v_left_[-1] = vsrc[- 1 - pic->uv_stride];
+    }
+    ImportLine(ysrc - 1, pic->y_stride,  it->y_left_, h,   16);
+    ImportLine(usrc - 1, pic->uv_stride, it->u_left_, uv_h, 8);
+    ImportLine(vsrc - 1, pic->uv_stride, it->v_left_, uv_h, 8);
+  }
+
+  it->y_top_  = tmp_32 + 0;
+  it->uv_top_ = tmp_32 + 16;
+  if (y == 0) {
+    memset(tmp_32, 127, 32 * sizeof(*tmp_32));
+  } else {
+    ImportLine(ysrc - pic->y_stride,  1, tmp_32,          w,   16);
+    ImportLine(usrc - pic->uv_stride, 1, tmp_32 + 16,     uv_w, 8);
+    ImportLine(vsrc - pic->uv_stride, 1, tmp_32 + 16 + 8, uv_w, 8);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Copy back the compressed samples into user space if requested.
+
+static void ExportBlock(const uint8_t* src, uint8_t* dst, int dst_stride,
+                        int w, int h) {
+  while (h-- > 0) {
+    memcpy(dst, src, w);
+    dst += dst_stride;
+    src += BPS;
+  }
+}
+
+void VP8IteratorExport(const VP8EncIterator* const it) {
+  const VP8Encoder* const enc = it->enc_;
+  if (enc->config_->show_compressed) {
+    const int x = it->x_, y = it->y_;
+    const uint8_t* const ysrc = it->yuv_out_ + Y_OFF;
+    const uint8_t* const usrc = it->yuv_out_ + U_OFF;
+    const uint8_t* const vsrc = it->yuv_out_ + V_OFF;
+    const WebPPicture* const pic = enc->pic_;
+    uint8_t* const ydst = pic->y + (y * pic->y_stride + x) * 16;
+    uint8_t* const udst = pic->u + (y * pic->uv_stride + x) * 8;
+    uint8_t* const vdst = pic->v + (y * pic->uv_stride + x) * 8;
+    int w = (pic->width - x * 16);
+    int h = (pic->height - y * 16);
+
+    if (w > 16) w = 16;
+    if (h > 16) h = 16;
+
+    // Luma plane
+    ExportBlock(ysrc, ydst, pic->y_stride, w, h);
+
+    {   // U/V planes
+      const int uv_w = (w + 1) >> 1;
+      const int uv_h = (h + 1) >> 1;
+      ExportBlock(usrc, udst, pic->uv_stride, uv_w, uv_h);
+      ExportBlock(vsrc, vdst, pic->uv_stride, uv_w, uv_h);
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+// Non-zero contexts setup/teardown
+
+// Nz bits:
+//  0  1  2  3  Y
+//  4  5  6  7
+//  8  9 10 11
+// 12 13 14 15
+// 16 17        U
+// 18 19
+// 20 21        V
+// 22 23
+// 24           DC-intra16
+
+// Convert packed context to byte array
+#define BIT(nz, n) (!!((nz) & (1 << (n))))
+
+void VP8IteratorNzToBytes(VP8EncIterator* const it) {
+  const int tnz = it->nz_[0], lnz = it->nz_[-1];
+  int* const top_nz = it->top_nz_;
+  int* const left_nz = it->left_nz_;
+
+  // Top-Y
+  top_nz[0] = BIT(tnz, 12);
+  top_nz[1] = BIT(tnz, 13);
+  top_nz[2] = BIT(tnz, 14);
+  top_nz[3] = BIT(tnz, 15);
+  // Top-U
+  top_nz[4] = BIT(tnz, 18);
+  top_nz[5] = BIT(tnz, 19);
+  // Top-V
+  top_nz[6] = BIT(tnz, 22);
+  top_nz[7] = BIT(tnz, 23);
+  // DC
+  top_nz[8] = BIT(tnz, 24);
+
+  // left-Y
+  left_nz[0] = BIT(lnz,  3);
+  left_nz[1] = BIT(lnz,  7);
+  left_nz[2] = BIT(lnz, 11);
+  left_nz[3] = BIT(lnz, 15);
+  // left-U
+  left_nz[4] = BIT(lnz, 17);
+  left_nz[5] = BIT(lnz, 19);
+  // left-V
+  left_nz[6] = BIT(lnz, 21);
+  left_nz[7] = BIT(lnz, 23);
+  // left-DC is special, iterated separately
+}
+
+void VP8IteratorBytesToNz(VP8EncIterator* const it) {
+  uint32_t nz = 0;
+  const int* const top_nz = it->top_nz_;
+  const int* const left_nz = it->left_nz_;
+  // top
+  nz |= (top_nz[0] << 12) | (top_nz[1] << 13);
+  nz |= (top_nz[2] << 14) | (top_nz[3] << 15);
+  nz |= (top_nz[4] << 18) | (top_nz[5] << 19);
+  nz |= (top_nz[6] << 22) | (top_nz[7] << 23);
+  nz |= (top_nz[8] << 24);  // we propagate the _top_ bit, esp. for intra4
+  // left
+  nz |= (left_nz[0] << 3) | (left_nz[1] << 7);
+  nz |= (left_nz[2] << 11);
+  nz |= (left_nz[4] << 17) | (left_nz[6] << 21);
+
+  *it->nz_ = nz;
+}
+
+#undef BIT
+
+//------------------------------------------------------------------------------
+// Advance to the next position, doing the bookkeeping.
+
+void VP8IteratorSaveBoundary(VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
+  const int x = it->x_, y = it->y_;
+  const uint8_t* const ysrc = it->yuv_out_ + Y_OFF;
+  const uint8_t* const uvsrc = it->yuv_out_ + U_OFF;
+  if (x < enc->mb_w_ - 1) {   // left
+    int i;
+    for (i = 0; i < 16; ++i) {
+      it->y_left_[i] = ysrc[15 + i * BPS];
+    }
+    for (i = 0; i < 8; ++i) {
+      it->u_left_[i] = uvsrc[7 + i * BPS];
+      it->v_left_[i] = uvsrc[15 + i * BPS];
+    }
+    // top-left (before 'top'!)
+    it->y_left_[-1] = it->y_top_[15];
+    it->u_left_[-1] = it->uv_top_[0 + 7];
+    it->v_left_[-1] = it->uv_top_[8 + 7];
+  }
+  if (y < enc->mb_h_ - 1) {  // top
+    memcpy(it->y_top_, ysrc + 15 * BPS, 16);
+    memcpy(it->uv_top_, uvsrc + 7 * BPS, 8 + 8);
+  }
+}
+
+int VP8IteratorNext(VP8EncIterator* const it) {
+  it->preds_ += 4;
+  it->mb_ += 1;
+  it->nz_ += 1;
+  it->y_top_ += 16;
+  it->uv_top_ += 16;
+  it->x_ += 1;
+  if (it->x_ == it->enc_->mb_w_) {
+    VP8IteratorSetRow(it, ++it->y_);
+  }
+  return (0 < --it->count_down_);
+}
+
+//------------------------------------------------------------------------------
+// Helper function to set mode properties
+
+void VP8SetIntra16Mode(const VP8EncIterator* const it, int mode) {
+  uint8_t* preds = it->preds_;
+  int y;
+  for (y = 0; y < 4; ++y) {
+    memset(preds, mode, 4);
+    preds += it->enc_->preds_w_;
+  }
+  it->mb_->type_ = 1;
+}
+
+void VP8SetIntra4Mode(const VP8EncIterator* const it, const uint8_t* modes) {
+  uint8_t* preds = it->preds_;
+  int y;
+  for (y = 4; y > 0; --y) {
+    memcpy(preds, modes, 4 * sizeof(*modes));
+    preds += it->enc_->preds_w_;
+    modes += 4;
+  }
+  it->mb_->type_ = 0;
+}
+
+void VP8SetIntraUVMode(const VP8EncIterator* const it, int mode) {
+  it->mb_->uv_mode_ = mode;
+}
+
+void VP8SetSkip(const VP8EncIterator* const it, int skip) {
+  it->mb_->skip_ = skip;
+}
+
+void VP8SetSegment(const VP8EncIterator* const it, int segment) {
+  it->mb_->segment_ = segment;
+}
+
+//------------------------------------------------------------------------------
+// Intra4x4 sub-blocks iteration
+//
+//  We store and update the boundary samples into an array of 37 pixels. They
+//  are updated as we iterate and reconstructs each intra4x4 blocks in turn.
+//  The position of the samples has the following snake pattern:
+//
+// 16|17 18 19 20|21 22 23 24|25 26 27 28|29 30 31 32|33 34 35 36  <- Top-right
+// --+-----------+-----------+-----------+-----------+
+// 15|         19|         23|         27|         31|
+// 14|         18|         22|         26|         30|
+// 13|         17|         21|         25|         29|
+// 12|13 14 15 16|17 18 19 20|21 22 23 24|25 26 27 28|
+// --+-----------+-----------+-----------+-----------+
+// 11|         15|         19|         23|         27|
+// 10|         14|         18|         22|         26|
+//  9|         13|         17|         21|         25|
+//  8| 9 10 11 12|13 14 15 16|17 18 19 20|21 22 23 24|
+// --+-----------+-----------+-----------+-----------+
+//  7|         11|         15|         19|         23|
+//  6|         10|         14|         18|         22|
+//  5|          9|         13|         17|         21|
+//  4| 5  6  7  8| 9 10 11 12|13 14 15 16|17 18 19 20|
+// --+-----------+-----------+-----------+-----------+
+//  3|          7|         11|         15|         19|
+//  2|          6|         10|         14|         18|
+//  1|          5|          9|         13|         17|
+//  0| 1  2  3  4| 5  6  7  8| 9 10 11 12|13 14 15 16|
+// --+-----------+-----------+-----------+-----------+
+
+// Array to record the position of the top sample to pass to the prediction
+// functions in dsp.c.
+static const uint8_t VP8TopLeftI4[16] = {
+  17, 21, 25, 29,
+  13, 17, 21, 25,
+  9,  13, 17, 21,
+  5,   9, 13, 17
+};
+
+void VP8IteratorStartI4(VP8EncIterator* const it) {
+  const VP8Encoder* const enc = it->enc_;
+  int i;
+
+  it->i4_ = 0;    // first 4x4 sub-block
+  it->i4_top_ = it->i4_boundary_ + VP8TopLeftI4[0];
+
+  // Import the boundary samples
+  for (i = 0; i < 17; ++i) {    // left
+    it->i4_boundary_[i] = it->y_left_[15 - i];
+  }
+  for (i = 0; i < 16; ++i) {    // top
+    it->i4_boundary_[17 + i] = it->y_top_[i];
+  }
+  // top-right samples have a special case on the far right of the picture
+  if (it->x_ < enc->mb_w_ - 1) {
+    for (i = 16; i < 16 + 4; ++i) {
+      it->i4_boundary_[17 + i] = it->y_top_[i];
+    }
+  } else {    // else, replicate the last valid pixel four times
+    for (i = 16; i < 16 + 4; ++i) {
+      it->i4_boundary_[17 + i] = it->i4_boundary_[17 + 15];
+    }
+  }
+  VP8IteratorNzToBytes(it);  // import the non-zero context
+}
+
+int VP8IteratorRotateI4(VP8EncIterator* const it,
+                        const uint8_t* const yuv_out) {
+  const uint8_t* const blk = yuv_out + VP8Scan[it->i4_];
+  uint8_t* const top = it->i4_top_;
+  int i;
+
+  // Update the cache with 7 fresh samples
+  for (i = 0; i <= 3; ++i) {
+    top[-4 + i] = blk[i + 3 * BPS];   // store future top samples
+  }
+  if ((it->i4_ & 3) != 3) {  // if not on the right sub-blocks #3, #7, #11, #15
+    for (i = 0; i <= 2; ++i) {        // store future left samples
+      top[i] = blk[3 + (2 - i) * BPS];
+    }
+  } else {  // else replicate top-right samples, as says the specs.
+    for (i = 0; i <= 3; ++i) {
+      top[i] = top[i + 4];
+    }
+  }
+  // move pointers to next sub-block
+  ++it->i4_;
+  if (it->i4_ == 16) {    // we're done
+    return 0;
+  }
+
+  it->i4_top_ = it->i4_boundary_ + VP8TopLeftI4[it->i4_];
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/enc/layer.c b/drivers/webp/enc/layer.c
new file mode 100644
index 0000000000..2402362359
--- /dev/null
+++ b/drivers/webp/enc/layer.c
@@ -0,0 +1,44 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Enhancement layer (for YUV444/422)
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+
+#include "./vp8enci.h"
+
+//------------------------------------------------------------------------------
+
+void VP8EncInitLayer(VP8Encoder* const enc) {
+  enc->use_layer_ = (enc->pic_->u0 != NULL);
+  enc->layer_data_size_ = 0;
+  enc->layer_data_ = NULL;
+  if (enc->use_layer_) {
+    VP8BitWriterInit(&enc->layer_bw_, enc->mb_w_ * enc->mb_h_ * 3);
+  }
+}
+
+void VP8EncCodeLayerBlock(VP8EncIterator* it) {
+  (void)it;   // remove a warning
+}
+
+int VP8EncFinishLayer(VP8Encoder* const enc) {
+  if (enc->use_layer_) {
+    enc->layer_data_ = VP8BitWriterFinish(&enc->layer_bw_);
+    enc->layer_data_size_ = VP8BitWriterSize(&enc->layer_bw_);
+  }
+  return 1;
+}
+
+void VP8EncDeleteLayer(VP8Encoder* enc) {
+  free(enc->layer_data_);
+}
+
diff --git a/drivers/webp/enc/picture.c b/drivers/webp/enc/picture.c
new file mode 100644
index 0000000000..011690d065
--- /dev/null
+++ b/drivers/webp/enc/picture.c
@@ -0,0 +1,1324 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// WebPPicture utils: colorspace conversion, crop, ...
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <math.h>
+
+#include "./vp8enci.h"
+#include "../utils/alpha_processing.h"
+#include "../utils/random.h"
+#include "../utils/rescaler.h"
+#include "../utils/utils.h"
+#include "../dsp/dsp.h"
+#include "../dsp/yuv.h"
+
+// Uncomment to disable gamma-compression during RGB->U/V averaging
+#define USE_GAMMA_COMPRESSION
+
+#define HALVE(x) (((x) + 1) >> 1)
+#define IS_YUV_CSP(csp, YUV_CSP) (((csp) & WEBP_CSP_UV_MASK) == (YUV_CSP))
+
+static const union {
+  uint32_t argb;
+  uint8_t  bytes[4];
+} test_endian = { 0xff000000u };
+#define ALPHA_IS_LAST (test_endian.bytes[3] == 0xff)
+
+static WEBP_INLINE uint32_t MakeARGB32(int r, int g, int b) {
+  return (0xff000000u | (r << 16) | (g << 8) | b);
+}
+
+//------------------------------------------------------------------------------
+// WebPPicture
+//------------------------------------------------------------------------------
+
+int WebPPictureAlloc(WebPPicture* picture) {
+  if (picture != NULL) {
+    const WebPEncCSP uv_csp = picture->colorspace & WEBP_CSP_UV_MASK;
+    const int has_alpha = picture->colorspace & WEBP_CSP_ALPHA_BIT;
+    const int width = picture->width;
+    const int height = picture->height;
+
+    if (!picture->use_argb) {
+      const int y_stride = width;
+      const int uv_width = HALVE(width);
+      const int uv_height = HALVE(height);
+      const int uv_stride = uv_width;
+      int uv0_stride = 0;
+      int a_width, a_stride;
+      uint64_t y_size, uv_size, uv0_size, a_size, total_size;
+      uint8_t* mem;
+
+      // U/V
+      switch (uv_csp) {
+        case WEBP_YUV420:
+          break;
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+        case WEBP_YUV400:    // for now, we'll just reset the U/V samples
+          break;
+        case WEBP_YUV422:
+          uv0_stride = uv_width;
+          break;
+        case WEBP_YUV444:
+          uv0_stride = width;
+          break;
+#endif
+        default:
+          return 0;
+      }
+      uv0_size = height * uv0_stride;
+
+      // alpha
+      a_width = has_alpha ? width : 0;
+      a_stride = a_width;
+      y_size = (uint64_t)y_stride * height;
+      uv_size = (uint64_t)uv_stride * uv_height;
+      a_size =  (uint64_t)a_stride * height;
+
+      total_size = y_size + a_size + 2 * uv_size + 2 * uv0_size;
+
+      // Security and validation checks
+      if (width <= 0 || height <= 0 ||         // luma/alpha param error
+          uv_width < 0 || uv_height < 0) {     // u/v param error
+        return 0;
+      }
+      // Clear previous buffer and allocate a new one.
+      WebPPictureFree(picture);   // erase previous buffer
+      mem = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*mem));
+      if (mem == NULL) return 0;
+
+      // From now on, we're in the clear, we can no longer fail...
+      picture->memory_ = (void*)mem;
+      picture->y_stride  = y_stride;
+      picture->uv_stride = uv_stride;
+      picture->a_stride  = a_stride;
+      picture->uv0_stride = uv0_stride;
+      // TODO(skal): we could align the y/u/v planes and adjust stride.
+      picture->y = mem;
+      mem += y_size;
+
+      picture->u = mem;
+      mem += uv_size;
+      picture->v = mem;
+      mem += uv_size;
+
+      if (a_size) {
+        picture->a = mem;
+        mem += a_size;
+      }
+      if (uv0_size) {
+        picture->u0 = mem;
+        mem += uv0_size;
+        picture->v0 = mem;
+        mem += uv0_size;
+      }
+      (void)mem;  // makes the static analyzer happy
+    } else {
+      void* memory;
+      const uint64_t argb_size = (uint64_t)width * height;
+      if (width <= 0 || height <= 0) {
+        return 0;
+      }
+      // Clear previous buffer and allocate a new one.
+      WebPPictureFree(picture);   // erase previous buffer
+      memory = WebPSafeMalloc(argb_size, sizeof(*picture->argb));
+      if (memory == NULL) return 0;
+
+      // TODO(skal): align plane to cache line?
+      picture->memory_argb_ = memory;
+      picture->argb = (uint32_t*)memory;
+      picture->argb_stride = width;
+    }
+  }
+  return 1;
+}
+
+// Remove reference to the ARGB buffer (doesn't free anything).
+static void PictureResetARGB(WebPPicture* const picture) {
+  picture->memory_argb_ = NULL;
+  picture->argb = NULL;
+  picture->argb_stride = 0;
+}
+
+// Remove reference to the YUVA buffer (doesn't free anything).
+static void PictureResetYUVA(WebPPicture* const picture) {
+  picture->memory_ = NULL;
+  picture->y = picture->u = picture->v = picture->a = NULL;
+  picture->u0 = picture->v0 = NULL;
+  picture->y_stride = picture->uv_stride = 0;
+  picture->a_stride = 0;
+  picture->uv0_stride = 0;
+}
+
+// Grab the 'specs' (writer, *opaque, width, height...) from 'src' and copy them
+// into 'dst'. Mark 'dst' as not owning any memory.
+static void WebPPictureGrabSpecs(const WebPPicture* const src,
+                                 WebPPicture* const dst) {
+  assert(src != NULL && dst != NULL);
+  *dst = *src;
+  PictureResetYUVA(dst);
+  PictureResetARGB(dst);
+}
+
+// Allocate a new argb buffer, discarding any existing one and preserving
+// the other YUV(A) buffer.
+static int PictureAllocARGB(WebPPicture* const picture) {
+  WebPPicture tmp;
+  free(picture->memory_argb_);
+  PictureResetARGB(picture);
+  picture->use_argb = 1;
+  WebPPictureGrabSpecs(picture, &tmp);
+  if (!WebPPictureAlloc(&tmp)) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+  }
+  picture->memory_argb_ = tmp.memory_argb_;
+  picture->argb = tmp.argb;
+  picture->argb_stride = tmp.argb_stride;
+  return 1;
+}
+
+// Release memory owned by 'picture' (both YUV and ARGB buffers).
+void WebPPictureFree(WebPPicture* picture) {
+  if (picture != NULL) {
+    free(picture->memory_);
+    free(picture->memory_argb_);
+    PictureResetYUVA(picture);
+    PictureResetARGB(picture);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Picture copying
+
+// Not worth moving to dsp/enc.c (only used here).
+static void CopyPlane(const uint8_t* src, int src_stride,
+                      uint8_t* dst, int dst_stride, int width, int height) {
+  while (height-- > 0) {
+    memcpy(dst, src, width);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+// Adjust top-left corner to chroma sample position.
+static void SnapTopLeftPosition(const WebPPicture* const pic,
+                                int* const left, int* const top) {
+  if (!pic->use_argb) {
+    const int is_yuv422 = IS_YUV_CSP(pic->colorspace, WEBP_YUV422);
+    if (IS_YUV_CSP(pic->colorspace, WEBP_YUV420) || is_yuv422) {
+      *left &= ~1;
+      if (!is_yuv422) *top &= ~1;
+    }
+  }
+}
+
+// Adjust top-left corner and verify that the sub-rectangle is valid.
+static int AdjustAndCheckRectangle(const WebPPicture* const pic,
+                                   int* const left, int* const top,
+                                   int width, int height) {
+  SnapTopLeftPosition(pic, left, top);
+  if ((*left) < 0 || (*top) < 0) return 0;
+  if (width <= 0 || height <= 0) return 0;
+  if ((*left) + width > pic->width) return 0;
+  if ((*top) + height > pic->height) return 0;
+  return 1;
+}
+
+int WebPPictureCopy(const WebPPicture* src, WebPPicture* dst) {
+  if (src == NULL || dst == NULL) return 0;
+  if (src == dst) return 1;
+
+  WebPPictureGrabSpecs(src, dst);
+  if (!WebPPictureAlloc(dst)) return 0;
+
+  if (!src->use_argb) {
+    CopyPlane(src->y, src->y_stride,
+              dst->y, dst->y_stride, dst->width, dst->height);
+    CopyPlane(src->u, src->uv_stride,
+              dst->u, dst->uv_stride, HALVE(dst->width), HALVE(dst->height));
+    CopyPlane(src->v, src->uv_stride,
+              dst->v, dst->uv_stride, HALVE(dst->width), HALVE(dst->height));
+    if (dst->a != NULL)  {
+      CopyPlane(src->a, src->a_stride,
+                dst->a, dst->a_stride, dst->width, dst->height);
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (dst->u0 != NULL)  {
+      int uv0_width = src->width;
+      if (IS_YUV_CSP(dst->colorspace, WEBP_YUV422)) {
+        uv0_width = HALVE(uv0_width);
+      }
+      CopyPlane(src->u0, src->uv0_stride,
+                dst->u0, dst->uv0_stride, uv0_width, dst->height);
+      CopyPlane(src->v0, src->uv0_stride,
+                dst->v0, dst->uv0_stride, uv0_width, dst->height);
+    }
+#endif
+  } else {
+    CopyPlane((const uint8_t*)src->argb, 4 * src->argb_stride,
+              (uint8_t*)dst->argb, 4 * dst->argb_stride,
+              4 * dst->width, dst->height);
+  }
+  return 1;
+}
+
+int WebPPictureIsView(const WebPPicture* picture) {
+  if (picture == NULL) return 0;
+  if (picture->use_argb) {
+    return (picture->memory_argb_ == NULL);
+  }
+  return (picture->memory_ == NULL);
+}
+
+int WebPPictureView(const WebPPicture* src,
+                    int left, int top, int width, int height,
+                    WebPPicture* dst) {
+  if (src == NULL || dst == NULL) return 0;
+
+  // verify rectangle position.
+  if (!AdjustAndCheckRectangle(src, &left, &top, width, height)) return 0;
+
+  if (src != dst) {  // beware of aliasing! We don't want to leak 'memory_'.
+    WebPPictureGrabSpecs(src, dst);
+  }
+  dst->width = width;
+  dst->height = height;
+  if (!src->use_argb) {
+    dst->y = src->y + top * src->y_stride + left;
+    dst->u = src->u + (top >> 1) * src->uv_stride + (left >> 1);
+    dst->v = src->v + (top >> 1) * src->uv_stride + (left >> 1);
+    dst->y_stride = src->y_stride;
+    dst->uv_stride = src->uv_stride;
+    if (src->a != NULL) {
+      dst->a = src->a + top * src->a_stride + left;
+      dst->a_stride = src->a_stride;
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (src->u0 != NULL) {
+      const int left_pos =
+          IS_YUV_CSP(dst->colorspace, WEBP_YUV422) ? (left >> 1) : left;
+      dst->u0 = src->u0 + top * src->uv0_stride + left_pos;
+      dst->v0 = src->v0 + top * src->uv0_stride + left_pos;
+      dst->uv0_stride = src->uv0_stride;
+    }
+#endif
+  } else {
+    dst->argb = src->argb + top * src->argb_stride + left;
+    dst->argb_stride = src->argb_stride;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Picture cropping
+
+int WebPPictureCrop(WebPPicture* pic,
+                    int left, int top, int width, int height) {
+  WebPPicture tmp;
+
+  if (pic == NULL) return 0;
+  if (!AdjustAndCheckRectangle(pic, &left, &top, width, height)) return 0;
+
+  WebPPictureGrabSpecs(pic, &tmp);
+  tmp.width = width;
+  tmp.height = height;
+  if (!WebPPictureAlloc(&tmp)) return 0;
+
+  if (!pic->use_argb) {
+    const int y_offset = top * pic->y_stride + left;
+    const int uv_offset = (top / 2) * pic->uv_stride + left / 2;
+    CopyPlane(pic->y + y_offset, pic->y_stride,
+              tmp.y, tmp.y_stride, width, height);
+    CopyPlane(pic->u + uv_offset, pic->uv_stride,
+              tmp.u, tmp.uv_stride, HALVE(width), HALVE(height));
+    CopyPlane(pic->v + uv_offset, pic->uv_stride,
+              tmp.v, tmp.uv_stride, HALVE(width), HALVE(height));
+
+    if (tmp.a != NULL) {
+      const int a_offset = top * pic->a_stride + left;
+      CopyPlane(pic->a + a_offset, pic->a_stride,
+                tmp.a, tmp.a_stride, width, height);
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (tmp.u0 != NULL) {
+      int w = width;
+      int left_pos = left;
+      if (IS_YUV_CSP(tmp.colorspace, WEBP_YUV422)) {
+        w = HALVE(w);
+        left_pos = HALVE(left_pos);
+      }
+      CopyPlane(pic->u0 + top * pic->uv0_stride + left_pos, pic->uv0_stride,
+                tmp.u0, tmp.uv0_stride, w, height);
+      CopyPlane(pic->v0 + top * pic->uv0_stride + left_pos, pic->uv0_stride,
+                tmp.v0, tmp.uv0_stride, w, height);
+    }
+#endif
+  } else {
+    const uint8_t* const src =
+        (const uint8_t*)(pic->argb + top * pic->argb_stride + left);
+    CopyPlane(src, pic->argb_stride * 4,
+              (uint8_t*)tmp.argb, tmp.argb_stride * 4,
+              width * 4, height);
+  }
+  WebPPictureFree(pic);
+  *pic = tmp;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Simple picture rescaler
+
+static void RescalePlane(const uint8_t* src,
+                         int src_width, int src_height, int src_stride,
+                         uint8_t* dst,
+                         int dst_width, int dst_height, int dst_stride,
+                         int32_t* const work,
+                         int num_channels) {
+  WebPRescaler rescaler;
+  int y = 0;
+  WebPRescalerInit(&rescaler, src_width, src_height,
+                   dst, dst_width, dst_height, dst_stride,
+                   num_channels,
+                   src_width, dst_width,
+                   src_height, dst_height,
+                   work);
+  memset(work, 0, 2 * dst_width * num_channels * sizeof(*work));
+  while (y < src_height) {
+    y += WebPRescalerImport(&rescaler, src_height - y,
+                            src + y * src_stride, src_stride);
+    WebPRescalerExport(&rescaler);
+  }
+}
+
+static void AlphaMultiplyARGB(WebPPicture* const pic, int inverse) {
+  uint32_t* ptr = pic->argb;
+  int y;
+  for (y = 0; y < pic->height; ++y) {
+    WebPMultARGBRow(ptr, pic->width, inverse);
+    ptr += pic->argb_stride;
+  }
+}
+
+static void AlphaMultiplyY(WebPPicture* const pic, int inverse) {
+  const uint8_t* ptr_a = pic->a;
+  if (ptr_a != NULL) {
+    uint8_t* ptr_y = pic->y;
+    int y;
+    for (y = 0; y < pic->height; ++y) {
+      WebPMultRow(ptr_y, ptr_a, pic->width, inverse);
+      ptr_y += pic->y_stride;
+      ptr_a += pic->a_stride;
+    }
+  }
+}
+
+int WebPPictureRescale(WebPPicture* pic, int width, int height) {
+  WebPPicture tmp;
+  int prev_width, prev_height;
+  int32_t* work;
+
+  if (pic == NULL) return 0;
+  prev_width = pic->width;
+  prev_height = pic->height;
+  // if width is unspecified, scale original proportionally to height ratio.
+  if (width == 0) {
+    width = (prev_width * height + prev_height / 2) / prev_height;
+  }
+  // if height is unspecified, scale original proportionally to width ratio.
+  if (height == 0) {
+    height = (prev_height * width + prev_width / 2) / prev_width;
+  }
+  // Check if the overall dimensions still make sense.
+  if (width <= 0 || height <= 0) return 0;
+
+  WebPPictureGrabSpecs(pic, &tmp);
+  tmp.width = width;
+  tmp.height = height;
+  if (!WebPPictureAlloc(&tmp)) return 0;
+
+  if (!pic->use_argb) {
+    work = (int32_t*)WebPSafeMalloc(2ULL * width, sizeof(*work));
+    if (work == NULL) {
+      WebPPictureFree(&tmp);
+      return 0;
+    }
+    // If present, we need to rescale alpha first (for AlphaMultiplyY).
+    if (pic->a != NULL) {
+      RescalePlane(pic->a, prev_width, prev_height, pic->a_stride,
+                   tmp.a, width, height, tmp.a_stride, work, 1);
+    }
+
+    // We take transparency into account on the luma plane only. That's not
+    // totally exact blending, but still is a good approximation.
+    AlphaMultiplyY(pic, 0);
+    RescalePlane(pic->y, prev_width, prev_height, pic->y_stride,
+                 tmp.y, width, height, tmp.y_stride, work, 1);
+    AlphaMultiplyY(&tmp, 1);
+
+    RescalePlane(pic->u,
+                 HALVE(prev_width), HALVE(prev_height), pic->uv_stride,
+                 tmp.u,
+                 HALVE(width), HALVE(height), tmp.uv_stride, work, 1);
+    RescalePlane(pic->v,
+                 HALVE(prev_width), HALVE(prev_height), pic->uv_stride,
+                 tmp.v,
+                 HALVE(width), HALVE(height), tmp.uv_stride, work, 1);
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (tmp.u0 != NULL) {
+      const int s = IS_YUV_CSP(tmp.colorspace, WEBP_YUV422) ? 2 : 1;
+      RescalePlane(
+          pic->u0, (prev_width + s / 2) / s, prev_height, pic->uv0_stride,
+          tmp.u0, (width + s / 2) / s, height, tmp.uv0_stride, work, 1);
+      RescalePlane(
+          pic->v0, (prev_width + s / 2) / s, prev_height, pic->uv0_stride,
+          tmp.v0, (width + s / 2) / s, height, tmp.uv0_stride, work, 1);
+    }
+#endif
+  } else {
+    work = (int32_t*)WebPSafeMalloc(2ULL * width * 4, sizeof(*work));
+    if (work == NULL) {
+      WebPPictureFree(&tmp);
+      return 0;
+    }
+    // In order to correctly interpolate colors, we need to apply the alpha
+    // weighting first (black-matting), scale the RGB values, and remove
+    // the premultiplication afterward (while preserving the alpha channel).
+    AlphaMultiplyARGB(pic, 0);
+    RescalePlane((const uint8_t*)pic->argb, prev_width, prev_height,
+                 pic->argb_stride * 4,
+                 (uint8_t*)tmp.argb, width, height,
+                 tmp.argb_stride * 4,
+                 work, 4);
+    AlphaMultiplyARGB(&tmp, 1);
+  }
+  WebPPictureFree(pic);
+  free(work);
+  *pic = tmp;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// WebPMemoryWriter: Write-to-memory
+
+void WebPMemoryWriterInit(WebPMemoryWriter* writer) {
+  writer->mem = NULL;
+  writer->size = 0;
+  writer->max_size = 0;
+}
+
+int WebPMemoryWrite(const uint8_t* data, size_t data_size,
+                    const WebPPicture* picture) {
+  WebPMemoryWriter* const w = (WebPMemoryWriter*)picture->custom_ptr;
+  uint64_t next_size;
+  if (w == NULL) {
+    return 1;
+  }
+  next_size = (uint64_t)w->size + data_size;
+  if (next_size > w->max_size) {
+    uint8_t* new_mem;
+    uint64_t next_max_size = 2ULL * w->max_size;
+    if (next_max_size < next_size) next_max_size = next_size;
+    if (next_max_size < 8192ULL) next_max_size = 8192ULL;
+    new_mem = (uint8_t*)WebPSafeMalloc(next_max_size, 1);
+    if (new_mem == NULL) {
+      return 0;
+    }
+    if (w->size > 0) {
+      memcpy(new_mem, w->mem, w->size);
+    }
+    free(w->mem);
+    w->mem = new_mem;
+    // down-cast is ok, thanks to WebPSafeMalloc
+    w->max_size = (size_t)next_max_size;
+  }
+  if (data_size > 0) {
+    memcpy(w->mem + w->size, data, data_size);
+    w->size += data_size;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Detection of non-trivial transparency
+
+// Returns true if alpha[] has non-0xff values.
+static int CheckNonOpaque(const uint8_t* alpha, int width, int height,
+                          int x_step, int y_step) {
+  if (alpha == NULL) return 0;
+  while (height-- > 0) {
+    int x;
+    for (x = 0; x < width * x_step; x += x_step) {
+      if (alpha[x] != 0xff) return 1;  // TODO(skal): check 4/8 bytes at a time.
+    }
+    alpha += y_step;
+  }
+  return 0;
+}
+
+// Checking for the presence of non-opaque alpha.
+int WebPPictureHasTransparency(const WebPPicture* picture) {
+  if (picture == NULL) return 0;
+  if (!picture->use_argb) {
+    return CheckNonOpaque(picture->a, picture->width, picture->height,
+                          1, picture->a_stride);
+  } else {
+    int x, y;
+    const uint32_t* argb = picture->argb;
+    if (argb == NULL) return 0;
+    for (y = 0; y < picture->height; ++y) {
+      for (x = 0; x < picture->width; ++x) {
+        if (argb[x] < 0xff000000u) return 1;   // test any alpha values != 0xff
+      }
+      argb += picture->argb_stride;
+    }
+  }
+  return 0;
+}
+
+//------------------------------------------------------------------------------
+// RGB -> YUV conversion
+
+static int RGBToY(int r, int g, int b, VP8Random* const rg) {
+  return VP8RGBToY(r, g, b, VP8RandomBits(rg, YUV_FIX));
+}
+
+static int RGBToU(int r, int g, int b, VP8Random* const rg) {
+  return VP8RGBToU(r, g, b, VP8RandomBits(rg, YUV_FIX + 2));
+}
+
+static int RGBToV(int r, int g, int b, VP8Random* const rg) {
+  return VP8RGBToV(r, g, b, VP8RandomBits(rg, YUV_FIX + 2));
+}
+
+//------------------------------------------------------------------------------
+
+#if defined(USE_GAMMA_COMPRESSION)
+
+// gamma-compensates loss of resolution during chroma subsampling
+#define kGamma 0.80
+#define kGammaFix 12     // fixed-point precision for linear values
+#define kGammaScale ((1 << kGammaFix) - 1)
+#define kGammaTabFix 7   // fixed-point fractional bits precision
+#define kGammaTabScale (1 << kGammaTabFix)
+#define kGammaTabRounder (kGammaTabScale >> 1)
+#define kGammaTabSize (1 << (kGammaFix - kGammaTabFix))
+
+static int kLinearToGammaTab[kGammaTabSize + 1];
+static uint16_t kGammaToLinearTab[256];
+static int kGammaTablesOk = 0;
+
+static void InitGammaTables(void) {
+  if (!kGammaTablesOk) {
+    int v;
+    const double scale = 1. / kGammaScale;
+    for (v = 0; v <= 255; ++v) {
+      kGammaToLinearTab[v] =
+          (uint16_t)(pow(v / 255., kGamma) * kGammaScale + .5);
+    }
+    for (v = 0; v <= kGammaTabSize; ++v) {
+      const double x = scale * (v << kGammaTabFix);
+      kLinearToGammaTab[v] = (int)(pow(x, 1. / kGamma) * 255. + .5);
+    }
+    kGammaTablesOk = 1;
+  }
+}
+
+static WEBP_INLINE uint32_t GammaToLinear(uint8_t v) {
+  return kGammaToLinearTab[v];
+}
+
+// Convert a linear value 'v' to YUV_FIX+2 fixed-point precision
+// U/V value, suitable for RGBToU/V calls.
+static WEBP_INLINE int LinearToGamma(uint32_t base_value, int shift) {
+  const int v = base_value << shift;              // final uplifted value
+  const int tab_pos = v >> (kGammaTabFix + 2);    // integer part
+  const int x = v & ((kGammaTabScale << 2) - 1);  // fractional part
+  const int v0 = kLinearToGammaTab[tab_pos];
+  const int v1 = kLinearToGammaTab[tab_pos + 1];
+  const int y = v1 * x + v0 * ((kGammaTabScale << 2) - x);   // interpolate
+  return (y + kGammaTabRounder) >> kGammaTabFix;             // descale
+}
+
+#else
+
+static void InitGammaTables(void) {}
+static WEBP_INLINE uint32_t GammaToLinear(uint8_t v) { return v; }
+static WEBP_INLINE int LinearToGamma(uint32_t base_value, int shift) {
+  (void)shift;
+  return v;
+}
+
+#endif    // USE_GAMMA_COMPRESSION
+
+//------------------------------------------------------------------------------
+
+#define SUM4(ptr) LinearToGamma(                         \
+    GammaToLinear((ptr)[0]) +                            \
+    GammaToLinear((ptr)[step]) +                         \
+    GammaToLinear((ptr)[rgb_stride]) +                   \
+    GammaToLinear((ptr)[rgb_stride + step]), 0)          \
+
+#define SUM2H(ptr) \
+    LinearToGamma(GammaToLinear((ptr)[0]) + GammaToLinear((ptr)[step]), 1)
+#define SUM2V(ptr) \
+    LinearToGamma(GammaToLinear((ptr)[0]) + GammaToLinear((ptr)[rgb_stride]), 1)
+#define SUM1(ptr)  \
+    LinearToGamma(GammaToLinear((ptr)[0]), 2)
+
+#define RGB_TO_UV(x, y, SUM) {                           \
+  const int src = (2 * (step * (x) + (y) * rgb_stride)); \
+  const int dst = (x) + (y) * picture->uv_stride;        \
+  const int r = SUM(r_ptr + src);                        \
+  const int g = SUM(g_ptr + src);                        \
+  const int b = SUM(b_ptr + src);                        \
+  picture->u[dst] = RGBToU(r, g, b, &rg);                \
+  picture->v[dst] = RGBToV(r, g, b, &rg);                \
+}
+
+#define RGB_TO_UV0(x_in, x_out, y, SUM) {                \
+  const int src = (step * (x_in) + (y) * rgb_stride);    \
+  const int dst = (x_out) + (y) * picture->uv0_stride;   \
+  const int r = SUM(r_ptr + src);                        \
+  const int g = SUM(g_ptr + src);                        \
+  const int b = SUM(b_ptr + src);                        \
+  picture->u0[dst] = RGBToU(r, g, b, &rg);               \
+  picture->v0[dst] = RGBToV(r, g, b, &rg);               \
+}
+
+static void MakeGray(WebPPicture* const picture) {
+  int y;
+  const int uv_width = HALVE(picture->width);
+  const int uv_height = HALVE(picture->height);
+  for (y = 0; y < uv_height; ++y) {
+    memset(picture->u + y * picture->uv_stride, 128, uv_width);
+    memset(picture->v + y * picture->uv_stride, 128, uv_width);
+  }
+}
+
+static int ImportYUVAFromRGBA(const uint8_t* const r_ptr,
+                              const uint8_t* const g_ptr,
+                              const uint8_t* const b_ptr,
+                              const uint8_t* const a_ptr,
+                              int step,         // bytes per pixel
+                              int rgb_stride,   // bytes per scanline
+                              float dithering,
+                              WebPPicture* const picture) {
+  const WebPEncCSP uv_csp = picture->colorspace & WEBP_CSP_UV_MASK;
+  int x, y;
+  const int width = picture->width;
+  const int height = picture->height;
+  const int has_alpha = CheckNonOpaque(a_ptr, width, height, step, rgb_stride);
+  VP8Random rg;
+
+  picture->colorspace = uv_csp;
+  picture->use_argb = 0;
+  if (has_alpha) {
+    picture->colorspace |= WEBP_CSP_ALPHA_BIT;
+  }
+  if (!WebPPictureAlloc(picture)) return 0;
+
+  VP8InitRandom(&rg, dithering);
+  InitGammaTables();
+
+  // Import luma plane
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      const int offset = step * x + y * rgb_stride;
+      picture->y[x + y * picture->y_stride] =
+          RGBToY(r_ptr[offset], g_ptr[offset], b_ptr[offset], &rg);
+    }
+  }
+
+  // Downsample U/V plane
+  if (uv_csp != WEBP_YUV400) {
+    for (y = 0; y < (height >> 1); ++y) {
+      for (x = 0; x < (width >> 1); ++x) {
+        RGB_TO_UV(x, y, SUM4);
+      }
+      if (width & 1) {
+        RGB_TO_UV(x, y, SUM2V);
+      }
+    }
+    if (height & 1) {
+      for (x = 0; x < (width >> 1); ++x) {
+        RGB_TO_UV(x, y, SUM2H);
+      }
+      if (width & 1) {
+        RGB_TO_UV(x, y, SUM1);
+      }
+    }
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    // Store original U/V samples too
+    if (uv_csp == WEBP_YUV422) {
+      for (y = 0; y < height; ++y) {
+        for (x = 0; x < (width >> 1); ++x) {
+          RGB_TO_UV0(2 * x, x, y, SUM2H);
+        }
+        if (width & 1) {
+          RGB_TO_UV0(2 * x, x, y, SUM1);
+        }
+      }
+    } else if (uv_csp == WEBP_YUV444) {
+      for (y = 0; y < height; ++y) {
+        for (x = 0; x < width; ++x) {
+          RGB_TO_UV0(x, x, y, SUM1);
+        }
+      }
+    }
+#endif
+  } else {
+    MakeGray(picture);
+  }
+
+  if (has_alpha) {
+    assert(step >= 4);
+    assert(picture->a != NULL);
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        picture->a[x + y * picture->a_stride] =
+            a_ptr[step * x + y * rgb_stride];
+      }
+    }
+  }
+  return 1;
+}
+
+static int Import(WebPPicture* const picture,
+                  const uint8_t* const rgb, int rgb_stride,
+                  int step, int swap_rb, int import_alpha) {
+  const uint8_t* const r_ptr = rgb + (swap_rb ? 2 : 0);
+  const uint8_t* const g_ptr = rgb + 1;
+  const uint8_t* const b_ptr = rgb + (swap_rb ? 0 : 2);
+  const uint8_t* const a_ptr = import_alpha ? rgb + 3 : NULL;
+  const int width = picture->width;
+  const int height = picture->height;
+
+  if (!picture->use_argb) {
+    return ImportYUVAFromRGBA(r_ptr, g_ptr, b_ptr, a_ptr, step, rgb_stride,
+                              0.f /* no dithering */, picture);
+  }
+  if (import_alpha) {
+    picture->colorspace |= WEBP_CSP_ALPHA_BIT;
+  } else {
+    picture->colorspace &= ~WEBP_CSP_ALPHA_BIT;
+  }
+  if (!WebPPictureAlloc(picture)) return 0;
+
+  if (!import_alpha) {
+    int x, y;
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const int offset = step * x + y * rgb_stride;
+        const uint32_t argb =
+            MakeARGB32(r_ptr[offset], g_ptr[offset], b_ptr[offset]);
+        picture->argb[x + y * picture->argb_stride] = argb;
+      }
+    }
+  } else {
+    int x, y;
+    assert(step >= 4);
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const int offset = step * x + y * rgb_stride;
+        const uint32_t argb = ((uint32_t)a_ptr[offset] << 24) |
+                              (r_ptr[offset] << 16) |
+                              (g_ptr[offset] <<  8) |
+                              (b_ptr[offset]);
+        picture->argb[x + y * picture->argb_stride] = argb;
+      }
+    }
+  }
+  return 1;
+}
+#undef SUM4
+#undef SUM2V
+#undef SUM2H
+#undef SUM1
+#undef RGB_TO_UV
+
+int WebPPictureImportRGB(WebPPicture* picture,
+                         const uint8_t* rgb, int rgb_stride) {
+  return Import(picture, rgb, rgb_stride, 3, 0, 0);
+}
+
+int WebPPictureImportBGR(WebPPicture* picture,
+                         const uint8_t* rgb, int rgb_stride) {
+  return Import(picture, rgb, rgb_stride, 3, 1, 0);
+}
+
+int WebPPictureImportRGBA(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return Import(picture, rgba, rgba_stride, 4, 0, 1);
+}
+
+int WebPPictureImportBGRA(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return Import(picture, rgba, rgba_stride, 4, 1, 1);
+}
+
+int WebPPictureImportRGBX(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return Import(picture, rgba, rgba_stride, 4, 0, 0);
+}
+
+int WebPPictureImportBGRX(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return Import(picture, rgba, rgba_stride, 4, 1, 0);
+}
+
+//------------------------------------------------------------------------------
+// Automatic YUV <-> ARGB conversions.
+
+int WebPPictureYUVAToARGB(WebPPicture* picture) {
+  if (picture == NULL) return 0;
+  if (picture->y == NULL || picture->u == NULL || picture->v == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_NULL_PARAMETER);
+  }
+  if ((picture->colorspace & WEBP_CSP_ALPHA_BIT) && picture->a == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_NULL_PARAMETER);
+  }
+  if ((picture->colorspace & WEBP_CSP_UV_MASK) != WEBP_YUV420) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_INVALID_CONFIGURATION);
+  }
+  // Allocate a new argb buffer (discarding the previous one).
+  if (!PictureAllocARGB(picture)) return 0;
+
+  // Convert
+  {
+    int y;
+    const int width = picture->width;
+    const int height = picture->height;
+    const int argb_stride = 4 * picture->argb_stride;
+    uint8_t* dst = (uint8_t*)picture->argb;
+    const uint8_t *cur_u = picture->u, *cur_v = picture->v, *cur_y = picture->y;
+    WebPUpsampleLinePairFunc upsample = WebPGetLinePairConverter(ALPHA_IS_LAST);
+
+    // First row, with replicated top samples.
+    upsample(cur_y, NULL, cur_u, cur_v, cur_u, cur_v, dst, NULL, width);
+    cur_y += picture->y_stride;
+    dst += argb_stride;
+    // Center rows.
+    for (y = 1; y + 1 < height; y += 2) {
+      const uint8_t* const top_u = cur_u;
+      const uint8_t* const top_v = cur_v;
+      cur_u += picture->uv_stride;
+      cur_v += picture->uv_stride;
+      upsample(cur_y, cur_y + picture->y_stride, top_u, top_v, cur_u, cur_v,
+               dst, dst + argb_stride, width);
+      cur_y += 2 * picture->y_stride;
+      dst += 2 * argb_stride;
+    }
+    // Last row (if needed), with replicated bottom samples.
+    if (height > 1 && !(height & 1)) {
+      upsample(cur_y, NULL, cur_u, cur_v, cur_u, cur_v, dst, NULL, width);
+    }
+    // Insert alpha values if needed, in replacement for the default 0xff ones.
+    if (picture->colorspace & WEBP_CSP_ALPHA_BIT) {
+      for (y = 0; y < height; ++y) {
+        uint32_t* const argb_dst = picture->argb + y * picture->argb_stride;
+        const uint8_t* const src = picture->a + y * picture->a_stride;
+        int x;
+        for (x = 0; x < width; ++x) {
+          argb_dst[x] = (argb_dst[x] & 0x00ffffffu) | ((uint32_t)src[x] << 24);
+        }
+      }
+    }
+  }
+  return 1;
+}
+
+int WebPPictureARGBToYUVADithered(WebPPicture* picture, WebPEncCSP colorspace,
+                                  float dithering) {
+  if (picture == NULL) return 0;
+  if (picture->argb == NULL) {
+    return WebPEncodingSetError(picture, VP8_ENC_ERROR_NULL_PARAMETER);
+  } else {
+    const uint8_t* const argb = (const uint8_t*)picture->argb;
+    const uint8_t* const r = ALPHA_IS_LAST ? argb + 2 : argb + 1;
+    const uint8_t* const g = ALPHA_IS_LAST ? argb + 1 : argb + 2;
+    const uint8_t* const b = ALPHA_IS_LAST ? argb + 0 : argb + 3;
+    const uint8_t* const a = ALPHA_IS_LAST ? argb + 3 : argb + 0;
+    // We work on a tmp copy of 'picture', because ImportYUVAFromRGBA()
+    // would be calling WebPPictureFree(picture) otherwise.
+    WebPPicture tmp = *picture;
+    PictureResetARGB(&tmp);  // reset ARGB buffer so that it's not free()'d.
+    tmp.use_argb = 0;
+    tmp.colorspace = colorspace & WEBP_CSP_UV_MASK;
+    if (!ImportYUVAFromRGBA(r, g, b, a, 4, 4 * picture->argb_stride, dithering,
+                            &tmp)) {
+      return WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    }
+    // Copy back the YUV specs into 'picture'.
+    tmp.argb = picture->argb;
+    tmp.argb_stride = picture->argb_stride;
+    tmp.memory_argb_ = picture->memory_argb_;
+    *picture = tmp;
+  }
+  return 1;
+}
+
+int WebPPictureARGBToYUVA(WebPPicture* picture, WebPEncCSP colorspace) {
+  return WebPPictureARGBToYUVADithered(picture, colorspace, 0.f);
+}
+
+//------------------------------------------------------------------------------
+// Helper: clean up fully transparent area to help compressibility.
+
+#define SIZE 8
+#define SIZE2 (SIZE / 2)
+static int is_transparent_area(const uint8_t* ptr, int stride, int size) {
+  int y, x;
+  for (y = 0; y < size; ++y) {
+    for (x = 0; x < size; ++x) {
+      if (ptr[x]) {
+        return 0;
+      }
+    }
+    ptr += stride;
+  }
+  return 1;
+}
+
+static WEBP_INLINE void flatten(uint8_t* ptr, int v, int stride, int size) {
+  int y;
+  for (y = 0; y < size; ++y) {
+    memset(ptr, v, size);
+    ptr += stride;
+  }
+}
+
+void WebPCleanupTransparentArea(WebPPicture* pic) {
+  int x, y, w, h;
+  const uint8_t* a_ptr;
+  int values[3] = { 0 };
+
+  if (pic == NULL) return;
+
+  a_ptr = pic->a;
+  if (a_ptr == NULL) return;    // nothing to do
+
+  w = pic->width / SIZE;
+  h = pic->height / SIZE;
+  for (y = 0; y < h; ++y) {
+    int need_reset = 1;
+    for (x = 0; x < w; ++x) {
+      const int off_a = (y * pic->a_stride + x) * SIZE;
+      const int off_y = (y * pic->y_stride + x) * SIZE;
+      const int off_uv = (y * pic->uv_stride + x) * SIZE2;
+      if (is_transparent_area(a_ptr + off_a, pic->a_stride, SIZE)) {
+        if (need_reset) {
+          values[0] = pic->y[off_y];
+          values[1] = pic->u[off_uv];
+          values[2] = pic->v[off_uv];
+          need_reset = 0;
+        }
+        flatten(pic->y + off_y, values[0], pic->y_stride, SIZE);
+        flatten(pic->u + off_uv, values[1], pic->uv_stride, SIZE2);
+        flatten(pic->v + off_uv, values[2], pic->uv_stride, SIZE2);
+      } else {
+        need_reset = 1;
+      }
+    }
+    // ignore the left-overs on right/bottom
+  }
+}
+
+#undef SIZE
+#undef SIZE2
+
+//------------------------------------------------------------------------------
+// Blend color and remove transparency info
+
+#define BLEND(V0, V1, ALPHA) \
+    ((((V0) * (255 - (ALPHA)) + (V1) * (ALPHA)) * 0x101) >> 16)
+#define BLEND_10BIT(V0, V1, ALPHA) \
+    ((((V0) * (1020 - (ALPHA)) + (V1) * (ALPHA)) * 0x101) >> 18)
+
+void WebPBlendAlpha(WebPPicture* pic, uint32_t background_rgb) {
+  const int red = (background_rgb >> 16) & 0xff;
+  const int green = (background_rgb >> 8) & 0xff;
+  const int blue = (background_rgb >> 0) & 0xff;
+  VP8Random rg;
+  int x, y;
+  if (pic == NULL) return;
+  VP8InitRandom(&rg, 0.f);
+  if (!pic->use_argb) {
+    const int uv_width = (pic->width >> 1);  // omit last pixel during u/v loop
+    const int Y0 = RGBToY(red, green, blue, &rg);
+    // VP8RGBToU/V expects the u/v values summed over four pixels
+    const int U0 = RGBToU(4 * red, 4 * green, 4 * blue, &rg);
+    const int V0 = RGBToV(4 * red, 4 * green, 4 * blue, &rg);
+    const int has_alpha = pic->colorspace & WEBP_CSP_ALPHA_BIT;
+    if (!has_alpha || pic->a == NULL) return;    // nothing to do
+    for (y = 0; y < pic->height; ++y) {
+      // Luma blending
+      uint8_t* const y_ptr = pic->y + y * pic->y_stride;
+      uint8_t* const a_ptr = pic->a + y * pic->a_stride;
+      for (x = 0; x < pic->width; ++x) {
+        const int alpha = a_ptr[x];
+        if (alpha < 0xff) {
+          y_ptr[x] = BLEND(Y0, y_ptr[x], a_ptr[x]);
+        }
+      }
+      // Chroma blending every even line
+      if ((y & 1) == 0) {
+        uint8_t* const u = pic->u + (y >> 1) * pic->uv_stride;
+        uint8_t* const v = pic->v + (y >> 1) * pic->uv_stride;
+        uint8_t* const a_ptr2 =
+            (y + 1 == pic->height) ? a_ptr : a_ptr + pic->a_stride;
+        for (x = 0; x < uv_width; ++x) {
+          // Average four alpha values into a single blending weight.
+          // TODO(skal): might lead to visible contouring. Can we do better?
+          const int alpha =
+              a_ptr[2 * x + 0] + a_ptr[2 * x + 1] +
+              a_ptr2[2 * x + 0] + a_ptr2[2 * x + 1];
+          u[x] = BLEND_10BIT(U0, u[x], alpha);
+          v[x] = BLEND_10BIT(V0, v[x], alpha);
+        }
+        if (pic->width & 1) {   // rightmost pixel
+          const int alpha = 2 * (a_ptr[2 * x + 0] + a_ptr2[2 * x + 0]);
+          u[x] = BLEND_10BIT(U0, u[x], alpha);
+          v[x] = BLEND_10BIT(V0, v[x], alpha);
+        }
+      }
+      memset(a_ptr, 0xff, pic->width);
+    }
+  } else {
+    uint32_t* argb = pic->argb;
+    const uint32_t background = MakeARGB32(red, green, blue);
+    for (y = 0; y < pic->height; ++y) {
+      for (x = 0; x < pic->width; ++x) {
+        const int alpha = (argb[x] >> 24) & 0xff;
+        if (alpha != 0xff) {
+          if (alpha > 0) {
+            int r = (argb[x] >> 16) & 0xff;
+            int g = (argb[x] >>  8) & 0xff;
+            int b = (argb[x] >>  0) & 0xff;
+            r = BLEND(red, r, alpha);
+            g = BLEND(green, g, alpha);
+            b = BLEND(blue, b, alpha);
+            argb[x] = MakeARGB32(r, g, b);
+          } else {
+            argb[x] = background;
+          }
+        }
+      }
+      argb += pic->argb_stride;
+    }
+  }
+}
+
+#undef BLEND
+#undef BLEND_10BIT
+
+//------------------------------------------------------------------------------
+// local-min distortion
+//
+// For every pixel in the *reference* picture, we search for the local best
+// match in the compressed image. This is not a symmetrical measure.
+
+// search radius. Shouldn't be too large.
+#define RADIUS 2
+
+static float AccumulateLSIM(const uint8_t* src, int src_stride,
+                            const uint8_t* ref, int ref_stride,
+                            int w, int h) {
+  int x, y;
+  double total_sse = 0.;
+  for (y = 0; y < h; ++y) {
+    const int y_0 = (y - RADIUS < 0) ? 0 : y - RADIUS;
+    const int y_1 = (y + RADIUS + 1 >= h) ? h : y + RADIUS + 1;
+    for (x = 0; x < w; ++x) {
+      const int x_0 = (x - RADIUS < 0) ? 0 : x - RADIUS;
+      const int x_1 = (x + RADIUS + 1 >= w) ? w : x + RADIUS + 1;
+      double best_sse = 255. * 255.;
+      const double value = (double)ref[y * ref_stride + x];
+      int i, j;
+      for (j = y_0; j < y_1; ++j) {
+        const uint8_t* s = src + j * src_stride;
+        for (i = x_0; i < x_1; ++i) {
+          const double sse = (double)(s[i] - value) * (s[i] - value);
+          if (sse < best_sse) best_sse = sse;
+        }
+      }
+      total_sse += best_sse;
+    }
+  }
+  return (float)total_sse;
+}
+#undef RADIUS
+
+//------------------------------------------------------------------------------
+// Distortion
+
+// Max value returned in case of exact similarity.
+static const double kMinDistortion_dB = 99.;
+static float GetPSNR(const double v) {
+  return (float)((v > 0.) ? -4.3429448 * log(v / (255 * 255.))
+                          : kMinDistortion_dB);
+}
+
+int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
+                          int type, float result[5]) {
+  DistoStats stats[5];
+  int has_alpha;
+  int uv_w, uv_h;
+
+  if (src == NULL || ref == NULL ||
+      src->width != ref->width || src->height != ref->height ||
+      src->y == NULL || ref->y == NULL ||
+      src->u == NULL || ref->u == NULL ||
+      src->v == NULL || ref->v == NULL ||
+      result == NULL) {
+    return 0;
+  }
+  // TODO(skal): provide distortion for ARGB too.
+  if (src->use_argb == 1 || src->use_argb != ref->use_argb) {
+    return 0;
+  }
+
+  has_alpha = !!(src->colorspace & WEBP_CSP_ALPHA_BIT);
+  if (has_alpha != !!(ref->colorspace & WEBP_CSP_ALPHA_BIT) ||
+      (has_alpha && (src->a == NULL || ref->a == NULL))) {
+    return 0;
+  }
+
+  memset(stats, 0, sizeof(stats));
+
+  uv_w = HALVE(src->width);
+  uv_h = HALVE(src->height);
+  if (type >= 2) {
+    float sse[4];
+    sse[0] = AccumulateLSIM(src->y, src->y_stride,
+                            ref->y, ref->y_stride, src->width, src->height);
+    sse[1] = AccumulateLSIM(src->u, src->uv_stride,
+                            ref->u, ref->uv_stride, uv_w, uv_h);
+    sse[2] = AccumulateLSIM(src->v, src->uv_stride,
+                            ref->v, ref->uv_stride, uv_w, uv_h);
+    sse[3] = has_alpha ? AccumulateLSIM(src->a, src->a_stride,
+                                        ref->a, ref->a_stride,
+                                        src->width, src->height)
+                       : 0.f;
+    result[0] = GetPSNR(sse[0] / (src->width * src->height));
+    result[1] = GetPSNR(sse[1] / (uv_w * uv_h));
+    result[2] = GetPSNR(sse[2] / (uv_w * uv_h));
+    result[3] = GetPSNR(sse[3] / (src->width * src->height));
+    {
+      double total_sse = sse[0] + sse[1] + sse[2];
+      int total_pixels = src->width * src->height + 2 * uv_w * uv_h;
+      if (has_alpha) {
+        total_pixels += src->width * src->height;
+        total_sse += sse[3];
+      }
+      result[4] = GetPSNR(total_sse / total_pixels);
+    }
+  } else {
+    int c;
+    VP8SSIMAccumulatePlane(src->y, src->y_stride,
+                           ref->y, ref->y_stride,
+                           src->width, src->height, &stats[0]);
+    VP8SSIMAccumulatePlane(src->u, src->uv_stride,
+                           ref->u, ref->uv_stride,
+                           uv_w, uv_h, &stats[1]);
+    VP8SSIMAccumulatePlane(src->v, src->uv_stride,
+                           ref->v, ref->uv_stride,
+                           uv_w, uv_h, &stats[2]);
+    if (has_alpha) {
+      VP8SSIMAccumulatePlane(src->a, src->a_stride,
+                             ref->a, ref->a_stride,
+                             src->width, src->height, &stats[3]);
+    }
+    for (c = 0; c <= 4; ++c) {
+      if (type == 1) {
+        const double v = VP8SSIMGet(&stats[c]);
+        result[c] = (float)((v < 1.) ? -10.0 * log10(1. - v)
+                                     : kMinDistortion_dB);
+      } else {
+        const double v = VP8SSIMGetSquaredError(&stats[c]);
+        result[c] = GetPSNR(v);
+      }
+      // Accumulate forward
+      if (c < 4) VP8SSIMAddStats(&stats[c], &stats[4]);
+    }
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// Simplest high-level calls:
+
+typedef int (*Importer)(WebPPicture* const, const uint8_t* const, int);
+
+static size_t Encode(const uint8_t* rgba, int width, int height, int stride,
+                     Importer import, float quality_factor, int lossless,
+                     uint8_t** output) {
+  WebPPicture pic;
+  WebPConfig config;
+  WebPMemoryWriter wrt;
+  int ok;
+
+  if (!WebPConfigPreset(&config, WEBP_PRESET_DEFAULT, quality_factor) ||
+      !WebPPictureInit(&pic)) {
+    return 0;  // shouldn't happen, except if system installation is broken
+  }
+
+  config.lossless = !!lossless;
+  pic.use_argb = !!lossless;
+  pic.width = width;
+  pic.height = height;
+  pic.writer = WebPMemoryWrite;
+  pic.custom_ptr = &wrt;
+  WebPMemoryWriterInit(&wrt);
+
+  ok = import(&pic, rgba, stride) && WebPEncode(&config, &pic);
+  WebPPictureFree(&pic);
+  if (!ok) {
+    free(wrt.mem);
+    *output = NULL;
+    return 0;
+  }
+  *output = wrt.mem;
+  return wrt.size;
+}
+
+#define ENCODE_FUNC(NAME, IMPORTER)                                     \
+size_t NAME(const uint8_t* in, int w, int h, int bps, float q,          \
+            uint8_t** out) {                                            \
+  return Encode(in, w, h, bps, IMPORTER, q, 0, out);                    \
+}
+
+ENCODE_FUNC(WebPEncodeRGB, WebPPictureImportRGB)
+ENCODE_FUNC(WebPEncodeBGR, WebPPictureImportBGR)
+ENCODE_FUNC(WebPEncodeRGBA, WebPPictureImportRGBA)
+ENCODE_FUNC(WebPEncodeBGRA, WebPPictureImportBGRA)
+
+#undef ENCODE_FUNC
+
+#define LOSSLESS_DEFAULT_QUALITY 70.
+#define LOSSLESS_ENCODE_FUNC(NAME, IMPORTER)                                 \
+size_t NAME(const uint8_t* in, int w, int h, int bps, uint8_t** out) {       \
+  return Encode(in, w, h, bps, IMPORTER, LOSSLESS_DEFAULT_QUALITY, 1, out);  \
+}
+
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGB, WebPPictureImportRGB)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGR, WebPPictureImportBGR)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGBA, WebPPictureImportRGBA)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGRA, WebPPictureImportBGRA)
+
+#undef LOSSLESS_ENCODE_FUNC
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/enc/quant.c b/drivers/webp/enc/quant.c
new file mode 100644
index 0000000000..e1d202b5a3
--- /dev/null
+++ b/drivers/webp/enc/quant.c
@@ -0,0 +1,1156 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//   Quantization
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <math.h>
+#include <stdlib.h>  // for abs()
+
+#include "./vp8enci.h"
+#include "./cost.h"
+
+#define DO_TRELLIS_I4  1
+#define DO_TRELLIS_I16 1   // not a huge gain, but ok at low bitrate.
+#define DO_TRELLIS_UV  0   // disable trellis for UV. Risky. Not worth.
+#define USE_TDISTO 1
+
+#define MID_ALPHA 64      // neutral value for susceptibility
+#define MIN_ALPHA 30      // lowest usable value for susceptibility
+#define MAX_ALPHA 100     // higher meaningful value for susceptibility
+
+#define SNS_TO_DQ 0.9     // Scaling constant between the sns value and the QP
+                          // power-law modulation. Must be strictly less than 1.
+
+#define I4_PENALTY 4000   // Rate-penalty for quick i4/i16 decision
+
+// number of non-zero coeffs below which we consider the block very flat
+// (and apply a penalty to complex predictions)
+#define FLATNESS_LIMIT_I16 10      // I16 mode
+#define FLATNESS_LIMIT_I4  3       // I4 mode
+#define FLATNESS_LIMIT_UV  2       // UV mode
+#define FLATNESS_PENALTY   140     // roughly ~1bit per block
+
+#define MULT_8B(a, b) (((a) * (b) + 128) >> 8)
+
+// #define DEBUG_BLOCK
+
+//------------------------------------------------------------------------------
+
+#if defined(DEBUG_BLOCK)
+
+#include <stdio.h>
+#include <stdlib.h>
+
+static void PrintBlockInfo(const VP8EncIterator* const it,
+                           const VP8ModeScore* const rd) {
+  int i, j;
+  const int is_i16 = (it->mb_->type_ == 1);
+  printf("SOURCE / OUTPUT / ABS DELTA\n");
+  for (j = 0; j < 24; ++j) {
+    if (j == 16) printf("\n");   // newline before the U/V block
+    for (i = 0; i < 16; ++i) printf("%3d ", it->yuv_in_[i + j * BPS]);
+    printf("     ");
+    for (i = 0; i < 16; ++i) printf("%3d ", it->yuv_out_[i + j * BPS]);
+    printf("     ");
+    for (i = 0; i < 16; ++i) {
+      printf("%1d ", abs(it->yuv_out_[i + j * BPS] - it->yuv_in_[i + j * BPS]));
+    }
+    printf("\n");
+  }
+  printf("\nD:%d SD:%d R:%d H:%d nz:0x%x score:%d\n",
+    (int)rd->D, (int)rd->SD, (int)rd->R, (int)rd->H, (int)rd->nz,
+    (int)rd->score);
+  if (is_i16) {
+    printf("Mode: %d\n", rd->mode_i16);
+    printf("y_dc_levels:");
+    for (i = 0; i < 16; ++i) printf("%3d ", rd->y_dc_levels[i]);
+    printf("\n");
+  } else {
+    printf("Modes[16]: ");
+    for (i = 0; i < 16; ++i) printf("%d ", rd->modes_i4[i]);
+    printf("\n");
+  }
+  printf("y_ac_levels:\n");
+  for (j = 0; j < 16; ++j) {
+    for (i = is_i16 ? 1 : 0; i < 16; ++i) {
+      printf("%4d ", rd->y_ac_levels[j][i]);
+    }
+    printf("\n");
+  }
+  printf("\n");
+  printf("uv_levels (mode=%d):\n", rd->mode_uv);
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 16; ++i) {
+      printf("%4d ", rd->uv_levels[j][i]);
+    }
+    printf("\n");
+  }
+}
+
+#endif   // DEBUG_BLOCK
+
+//------------------------------------------------------------------------------
+
+static WEBP_INLINE int clip(int v, int m, int M) {
+  return v < m ? m : v > M ? M : v;
+}
+
+static const uint8_t kZigzag[16] = {
+  0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
+};
+
+static const uint8_t kDcTable[128] = {
+  4,     5,   6,   7,   8,   9,  10,  10,
+  11,   12,  13,  14,  15,  16,  17,  17,
+  18,   19,  20,  20,  21,  21,  22,  22,
+  23,   23,  24,  25,  25,  26,  27,  28,
+  29,   30,  31,  32,  33,  34,  35,  36,
+  37,   37,  38,  39,  40,  41,  42,  43,
+  44,   45,  46,  46,  47,  48,  49,  50,
+  51,   52,  53,  54,  55,  56,  57,  58,
+  59,   60,  61,  62,  63,  64,  65,  66,
+  67,   68,  69,  70,  71,  72,  73,  74,
+  75,   76,  76,  77,  78,  79,  80,  81,
+  82,   83,  84,  85,  86,  87,  88,  89,
+  91,   93,  95,  96,  98, 100, 101, 102,
+  104, 106, 108, 110, 112, 114, 116, 118,
+  122, 124, 126, 128, 130, 132, 134, 136,
+  138, 140, 143, 145, 148, 151, 154, 157
+};
+
+static const uint16_t kAcTable[128] = {
+  4,     5,   6,   7,   8,   9,  10,  11,
+  12,   13,  14,  15,  16,  17,  18,  19,
+  20,   21,  22,  23,  24,  25,  26,  27,
+  28,   29,  30,  31,  32,  33,  34,  35,
+  36,   37,  38,  39,  40,  41,  42,  43,
+  44,   45,  46,  47,  48,  49,  50,  51,
+  52,   53,  54,  55,  56,  57,  58,  60,
+  62,   64,  66,  68,  70,  72,  74,  76,
+  78,   80,  82,  84,  86,  88,  90,  92,
+  94,   96,  98, 100, 102, 104, 106, 108,
+  110, 112, 114, 116, 119, 122, 125, 128,
+  131, 134, 137, 140, 143, 146, 149, 152,
+  155, 158, 161, 164, 167, 170, 173, 177,
+  181, 185, 189, 193, 197, 201, 205, 209,
+  213, 217, 221, 225, 229, 234, 239, 245,
+  249, 254, 259, 264, 269, 274, 279, 284
+};
+
+static const uint16_t kAcTable2[128] = {
+  8,     8,   9,  10,  12,  13,  15,  17,
+  18,   20,  21,  23,  24,  26,  27,  29,
+  31,   32,  34,  35,  37,  38,  40,  41,
+  43,   44,  46,  48,  49,  51,  52,  54,
+  55,   57,  58,  60,  62,  63,  65,  66,
+  68,   69,  71,  72,  74,  75,  77,  79,
+  80,   82,  83,  85,  86,  88,  89,  93,
+  96,   99, 102, 105, 108, 111, 114, 117,
+  120, 124, 127, 130, 133, 136, 139, 142,
+  145, 148, 151, 155, 158, 161, 164, 167,
+  170, 173, 176, 179, 184, 189, 193, 198,
+  203, 207, 212, 217, 221, 226, 230, 235,
+  240, 244, 249, 254, 258, 263, 268, 274,
+  280, 286, 292, 299, 305, 311, 317, 323,
+  330, 336, 342, 348, 354, 362, 370, 379,
+  385, 393, 401, 409, 416, 424, 432, 440
+};
+
+static const uint8_t kBiasMatrices[3][2] = {  // [luma-ac,luma-dc,chroma][dc,ac]
+  { 96, 110 }, { 96, 108 }, { 110, 115 }
+};
+
+// Sharpening by (slightly) raising the hi-frequency coeffs.
+// Hack-ish but helpful for mid-bitrate range. Use with care.
+#define SHARPEN_BITS 11  // number of descaling bits for sharpening bias
+static const uint8_t kFreqSharpening[16] = {
+  0,  30, 60, 90,
+  30, 60, 90, 90,
+  60, 90, 90, 90,
+  90, 90, 90, 90
+};
+
+//------------------------------------------------------------------------------
+// Initialize quantization parameters in VP8Matrix
+
+// Returns the average quantizer
+static int ExpandMatrix(VP8Matrix* const m, int type) {
+  int i, sum;
+  for (i = 0; i < 2; ++i) {
+    const int is_ac_coeff = (i > 0);
+    const int bias = kBiasMatrices[type][is_ac_coeff];
+    m->iq_[i] = (1 << QFIX) / m->q_[i];
+    m->bias_[i] = BIAS(bias);
+    // zthresh_ is the exact value such that QUANTDIV(coeff, iQ, B) is:
+    //   * zero if coeff <= zthresh
+    //   * non-zero if coeff > zthresh
+    m->zthresh_[i] = ((1 << QFIX) - 1 - m->bias_[i]) / m->iq_[i];
+  }
+  for (i = 2; i < 16; ++i) {
+    m->q_[i] = m->q_[1];
+    m->iq_[i] = m->iq_[1];
+    m->bias_[i] = m->bias_[1];
+    m->zthresh_[i] = m->zthresh_[1];
+  }
+  for (sum = 0, i = 0; i < 16; ++i) {
+    if (type == 0) {  // we only use sharpening for AC luma coeffs
+      m->sharpen_[i] = (kFreqSharpening[i] * m->q_[i]) >> SHARPEN_BITS;
+    } else {
+      m->sharpen_[i] = 0;
+    }
+    sum += m->q_[i];
+  }
+  return (sum + 8) >> 4;
+}
+
+static void SetupMatrices(VP8Encoder* enc) {
+  int i;
+  const int tlambda_scale =
+    (enc->method_ >= 4) ? enc->config_->sns_strength
+                        : 0;
+  const int num_segments = enc->segment_hdr_.num_segments_;
+  for (i = 0; i < num_segments; ++i) {
+    VP8SegmentInfo* const m = &enc->dqm_[i];
+    const int q = m->quant_;
+    int q4, q16, quv;
+    m->y1_.q_[0] = kDcTable[clip(q + enc->dq_y1_dc_, 0, 127)];
+    m->y1_.q_[1] = kAcTable[clip(q,                  0, 127)];
+
+    m->y2_.q_[0] = kDcTable[ clip(q + enc->dq_y2_dc_, 0, 127)] * 2;
+    m->y2_.q_[1] = kAcTable2[clip(q + enc->dq_y2_ac_, 0, 127)];
+
+    m->uv_.q_[0] = kDcTable[clip(q + enc->dq_uv_dc_, 0, 117)];
+    m->uv_.q_[1] = kAcTable[clip(q + enc->dq_uv_ac_, 0, 127)];
+
+    q4  = ExpandMatrix(&m->y1_, 0);
+    q16 = ExpandMatrix(&m->y2_, 1);
+    quv = ExpandMatrix(&m->uv_, 2);
+
+    m->lambda_i4_          = (3 * q4 * q4) >> 7;
+    m->lambda_i16_         = (3 * q16 * q16);
+    m->lambda_uv_          = (3 * quv * quv) >> 6;
+    m->lambda_mode_        = (1 * q4 * q4) >> 7;
+    m->lambda_trellis_i4_  = (7 * q4 * q4) >> 3;
+    m->lambda_trellis_i16_ = (q16 * q16) >> 2;
+    m->lambda_trellis_uv_  = (quv *quv) << 1;
+    m->tlambda_            = (tlambda_scale * q4) >> 5;
+
+    m->min_disto_ = 10 * m->y1_.q_[0];   // quantization-aware min disto
+    m->max_edge_  = 0;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Initialize filtering parameters
+
+// Very small filter-strength values have close to no visual effect. So we can
+// save a little decoding-CPU by turning filtering off for these.
+#define FSTRENGTH_CUTOFF 2
+
+static void SetupFilterStrength(VP8Encoder* const enc) {
+  int i;
+  // level0 is in [0..500]. Using '-f 50' as filter_strength is mid-filtering.
+  const int level0 = 5 * enc->config_->filter_strength;
+  for (i = 0; i < NUM_MB_SEGMENTS; ++i) {
+    VP8SegmentInfo* const m = &enc->dqm_[i];
+    // We focus on the quantization of AC coeffs.
+    const int qstep = kAcTable[clip(m->quant_, 0, 127)] >> 2;
+    const int base_strength =
+        VP8FilterStrengthFromDelta(enc->filter_hdr_.sharpness_, qstep);
+    // Segments with lower complexity ('beta') will be less filtered.
+    const int f = base_strength * level0 / (256 + m->beta_);
+    m->fstrength_ = (f < FSTRENGTH_CUTOFF) ? 0 : (f > 63) ? 63 : f;
+  }
+  // We record the initial strength (mainly for the case of 1-segment only).
+  enc->filter_hdr_.level_ = enc->dqm_[0].fstrength_;
+  enc->filter_hdr_.simple_ = (enc->config_->filter_type == 0);
+  enc->filter_hdr_.sharpness_ = enc->config_->filter_sharpness;
+}
+
+//------------------------------------------------------------------------------
+
+// Note: if you change the values below, remember that the max range
+// allowed by the syntax for DQ_UV is [-16,16].
+#define MAX_DQ_UV (6)
+#define MIN_DQ_UV (-4)
+
+// We want to emulate jpeg-like behaviour where the expected "good" quality
+// is around q=75. Internally, our "good" middle is around c=50. So we
+// map accordingly using linear piece-wise function
+static double QualityToCompression(double c) {
+  const double linear_c = (c < 0.75) ? c * (2. / 3.) : 2. * c - 1.;
+  // The file size roughly scales as pow(quantizer, 3.). Actually, the
+  // exponent is somewhere between 2.8 and 3.2, but we're mostly interested
+  // in the mid-quant range. So we scale the compressibility inversely to
+  // this power-law: quant ~= compression ^ 1/3. This law holds well for
+  // low quant. Finer modeling for high-quant would make use of kAcTable[]
+  // more explicitly.
+  const double v = pow(linear_c, 1 / 3.);
+  return v;
+}
+
+static double QualityToJPEGCompression(double c, double alpha) {
+  // We map the complexity 'alpha' and quality setting 'c' to a compression
+  // exponent empirically matched to the compression curve of libjpeg6b.
+  // On average, the WebP output size will be roughly similar to that of a
+  // JPEG file compressed with same quality factor.
+  const double amin = 0.30;
+  const double amax = 0.85;
+  const double exp_min = 0.4;
+  const double exp_max = 0.9;
+  const double slope = (exp_min - exp_max) / (amax - amin);
+  // Linearly interpolate 'expn' from exp_min to exp_max
+  // in the [amin, amax] range.
+  const double expn = (alpha > amax) ? exp_min
+                    : (alpha < amin) ? exp_max
+                    : exp_max + slope * (alpha - amin);
+  const double v = pow(c, expn);
+  return v;
+}
+
+static int SegmentsAreEquivalent(const VP8SegmentInfo* const S1,
+                                 const VP8SegmentInfo* const S2) {
+  return (S1->quant_ == S2->quant_) && (S1->fstrength_ == S2->fstrength_);
+}
+
+static void SimplifySegments(VP8Encoder* const enc) {
+  int map[NUM_MB_SEGMENTS] = { 0, 1, 2, 3 };
+  const int num_segments = enc->segment_hdr_.num_segments_;
+  int num_final_segments = 1;
+  int s1, s2;
+  for (s1 = 1; s1 < num_segments; ++s1) {    // find similar segments
+    const VP8SegmentInfo* const S1 = &enc->dqm_[s1];
+    int found = 0;
+    // check if we already have similar segment
+    for (s2 = 0; s2 < num_final_segments; ++s2) {
+      const VP8SegmentInfo* const S2 = &enc->dqm_[s2];
+      if (SegmentsAreEquivalent(S1, S2)) {
+        found = 1;
+        break;
+      }
+    }
+    map[s1] = s2;
+    if (!found) {
+      if (num_final_segments != s1) {
+        enc->dqm_[num_final_segments] = enc->dqm_[s1];
+      }
+      ++num_final_segments;
+    }
+  }
+  if (num_final_segments < num_segments) {  // Remap
+    int i = enc->mb_w_ * enc->mb_h_;
+    while (i-- > 0) enc->mb_info_[i].segment_ = map[enc->mb_info_[i].segment_];
+    enc->segment_hdr_.num_segments_ = num_final_segments;
+    // Replicate the trailing segment infos (it's mostly cosmetics)
+    for (i = num_final_segments; i < num_segments; ++i) {
+      enc->dqm_[i] = enc->dqm_[num_final_segments - 1];
+    }
+  }
+}
+
+void VP8SetSegmentParams(VP8Encoder* const enc, float quality) {
+  int i;
+  int dq_uv_ac, dq_uv_dc;
+  const int num_segments = enc->segment_hdr_.num_segments_;
+  const double amp = SNS_TO_DQ * enc->config_->sns_strength / 100. / 128.;
+  const double Q = quality / 100.;
+  const double c_base = enc->config_->emulate_jpeg_size ?
+      QualityToJPEGCompression(Q, enc->alpha_ / 255.) :
+      QualityToCompression(Q);
+  for (i = 0; i < num_segments; ++i) {
+    // We modulate the base coefficient to accommodate for the quantization
+    // susceptibility and allow denser segments to be quantized more.
+    const double expn = 1. - amp * enc->dqm_[i].alpha_;
+    const double c = pow(c_base, expn);
+    const int q = (int)(127. * (1. - c));
+    assert(expn > 0.);
+    enc->dqm_[i].quant_ = clip(q, 0, 127);
+  }
+
+  // purely indicative in the bitstream (except for the 1-segment case)
+  enc->base_quant_ = enc->dqm_[0].quant_;
+
+  // fill-in values for the unused segments (required by the syntax)
+  for (i = num_segments; i < NUM_MB_SEGMENTS; ++i) {
+    enc->dqm_[i].quant_ = enc->base_quant_;
+  }
+
+  // uv_alpha_ is normally spread around ~60. The useful range is
+  // typically ~30 (quite bad) to ~100 (ok to decimate UV more).
+  // We map it to the safe maximal range of MAX/MIN_DQ_UV for dq_uv.
+  dq_uv_ac = (enc->uv_alpha_ - MID_ALPHA) * (MAX_DQ_UV - MIN_DQ_UV)
+                                          / (MAX_ALPHA - MIN_ALPHA);
+  // we rescale by the user-defined strength of adaptation
+  dq_uv_ac = dq_uv_ac * enc->config_->sns_strength / 100;
+  // and make it safe.
+  dq_uv_ac = clip(dq_uv_ac, MIN_DQ_UV, MAX_DQ_UV);
+  // We also boost the dc-uv-quant a little, based on sns-strength, since
+  // U/V channels are quite more reactive to high quants (flat DC-blocks
+  // tend to appear, and are displeasant).
+  dq_uv_dc = -4 * enc->config_->sns_strength / 100;
+  dq_uv_dc = clip(dq_uv_dc, -15, 15);   // 4bit-signed max allowed
+
+  enc->dq_y1_dc_ = 0;       // TODO(skal): dq-lum
+  enc->dq_y2_dc_ = 0;
+  enc->dq_y2_ac_ = 0;
+  enc->dq_uv_dc_ = dq_uv_dc;
+  enc->dq_uv_ac_ = dq_uv_ac;
+
+  SetupFilterStrength(enc);   // initialize segments' filtering, eventually
+
+  if (num_segments > 1) SimplifySegments(enc);
+
+  SetupMatrices(enc);         // finalize quantization matrices
+}
+
+//------------------------------------------------------------------------------
+// Form the predictions in cache
+
+// Must be ordered using {DC_PRED, TM_PRED, V_PRED, H_PRED} as index
+const int VP8I16ModeOffsets[4] = { I16DC16, I16TM16, I16VE16, I16HE16 };
+const int VP8UVModeOffsets[4] = { C8DC8, C8TM8, C8VE8, C8HE8 };
+
+// Must be indexed using {B_DC_PRED -> B_HU_PRED} as index
+const int VP8I4ModeOffsets[NUM_BMODES] = {
+  I4DC4, I4TM4, I4VE4, I4HE4, I4RD4, I4VR4, I4LD4, I4VL4, I4HD4, I4HU4
+};
+
+void VP8MakeLuma16Preds(const VP8EncIterator* const it) {
+  const uint8_t* const left = it->x_ ? it->y_left_ : NULL;
+  const uint8_t* const top = it->y_ ? it->y_top_ : NULL;
+  VP8EncPredLuma16(it->yuv_p_, left, top);
+}
+
+void VP8MakeChroma8Preds(const VP8EncIterator* const it) {
+  const uint8_t* const left = it->x_ ? it->u_left_ : NULL;
+  const uint8_t* const top = it->y_ ? it->uv_top_ : NULL;
+  VP8EncPredChroma8(it->yuv_p_, left, top);
+}
+
+void VP8MakeIntra4Preds(const VP8EncIterator* const it) {
+  VP8EncPredLuma4(it->yuv_p_, it->i4_top_);
+}
+
+//------------------------------------------------------------------------------
+// Quantize
+
+// Layout:
+// +----+
+// |YYYY| 0
+// |YYYY| 4
+// |YYYY| 8
+// |YYYY| 12
+// +----+
+// |UUVV| 16
+// |UUVV| 20
+// +----+
+
+const int VP8Scan[16 + 4 + 4] = {
+  // Luma
+  0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
+  0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
+  0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
+  0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS,
+
+  0 + 0 * BPS,   4 + 0 * BPS, 0 + 4 * BPS,  4 + 4 * BPS,    // U
+  8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
+};
+
+//------------------------------------------------------------------------------
+// Distortion measurement
+
+static const uint16_t kWeightY[16] = {
+  38, 32, 20, 9, 32, 28, 17, 7, 20, 17, 10, 4, 9, 7, 4, 2
+};
+
+static const uint16_t kWeightTrellis[16] = {
+#if USE_TDISTO == 0
+  16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16
+#else
+  30, 27, 19, 11,
+  27, 24, 17, 10,
+  19, 17, 12,  8,
+  11, 10,  8,  6
+#endif
+};
+
+// Init/Copy the common fields in score.
+static void InitScore(VP8ModeScore* const rd) {
+  rd->D  = 0;
+  rd->SD = 0;
+  rd->R  = 0;
+  rd->H  = 0;
+  rd->nz = 0;
+  rd->score = MAX_COST;
+}
+
+static void CopyScore(VP8ModeScore* const dst, const VP8ModeScore* const src) {
+  dst->D  = src->D;
+  dst->SD = src->SD;
+  dst->R  = src->R;
+  dst->H  = src->H;
+  dst->nz = src->nz;      // note that nz is not accumulated, but just copied.
+  dst->score = src->score;
+}
+
+static void AddScore(VP8ModeScore* const dst, const VP8ModeScore* const src) {
+  dst->D  += src->D;
+  dst->SD += src->SD;
+  dst->R  += src->R;
+  dst->H  += src->H;
+  dst->nz |= src->nz;     // here, new nz bits are accumulated.
+  dst->score += src->score;
+}
+
+//------------------------------------------------------------------------------
+// Performs trellis-optimized quantization.
+
+// Trellis
+
+typedef struct {
+  int prev;        // best previous
+  int level;       // level
+  int sign;        // sign of coeff_i
+  score_t cost;    // bit cost
+  score_t error;   // distortion = sum of (|coeff_i| - level_i * Q_i)^2
+  int ctx;         // context (only depends on 'level'. Could be spared.)
+} Node;
+
+// If a coefficient was quantized to a value Q (using a neutral bias),
+// we test all alternate possibilities between [Q-MIN_DELTA, Q+MAX_DELTA]
+// We don't test negative values though.
+#define MIN_DELTA 0   // how much lower level to try
+#define MAX_DELTA 1   // how much higher
+#define NUM_NODES (MIN_DELTA + 1 + MAX_DELTA)
+#define NODE(n, l) (nodes[(n) + 1][(l) + MIN_DELTA])
+
+static WEBP_INLINE void SetRDScore(int lambda, VP8ModeScore* const rd) {
+  // TODO: incorporate the "* 256" in the tables?
+  rd->score = (rd->R + rd->H) * lambda + 256 * (rd->D + rd->SD);
+}
+
+static WEBP_INLINE score_t RDScoreTrellis(int lambda, score_t rate,
+                                          score_t distortion) {
+  return rate * lambda + 256 * distortion;
+}
+
+static int TrellisQuantizeBlock(const VP8EncIterator* const it,
+                                int16_t in[16], int16_t out[16],
+                                int ctx0, int coeff_type,
+                                const VP8Matrix* const mtx,
+                                int lambda) {
+  ProbaArray* const last_costs = it->enc_->proba_.coeffs_[coeff_type];
+  CostArray* const costs = it->enc_->proba_.level_cost_[coeff_type];
+  const int first = (coeff_type == 0) ? 1 : 0;
+  Node nodes[17][NUM_NODES];
+  int best_path[3] = {-1, -1, -1};   // store best-last/best-level/best-previous
+  score_t best_score;
+  int best_node;
+  int last = first - 1;
+  int n, m, p, nz;
+
+  {
+    score_t cost;
+    score_t max_error;
+    const int thresh = mtx->q_[1] * mtx->q_[1] / 4;
+    const int last_proba = last_costs[VP8EncBands[first]][ctx0][0];
+
+    // compute maximal distortion.
+    max_error = 0;
+    for (n = first; n < 16; ++n) {
+      const int j  = kZigzag[n];
+      const int err = in[j] * in[j];
+      max_error += kWeightTrellis[j] * err;
+      if (err > thresh) last = n;
+    }
+    // we don't need to go inspect up to n = 16 coeffs. We can just go up
+    // to last + 1 (inclusive) without losing much.
+    if (last < 15) ++last;
+
+    // compute 'skip' score. This is the max score one can do.
+    cost = VP8BitCost(0, last_proba);
+    best_score = RDScoreTrellis(lambda, cost, max_error);
+
+    // initialize source node.
+    n = first - 1;
+    for (m = -MIN_DELTA; m <= MAX_DELTA; ++m) {
+      NODE(n, m).cost = 0;
+      NODE(n, m).error = max_error;
+      NODE(n, m).ctx = ctx0;
+    }
+  }
+
+  // traverse trellis.
+  for (n = first; n <= last; ++n) {
+    const int j  = kZigzag[n];
+    const int Q  = mtx->q_[j];
+    const int iQ = mtx->iq_[j];
+    const int B = BIAS(0x00);     // neutral bias
+    // note: it's important to take sign of the _original_ coeff,
+    // so we don't have to consider level < 0 afterward.
+    const int sign = (in[j] < 0);
+    const int coeff0 = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
+    int level0 = QUANTDIV(coeff0, iQ, B);
+    if (level0 > MAX_LEVEL) level0 = MAX_LEVEL;
+
+    // test all alternate level values around level0.
+    for (m = -MIN_DELTA; m <= MAX_DELTA; ++m) {
+      Node* const cur = &NODE(n, m);
+      int delta_error, new_error;
+      score_t cur_score = MAX_COST;
+      int level = level0 + m;
+      int last_proba;
+
+      cur->sign = sign;
+      cur->level = level;
+      cur->ctx = (level == 0) ? 0 : (level == 1) ? 1 : 2;
+      if (level > MAX_LEVEL || level < 0) {   // node is dead?
+        cur->cost = MAX_COST;
+        continue;
+      }
+      last_proba = last_costs[VP8EncBands[n + 1]][cur->ctx][0];
+
+      // Compute delta_error = how much coding this level will
+      // subtract as distortion to max_error
+      new_error = coeff0 - level * Q;
+      delta_error =
+        kWeightTrellis[j] * (coeff0 * coeff0 - new_error * new_error);
+
+      // Inspect all possible non-dead predecessors. Retain only the best one.
+      for (p = -MIN_DELTA; p <= MAX_DELTA; ++p) {
+        const Node* const prev = &NODE(n - 1, p);
+        const int prev_ctx = prev->ctx;
+        const uint16_t* const tcost = costs[VP8EncBands[n]][prev_ctx];
+        const score_t total_error = prev->error - delta_error;
+        score_t cost, base_cost, score;
+
+        if (prev->cost >= MAX_COST) {   // dead node?
+          continue;
+        }
+
+        // Base cost of both terminal/non-terminal
+        base_cost = prev->cost + VP8LevelCost(tcost, level);
+
+        // Examine node assuming it's a non-terminal one.
+        cost = base_cost;
+        if (level && n < 15) {
+          cost += VP8BitCost(1, last_proba);
+        }
+        score = RDScoreTrellis(lambda, cost, total_error);
+        if (score < cur_score) {
+          cur_score = score;
+          cur->cost  = cost;
+          cur->error = total_error;
+          cur->prev  = p;
+        }
+
+        // Now, record best terminal node (and thus best entry in the graph).
+        if (level) {
+          cost = base_cost;
+          if (n < 15) cost += VP8BitCost(0, last_proba);
+          score = RDScoreTrellis(lambda, cost, total_error);
+          if (score < best_score) {
+            best_score = score;
+            best_path[0] = n;   // best eob position
+            best_path[1] = m;   // best level
+            best_path[2] = p;   // best predecessor
+          }
+        }
+      }
+    }
+  }
+
+  // Fresh start
+  memset(in + first, 0, (16 - first) * sizeof(*in));
+  memset(out + first, 0, (16 - first) * sizeof(*out));
+  if (best_path[0] == -1) {
+    return 0;   // skip!
+  }
+
+  // Unwind the best path.
+  // Note: best-prev on terminal node is not necessarily equal to the
+  // best_prev for non-terminal. So we patch best_path[2] in.
+  n = best_path[0];
+  best_node = best_path[1];
+  NODE(n, best_node).prev = best_path[2];   // force best-prev for terminal
+  nz = 0;
+
+  for (; n >= first; --n) {
+    const Node* const node = &NODE(n, best_node);
+    const int j = kZigzag[n];
+    out[n] = node->sign ? -node->level : node->level;
+    nz |= (node->level != 0);
+    in[j] = out[n] * mtx->q_[j];
+    best_node = node->prev;
+  }
+  return nz;
+}
+
+#undef NODE
+
+//------------------------------------------------------------------------------
+// Performs: difference, transform, quantize, back-transform, add
+// all at once. Output is the reconstructed block in *yuv_out, and the
+// quantized levels in *levels.
+
+static int ReconstructIntra16(VP8EncIterator* const it,
+                              VP8ModeScore* const rd,
+                              uint8_t* const yuv_out,
+                              int mode) {
+  VP8Encoder* const enc = it->enc_;
+  const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
+  const uint8_t* const src = it->yuv_in_ + Y_OFF;
+  VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  int nz = 0;
+  int n;
+  int16_t tmp[16][16], dc_tmp[16];
+
+  for (n = 0; n < 16; ++n) {
+    VP8FTransform(src + VP8Scan[n], ref + VP8Scan[n], tmp[n]);
+  }
+  VP8FTransformWHT(tmp[0], dc_tmp);
+  nz |= VP8EncQuantizeBlockWHT(dc_tmp, rd->y_dc_levels, &dqm->y2_) << 24;
+
+  if (DO_TRELLIS_I16 && it->do_trellis_) {
+    int x, y;
+    VP8IteratorNzToBytes(it);
+    for (y = 0, n = 0; y < 4; ++y) {
+      for (x = 0; x < 4; ++x, ++n) {
+        const int ctx = it->top_nz_[x] + it->left_nz_[y];
+        const int non_zero =
+           TrellisQuantizeBlock(it, tmp[n], rd->y_ac_levels[n], ctx, 0,
+                                &dqm->y1_, dqm->lambda_trellis_i16_);
+        it->top_nz_[x] = it->left_nz_[y] = non_zero;
+        nz |= non_zero << n;
+      }
+    }
+  } else {
+    for (n = 0; n < 16; ++n) {
+      nz |= VP8EncQuantizeBlock(tmp[n], rd->y_ac_levels[n], 1, &dqm->y1_) << n;
+    }
+  }
+
+  // Transform back
+  VP8ITransformWHT(dc_tmp, tmp[0]);
+  for (n = 0; n < 16; n += 2) {
+    VP8ITransform(ref + VP8Scan[n], tmp[n], yuv_out + VP8Scan[n], 1);
+  }
+
+  return nz;
+}
+
+static int ReconstructIntra4(VP8EncIterator* const it,
+                             int16_t levels[16],
+                             const uint8_t* const src,
+                             uint8_t* const yuv_out,
+                             int mode) {
+  const VP8Encoder* const enc = it->enc_;
+  const uint8_t* const ref = it->yuv_p_ + VP8I4ModeOffsets[mode];
+  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  int nz = 0;
+  int16_t tmp[16];
+
+  VP8FTransform(src, ref, tmp);
+  if (DO_TRELLIS_I4 && it->do_trellis_) {
+    const int x = it->i4_ & 3, y = it->i4_ >> 2;
+    const int ctx = it->top_nz_[x] + it->left_nz_[y];
+    nz = TrellisQuantizeBlock(it, tmp, levels, ctx, 3, &dqm->y1_,
+                              dqm->lambda_trellis_i4_);
+  } else {
+    nz = VP8EncQuantizeBlock(tmp, levels, 0, &dqm->y1_);
+  }
+  VP8ITransform(ref, tmp, yuv_out, 0);
+  return nz;
+}
+
+static int ReconstructUV(VP8EncIterator* const it, VP8ModeScore* const rd,
+                         uint8_t* const yuv_out, int mode) {
+  const VP8Encoder* const enc = it->enc_;
+  const uint8_t* const ref = it->yuv_p_ + VP8UVModeOffsets[mode];
+  const uint8_t* const src = it->yuv_in_ + U_OFF;
+  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  int nz = 0;
+  int n;
+  int16_t tmp[8][16];
+
+  for (n = 0; n < 8; ++n) {
+    VP8FTransform(src + VP8Scan[16 + n], ref + VP8Scan[16 + n], tmp[n]);
+  }
+  if (DO_TRELLIS_UV && it->do_trellis_) {
+    int ch, x, y;
+    for (ch = 0, n = 0; ch <= 2; ch += 2) {
+      for (y = 0; y < 2; ++y) {
+        for (x = 0; x < 2; ++x, ++n) {
+          const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
+          const int non_zero =
+            TrellisQuantizeBlock(it, tmp[n], rd->uv_levels[n], ctx, 2,
+                                 &dqm->uv_, dqm->lambda_trellis_uv_);
+          it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = non_zero;
+          nz |= non_zero << n;
+        }
+      }
+    }
+  } else {
+    for (n = 0; n < 8; ++n) {
+      nz |= VP8EncQuantizeBlock(tmp[n], rd->uv_levels[n], 0, &dqm->uv_) << n;
+    }
+  }
+
+  for (n = 0; n < 8; n += 2) {
+    VP8ITransform(ref + VP8Scan[16 + n], tmp[n], yuv_out + VP8Scan[16 + n], 1);
+  }
+  return (nz << 16);
+}
+
+//------------------------------------------------------------------------------
+// RD-opt decision. Reconstruct each modes, evalue distortion and bit-cost.
+// Pick the mode is lower RD-cost = Rate + lambda * Distortion.
+
+static void StoreMaxDelta(VP8SegmentInfo* const dqm, const int16_t DCs[16]) {
+  // We look at the first three AC coefficients to determine what is the average
+  // delta between each sub-4x4 block.
+  const int v0 = abs(DCs[1]);
+  const int v1 = abs(DCs[4]);
+  const int v2 = abs(DCs[5]);
+  int max_v = (v0 > v1) ? v1 : v0;
+  max_v = (v2 > max_v) ? v2 : max_v;
+  if (max_v > dqm->max_edge_) dqm->max_edge_ = max_v;
+}
+
+static void SwapPtr(uint8_t** a, uint8_t** b) {
+  uint8_t* const tmp = *a;
+  *a = *b;
+  *b = tmp;
+}
+
+static void SwapOut(VP8EncIterator* const it) {
+  SwapPtr(&it->yuv_out_, &it->yuv_out2_);
+}
+
+static score_t IsFlat(const int16_t* levels, int num_blocks, score_t thresh) {
+  score_t score = 0;
+  while (num_blocks-- > 0) {      // TODO(skal): refine positional scoring?
+    int i;
+    for (i = 1; i < 16; ++i) {    // omit DC, we're only interested in AC
+      score += (levels[i] != 0);
+      if (score > thresh) return 0;
+    }
+    levels += 16;
+  }
+  return 1;
+}
+
+static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* const rd) {
+  const int kNumBlocks = 16;
+  VP8Encoder* const enc = it->enc_;
+  VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  const int lambda = dqm->lambda_i16_;
+  const int tlambda = dqm->tlambda_;
+  const uint8_t* const src = it->yuv_in_ + Y_OFF;
+  VP8ModeScore rd16;
+  int mode;
+
+  rd->mode_i16 = -1;
+  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
+    uint8_t* const tmp_dst = it->yuv_out2_ + Y_OFF;  // scratch buffer
+    int nz;
+
+    // Reconstruct
+    nz = ReconstructIntra16(it, &rd16, tmp_dst, mode);
+
+    // Measure RD-score
+    rd16.D = VP8SSE16x16(src, tmp_dst);
+    rd16.SD = tlambda ? MULT_8B(tlambda, VP8TDisto16x16(src, tmp_dst, kWeightY))
+            : 0;
+    rd16.H = VP8FixedCostsI16[mode];
+    rd16.R = VP8GetCostLuma16(it, &rd16);
+    if (mode > 0 &&
+        IsFlat(rd16.y_ac_levels[0], kNumBlocks, FLATNESS_LIMIT_I16)) {
+      // penalty to avoid flat area to be mispredicted by complex mode
+      rd16.R += FLATNESS_PENALTY * kNumBlocks;
+    }
+
+    // Since we always examine Intra16 first, we can overwrite *rd directly.
+    SetRDScore(lambda, &rd16);
+    if (mode == 0 || rd16.score < rd->score) {
+      CopyScore(rd, &rd16);
+      rd->mode_i16 = mode;
+      rd->nz = nz;
+      memcpy(rd->y_ac_levels, rd16.y_ac_levels, sizeof(rd16.y_ac_levels));
+      memcpy(rd->y_dc_levels, rd16.y_dc_levels, sizeof(rd16.y_dc_levels));
+      SwapOut(it);
+    }
+  }
+  SetRDScore(dqm->lambda_mode_, rd);   // finalize score for mode decision.
+  VP8SetIntra16Mode(it, rd->mode_i16);
+
+  // we have a blocky macroblock (only DCs are non-zero) with fairly high
+  // distortion, record max delta so we can later adjust the minimal filtering
+  // strength needed to smooth these blocks out.
+  if ((rd->nz & 0xffff) == 0 && rd->D > dqm->min_disto_) {
+    StoreMaxDelta(dqm, rd->y_dc_levels);
+  }
+}
+
+//------------------------------------------------------------------------------
+
+// return the cost array corresponding to the surrounding prediction modes.
+static const uint16_t* GetCostModeI4(VP8EncIterator* const it,
+                                     const uint8_t modes[16]) {
+  const int preds_w = it->enc_->preds_w_;
+  const int x = (it->i4_ & 3), y = it->i4_ >> 2;
+  const int left = (x == 0) ? it->preds_[y * preds_w - 1] : modes[it->i4_ - 1];
+  const int top = (y == 0) ? it->preds_[-preds_w + x] : modes[it->i4_ - 4];
+  return VP8FixedCostsI4[top][left];
+}
+
+static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
+  const VP8Encoder* const enc = it->enc_;
+  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  const int lambda = dqm->lambda_i4_;
+  const int tlambda = dqm->tlambda_;
+  const uint8_t* const src0 = it->yuv_in_ + Y_OFF;
+  uint8_t* const best_blocks = it->yuv_out2_ + Y_OFF;
+  int total_header_bits = 0;
+  VP8ModeScore rd_best;
+
+  if (enc->max_i4_header_bits_ == 0) {
+    return 0;
+  }
+
+  InitScore(&rd_best);
+  rd_best.H = 211;  // '211' is the value of VP8BitCost(0, 145)
+  SetRDScore(dqm->lambda_mode_, &rd_best);
+  VP8IteratorStartI4(it);
+  do {
+    const int kNumBlocks = 1;
+    VP8ModeScore rd_i4;
+    int mode;
+    int best_mode = -1;
+    const uint8_t* const src = src0 + VP8Scan[it->i4_];
+    const uint16_t* const mode_costs = GetCostModeI4(it, rd->modes_i4);
+    uint8_t* best_block = best_blocks + VP8Scan[it->i4_];
+    uint8_t* tmp_dst = it->yuv_p_ + I4TMP;    // scratch buffer.
+
+    InitScore(&rd_i4);
+    VP8MakeIntra4Preds(it);
+    for (mode = 0; mode < NUM_BMODES; ++mode) {
+      VP8ModeScore rd_tmp;
+      int16_t tmp_levels[16];
+
+      // Reconstruct
+      rd_tmp.nz =
+          ReconstructIntra4(it, tmp_levels, src, tmp_dst, mode) << it->i4_;
+
+      // Compute RD-score
+      rd_tmp.D = VP8SSE4x4(src, tmp_dst);
+      rd_tmp.SD =
+          tlambda ? MULT_8B(tlambda, VP8TDisto4x4(src, tmp_dst, kWeightY))
+                  : 0;
+      rd_tmp.H = mode_costs[mode];
+      rd_tmp.R = VP8GetCostLuma4(it, tmp_levels);
+      if (mode > 0 && IsFlat(tmp_levels, kNumBlocks, FLATNESS_LIMIT_I4)) {
+        rd_tmp.R += FLATNESS_PENALTY * kNumBlocks;
+      }
+
+      SetRDScore(lambda, &rd_tmp);
+      if (best_mode < 0 || rd_tmp.score < rd_i4.score) {
+        CopyScore(&rd_i4, &rd_tmp);
+        best_mode = mode;
+        SwapPtr(&tmp_dst, &best_block);
+        memcpy(rd_best.y_ac_levels[it->i4_], tmp_levels, sizeof(tmp_levels));
+      }
+    }
+    SetRDScore(dqm->lambda_mode_, &rd_i4);
+    AddScore(&rd_best, &rd_i4);
+    if (rd_best.score >= rd->score) {
+      return 0;
+    }
+    total_header_bits += (int)rd_i4.H;   // <- equal to mode_costs[best_mode];
+    if (total_header_bits > enc->max_i4_header_bits_) {
+      return 0;
+    }
+    // Copy selected samples if not in the right place already.
+    if (best_block != best_blocks + VP8Scan[it->i4_]) {
+      VP8Copy4x4(best_block, best_blocks + VP8Scan[it->i4_]);
+    }
+    rd->modes_i4[it->i4_] = best_mode;
+    it->top_nz_[it->i4_ & 3] = it->left_nz_[it->i4_ >> 2] = (rd_i4.nz ? 1 : 0);
+  } while (VP8IteratorRotateI4(it, best_blocks));
+
+  // finalize state
+  CopyScore(rd, &rd_best);
+  VP8SetIntra4Mode(it, rd->modes_i4);
+  SwapOut(it);
+  memcpy(rd->y_ac_levels, rd_best.y_ac_levels, sizeof(rd->y_ac_levels));
+  return 1;   // select intra4x4 over intra16x16
+}
+
+//------------------------------------------------------------------------------
+
+static void PickBestUV(VP8EncIterator* const it, VP8ModeScore* const rd) {
+  const int kNumBlocks = 8;
+  const VP8Encoder* const enc = it->enc_;
+  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  const int lambda = dqm->lambda_uv_;
+  const uint8_t* const src = it->yuv_in_ + U_OFF;
+  uint8_t* const tmp_dst = it->yuv_out2_ + U_OFF;  // scratch buffer
+  uint8_t* const dst0 = it->yuv_out_ + U_OFF;
+  VP8ModeScore rd_best;
+  int mode;
+
+  rd->mode_uv = -1;
+  InitScore(&rd_best);
+  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
+    VP8ModeScore rd_uv;
+
+    // Reconstruct
+    rd_uv.nz = ReconstructUV(it, &rd_uv, tmp_dst, mode);
+
+    // Compute RD-score
+    rd_uv.D  = VP8SSE16x8(src, tmp_dst);
+    rd_uv.SD = 0;    // TODO: should we call TDisto? it tends to flatten areas.
+    rd_uv.H  = VP8FixedCostsUV[mode];
+    rd_uv.R  = VP8GetCostUV(it, &rd_uv);
+    if (mode > 0 && IsFlat(rd_uv.uv_levels[0], kNumBlocks, FLATNESS_LIMIT_UV)) {
+      rd_uv.R += FLATNESS_PENALTY * kNumBlocks;
+    }
+
+    SetRDScore(lambda, &rd_uv);
+    if (mode == 0 || rd_uv.score < rd_best.score) {
+      CopyScore(&rd_best, &rd_uv);
+      rd->mode_uv = mode;
+      memcpy(rd->uv_levels, rd_uv.uv_levels, sizeof(rd->uv_levels));
+      memcpy(dst0, tmp_dst, UV_SIZE);   //  TODO: SwapUVOut() ?
+    }
+  }
+  VP8SetIntraUVMode(it, rd->mode_uv);
+  AddScore(rd, &rd_best);
+}
+
+//------------------------------------------------------------------------------
+// Final reconstruction and quantization.
+
+static void SimpleQuantize(VP8EncIterator* const it, VP8ModeScore* const rd) {
+  const VP8Encoder* const enc = it->enc_;
+  const int is_i16 = (it->mb_->type_ == 1);
+  int nz = 0;
+
+  if (is_i16) {
+    nz = ReconstructIntra16(it, rd, it->yuv_out_ + Y_OFF, it->preds_[0]);
+  } else {
+    VP8IteratorStartI4(it);
+    do {
+      const int mode =
+          it->preds_[(it->i4_ & 3) + (it->i4_ >> 2) * enc->preds_w_];
+      const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
+      uint8_t* const dst = it->yuv_out_ + Y_OFF + VP8Scan[it->i4_];
+      VP8MakeIntra4Preds(it);
+      nz |= ReconstructIntra4(it, rd->y_ac_levels[it->i4_],
+                              src, dst, mode) << it->i4_;
+    } while (VP8IteratorRotateI4(it, it->yuv_out_ + Y_OFF));
+  }
+
+  nz |= ReconstructUV(it, rd, it->yuv_out_ + U_OFF, it->mb_->uv_mode_);
+  rd->nz = nz;
+}
+
+// Refine intra16/intra4 sub-modes based on distortion only (not rate).
+static void DistoRefine(VP8EncIterator* const it, int try_both_i4_i16) {
+  const int is_i16 = (it->mb_->type_ == 1);
+  score_t best_score = MAX_COST;
+
+  if (try_both_i4_i16 || is_i16) {
+    int mode;
+    int best_mode = -1;
+    for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
+      const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
+      const uint8_t* const src = it->yuv_in_ + Y_OFF;
+      const score_t score = VP8SSE16x16(src, ref);
+      if (score < best_score) {
+        best_mode = mode;
+        best_score = score;
+      }
+    }
+    VP8SetIntra16Mode(it, best_mode);
+  }
+  if (try_both_i4_i16 || !is_i16) {
+    uint8_t modes_i4[16];
+    // We don't evaluate the rate here, but just account for it through a
+    // constant penalty (i4 mode usually needs more bits compared to i16).
+    score_t score_i4 = (score_t)I4_PENALTY;
+
+    VP8IteratorStartI4(it);
+    do {
+      int mode;
+      int best_sub_mode = -1;
+      score_t best_sub_score = MAX_COST;
+      const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
+
+      // TODO(skal): we don't really need the prediction pixels here,
+      // but just the distortion against 'src'.
+      VP8MakeIntra4Preds(it);
+      for (mode = 0; mode < NUM_BMODES; ++mode) {
+        const uint8_t* const ref = it->yuv_p_ + VP8I4ModeOffsets[mode];
+        const score_t score = VP8SSE4x4(src, ref);
+        if (score < best_sub_score) {
+          best_sub_mode = mode;
+          best_sub_score = score;
+        }
+      }
+      modes_i4[it->i4_] = best_sub_mode;
+      score_i4 += best_sub_score;
+      if (score_i4 >= best_score) break;
+    } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF));
+    if (score_i4 < best_score) {
+      VP8SetIntra4Mode(it, modes_i4);
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+// Entry point
+
+int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd,
+                VP8RDLevel rd_opt) {
+  int is_skipped;
+  const int method = it->enc_->method_;
+
+  InitScore(rd);
+
+  // We can perform predictions for Luma16x16 and Chroma8x8 already.
+  // Luma4x4 predictions needs to be done as-we-go.
+  VP8MakeLuma16Preds(it);
+  VP8MakeChroma8Preds(it);
+
+  if (rd_opt > RD_OPT_NONE) {
+    it->do_trellis_ = (rd_opt >= RD_OPT_TRELLIS_ALL);
+    PickBestIntra16(it, rd);
+    if (method >= 2) {
+      PickBestIntra4(it, rd);
+    }
+    PickBestUV(it, rd);
+    if (rd_opt == RD_OPT_TRELLIS) {   // finish off with trellis-optim now
+      it->do_trellis_ = 1;
+      SimpleQuantize(it, rd);
+    }
+  } else {
+    // For method == 2, pick the best intra4/intra16 based on SSE (~tad slower).
+    // For method <= 1, we refine intra4 or intra16 (but don't re-examine mode).
+    DistoRefine(it, (method >= 2));
+    SimpleQuantize(it, rd);
+  }
+  is_skipped = (rd->nz == 0);
+  VP8SetSkip(it, is_skipped);
+  return is_skipped;
+}
+
diff --git a/drivers/webp/enc/syntax.c b/drivers/webp/enc/syntax.c
new file mode 100644
index 0000000000..08cfe79ece
--- /dev/null
+++ b/drivers/webp/enc/syntax.c
@@ -0,0 +1,423 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Header syntax writing
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+
+#include "../utils/utils.h"
+#include "../webp/format_constants.h"  // RIFF constants
+#include "../webp/mux_types.h"         // ALPHA_FLAG
+#include "./vp8enci.h"
+
+//------------------------------------------------------------------------------
+// Helper functions
+
+static int IsVP8XNeeded(const VP8Encoder* const enc) {
+  return !!enc->has_alpha_;  // Currently the only case when VP8X is needed.
+                             // This could change in the future.
+}
+
+static int PutPaddingByte(const WebPPicture* const pic) {
+  const uint8_t pad_byte[1] = { 0 };
+  return !!pic->writer(pad_byte, 1, pic);
+}
+
+//------------------------------------------------------------------------------
+// Writers for header's various pieces (in order of appearance)
+
+static WebPEncodingError PutRIFFHeader(const VP8Encoder* const enc,
+                                       size_t riff_size) {
+  const WebPPicture* const pic = enc->pic_;
+  uint8_t riff[RIFF_HEADER_SIZE] = {
+    'R', 'I', 'F', 'F', 0, 0, 0, 0, 'W', 'E', 'B', 'P'
+  };
+  assert(riff_size == (uint32_t)riff_size);
+  PutLE32(riff + TAG_SIZE, (uint32_t)riff_size);
+  if (!pic->writer(riff, sizeof(riff), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static WebPEncodingError PutVP8XHeader(const VP8Encoder* const enc) {
+  const WebPPicture* const pic = enc->pic_;
+  uint8_t vp8x[CHUNK_HEADER_SIZE + VP8X_CHUNK_SIZE] = {
+    'V', 'P', '8', 'X'
+  };
+  uint32_t flags = 0;
+
+  assert(IsVP8XNeeded(enc));
+  assert(pic->width >= 1 && pic->height >= 1);
+  assert(pic->width <= MAX_CANVAS_SIZE && pic->height <= MAX_CANVAS_SIZE);
+
+  if (enc->has_alpha_) {
+    flags |= ALPHA_FLAG;
+  }
+
+  PutLE32(vp8x + TAG_SIZE,              VP8X_CHUNK_SIZE);
+  PutLE32(vp8x + CHUNK_HEADER_SIZE,     flags);
+  PutLE24(vp8x + CHUNK_HEADER_SIZE + 4, pic->width - 1);
+  PutLE24(vp8x + CHUNK_HEADER_SIZE + 7, pic->height - 1);
+  if (!pic->writer(vp8x, sizeof(vp8x), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static WebPEncodingError PutAlphaChunk(const VP8Encoder* const enc) {
+  const WebPPicture* const pic = enc->pic_;
+  uint8_t alpha_chunk_hdr[CHUNK_HEADER_SIZE] = {
+    'A', 'L', 'P', 'H'
+  };
+
+  assert(enc->has_alpha_);
+
+  // Alpha chunk header.
+  PutLE32(alpha_chunk_hdr + TAG_SIZE, enc->alpha_data_size_);
+  if (!pic->writer(alpha_chunk_hdr, sizeof(alpha_chunk_hdr), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+
+  // Alpha chunk data.
+  if (!pic->writer(enc->alpha_data_, enc->alpha_data_size_, pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+
+  // Padding.
+  if ((enc->alpha_data_size_ & 1) && !PutPaddingByte(pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static WebPEncodingError PutVP8Header(const WebPPicture* const pic,
+                                      size_t vp8_size) {
+  uint8_t vp8_chunk_hdr[CHUNK_HEADER_SIZE] = {
+    'V', 'P', '8', ' '
+  };
+  assert(vp8_size == (uint32_t)vp8_size);
+  PutLE32(vp8_chunk_hdr + TAG_SIZE, (uint32_t)vp8_size);
+  if (!pic->writer(vp8_chunk_hdr, sizeof(vp8_chunk_hdr), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static WebPEncodingError PutVP8FrameHeader(const WebPPicture* const pic,
+                                           int profile, size_t size0) {
+  uint8_t vp8_frm_hdr[VP8_FRAME_HEADER_SIZE];
+  uint32_t bits;
+
+  if (size0 >= VP8_MAX_PARTITION0_SIZE) {  // partition #0 is too big to fit
+    return VP8_ENC_ERROR_PARTITION0_OVERFLOW;
+  }
+
+  // Paragraph 9.1.
+  bits = 0                         // keyframe (1b)
+       | (profile << 1)            // profile (3b)
+       | (1 << 4)                  // visible (1b)
+       | ((uint32_t)size0 << 5);   // partition length (19b)
+  vp8_frm_hdr[0] = (bits >>  0) & 0xff;
+  vp8_frm_hdr[1] = (bits >>  8) & 0xff;
+  vp8_frm_hdr[2] = (bits >> 16) & 0xff;
+  // signature
+  vp8_frm_hdr[3] = (VP8_SIGNATURE >> 16) & 0xff;
+  vp8_frm_hdr[4] = (VP8_SIGNATURE >>  8) & 0xff;
+  vp8_frm_hdr[5] = (VP8_SIGNATURE >>  0) & 0xff;
+  // dimensions
+  vp8_frm_hdr[6] = pic->width & 0xff;
+  vp8_frm_hdr[7] = pic->width >> 8;
+  vp8_frm_hdr[8] = pic->height & 0xff;
+  vp8_frm_hdr[9] = pic->height >> 8;
+
+  if (!pic->writer(vp8_frm_hdr, sizeof(vp8_frm_hdr), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+// WebP Headers.
+static int PutWebPHeaders(const VP8Encoder* const enc, size_t size0,
+                          size_t vp8_size, size_t riff_size) {
+  WebPPicture* const pic = enc->pic_;
+  WebPEncodingError err = VP8_ENC_OK;
+
+  // RIFF header.
+  err = PutRIFFHeader(enc, riff_size);
+  if (err != VP8_ENC_OK) goto Error;
+
+  // VP8X.
+  if (IsVP8XNeeded(enc)) {
+    err = PutVP8XHeader(enc);
+    if (err != VP8_ENC_OK) goto Error;
+  }
+
+  // Alpha.
+  if (enc->has_alpha_) {
+    err = PutAlphaChunk(enc);
+    if (err != VP8_ENC_OK) goto Error;
+  }
+
+  // VP8 header.
+  err = PutVP8Header(pic, vp8_size);
+  if (err != VP8_ENC_OK) goto Error;
+
+  // VP8 frame header.
+  err = PutVP8FrameHeader(pic, enc->profile_, size0);
+  if (err != VP8_ENC_OK) goto Error;
+
+  // All OK.
+  return 1;
+
+  // Error.
+ Error:
+  return WebPEncodingSetError(pic, err);
+}
+
+// Segmentation header
+static void PutSegmentHeader(VP8BitWriter* const bw,
+                             const VP8Encoder* const enc) {
+  const VP8SegmentHeader* const hdr = &enc->segment_hdr_;
+  const VP8Proba* const proba = &enc->proba_;
+  if (VP8PutBitUniform(bw, (hdr->num_segments_ > 1))) {
+    // We always 'update' the quant and filter strength values
+    const int update_data = 1;
+    int s;
+    VP8PutBitUniform(bw, hdr->update_map_);
+    if (VP8PutBitUniform(bw, update_data)) {
+      // we always use absolute values, not relative ones
+      VP8PutBitUniform(bw, 1);   // (segment_feature_mode = 1. Paragraph 9.3.)
+      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+        VP8PutSignedValue(bw, enc->dqm_[s].quant_, 7);
+      }
+      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+        VP8PutSignedValue(bw, enc->dqm_[s].fstrength_, 6);
+      }
+    }
+    if (hdr->update_map_) {
+      for (s = 0; s < 3; ++s) {
+        if (VP8PutBitUniform(bw, (proba->segments_[s] != 255u))) {
+          VP8PutValue(bw, proba->segments_[s], 8);
+        }
+      }
+    }
+  }
+}
+
+// Filtering parameters header
+static void PutFilterHeader(VP8BitWriter* const bw,
+                            const VP8FilterHeader* const hdr) {
+  const int use_lf_delta = (hdr->i4x4_lf_delta_ != 0);
+  VP8PutBitUniform(bw, hdr->simple_);
+  VP8PutValue(bw, hdr->level_, 6);
+  VP8PutValue(bw, hdr->sharpness_, 3);
+  if (VP8PutBitUniform(bw, use_lf_delta)) {
+    // '0' is the default value for i4x4_lf_delta_ at frame #0.
+    const int need_update = (hdr->i4x4_lf_delta_ != 0);
+    if (VP8PutBitUniform(bw, need_update)) {
+      // we don't use ref_lf_delta => emit four 0 bits
+      VP8PutValue(bw, 0, 4);
+      // we use mode_lf_delta for i4x4
+      VP8PutSignedValue(bw, hdr->i4x4_lf_delta_, 6);
+      VP8PutValue(bw, 0, 3);    // all others unused
+    }
+  }
+}
+
+// Nominal quantization parameters
+static void PutQuant(VP8BitWriter* const bw,
+                     const VP8Encoder* const enc) {
+  VP8PutValue(bw, enc->base_quant_, 7);
+  VP8PutSignedValue(bw, enc->dq_y1_dc_, 4);
+  VP8PutSignedValue(bw, enc->dq_y2_dc_, 4);
+  VP8PutSignedValue(bw, enc->dq_y2_ac_, 4);
+  VP8PutSignedValue(bw, enc->dq_uv_dc_, 4);
+  VP8PutSignedValue(bw, enc->dq_uv_ac_, 4);
+}
+
+// Partition sizes
+static int EmitPartitionsSize(const VP8Encoder* const enc,
+                              WebPPicture* const pic) {
+  uint8_t buf[3 * (MAX_NUM_PARTITIONS - 1)];
+  int p;
+  for (p = 0; p < enc->num_parts_ - 1; ++p) {
+    const size_t part_size = VP8BitWriterSize(enc->parts_ + p);
+    if (part_size >= VP8_MAX_PARTITION_SIZE) {
+      return WebPEncodingSetError(pic, VP8_ENC_ERROR_PARTITION_OVERFLOW);
+    }
+    buf[3 * p + 0] = (part_size >>  0) & 0xff;
+    buf[3 * p + 1] = (part_size >>  8) & 0xff;
+    buf[3 * p + 2] = (part_size >> 16) & 0xff;
+  }
+  return p ? pic->writer(buf, 3 * p, pic) : 1;
+}
+
+//------------------------------------------------------------------------------
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+
+#define KTRAILER_SIZE 8
+
+static int WriteExtensions(VP8Encoder* const enc) {
+  uint8_t buffer[KTRAILER_SIZE];
+  VP8BitWriter* const bw = &enc->bw_;
+  WebPPicture* const pic = enc->pic_;
+
+  // Layer (bytes 0..3)
+  PutLE24(buffer + 0, enc->layer_data_size_);
+  buffer[3] = enc->pic_->colorspace & WEBP_CSP_UV_MASK;
+  if (enc->layer_data_size_ > 0) {
+    assert(enc->use_layer_);
+    // append layer data to last partition
+    if (!VP8BitWriterAppend(&enc->parts_[enc->num_parts_ - 1],
+                            enc->layer_data_, enc->layer_data_size_)) {
+      return WebPEncodingSetError(pic, VP8_ENC_ERROR_BITSTREAM_OUT_OF_MEMORY);
+    }
+  }
+
+  buffer[KTRAILER_SIZE - 1] = 0x01;  // marker
+  if (!VP8BitWriterAppend(bw, buffer, KTRAILER_SIZE)) {
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_BITSTREAM_OUT_OF_MEMORY);
+  }
+  return 1;
+}
+
+#endif    /* WEBP_EXPERIMENTAL_FEATURES */
+
+//------------------------------------------------------------------------------
+
+static size_t GeneratePartition0(VP8Encoder* const enc) {
+  VP8BitWriter* const bw = &enc->bw_;
+  const int mb_size = enc->mb_w_ * enc->mb_h_;
+  uint64_t pos1, pos2, pos3;
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  const int need_extensions = enc->use_layer_;
+#endif
+
+  pos1 = VP8BitWriterPos(bw);
+  VP8BitWriterInit(bw, mb_size * 7 / 8);        // ~7 bits per macroblock
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  VP8PutBitUniform(bw, need_extensions);   // extensions
+#else
+  VP8PutBitUniform(bw, 0);   // colorspace
+#endif
+  VP8PutBitUniform(bw, 0);   // clamp type
+
+  PutSegmentHeader(bw, enc);
+  PutFilterHeader(bw, &enc->filter_hdr_);
+  VP8PutValue(bw, enc->num_parts_ == 8 ? 3 :
+                  enc->num_parts_ == 4 ? 2 :
+                  enc->num_parts_ == 2 ? 1 : 0, 2);
+  PutQuant(bw, enc);
+  VP8PutBitUniform(bw, 0);   // no proba update
+  VP8WriteProbas(bw, &enc->proba_);
+  pos2 = VP8BitWriterPos(bw);
+  VP8CodeIntraModes(enc);
+  VP8BitWriterFinish(bw);
+
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  if (need_extensions && !WriteExtensions(enc)) {
+    return 0;
+  }
+#endif
+
+  pos3 = VP8BitWriterPos(bw);
+
+  if (enc->pic_->stats) {
+    enc->pic_->stats->header_bytes[0] = (int)((pos2 - pos1 + 7) >> 3);
+    enc->pic_->stats->header_bytes[1] = (int)((pos3 - pos2 + 7) >> 3);
+    enc->pic_->stats->alpha_data_size = (int)enc->alpha_data_size_;
+    enc->pic_->stats->layer_data_size = (int)enc->layer_data_size_;
+  }
+  return !bw->error_;
+}
+
+void VP8EncFreeBitWriters(VP8Encoder* const enc) {
+  int p;
+  VP8BitWriterWipeOut(&enc->bw_);
+  for (p = 0; p < enc->num_parts_; ++p) {
+    VP8BitWriterWipeOut(enc->parts_ + p);
+  }
+}
+
+int VP8EncWrite(VP8Encoder* const enc) {
+  WebPPicture* const pic = enc->pic_;
+  VP8BitWriter* const bw = &enc->bw_;
+  const int task_percent = 19;
+  const int percent_per_part = task_percent / enc->num_parts_;
+  const int final_percent = enc->percent_ + task_percent;
+  int ok = 0;
+  size_t vp8_size, pad, riff_size;
+  int p;
+
+  // Partition #0 with header and partition sizes
+  ok = !!GeneratePartition0(enc);
+
+  // Compute VP8 size
+  vp8_size = VP8_FRAME_HEADER_SIZE +
+             VP8BitWriterSize(bw) +
+             3 * (enc->num_parts_ - 1);
+  for (p = 0; p < enc->num_parts_; ++p) {
+    vp8_size += VP8BitWriterSize(enc->parts_ + p);
+  }
+  pad = vp8_size & 1;
+  vp8_size += pad;
+
+  // Compute RIFF size
+  // At the minimum it is: "WEBPVP8 nnnn" + VP8 data size.
+  riff_size = TAG_SIZE + CHUNK_HEADER_SIZE + vp8_size;
+  if (IsVP8XNeeded(enc)) {  // Add size for: VP8X header + data.
+    riff_size += CHUNK_HEADER_SIZE + VP8X_CHUNK_SIZE;
+  }
+  if (enc->has_alpha_) {  // Add size for: ALPH header + data.
+    const uint32_t padded_alpha_size = enc->alpha_data_size_ +
+                                       (enc->alpha_data_size_ & 1);
+    riff_size += CHUNK_HEADER_SIZE + padded_alpha_size;
+  }
+  // Sanity check.
+  if (riff_size > 0xfffffffeU) {
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_FILE_TOO_BIG);
+  }
+
+  // Emit headers and partition #0
+  {
+    const uint8_t* const part0 = VP8BitWriterBuf(bw);
+    const size_t size0 = VP8BitWriterSize(bw);
+    ok = ok && PutWebPHeaders(enc, size0, vp8_size, riff_size)
+            && pic->writer(part0, size0, pic)
+            && EmitPartitionsSize(enc, pic);
+    VP8BitWriterWipeOut(bw);    // will free the internal buffer.
+  }
+
+  // Token partitions
+  for (p = 0; p < enc->num_parts_; ++p) {
+    const uint8_t* const buf = VP8BitWriterBuf(enc->parts_ + p);
+    const size_t size = VP8BitWriterSize(enc->parts_ + p);
+    if (size)
+      ok = ok && pic->writer(buf, size, pic);
+    VP8BitWriterWipeOut(enc->parts_ + p);    // will free the internal buffer.
+    ok = ok && WebPReportProgress(pic, enc->percent_ + percent_per_part,
+                                  &enc->percent_);
+  }
+
+  // Padding byte
+  if (ok && pad) {
+    ok = PutPaddingByte(pic);
+  }
+
+  enc->coded_size_ = (int)(CHUNK_HEADER_SIZE + riff_size);
+  ok = ok && WebPReportProgress(pic, final_percent, &enc->percent_);
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/enc/token.c b/drivers/webp/enc/token.c
new file mode 100644
index 0000000000..e696642f16
--- /dev/null
+++ b/drivers/webp/enc/token.c
@@ -0,0 +1,273 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Paginated token buffer
+//
+//  A 'token' is a bit value associated with a probability, either fixed
+// or a later-to-be-determined after statistics have been collected.
+// For dynamic probability, we just record the slot id (idx) for the probability
+// value in the final probability array (uint8_t* probas in VP8EmitTokens).
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./cost.h"
+#include "./vp8enci.h"
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+// we use pages to reduce the number of memcpy()
+#define MAX_NUM_TOKEN 8192          // max number of token per page
+#define FIXED_PROBA_BIT (1u << 14)
+
+struct VP8Tokens {
+  uint16_t tokens_[MAX_NUM_TOKEN];  // bit#15: bit
+                                    // bit #14: constant proba or idx
+                                    // bits 0..13: slot or constant proba
+  VP8Tokens* next_;
+};
+
+//------------------------------------------------------------------------------
+
+void VP8TBufferInit(VP8TBuffer* const b) {
+  b->tokens_ = NULL;
+  b->pages_ = NULL;
+  b->last_page_ = &b->pages_;
+  b->left_ = 0;
+  b->error_ = 0;
+}
+
+void VP8TBufferClear(VP8TBuffer* const b) {
+  if (b != NULL) {
+    const VP8Tokens* p = b->pages_;
+    while (p != NULL) {
+      const VP8Tokens* const next = p->next_;
+      free((void*)p);
+      p = next;
+    }
+    VP8TBufferInit(b);
+  }
+}
+
+static int TBufferNewPage(VP8TBuffer* const b) {
+  VP8Tokens* const page = b->error_ ? NULL : (VP8Tokens*)malloc(sizeof(*page));
+  if (page == NULL) {
+    b->error_ = 1;
+    return 0;
+  }
+  *b->last_page_ = page;
+  b->last_page_ = &page->next_;
+  b->left_ = MAX_NUM_TOKEN;
+  b->tokens_ = page->tokens_;
+  page->next_ = NULL;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+#define TOKEN_ID(t, b, ctx, p) \
+    ((p) + NUM_PROBAS * ((ctx) + NUM_CTX * ((b) + NUM_BANDS * (t))))
+
+static WEBP_INLINE int AddToken(VP8TBuffer* const b,
+                                int bit, uint32_t proba_idx) {
+  assert(proba_idx < FIXED_PROBA_BIT);
+  assert(bit == 0 || bit == 1);
+  if (b->left_ > 0 || TBufferNewPage(b)) {
+    const int slot = --b->left_;
+    b->tokens_[slot] = (bit << 15) | proba_idx;
+  }
+  return bit;
+}
+
+static WEBP_INLINE void AddConstantToken(VP8TBuffer* const b,
+                                         int bit, int proba) {
+  assert(proba < 256);
+  assert(bit == 0 || bit == 1);
+  if (b->left_ > 0 || TBufferNewPage(b)) {
+    const int slot = --b->left_;
+    b->tokens_[slot] = (bit << 15) | FIXED_PROBA_BIT | proba;
+  }
+}
+
+int VP8RecordCoeffTokens(int ctx, int coeff_type, int first, int last,
+                         const int16_t* const coeffs,
+                         VP8TBuffer* const tokens) {
+  int n = first;
+  uint32_t base_id = TOKEN_ID(coeff_type, n, ctx, 0);
+  if (!AddToken(tokens, last >= 0, base_id + 0)) {
+    return 0;
+  }
+
+  while (n < 16) {
+    const int c = coeffs[n++];
+    const int sign = c < 0;
+    int v = sign ? -c : c;
+    if (!AddToken(tokens, v != 0, base_id + 1)) {
+      ctx = 0;
+      base_id = TOKEN_ID(coeff_type, VP8EncBands[n], ctx, 0);
+      continue;
+    }
+    if (!AddToken(tokens, v > 1, base_id + 2)) {
+      ctx = 1;
+    } else {
+      if (!AddToken(tokens, v > 4, base_id + 3)) {
+        if (AddToken(tokens, v != 2, base_id + 4))
+          AddToken(tokens, v == 4, base_id + 5);
+      } else if (!AddToken(tokens, v > 10, base_id + 6)) {
+        if (!AddToken(tokens, v > 6, base_id + 7)) {
+          AddConstantToken(tokens, v == 6, 159);
+        } else {
+          AddConstantToken(tokens, v >= 9, 165);
+          AddConstantToken(tokens, !(v & 1), 145);
+        }
+      } else {
+        int mask;
+        const uint8_t* tab;
+        if (v < 3 + (8 << 1)) {          // VP8Cat3  (3b)
+          AddToken(tokens, 0, base_id + 8);
+          AddToken(tokens, 0, base_id + 9);
+          v -= 3 + (8 << 0);
+          mask = 1 << 2;
+          tab = VP8Cat3;
+        } else if (v < 3 + (8 << 2)) {   // VP8Cat4  (4b)
+          AddToken(tokens, 0, base_id + 8);
+          AddToken(tokens, 1, base_id + 9);
+          v -= 3 + (8 << 1);
+          mask = 1 << 3;
+          tab = VP8Cat4;
+        } else if (v < 3 + (8 << 3)) {   // VP8Cat5  (5b)
+          AddToken(tokens, 1, base_id + 8);
+          AddToken(tokens, 0, base_id + 10);
+          v -= 3 + (8 << 2);
+          mask = 1 << 4;
+          tab = VP8Cat5;
+        } else {                         // VP8Cat6 (11b)
+          AddToken(tokens, 1, base_id + 8);
+          AddToken(tokens, 1, base_id + 10);
+          v -= 3 + (8 << 3);
+          mask = 1 << 10;
+          tab = VP8Cat6;
+        }
+        while (mask) {
+          AddConstantToken(tokens, !!(v & mask), *tab++);
+          mask >>= 1;
+        }
+      }
+      ctx = 2;
+    }
+    AddConstantToken(tokens, sign, 128);
+    base_id = TOKEN_ID(coeff_type, VP8EncBands[n], ctx, 0);
+    if (n == 16 || !AddToken(tokens, n <= last, base_id + 0)) {
+      return 1;   // EOB
+    }
+  }
+  return 1;
+}
+
+#undef TOKEN_ID
+
+//------------------------------------------------------------------------------
+// This function works, but isn't currently used. Saved for later.
+
+#if 0
+
+static void Record(int bit, proba_t* const stats) {
+  proba_t p = *stats;
+  if (p >= 0xffff0000u) {               // an overflow is inbound.
+    p = ((p + 1u) >> 1) & 0x7fff7fffu;  // -> divide the stats by 2.
+  }
+  // record bit count (lower 16 bits) and increment total count (upper 16 bits).
+  p += 0x00010000u + bit;
+  *stats = p;
+}
+
+void VP8TokenToStats(const VP8TBuffer* const b, proba_t* const stats) {
+  const VP8Tokens* p = b->pages_;
+  while (p != NULL) {
+    const int N = (p->next_ == NULL) ? b->left_ : 0;
+    int n = MAX_NUM_TOKEN;
+    while (n-- > N) {
+      const uint16_t token = p->tokens_[n];
+      if (!(token & FIXED_PROBA_BIT)) {
+        Record((token >> 15) & 1, stats + (token & 0x3fffu));
+      }
+    }
+    p = p->next_;
+  }
+}
+
+#endif   // 0
+
+//------------------------------------------------------------------------------
+// Final coding pass, with known probabilities
+
+int VP8EmitTokens(VP8TBuffer* const b, VP8BitWriter* const bw,
+                  const uint8_t* const probas, int final_pass) {
+  const VP8Tokens* p = b->pages_;
+  (void)final_pass;
+  if (b->error_) return 0;
+  while (p != NULL) {
+    const VP8Tokens* const next = p->next_;
+    const int N = (next == NULL) ? b->left_ : 0;
+    int n = MAX_NUM_TOKEN;
+    while (n-- > N) {
+      const uint16_t token = p->tokens_[n];
+      const int bit = (token >> 15) & 1;
+      if (token & FIXED_PROBA_BIT) {
+        VP8PutBit(bw, bit, token & 0xffu);  // constant proba
+      } else {
+        VP8PutBit(bw, bit, probas[token & 0x3fffu]);
+      }
+    }
+    if (final_pass) free((void*)p);
+    p = next;
+  }
+  if (final_pass) b->pages_ = NULL;
+  return 1;
+}
+
+// Size estimation
+size_t VP8EstimateTokenSize(VP8TBuffer* const b, const uint8_t* const probas) {
+  size_t size = 0;
+  const VP8Tokens* p = b->pages_;
+  if (b->error_) return 0;
+  while (p != NULL) {
+    const VP8Tokens* const next = p->next_;
+    const int N = (next == NULL) ? b->left_ : 0;
+    int n = MAX_NUM_TOKEN;
+    while (n-- > N) {
+      const uint16_t token = p->tokens_[n];
+      const int bit = token & (1 << 15);
+      if (token & FIXED_PROBA_BIT) {
+        size += VP8BitCost(bit, token & 0xffu);
+      } else {
+        size += VP8BitCost(bit, probas[token & 0x3fffu]);
+      }
+    }
+    p = next;
+  }
+  return size;
+}
+
+//------------------------------------------------------------------------------
+
+#else     // DISABLE_TOKEN_BUFFER
+
+void VP8TBufferInit(VP8TBuffer* const b) {
+  (void)b;
+}
+void VP8TBufferClear(VP8TBuffer* const b) {
+  (void)b;
+}
+
+#endif    // !DISABLE_TOKEN_BUFFER
+
diff --git a/drivers/webp/enc/tree.c b/drivers/webp/enc/tree.c
new file mode 100644
index 0000000000..e5d05e5221
--- /dev/null
+++ b/drivers/webp/enc/tree.c
@@ -0,0 +1,504 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Coding of token probabilities, intra modes and segments.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./vp8enci.h"
+
+//------------------------------------------------------------------------------
+// Default probabilities
+
+// Paragraph 13.5
+const uint8_t
+  VP8CoeffsProba0[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS] = {
+  { { { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 253, 136, 254, 255, 228, 219, 128, 128, 128, 128, 128 },
+      { 189, 129, 242, 255, 227, 213, 255, 219, 128, 128, 128 },
+      { 106, 126, 227, 252, 214, 209, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 98, 248, 255, 236, 226, 255, 255, 128, 128, 128 },
+      { 181, 133, 238, 254, 221, 234, 255, 154, 128, 128, 128 },
+      { 78, 134, 202, 247, 198, 180, 255, 219, 128, 128, 128 },
+    },
+    { { 1, 185, 249, 255, 243, 255, 128, 128, 128, 128, 128 },
+      { 184, 150, 247, 255, 236, 224, 128, 128, 128, 128, 128 },
+      { 77, 110, 216, 255, 236, 230, 128, 128, 128, 128, 128 },
+    },
+    { { 1, 101, 251, 255, 241, 255, 128, 128, 128, 128, 128 },
+      { 170, 139, 241, 252, 236, 209, 255, 255, 128, 128, 128 },
+      { 37, 116, 196, 243, 228, 255, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 204, 254, 255, 245, 255, 128, 128, 128, 128, 128 },
+      { 207, 160, 250, 255, 238, 128, 128, 128, 128, 128, 128 },
+      { 102, 103, 231, 255, 211, 171, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 152, 252, 255, 240, 255, 128, 128, 128, 128, 128 },
+      { 177, 135, 243, 255, 234, 225, 128, 128, 128, 128, 128 },
+      { 80, 129, 211, 255, 194, 224, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 246, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 255, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 198, 35, 237, 223, 193, 187, 162, 160, 145, 155, 62 },
+      { 131, 45, 198, 221, 172, 176, 220, 157, 252, 221, 1 },
+      { 68, 47, 146, 208, 149, 167, 221, 162, 255, 223, 128 }
+    },
+    { { 1, 149, 241, 255, 221, 224, 255, 255, 128, 128, 128 },
+      { 184, 141, 234, 253, 222, 220, 255, 199, 128, 128, 128 },
+      { 81, 99, 181, 242, 176, 190, 249, 202, 255, 255, 128 }
+    },
+    { { 1, 129, 232, 253, 214, 197, 242, 196, 255, 255, 128 },
+      { 99, 121, 210, 250, 201, 198, 255, 202, 128, 128, 128 },
+      { 23, 91, 163, 242, 170, 187, 247, 210, 255, 255, 128 }
+    },
+    { { 1, 200, 246, 255, 234, 255, 128, 128, 128, 128, 128 },
+      { 109, 178, 241, 255, 231, 245, 255, 255, 128, 128, 128 },
+      { 44, 130, 201, 253, 205, 192, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 132, 239, 251, 219, 209, 255, 165, 128, 128, 128 },
+      { 94, 136, 225, 251, 218, 190, 255, 255, 128, 128, 128 },
+      { 22, 100, 174, 245, 186, 161, 255, 199, 128, 128, 128 }
+    },
+    { { 1, 182, 249, 255, 232, 235, 128, 128, 128, 128, 128 },
+      { 124, 143, 241, 255, 227, 234, 128, 128, 128, 128, 128 },
+      { 35, 77, 181, 251, 193, 211, 255, 205, 128, 128, 128 }
+    },
+    { { 1, 157, 247, 255, 236, 231, 255, 255, 128, 128, 128 },
+      { 121, 141, 235, 255, 225, 227, 255, 255, 128, 128, 128 },
+      { 45, 99, 188, 251, 195, 217, 255, 224, 128, 128, 128 }
+    },
+    { { 1, 1, 251, 255, 213, 255, 128, 128, 128, 128, 128 },
+      { 203, 1, 248, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 137, 1, 177, 255, 224, 255, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 253, 9, 248, 251, 207, 208, 255, 192, 128, 128, 128 },
+      { 175, 13, 224, 243, 193, 185, 249, 198, 255, 255, 128 },
+      { 73, 17, 171, 221, 161, 179, 236, 167, 255, 234, 128 }
+    },
+    { { 1, 95, 247, 253, 212, 183, 255, 255, 128, 128, 128 },
+      { 239, 90, 244, 250, 211, 209, 255, 255, 128, 128, 128 },
+      { 155, 77, 195, 248, 188, 195, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 24, 239, 251, 218, 219, 255, 205, 128, 128, 128 },
+      { 201, 51, 219, 255, 196, 186, 128, 128, 128, 128, 128 },
+      { 69, 46, 190, 239, 201, 218, 255, 228, 128, 128, 128 }
+    },
+    { { 1, 191, 251, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 223, 165, 249, 255, 213, 255, 128, 128, 128, 128, 128 },
+      { 141, 124, 248, 255, 255, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 16, 248, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 190, 36, 230, 255, 236, 255, 128, 128, 128, 128, 128 },
+      { 149, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 226, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 247, 192, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 240, 128, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 1, 134, 252, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 213, 62, 250, 255, 255, 128, 128, 128, 128, 128, 128 },
+      { 55, 93, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    },
+    { { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  },
+  { { { 202, 24, 213, 235, 186, 191, 220, 160, 240, 175, 255 },
+      { 126, 38, 182, 232, 169, 184, 228, 174, 255, 187, 128 },
+      { 61, 46, 138, 219, 151, 178, 240, 170, 255, 216, 128 }
+    },
+    { { 1, 112, 230, 250, 199, 191, 247, 159, 255, 255, 128 },
+      { 166, 109, 228, 252, 211, 215, 255, 174, 128, 128, 128 },
+      { 39, 77, 162, 232, 172, 180, 245, 178, 255, 255, 128 }
+    },
+    { { 1, 52, 220, 246, 198, 199, 249, 220, 255, 255, 128 },
+      { 124, 74, 191, 243, 183, 193, 250, 221, 255, 255, 128 },
+      { 24, 71, 130, 219, 154, 170, 243, 182, 255, 255, 128 }
+    },
+    { { 1, 182, 225, 249, 219, 240, 255, 224, 128, 128, 128 },
+      { 149, 150, 226, 252, 216, 205, 255, 171, 128, 128, 128 },
+      { 28, 108, 170, 242, 183, 194, 254, 223, 255, 255, 128 }
+    },
+    { { 1, 81, 230, 252, 204, 203, 255, 192, 128, 128, 128 },
+      { 123, 102, 209, 247, 188, 196, 255, 233, 128, 128, 128 },
+      { 20, 95, 153, 243, 164, 173, 255, 203, 128, 128, 128 }
+    },
+    { { 1, 222, 248, 255, 216, 213, 128, 128, 128, 128, 128 },
+      { 168, 175, 246, 252, 235, 205, 255, 255, 128, 128, 128 },
+      { 47, 116, 215, 255, 211, 212, 255, 255, 128, 128, 128 }
+    },
+    { { 1, 121, 236, 253, 212, 214, 255, 255, 128, 128, 128 },
+      { 141, 84, 213, 252, 201, 202, 255, 219, 128, 128, 128 },
+      { 42, 80, 160, 240, 162, 185, 255, 205, 128, 128, 128 }
+    },
+    { { 1, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 244, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 238, 1, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+    }
+  }
+};
+
+void VP8DefaultProbas(VP8Encoder* const enc) {
+  VP8Proba* const probas = &enc->proba_;
+  probas->use_skip_proba_ = 0;
+  memset(probas->segments_, 255u, sizeof(probas->segments_));
+  memcpy(probas->coeffs_, VP8CoeffsProba0, sizeof(VP8CoeffsProba0));
+  // Note: we could hard-code the level_costs_ corresponding to VP8CoeffsProba0,
+  // but that's ~11k of static data. Better call VP8CalculateLevelCosts() later.
+  probas->dirty_ = 1;
+}
+
+// Paragraph 11.5.  900bytes.
+static const uint8_t kBModesProba[NUM_BMODES][NUM_BMODES][NUM_BMODES - 1] = {
+  { { 231, 120, 48, 89, 115, 113, 120, 152, 112 },
+    { 152, 179, 64, 126, 170, 118, 46, 70, 95 },
+    { 175, 69, 143, 80, 85, 82, 72, 155, 103 },
+    { 56, 58, 10, 171, 218, 189, 17, 13, 152 },
+    { 114, 26, 17, 163, 44, 195, 21, 10, 173 },
+    { 121, 24, 80, 195, 26, 62, 44, 64, 85 },
+    { 144, 71, 10, 38, 171, 213, 144, 34, 26 },
+    { 170, 46, 55, 19, 136, 160, 33, 206, 71 },
+    { 63, 20, 8, 114, 114, 208, 12, 9, 226 },
+    { 81, 40, 11, 96, 182, 84, 29, 16, 36 } },
+  { { 134, 183, 89, 137, 98, 101, 106, 165, 148 },
+    { 72, 187, 100, 130, 157, 111, 32, 75, 80 },
+    { 66, 102, 167, 99, 74, 62, 40, 234, 128 },
+    { 41, 53, 9, 178, 241, 141, 26, 8, 107 },
+    { 74, 43, 26, 146, 73, 166, 49, 23, 157 },
+    { 65, 38, 105, 160, 51, 52, 31, 115, 128 },
+    { 104, 79, 12, 27, 217, 255, 87, 17, 7 },
+    { 87, 68, 71, 44, 114, 51, 15, 186, 23 },
+    { 47, 41, 14, 110, 182, 183, 21, 17, 194 },
+    { 66, 45, 25, 102, 197, 189, 23, 18, 22 } },
+  { { 88, 88, 147, 150, 42, 46, 45, 196, 205 },
+    { 43, 97, 183, 117, 85, 38, 35, 179, 61 },
+    { 39, 53, 200, 87, 26, 21, 43, 232, 171 },
+    { 56, 34, 51, 104, 114, 102, 29, 93, 77 },
+    { 39, 28, 85, 171, 58, 165, 90, 98, 64 },
+    { 34, 22, 116, 206, 23, 34, 43, 166, 73 },
+    { 107, 54, 32, 26, 51, 1, 81, 43, 31 },
+    { 68, 25, 106, 22, 64, 171, 36, 225, 114 },
+    { 34, 19, 21, 102, 132, 188, 16, 76, 124 },
+    { 62, 18, 78, 95, 85, 57, 50, 48, 51 } },
+  { { 193, 101, 35, 159, 215, 111, 89, 46, 111 },
+    { 60, 148, 31, 172, 219, 228, 21, 18, 111 },
+    { 112, 113, 77, 85, 179, 255, 38, 120, 114 },
+    { 40, 42, 1, 196, 245, 209, 10, 25, 109 },
+    { 88, 43, 29, 140, 166, 213, 37, 43, 154 },
+    { 61, 63, 30, 155, 67, 45, 68, 1, 209 },
+    { 100, 80, 8, 43, 154, 1, 51, 26, 71 },
+    { 142, 78, 78, 16, 255, 128, 34, 197, 171 },
+    { 41, 40, 5, 102, 211, 183, 4, 1, 221 },
+    { 51, 50, 17, 168, 209, 192, 23, 25, 82 } },
+  { { 138, 31, 36, 171, 27, 166, 38, 44, 229 },
+    { 67, 87, 58, 169, 82, 115, 26, 59, 179 },
+    { 63, 59, 90, 180, 59, 166, 93, 73, 154 },
+    { 40, 40, 21, 116, 143, 209, 34, 39, 175 },
+    { 47, 15, 16, 183, 34, 223, 49, 45, 183 },
+    { 46, 17, 33, 183, 6, 98, 15, 32, 183 },
+    { 57, 46, 22, 24, 128, 1, 54, 17, 37 },
+    { 65, 32, 73, 115, 28, 128, 23, 128, 205 },
+    { 40, 3, 9, 115, 51, 192, 18, 6, 223 },
+    { 87, 37, 9, 115, 59, 77, 64, 21, 47 } },
+  { { 104, 55, 44, 218, 9, 54, 53, 130, 226 },
+    { 64, 90, 70, 205, 40, 41, 23, 26, 57 },
+    { 54, 57, 112, 184, 5, 41, 38, 166, 213 },
+    { 30, 34, 26, 133, 152, 116, 10, 32, 134 },
+    { 39, 19, 53, 221, 26, 114, 32, 73, 255 },
+    { 31, 9, 65, 234, 2, 15, 1, 118, 73 },
+    { 75, 32, 12, 51, 192, 255, 160, 43, 51 },
+    { 88, 31, 35, 67, 102, 85, 55, 186, 85 },
+    { 56, 21, 23, 111, 59, 205, 45, 37, 192 },
+    { 55, 38, 70, 124, 73, 102, 1, 34, 98 } },
+  { { 125, 98, 42, 88, 104, 85, 117, 175, 82 },
+    { 95, 84, 53, 89, 128, 100, 113, 101, 45 },
+    { 75, 79, 123, 47, 51, 128, 81, 171, 1 },
+    { 57, 17, 5, 71, 102, 57, 53, 41, 49 },
+    { 38, 33, 13, 121, 57, 73, 26, 1, 85 },
+    { 41, 10, 67, 138, 77, 110, 90, 47, 114 },
+    { 115, 21, 2, 10, 102, 255, 166, 23, 6 },
+    { 101, 29, 16, 10, 85, 128, 101, 196, 26 },
+    { 57, 18, 10, 102, 102, 213, 34, 20, 43 },
+    { 117, 20, 15, 36, 163, 128, 68, 1, 26 } },
+  { { 102, 61, 71, 37, 34, 53, 31, 243, 192 },
+    { 69, 60, 71, 38, 73, 119, 28, 222, 37 },
+    { 68, 45, 128, 34, 1, 47, 11, 245, 171 },
+    { 62, 17, 19, 70, 146, 85, 55, 62, 70 },
+    { 37, 43, 37, 154, 100, 163, 85, 160, 1 },
+    { 63, 9, 92, 136, 28, 64, 32, 201, 85 },
+    { 75, 15, 9, 9, 64, 255, 184, 119, 16 },
+    { 86, 6, 28, 5, 64, 255, 25, 248, 1 },
+    { 56, 8, 17, 132, 137, 255, 55, 116, 128 },
+    { 58, 15, 20, 82, 135, 57, 26, 121, 40 } },
+  { { 164, 50, 31, 137, 154, 133, 25, 35, 218 },
+    { 51, 103, 44, 131, 131, 123, 31, 6, 158 },
+    { 86, 40, 64, 135, 148, 224, 45, 183, 128 },
+    { 22, 26, 17, 131, 240, 154, 14, 1, 209 },
+    { 45, 16, 21, 91, 64, 222, 7, 1, 197 },
+    { 56, 21, 39, 155, 60, 138, 23, 102, 213 },
+    { 83, 12, 13, 54, 192, 255, 68, 47, 28 },
+    { 85, 26, 85, 85, 128, 128, 32, 146, 171 },
+    { 18, 11, 7, 63, 144, 171, 4, 4, 246 },
+    { 35, 27, 10, 146, 174, 171, 12, 26, 128 } },
+  { { 190, 80, 35, 99, 180, 80, 126, 54, 45 },
+    { 85, 126, 47, 87, 176, 51, 41, 20, 32 },
+    { 101, 75, 128, 139, 118, 146, 116, 128, 85 },
+    { 56, 41, 15, 176, 236, 85, 37, 9, 62 },
+    { 71, 30, 17, 119, 118, 255, 17, 18, 138 },
+    { 101, 38, 60, 138, 55, 70, 43, 26, 142 },
+    { 146, 36, 19, 30, 171, 255, 97, 27, 20 },
+    { 138, 45, 61, 62, 219, 1, 81, 188, 64 },
+    { 32, 41, 20, 117, 151, 142, 20, 21, 163 },
+    { 112, 19, 12, 61, 195, 128, 48, 4, 24 } }
+};
+
+static int PutI4Mode(VP8BitWriter* const bw, int mode,
+                     const uint8_t* const prob) {
+  if (VP8PutBit(bw, mode != B_DC_PRED, prob[0])) {
+    if (VP8PutBit(bw, mode != B_TM_PRED, prob[1])) {
+      if (VP8PutBit(bw, mode != B_VE_PRED, prob[2])) {
+        if (!VP8PutBit(bw, mode >= B_LD_PRED, prob[3])) {
+          if (VP8PutBit(bw, mode != B_HE_PRED, prob[4])) {
+            VP8PutBit(bw, mode != B_RD_PRED, prob[5]);
+          }
+        } else {
+          if (VP8PutBit(bw, mode != B_LD_PRED, prob[6])) {
+            if (VP8PutBit(bw, mode != B_VL_PRED, prob[7])) {
+              VP8PutBit(bw, mode != B_HD_PRED, prob[8]);
+            }
+          }
+        }
+      }
+    }
+  }
+  return mode;
+}
+
+static void PutI16Mode(VP8BitWriter* const bw, int mode) {
+  if (VP8PutBit(bw, (mode == TM_PRED || mode == H_PRED), 156)) {
+    VP8PutBit(bw, mode == TM_PRED, 128);    // TM or HE
+  } else {
+    VP8PutBit(bw, mode == V_PRED, 163);     // VE or DC
+  }
+}
+
+static void PutUVMode(VP8BitWriter* const bw, int uv_mode) {
+  if (VP8PutBit(bw, uv_mode != DC_PRED, 142)) {
+    if (VP8PutBit(bw, uv_mode != V_PRED, 114)) {
+      VP8PutBit(bw, uv_mode != H_PRED, 183);    // else: TM_PRED
+    }
+  }
+}
+
+static void PutSegment(VP8BitWriter* const bw, int s, const uint8_t* p) {
+  if (VP8PutBit(bw, s >= 2, p[0])) p += 1;
+  VP8PutBit(bw, s & 1, p[1]);
+}
+
+void VP8CodeIntraModes(VP8Encoder* const enc) {
+  VP8BitWriter* const bw = &enc->bw_;
+  VP8EncIterator it;
+  VP8IteratorInit(enc, &it);
+  do {
+    const VP8MBInfo* const mb = it.mb_;
+    const uint8_t* preds = it.preds_;
+    if (enc->segment_hdr_.update_map_) {
+      PutSegment(bw, mb->segment_, enc->proba_.segments_);
+    }
+    if (enc->proba_.use_skip_proba_) {
+      VP8PutBit(bw, mb->skip_, enc->proba_.skip_proba_);
+    }
+    if (VP8PutBit(bw, (mb->type_ != 0), 145)) {  // i16x16
+      PutI16Mode(bw, preds[0]);
+    } else {
+      const int preds_w = enc->preds_w_;
+      const uint8_t* top_pred = preds - preds_w;
+      int x, y;
+      for (y = 0; y < 4; ++y) {
+        int left = preds[-1];
+        for (x = 0; x < 4; ++x) {
+          const uint8_t* const probas = kBModesProba[top_pred[x]][left];
+          left = PutI4Mode(bw, preds[x], probas);
+        }
+        top_pred = preds;
+        preds += preds_w;
+      }
+    }
+    PutUVMode(bw, mb->uv_mode_);
+  } while (VP8IteratorNext(&it));
+}
+
+//------------------------------------------------------------------------------
+// Paragraph 13
+
+const uint8_t
+    VP8CoeffsUpdateProba[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS] = {
+  { { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 176, 246, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 223, 241, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 244, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 234, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 246, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 239, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 253, 255, 254, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 254, 255, 254, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 217, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 225, 252, 241, 253, 255, 255, 254, 255, 255, 255, 255 },
+      { 234, 250, 241, 250, 253, 255, 253, 254, 255, 255, 255 }
+    },
+    { { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 223, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 238, 253, 254, 254, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 247, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 186, 251, 250, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 234, 251, 244, 254, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 251, 243, 253, 254, 255, 254, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 236, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 251, 253, 253, 254, 254, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  },
+  { { { 248, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 254, 252, 254, 255, 255, 255, 255, 255, 255, 255 },
+      { 248, 254, 249, 253, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 246, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 254, 251, 254, 254, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 254, 252, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 248, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 255, 254, 254, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 245, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 253, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 251, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 252, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 252, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 249, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 253, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    },
+    { { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 },
+      { 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 }
+    }
+  }
+};
+
+void VP8WriteProbas(VP8BitWriter* const bw, const VP8Proba* const probas) {
+  int t, b, c, p;
+  for (t = 0; t < NUM_TYPES; ++t) {
+    for (b = 0; b < NUM_BANDS; ++b) {
+      for (c = 0; c < NUM_CTX; ++c) {
+        for (p = 0; p < NUM_PROBAS; ++p) {
+          const uint8_t p0 = probas->coeffs_[t][b][c][p];
+          const int update = (p0 != VP8CoeffsProba0[t][b][c][p]);
+          if (VP8PutBit(bw, update, VP8CoeffsUpdateProba[t][b][c][p])) {
+            VP8PutValue(bw, p0, 8);
+          }
+        }
+      }
+    }
+  }
+  if (VP8PutBitUniform(bw, probas->use_skip_proba_)) {
+    VP8PutValue(bw, probas->skip_proba_, 8);
+  }
+}
+
diff --git a/drivers/webp/enc/vp8enci.h b/drivers/webp/enc/vp8enci.h
new file mode 100644
index 0000000000..71adf6c38a
--- /dev/null
+++ b/drivers/webp/enc/vp8enci.h
@@ -0,0 +1,570 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//   WebP encoder: internal header.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_ENC_VP8ENCI_H_
+#define WEBP_ENC_VP8ENCI_H_
+
+#include <string.h>     // for memcpy()
+#include "../webp/encode.h"
+#include "../dsp/dsp.h"
+#include "../utils/bit_writer.h"
+#include "../utils/thread.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// Various defines and enums
+
+// version numbers
+#define ENC_MAJ_VERSION 0
+#define ENC_MIN_VERSION 4
+#define ENC_REV_VERSION 0
+
+// intra prediction modes
+enum { B_DC_PRED = 0,   // 4x4 modes
+       B_TM_PRED = 1,
+       B_VE_PRED = 2,
+       B_HE_PRED = 3,
+       B_RD_PRED = 4,
+       B_VR_PRED = 5,
+       B_LD_PRED = 6,
+       B_VL_PRED = 7,
+       B_HD_PRED = 8,
+       B_HU_PRED = 9,
+       NUM_BMODES = B_HU_PRED + 1 - B_DC_PRED,  // = 10
+
+       // Luma16 or UV modes
+       DC_PRED = B_DC_PRED, V_PRED = B_VE_PRED,
+       H_PRED = B_HE_PRED, TM_PRED = B_TM_PRED,
+       NUM_PRED_MODES = 4
+     };
+
+enum { NUM_MB_SEGMENTS = 4,
+       MAX_NUM_PARTITIONS = 8,
+       NUM_TYPES = 4,   // 0: i16-AC,  1: i16-DC,  2:chroma-AC,  3:i4-AC
+       NUM_BANDS = 8,
+       NUM_CTX = 3,
+       NUM_PROBAS = 11,
+       MAX_LF_LEVELS = 64,       // Maximum loop filter level
+       MAX_VARIABLE_LEVEL = 67,  // last (inclusive) level with variable cost
+       MAX_LEVEL = 2047          // max level (note: max codable is 2047 + 67)
+     };
+
+typedef enum {   // Rate-distortion optimization levels
+  RD_OPT_NONE        = 0,  // no rd-opt
+  RD_OPT_BASIC       = 1,  // basic scoring (no trellis)
+  RD_OPT_TRELLIS     = 2,  // perform trellis-quant on the final decision only
+  RD_OPT_TRELLIS_ALL = 3   // trellis-quant for every scoring (much slower)
+} VP8RDLevel;
+
+// YUV-cache parameters. Cache is 16-pixels wide.
+// The original or reconstructed samples can be accessed using VP8Scan[]
+// The predicted blocks can be accessed using offsets to yuv_p_ and
+// the arrays VP8*ModeOffsets[];
+//         +----+      YUV Samples area. See VP8Scan[] for accessing the blocks.
+//  Y_OFF  |YYYY| <- original samples  ('yuv_in_')
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+//  U_OFF  |UUVV| V_OFF  (=U_OFF + 8)
+//         |UUVV|
+//         +----+
+//  Y_OFF  |YYYY| <- compressed/decoded samples  ('yuv_out_')
+//         |YYYY|    There are two buffers like this ('yuv_out_'/'yuv_out2_')
+//         |YYYY|
+//         |YYYY|
+//  U_OFF  |UUVV| V_OFF
+//         |UUVV|
+//          x2 (for yuv_out2_)
+//         +----+     Prediction area ('yuv_p_', size = PRED_SIZE)
+// I16DC16 |YYYY|  Intra16 predictions (16x16 block each)
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+// I16TM16 |YYYY|
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+// I16VE16 |YYYY|
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+// I16HE16 |YYYY|
+//         |YYYY|
+//         |YYYY|
+//         |YYYY|
+//         +----+  Chroma U/V predictions (16x8 block each)
+// C8DC8   |UUVV|
+//         |UUVV|
+// C8TM8   |UUVV|
+//         |UUVV|
+// C8VE8   |UUVV|
+//         |UUVV|
+// C8HE8   |UUVV|
+//         |UUVV|
+//         +----+  Intra 4x4 predictions (4x4 block each)
+//         |YYYY| I4DC4 I4TM4 I4VE4 I4HE4
+//         |YYYY| I4RD4 I4VR4 I4LD4 I4VL4
+//         |YY..| I4HD4 I4HU4 I4TMP
+//         +----+
+#define BPS       16   // this is the common stride
+#define Y_SIZE   (BPS * 16)
+#define UV_SIZE  (BPS * 8)
+#define YUV_SIZE (Y_SIZE + UV_SIZE)
+#define PRED_SIZE (6 * 16 * BPS + 12 * BPS)
+#define Y_OFF    (0)
+#define U_OFF    (Y_SIZE)
+#define V_OFF    (U_OFF + 8)
+#define ALIGN_CST 15
+#define DO_ALIGN(PTR) ((uintptr_t)((PTR) + ALIGN_CST) & ~ALIGN_CST)
+
+extern const int VP8Scan[16 + 4 + 4];           // in quant.c
+extern const int VP8UVModeOffsets[4];           // in analyze.c
+extern const int VP8I16ModeOffsets[4];
+extern const int VP8I4ModeOffsets[NUM_BMODES];
+
+// Layout of prediction blocks
+// intra 16x16
+#define I16DC16 (0 * 16 * BPS)
+#define I16TM16 (1 * 16 * BPS)
+#define I16VE16 (2 * 16 * BPS)
+#define I16HE16 (3 * 16 * BPS)
+// chroma 8x8, two U/V blocks side by side (hence: 16x8 each)
+#define C8DC8 (4 * 16 * BPS)
+#define C8TM8 (4 * 16 * BPS + 8 * BPS)
+#define C8VE8 (5 * 16 * BPS)
+#define C8HE8 (5 * 16 * BPS + 8 * BPS)
+// intra 4x4
+#define I4DC4 (6 * 16 * BPS +  0)
+#define I4TM4 (6 * 16 * BPS +  4)
+#define I4VE4 (6 * 16 * BPS +  8)
+#define I4HE4 (6 * 16 * BPS + 12)
+#define I4RD4 (6 * 16 * BPS + 4 * BPS +  0)
+#define I4VR4 (6 * 16 * BPS + 4 * BPS +  4)
+#define I4LD4 (6 * 16 * BPS + 4 * BPS +  8)
+#define I4VL4 (6 * 16 * BPS + 4 * BPS + 12)
+#define I4HD4 (6 * 16 * BPS + 8 * BPS +  0)
+#define I4HU4 (6 * 16 * BPS + 8 * BPS +  4)
+#define I4TMP (6 * 16 * BPS + 8 * BPS +  8)
+
+typedef int64_t score_t;     // type used for scores, rate, distortion
+#define MAX_COST ((score_t)0x7fffffffffffffLL)
+
+#define QFIX 17
+#define BIAS(b)  ((b) << (QFIX - 8))
+// Fun fact: this is the _only_ line where we're actually being lossy and
+// discarding bits.
+static WEBP_INLINE int QUANTDIV(int n, int iQ, int B) {
+  return (n * iQ + B) >> QFIX;
+}
+
+// size of histogram used by CollectHistogram.
+#define MAX_COEFF_THRESH   31
+typedef struct VP8Histogram VP8Histogram;
+struct VP8Histogram {
+  // TODO(skal): we only need to store the max_value and last_non_zero actually.
+  int distribution[MAX_COEFF_THRESH + 1];
+};
+
+// Uncomment the following to remove token-buffer code:
+// #define DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+// Headers
+
+typedef uint32_t proba_t;   // 16b + 16b
+typedef uint8_t ProbaArray[NUM_CTX][NUM_PROBAS];
+typedef proba_t StatsArray[NUM_CTX][NUM_PROBAS];
+typedef uint16_t CostArray[NUM_CTX][MAX_VARIABLE_LEVEL + 1];
+typedef double LFStats[NUM_MB_SEGMENTS][MAX_LF_LEVELS];  // filter stats
+
+typedef struct VP8Encoder VP8Encoder;
+
+// segment features
+typedef struct {
+  int num_segments_;      // Actual number of segments. 1 segment only = unused.
+  int update_map_;        // whether to update the segment map or not.
+                          // must be 0 if there's only 1 segment.
+  int size_;              // bit-cost for transmitting the segment map
+} VP8SegmentHeader;
+
+// Struct collecting all frame-persistent probabilities.
+typedef struct {
+  uint8_t segments_[3];     // probabilities for segment tree
+  uint8_t skip_proba_;      // final probability of being skipped.
+  ProbaArray coeffs_[NUM_TYPES][NUM_BANDS];      // 924 bytes
+  StatsArray stats_[NUM_TYPES][NUM_BANDS];       // 4224 bytes
+  CostArray level_cost_[NUM_TYPES][NUM_BANDS];   // 11.4k
+  int dirty_;               // if true, need to call VP8CalculateLevelCosts()
+  int use_skip_proba_;      // Note: we always use skip_proba for now.
+  int nb_skip_;             // number of skipped blocks
+} VP8Proba;
+
+// Filter parameters. Not actually used in the code (we don't perform
+// the in-loop filtering), but filled from user's config
+typedef struct {
+  int simple_;             // filtering type: 0=complex, 1=simple
+  int level_;              // base filter level [0..63]
+  int sharpness_;          // [0..7]
+  int i4x4_lf_delta_;      // delta filter level for i4x4 relative to i16x16
+} VP8FilterHeader;
+
+//------------------------------------------------------------------------------
+// Informations about the macroblocks.
+
+typedef struct {
+  // block type
+  unsigned int type_:2;     // 0=i4x4, 1=i16x16
+  unsigned int uv_mode_:2;
+  unsigned int skip_:1;
+  unsigned int segment_:2;
+  uint8_t alpha_;      // quantization-susceptibility
+} VP8MBInfo;
+
+typedef struct VP8Matrix {
+  uint16_t q_[16];        // quantizer steps
+  uint16_t iq_[16];       // reciprocals, fixed point.
+  uint16_t bias_[16];     // rounding bias
+  uint16_t zthresh_[16];  // value under which a coefficient is zeroed
+  uint16_t sharpen_[16];  // frequency boosters for slight sharpening
+} VP8Matrix;
+
+typedef struct {
+  VP8Matrix y1_, y2_, uv_;  // quantization matrices
+  int alpha_;      // quant-susceptibility, range [-127,127]. Zero is neutral.
+                   // Lower values indicate a lower risk of blurriness.
+  int beta_;       // filter-susceptibility, range [0,255].
+  int quant_;      // final segment quantizer.
+  int fstrength_;  // final in-loop filtering strength
+  int max_edge_;   // max edge delta (for filtering strength)
+  int min_disto_;  // minimum distortion required to trigger filtering record
+  // reactivities
+  int lambda_i16_, lambda_i4_, lambda_uv_;
+  int lambda_mode_, lambda_trellis_, tlambda_;
+  int lambda_trellis_i16_, lambda_trellis_i4_, lambda_trellis_uv_;
+} VP8SegmentInfo;
+
+// Handy transient struct to accumulate score and info during RD-optimization
+// and mode evaluation.
+typedef struct {
+  score_t D, SD;              // Distortion, spectral distortion
+  score_t H, R, score;        // header bits, rate, score.
+  int16_t y_dc_levels[16];    // Quantized levels for luma-DC, luma-AC, chroma.
+  int16_t y_ac_levels[16][16];
+  int16_t uv_levels[4 + 4][16];
+  int mode_i16;               // mode number for intra16 prediction
+  uint8_t modes_i4[16];       // mode numbers for intra4 predictions
+  int mode_uv;                // mode number of chroma prediction
+  uint32_t nz;                // non-zero blocks
+} VP8ModeScore;
+
+// Iterator structure to iterate through macroblocks, pointing to the
+// right neighbouring data (samples, predictions, contexts, ...)
+typedef struct {
+  int x_, y_;                      // current macroblock
+  int y_stride_, uv_stride_;       // respective strides
+  uint8_t*      yuv_in_;           // input samples
+  uint8_t*      yuv_out_;          // output samples
+  uint8_t*      yuv_out2_;         // secondary buffer swapped with yuv_out_.
+  uint8_t*      yuv_p_;            // scratch buffer for prediction
+  VP8Encoder*   enc_;              // back-pointer
+  VP8MBInfo*    mb_;               // current macroblock
+  VP8BitWriter* bw_;               // current bit-writer
+  uint8_t*      preds_;            // intra mode predictors (4x4 blocks)
+  uint32_t*     nz_;               // non-zero pattern
+  uint8_t       i4_boundary_[37];  // 32+5 boundary samples needed by intra4x4
+  uint8_t*      i4_top_;           // pointer to the current top boundary sample
+  int           i4_;               // current intra4x4 mode being tested
+  int           top_nz_[9];        // top-non-zero context.
+  int           left_nz_[9];       // left-non-zero. left_nz[8] is independent.
+  uint64_t      bit_count_[4][3];  // bit counters for coded levels.
+  uint64_t      luma_bits_;        // macroblock bit-cost for luma
+  uint64_t      uv_bits_;          // macroblock bit-cost for chroma
+  LFStats*      lf_stats_;         // filter stats (borrowed from enc_)
+  int           do_trellis_;       // if true, perform extra level optimisation
+  int           count_down_;       // number of mb still to be processed
+  int           count_down0_;      // starting counter value (for progress)
+  int           percent0_;         // saved initial progress percent
+
+  uint8_t* y_left_;    // left luma samples (addressable from index -1 to 15).
+  uint8_t* u_left_;    // left u samples (addressable from index -1 to 7)
+  uint8_t* v_left_;    // left v samples (addressable from index -1 to 7)
+
+  uint8_t* y_top_;     // top luma samples at position 'x_'
+  uint8_t* uv_top_;    // top u/v samples at position 'x_', packed as 16 bytes
+
+  // memory for storing y/u/v_left_ and yuv_in_/out_*
+  uint8_t yuv_left_mem_[17 + 16 + 16 + 8 + ALIGN_CST];     // memory for *_left_
+  uint8_t yuv_mem_[3 * YUV_SIZE + PRED_SIZE + ALIGN_CST];  // memory for yuv_*
+} VP8EncIterator;
+
+  // in iterator.c
+// must be called first
+void VP8IteratorInit(VP8Encoder* const enc, VP8EncIterator* const it);
+// restart a scan
+void VP8IteratorReset(VP8EncIterator* const it);
+// reset iterator position to row 'y'
+void VP8IteratorSetRow(VP8EncIterator* const it, int y);
+// set count down (=number of iterations to go)
+void VP8IteratorSetCountDown(VP8EncIterator* const it, int count_down);
+// return true if iteration is finished
+int VP8IteratorIsDone(const VP8EncIterator* const it);
+// Import uncompressed samples from source.
+// If tmp_32 is not NULL, import boundary samples too.
+// tmp_32 is a 32-bytes scratch buffer that must be aligned in memory.
+void VP8IteratorImport(VP8EncIterator* const it, uint8_t* tmp_32);
+// export decimated samples
+void VP8IteratorExport(const VP8EncIterator* const it);
+// go to next macroblock. Returns false if not finished.
+int VP8IteratorNext(VP8EncIterator* const it);
+// save the yuv_out_ boundary values to top_/left_ arrays for next iterations.
+void VP8IteratorSaveBoundary(VP8EncIterator* const it);
+// Report progression based on macroblock rows. Return 0 for user-abort request.
+int VP8IteratorProgress(const VP8EncIterator* const it,
+                        int final_delta_percent);
+// Intra4x4 iterations
+void VP8IteratorStartI4(VP8EncIterator* const it);
+// returns true if not done.
+int VP8IteratorRotateI4(VP8EncIterator* const it,
+                        const uint8_t* const yuv_out);
+
+// Non-zero context setup/teardown
+void VP8IteratorNzToBytes(VP8EncIterator* const it);
+void VP8IteratorBytesToNz(VP8EncIterator* const it);
+
+// Helper functions to set mode properties
+void VP8SetIntra16Mode(const VP8EncIterator* const it, int mode);
+void VP8SetIntra4Mode(const VP8EncIterator* const it, const uint8_t* modes);
+void VP8SetIntraUVMode(const VP8EncIterator* const it, int mode);
+void VP8SetSkip(const VP8EncIterator* const it, int skip);
+void VP8SetSegment(const VP8EncIterator* const it, int segment);
+
+//------------------------------------------------------------------------------
+// Paginated token buffer
+
+typedef struct VP8Tokens VP8Tokens;  // struct details in token.c
+
+typedef struct {
+#if !defined(DISABLE_TOKEN_BUFFER)
+  VP8Tokens* pages_;        // first page
+  VP8Tokens** last_page_;   // last page
+  uint16_t* tokens_;        // set to (*last_page_)->tokens_
+  int left_;          // how many free tokens left before the page is full.
+#endif
+  int error_;         // true in case of malloc error
+} VP8TBuffer;
+
+void VP8TBufferInit(VP8TBuffer* const b);    // initialize an empty buffer
+void VP8TBufferClear(VP8TBuffer* const b);   // de-allocate pages memory
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+// Finalizes bitstream when probabilities are known.
+// Deletes the allocated token memory if final_pass is true.
+int VP8EmitTokens(VP8TBuffer* const b, VP8BitWriter* const bw,
+                  const uint8_t* const probas, int final_pass);
+
+// record the coding of coefficients without knowing the probabilities yet
+int VP8RecordCoeffTokens(int ctx, int coeff_type, int first, int last,
+                         const int16_t* const coeffs,
+                         VP8TBuffer* const tokens);
+
+// Estimate the final coded size given a set of 'probas'.
+size_t VP8EstimateTokenSize(VP8TBuffer* const b, const uint8_t* const probas);
+
+// unused for now
+void VP8TokenToStats(const VP8TBuffer* const b, proba_t* const stats);
+
+#endif  // !DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
+// VP8Encoder
+
+struct VP8Encoder {
+  const WebPConfig* config_;    // user configuration and parameters
+  WebPPicture* pic_;            // input / output picture
+
+  // headers
+  VP8FilterHeader   filter_hdr_;     // filtering information
+  VP8SegmentHeader  segment_hdr_;    // segment information
+
+  int profile_;                      // VP8's profile, deduced from Config.
+
+  // dimension, in macroblock units.
+  int mb_w_, mb_h_;
+  int preds_w_;   // stride of the *preds_ prediction plane (=4*mb_w + 1)
+
+  // number of partitions (1, 2, 4 or 8 = MAX_NUM_PARTITIONS)
+  int num_parts_;
+
+  // per-partition boolean decoders.
+  VP8BitWriter bw_;                         // part0
+  VP8BitWriter parts_[MAX_NUM_PARTITIONS];  // token partitions
+  VP8TBuffer tokens_;                       // token buffer
+
+  int percent_;                             // for progress
+
+  // transparency blob
+  int has_alpha_;
+  uint8_t* alpha_data_;       // non-NULL if transparency is present
+  uint32_t alpha_data_size_;
+  WebPWorker alpha_worker_;
+
+  // enhancement layer
+  int use_layer_;
+  VP8BitWriter layer_bw_;
+  uint8_t* layer_data_;
+  size_t layer_data_size_;
+
+  // quantization info (one set of DC/AC dequant factor per segment)
+  VP8SegmentInfo dqm_[NUM_MB_SEGMENTS];
+  int base_quant_;                 // nominal quantizer value. Only used
+                                   // for relative coding of segments' quant.
+  int alpha_;                      // global susceptibility (<=> complexity)
+  int uv_alpha_;                   // U/V quantization susceptibility
+  // global offset of quantizers, shared by all segments
+  int dq_y1_dc_;
+  int dq_y2_dc_, dq_y2_ac_;
+  int dq_uv_dc_, dq_uv_ac_;
+
+  // probabilities and statistics
+  VP8Proba proba_;
+  uint64_t sse_[4];        // sum of Y/U/V/A squared errors for all macroblocks
+  uint64_t sse_count_;     // pixel count for the sse_[] stats
+  int      coded_size_;
+  int      residual_bytes_[3][4];
+  int      block_count_[3];
+
+  // quality/speed settings
+  int method_;               // 0=fastest, 6=best/slowest.
+  VP8RDLevel rd_opt_level_;  // Deduced from method_.
+  int max_i4_header_bits_;   // partition #0 safeness factor
+  int thread_level_;         // derived from config->thread_level
+  int do_search_;            // derived from config->target_XXX
+  int use_tokens_;           // if true, use token buffer
+
+  // Memory
+  VP8MBInfo* mb_info_;   // contextual macroblock infos (mb_w_ + 1)
+  uint8_t*   preds_;     // predictions modes: (4*mb_w+1) * (4*mb_h+1)
+  uint32_t*  nz_;        // non-zero bit context: mb_w+1
+  uint8_t   *y_top_;     // top luma samples.
+  uint8_t   *uv_top_;    // top u/v samples.
+                         // U and V are packed into 16 bytes (8 U + 8 V)
+  LFStats   *lf_stats_;  // autofilter stats (if NULL, autofilter is off)
+};
+
+//------------------------------------------------------------------------------
+// internal functions. Not public.
+
+  // in tree.c
+extern const uint8_t VP8CoeffsProba0[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS];
+extern const uint8_t
+    VP8CoeffsUpdateProba[NUM_TYPES][NUM_BANDS][NUM_CTX][NUM_PROBAS];
+// Reset the token probabilities to their initial (default) values
+void VP8DefaultProbas(VP8Encoder* const enc);
+// Write the token probabilities
+void VP8WriteProbas(VP8BitWriter* const bw, const VP8Proba* const probas);
+// Writes the partition #0 modes (that is: all intra modes)
+void VP8CodeIntraModes(VP8Encoder* const enc);
+
+  // in syntax.c
+// Generates the final bitstream by coding the partition0 and headers,
+// and appending an assembly of all the pre-coded token partitions.
+// Return true if everything is ok.
+int VP8EncWrite(VP8Encoder* const enc);
+// Release memory allocated for bit-writing in VP8EncLoop & seq.
+void VP8EncFreeBitWriters(VP8Encoder* const enc);
+
+  // in frame.c
+extern const uint8_t VP8EncBands[16 + 1];
+extern const uint8_t VP8Cat3[];
+extern const uint8_t VP8Cat4[];
+extern const uint8_t VP8Cat5[];
+extern const uint8_t VP8Cat6[];
+
+// Form all the four Intra16x16 predictions in the yuv_p_ cache
+void VP8MakeLuma16Preds(const VP8EncIterator* const it);
+// Form all the four Chroma8x8 predictions in the yuv_p_ cache
+void VP8MakeChroma8Preds(const VP8EncIterator* const it);
+// Form all the ten Intra4x4 predictions in the yuv_p_ cache
+// for the 4x4 block it->i4_
+void VP8MakeIntra4Preds(const VP8EncIterator* const it);
+// Rate calculation
+int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd);
+int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]);
+int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd);
+// Main coding calls
+int VP8EncLoop(VP8Encoder* const enc);
+int VP8EncTokenLoop(VP8Encoder* const enc);
+
+  // in webpenc.c
+// Assign an error code to a picture. Return false for convenience.
+int WebPEncodingSetError(const WebPPicture* const pic, WebPEncodingError error);
+int WebPReportProgress(const WebPPicture* const pic,
+                       int percent, int* const percent_store);
+
+  // in analysis.c
+// Main analysis loop. Decides the segmentations and complexity.
+// Assigns a first guess for Intra16 and uvmode_ prediction modes.
+int VP8EncAnalyze(VP8Encoder* const enc);
+
+  // in quant.c
+// Sets up segment's quantization values, base_quant_ and filter strengths.
+void VP8SetSegmentParams(VP8Encoder* const enc, float quality);
+// Pick best modes and fills the levels. Returns true if skipped.
+int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd,
+                VP8RDLevel rd_opt);
+
+  // in alpha.c
+void VP8EncInitAlpha(VP8Encoder* const enc);    // initialize alpha compression
+int VP8EncStartAlpha(VP8Encoder* const enc);    // start alpha coding process
+int VP8EncFinishAlpha(VP8Encoder* const enc);   // finalize compressed data
+int VP8EncDeleteAlpha(VP8Encoder* const enc);   // delete compressed data
+
+  // in layer.c
+void VP8EncInitLayer(VP8Encoder* const enc);     // init everything
+void VP8EncCodeLayerBlock(VP8EncIterator* it);   // code one more macroblock
+int VP8EncFinishLayer(VP8Encoder* const enc);    // finalize coding
+void VP8EncDeleteLayer(VP8Encoder* enc);         // reclaim memory
+
+  // in filter.c
+
+// SSIM utils
+typedef struct {
+  double w, xm, ym, xxm, xym, yym;
+} DistoStats;
+void VP8SSIMAddStats(const DistoStats* const src, DistoStats* const dst);
+void VP8SSIMAccumulatePlane(const uint8_t* src1, int stride1,
+                            const uint8_t* src2, int stride2,
+                            int W, int H, DistoStats* const stats);
+double VP8SSIMGet(const DistoStats* const stats);
+double VP8SSIMGetSquaredError(const DistoStats* const stats);
+
+// autofilter
+void VP8InitFilter(VP8EncIterator* const it);
+void VP8StoreFilterStats(VP8EncIterator* const it);
+void VP8AdjustFilterStrength(VP8EncIterator* const it);
+
+// returns the approximate filtering strength needed to smooth a edge
+// step of 'delta', given a sharpness parameter 'sharpness'.
+int VP8FilterStrengthFromDelta(int sharpness, int delta);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_ENC_VP8ENCI_H_ */
diff --git a/drivers/webp/enc/vp8l.c b/drivers/webp/enc/vp8l.c
new file mode 100644
index 0000000000..15726318e2
--- /dev/null
+++ b/drivers/webp/enc/vp8l.c
@@ -0,0 +1,1168 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// main entry for the lossless encoder.
+//
+// Author: Vikas Arora (vikaas.arora@gmail.com)
+//
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "./backward_references.h"
+#include "./vp8enci.h"
+#include "./vp8li.h"
+#include "../dsp/lossless.h"
+#include "../utils/bit_writer.h"
+#include "../utils/huffman_encode.h"
+#include "../utils/utils.h"
+#include "../webp/format_constants.h"
+
+#define PALETTE_KEY_RIGHT_SHIFT   22  // Key for 1K buffer.
+#define MAX_HUFF_IMAGE_SIZE       (16 * 1024 * 1024)
+#define MAX_COLORS_FOR_GRAPH      64
+
+// -----------------------------------------------------------------------------
+// Palette
+
+static int CompareColors(const void* p1, const void* p2) {
+  const uint32_t a = *(const uint32_t*)p1;
+  const uint32_t b = *(const uint32_t*)p2;
+  assert(a != b);
+  return (a < b) ? -1 : 1;
+}
+
+// If number of colors in the image is less than or equal to MAX_PALETTE_SIZE,
+// creates a palette and returns true, else returns false.
+static int AnalyzeAndCreatePalette(const WebPPicture* const pic,
+                                   uint32_t palette[MAX_PALETTE_SIZE],
+                                   int* const palette_size) {
+  int i, x, y, key;
+  int num_colors = 0;
+  uint8_t in_use[MAX_PALETTE_SIZE * 4] = { 0 };
+  uint32_t colors[MAX_PALETTE_SIZE * 4];
+  static const uint32_t kHashMul = 0x1e35a7bd;
+  const uint32_t* argb = pic->argb;
+  const int width = pic->width;
+  const int height = pic->height;
+  uint32_t last_pix = ~argb[0];   // so we're sure that last_pix != argb[0]
+
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      if (argb[x] == last_pix) {
+        continue;
+      }
+      last_pix = argb[x];
+      key = (kHashMul * last_pix) >> PALETTE_KEY_RIGHT_SHIFT;
+      while (1) {
+        if (!in_use[key]) {
+          colors[key] = last_pix;
+          in_use[key] = 1;
+          ++num_colors;
+          if (num_colors > MAX_PALETTE_SIZE) {
+            return 0;
+          }
+          break;
+        } else if (colors[key] == last_pix) {
+          // The color is already there.
+          break;
+        } else {
+          // Some other color sits there.
+          // Do linear conflict resolution.
+          ++key;
+          key &= (MAX_PALETTE_SIZE * 4 - 1);  // key mask for 1K buffer.
+        }
+      }
+    }
+    argb += pic->argb_stride;
+  }
+
+  // TODO(skal): could we reuse in_use[] to speed up EncodePalette()?
+  num_colors = 0;
+  for (i = 0; i < (int)(sizeof(in_use) / sizeof(in_use[0])); ++i) {
+    if (in_use[i]) {
+      palette[num_colors] = colors[i];
+      ++num_colors;
+    }
+  }
+
+  qsort(palette, num_colors, sizeof(*palette), CompareColors);
+  *palette_size = num_colors;
+  return 1;
+}
+
+static int AnalyzeEntropy(const uint32_t* argb,
+                          int width, int height, int argb_stride,
+                          double* const nonpredicted_bits,
+                          double* const predicted_bits) {
+  int x, y;
+  const uint32_t* last_line = NULL;
+  uint32_t last_pix = argb[0];    // so we're sure that pix_diff == 0
+
+  VP8LHistogram* nonpredicted = NULL;
+  VP8LHistogram* predicted =
+      (VP8LHistogram*)malloc(2 * sizeof(*predicted));
+  if (predicted == NULL) return 0;
+  nonpredicted = predicted + 1;
+
+  VP8LHistogramInit(predicted, 0);
+  VP8LHistogramInit(nonpredicted, 0);
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      const uint32_t pix = argb[x];
+      const uint32_t pix_diff = VP8LSubPixels(pix, last_pix);
+      if (pix_diff == 0) continue;
+      if (last_line != NULL && pix == last_line[x]) {
+        continue;
+      }
+      last_pix = pix;
+      {
+        const PixOrCopy pix_token = PixOrCopyCreateLiteral(pix);
+        const PixOrCopy pix_diff_token = PixOrCopyCreateLiteral(pix_diff);
+        VP8LHistogramAddSinglePixOrCopy(nonpredicted, &pix_token);
+        VP8LHistogramAddSinglePixOrCopy(predicted, &pix_diff_token);
+      }
+    }
+    last_line = argb;
+    argb += argb_stride;
+  }
+  *nonpredicted_bits = VP8LHistogramEstimateBitsBulk(nonpredicted);
+  *predicted_bits = VP8LHistogramEstimateBitsBulk(predicted);
+  free(predicted);
+  return 1;
+}
+
+static int VP8LEncAnalyze(VP8LEncoder* const enc, WebPImageHint image_hint) {
+  const WebPPicture* const pic = enc->pic_;
+  assert(pic != NULL && pic->argb != NULL);
+
+  enc->use_palette_ =
+      AnalyzeAndCreatePalette(pic, enc->palette_, &enc->palette_size_);
+
+  if (image_hint == WEBP_HINT_GRAPH) {
+    if (enc->use_palette_ && enc->palette_size_ < MAX_COLORS_FOR_GRAPH) {
+      enc->use_palette_ = 0;
+    }
+  }
+
+  if (!enc->use_palette_) {
+    if (image_hint == WEBP_HINT_PHOTO) {
+      enc->use_predict_ = 1;
+      enc->use_cross_color_ = 1;
+    } else {
+      double non_pred_entropy, pred_entropy;
+      if (!AnalyzeEntropy(pic->argb, pic->width, pic->height, pic->argb_stride,
+                          &non_pred_entropy, &pred_entropy)) {
+        return 0;
+      }
+      if (pred_entropy < 0.95 * non_pred_entropy) {
+        enc->use_predict_ = 1;
+        enc->use_cross_color_ = 1;
+      }
+    }
+  }
+
+  return 1;
+}
+
+static int GetHuffBitLengthsAndCodes(
+    const VP8LHistogramSet* const histogram_image,
+    HuffmanTreeCode* const huffman_codes) {
+  int i, k;
+  int ok = 1;
+  uint64_t total_length_size = 0;
+  uint8_t* mem_buf = NULL;
+  const int histogram_image_size = histogram_image->size;
+
+  // Iterate over all histograms and get the aggregate number of codes used.
+  for (i = 0; i < histogram_image_size; ++i) {
+    const VP8LHistogram* const histo = histogram_image->histograms[i];
+    HuffmanTreeCode* const codes = &huffman_codes[5 * i];
+    for (k = 0; k < 5; ++k) {
+      const int num_symbols = (k == 0) ? VP8LHistogramNumCodes(histo)
+                            : (k == 4) ? NUM_DISTANCE_CODES
+                            : 256;
+      codes[k].num_symbols = num_symbols;
+      total_length_size += num_symbols;
+    }
+  }
+
+  // Allocate and Set Huffman codes.
+  {
+    uint16_t* codes;
+    uint8_t* lengths;
+    mem_buf = (uint8_t*)WebPSafeCalloc(total_length_size,
+                                       sizeof(*lengths) + sizeof(*codes));
+    if (mem_buf == NULL) {
+      ok = 0;
+      goto End;
+    }
+    codes = (uint16_t*)mem_buf;
+    lengths = (uint8_t*)&codes[total_length_size];
+    for (i = 0; i < 5 * histogram_image_size; ++i) {
+      const int bit_length = huffman_codes[i].num_symbols;
+      huffman_codes[i].codes = codes;
+      huffman_codes[i].code_lengths = lengths;
+      codes += bit_length;
+      lengths += bit_length;
+    }
+  }
+
+  // Create Huffman trees.
+  for (i = 0; ok && (i < histogram_image_size); ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[5 * i];
+    VP8LHistogram* const histo = histogram_image->histograms[i];
+    ok = ok && VP8LCreateHuffmanTree(histo->literal_, 15, codes + 0);
+    ok = ok && VP8LCreateHuffmanTree(histo->red_, 15, codes + 1);
+    ok = ok && VP8LCreateHuffmanTree(histo->blue_, 15, codes + 2);
+    ok = ok && VP8LCreateHuffmanTree(histo->alpha_, 15, codes + 3);
+    ok = ok && VP8LCreateHuffmanTree(histo->distance_, 15, codes + 4);
+  }
+
+ End:
+  if (!ok) {
+    free(mem_buf);
+    // If one VP8LCreateHuffmanTree() above fails, we need to clean up behind.
+    memset(huffman_codes, 0, 5 * histogram_image_size * sizeof(*huffman_codes));
+  }
+  return ok;
+}
+
+static void StoreHuffmanTreeOfHuffmanTreeToBitMask(
+    VP8LBitWriter* const bw, const uint8_t* code_length_bitdepth) {
+  // RFC 1951 will calm you down if you are worried about this funny sequence.
+  // This sequence is tuned from that, but more weighted for lower symbol count,
+  // and more spiking histograms.
+  static const uint8_t kStorageOrder[CODE_LENGTH_CODES] = {
+    17, 18, 0, 1, 2, 3, 4, 5, 16, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
+  };
+  int i;
+  // Throw away trailing zeros:
+  int codes_to_store = CODE_LENGTH_CODES;
+  for (; codes_to_store > 4; --codes_to_store) {
+    if (code_length_bitdepth[kStorageOrder[codes_to_store - 1]] != 0) {
+      break;
+    }
+  }
+  VP8LWriteBits(bw, 4, codes_to_store - 4);
+  for (i = 0; i < codes_to_store; ++i) {
+    VP8LWriteBits(bw, 3, code_length_bitdepth[kStorageOrder[i]]);
+  }
+}
+
+static void ClearHuffmanTreeIfOnlyOneSymbol(
+    HuffmanTreeCode* const huffman_code) {
+  int k;
+  int count = 0;
+  for (k = 0; k < huffman_code->num_symbols; ++k) {
+    if (huffman_code->code_lengths[k] != 0) {
+      ++count;
+      if (count > 1) return;
+    }
+  }
+  for (k = 0; k < huffman_code->num_symbols; ++k) {
+    huffman_code->code_lengths[k] = 0;
+    huffman_code->codes[k] = 0;
+  }
+}
+
+static void StoreHuffmanTreeToBitMask(
+    VP8LBitWriter* const bw,
+    const HuffmanTreeToken* const tokens, const int num_tokens,
+    const HuffmanTreeCode* const huffman_code) {
+  int i;
+  for (i = 0; i < num_tokens; ++i) {
+    const int ix = tokens[i].code;
+    const int extra_bits = tokens[i].extra_bits;
+    VP8LWriteBits(bw, huffman_code->code_lengths[ix], huffman_code->codes[ix]);
+    switch (ix) {
+      case 16:
+        VP8LWriteBits(bw, 2, extra_bits);
+        break;
+      case 17:
+        VP8LWriteBits(bw, 3, extra_bits);
+        break;
+      case 18:
+        VP8LWriteBits(bw, 7, extra_bits);
+        break;
+    }
+  }
+}
+
+static int StoreFullHuffmanCode(VP8LBitWriter* const bw,
+                                const HuffmanTreeCode* const tree) {
+  int ok = 0;
+  uint8_t code_length_bitdepth[CODE_LENGTH_CODES] = { 0 };
+  uint16_t code_length_bitdepth_symbols[CODE_LENGTH_CODES] = { 0 };
+  const int max_tokens = tree->num_symbols;
+  int num_tokens;
+  HuffmanTreeCode huffman_code;
+  HuffmanTreeToken* const tokens =
+      (HuffmanTreeToken*)WebPSafeMalloc((uint64_t)max_tokens, sizeof(*tokens));
+  if (tokens == NULL) return 0;
+
+  huffman_code.num_symbols = CODE_LENGTH_CODES;
+  huffman_code.code_lengths = code_length_bitdepth;
+  huffman_code.codes = code_length_bitdepth_symbols;
+
+  VP8LWriteBits(bw, 1, 0);
+  num_tokens = VP8LCreateCompressedHuffmanTree(tree, tokens, max_tokens);
+  {
+    int histogram[CODE_LENGTH_CODES] = { 0 };
+    int i;
+    for (i = 0; i < num_tokens; ++i) {
+      ++histogram[tokens[i].code];
+    }
+
+    if (!VP8LCreateHuffmanTree(histogram, 7, &huffman_code)) {
+      goto End;
+    }
+  }
+
+  StoreHuffmanTreeOfHuffmanTreeToBitMask(bw, code_length_bitdepth);
+  ClearHuffmanTreeIfOnlyOneSymbol(&huffman_code);
+  {
+    int trailing_zero_bits = 0;
+    int trimmed_length = num_tokens;
+    int write_trimmed_length;
+    int length;
+    int i = num_tokens;
+    while (i-- > 0) {
+      const int ix = tokens[i].code;
+      if (ix == 0 || ix == 17 || ix == 18) {
+        --trimmed_length;   // discount trailing zeros
+        trailing_zero_bits += code_length_bitdepth[ix];
+        if (ix == 17) {
+          trailing_zero_bits += 3;
+        } else if (ix == 18) {
+          trailing_zero_bits += 7;
+        }
+      } else {
+        break;
+      }
+    }
+    write_trimmed_length = (trimmed_length > 1 && trailing_zero_bits > 12);
+    length = write_trimmed_length ? trimmed_length : num_tokens;
+    VP8LWriteBits(bw, 1, write_trimmed_length);
+    if (write_trimmed_length) {
+      const int nbits = VP8LBitsLog2Ceiling(trimmed_length - 1);
+      const int nbitpairs = (nbits == 0) ? 1 : (nbits + 1) / 2;
+      VP8LWriteBits(bw, 3, nbitpairs - 1);
+      assert(trimmed_length >= 2);
+      VP8LWriteBits(bw, nbitpairs * 2, trimmed_length - 2);
+    }
+    StoreHuffmanTreeToBitMask(bw, tokens, length, &huffman_code);
+  }
+  ok = 1;
+ End:
+  free(tokens);
+  return ok;
+}
+
+static int StoreHuffmanCode(VP8LBitWriter* const bw,
+                            const HuffmanTreeCode* const huffman_code) {
+  int i;
+  int count = 0;
+  int symbols[2] = { 0, 0 };
+  const int kMaxBits = 8;
+  const int kMaxSymbol = 1 << kMaxBits;
+
+  // Check whether it's a small tree.
+  for (i = 0; i < huffman_code->num_symbols && count < 3; ++i) {
+    if (huffman_code->code_lengths[i] != 0) {
+      if (count < 2) symbols[count] = i;
+      ++count;
+    }
+  }
+
+  if (count == 0) {   // emit minimal tree for empty cases
+    // bits: small tree marker: 1, count-1: 0, large 8-bit code: 0, code: 0
+    VP8LWriteBits(bw, 4, 0x01);
+    return 1;
+  } else if (count <= 2 && symbols[0] < kMaxSymbol && symbols[1] < kMaxSymbol) {
+    VP8LWriteBits(bw, 1, 1);  // Small tree marker to encode 1 or 2 symbols.
+    VP8LWriteBits(bw, 1, count - 1);
+    if (symbols[0] <= 1) {
+      VP8LWriteBits(bw, 1, 0);  // Code bit for small (1 bit) symbol value.
+      VP8LWriteBits(bw, 1, symbols[0]);
+    } else {
+      VP8LWriteBits(bw, 1, 1);
+      VP8LWriteBits(bw, 8, symbols[0]);
+    }
+    if (count == 2) {
+      VP8LWriteBits(bw, 8, symbols[1]);
+    }
+    return 1;
+  } else {
+    return StoreFullHuffmanCode(bw, huffman_code);
+  }
+}
+
+static void WriteHuffmanCode(VP8LBitWriter* const bw,
+                             const HuffmanTreeCode* const code,
+                             int code_index) {
+  const int depth = code->code_lengths[code_index];
+  const int symbol = code->codes[code_index];
+  VP8LWriteBits(bw, depth, symbol);
+}
+
+static void StoreImageToBitMask(
+    VP8LBitWriter* const bw, int width, int histo_bits,
+    const VP8LBackwardRefs* const refs,
+    const uint16_t* histogram_symbols,
+    const HuffmanTreeCode* const huffman_codes) {
+  // x and y trace the position in the image.
+  int x = 0;
+  int y = 0;
+  const int histo_xsize = histo_bits ? VP8LSubSampleSize(width, histo_bits) : 1;
+  int i;
+  for (i = 0; i < refs->size; ++i) {
+    const PixOrCopy* const v = &refs->refs[i];
+    const int histogram_ix = histogram_symbols[histo_bits ?
+                                               (y >> histo_bits) * histo_xsize +
+                                               (x >> histo_bits) : 0];
+    const HuffmanTreeCode* const codes = huffman_codes + 5 * histogram_ix;
+    if (PixOrCopyIsCacheIdx(v)) {
+      const int code = PixOrCopyCacheIdx(v);
+      const int literal_ix = 256 + NUM_LENGTH_CODES + code;
+      WriteHuffmanCode(bw, codes, literal_ix);
+    } else if (PixOrCopyIsLiteral(v)) {
+      static const int order[] = { 1, 2, 0, 3 };
+      int k;
+      for (k = 0; k < 4; ++k) {
+        const int code = PixOrCopyLiteral(v, order[k]);
+        WriteHuffmanCode(bw, codes + k, code);
+      }
+    } else {
+      int bits, n_bits;
+      int code, distance;
+
+      VP8LPrefixEncode(v->len, &code, &n_bits, &bits);
+      WriteHuffmanCode(bw, codes, 256 + code);
+      VP8LWriteBits(bw, n_bits, bits);
+
+      distance = PixOrCopyDistance(v);
+      VP8LPrefixEncode(distance, &code, &n_bits, &bits);
+      WriteHuffmanCode(bw, codes + 4, code);
+      VP8LWriteBits(bw, n_bits, bits);
+    }
+    x += PixOrCopyLength(v);
+    while (x >= width) {
+      x -= width;
+      ++y;
+    }
+  }
+}
+
+// Special case of EncodeImageInternal() for cache-bits=0, histo_bits=31
+static int EncodeImageNoHuffman(VP8LBitWriter* const bw,
+                                const uint32_t* const argb,
+                                int width, int height, int quality) {
+  int i;
+  int ok = 0;
+  VP8LBackwardRefs refs;
+  HuffmanTreeCode huffman_codes[5] = { { 0, NULL, NULL } };
+  const uint16_t histogram_symbols[1] = { 0 };    // only one tree, one symbol
+  VP8LHistogramSet* const histogram_image = VP8LAllocateHistogramSet(1, 0);
+  if (histogram_image == NULL) return 0;
+
+  // Calculate backward references from ARGB image.
+  if (!VP8LGetBackwardReferences(width, height, argb, quality, 0, 1, &refs)) {
+    goto Error;
+  }
+  // Build histogram image and symbols from backward references.
+  VP8LHistogramStoreRefs(&refs, histogram_image->histograms[0]);
+
+  // Create Huffman bit lengths and codes for each histogram image.
+  assert(histogram_image->size == 1);
+  if (!GetHuffBitLengthsAndCodes(histogram_image, huffman_codes)) {
+    goto Error;
+  }
+
+  // No color cache, no Huffman image.
+  VP8LWriteBits(bw, 1, 0);
+
+  // Store Huffman codes.
+  for (i = 0; i < 5; ++i) {
+    HuffmanTreeCode* const codes = &huffman_codes[i];
+    if (!StoreHuffmanCode(bw, codes)) {
+      goto Error;
+    }
+    ClearHuffmanTreeIfOnlyOneSymbol(codes);
+  }
+
+  // Store actual literals.
+  StoreImageToBitMask(bw, width, 0, &refs, histogram_symbols, huffman_codes);
+  ok = 1;
+
+ Error:
+  free(histogram_image);
+  VP8LClearBackwardRefs(&refs);
+  free(huffman_codes[0].codes);
+  return ok;
+}
+
+static int EncodeImageInternal(VP8LBitWriter* const bw,
+                               const uint32_t* const argb,
+                               int width, int height, int quality,
+                               int cache_bits, int histogram_bits) {
+  int ok = 0;
+  const int use_2d_locality = 1;
+  const int use_color_cache = (cache_bits > 0);
+  const uint32_t histogram_image_xysize =
+      VP8LSubSampleSize(width, histogram_bits) *
+      VP8LSubSampleSize(height, histogram_bits);
+  VP8LHistogramSet* histogram_image =
+      VP8LAllocateHistogramSet(histogram_image_xysize, 0);
+  int histogram_image_size = 0;
+  size_t bit_array_size = 0;
+  HuffmanTreeCode* huffman_codes = NULL;
+  VP8LBackwardRefs refs;
+  uint16_t* const histogram_symbols =
+      (uint16_t*)WebPSafeMalloc((uint64_t)histogram_image_xysize,
+                                sizeof(*histogram_symbols));
+  assert(histogram_bits >= MIN_HUFFMAN_BITS);
+  assert(histogram_bits <= MAX_HUFFMAN_BITS);
+
+  if (histogram_image == NULL || histogram_symbols == NULL) {
+    free(histogram_image);
+    free(histogram_symbols);
+    return 0;
+  }
+
+  // Calculate backward references from ARGB image.
+  if (!VP8LGetBackwardReferences(width, height, argb, quality, cache_bits,
+                                 use_2d_locality, &refs)) {
+    goto Error;
+  }
+  // Build histogram image and symbols from backward references.
+  if (!VP8LGetHistoImageSymbols(width, height, &refs,
+                                quality, histogram_bits, cache_bits,
+                                histogram_image,
+                                histogram_symbols)) {
+    goto Error;
+  }
+  // Create Huffman bit lengths and codes for each histogram image.
+  histogram_image_size = histogram_image->size;
+  bit_array_size = 5 * histogram_image_size;
+  huffman_codes = (HuffmanTreeCode*)WebPSafeCalloc(bit_array_size,
+                                                   sizeof(*huffman_codes));
+  if (huffman_codes == NULL ||
+      !GetHuffBitLengthsAndCodes(histogram_image, huffman_codes)) {
+    goto Error;
+  }
+  // Free combined histograms.
+  free(histogram_image);
+  histogram_image = NULL;
+
+  // Color Cache parameters.
+  VP8LWriteBits(bw, 1, use_color_cache);
+  if (use_color_cache) {
+    VP8LWriteBits(bw, 4, cache_bits);
+  }
+
+  // Huffman image + meta huffman.
+  {
+    const int write_histogram_image = (histogram_image_size > 1);
+    VP8LWriteBits(bw, 1, write_histogram_image);
+    if (write_histogram_image) {
+      uint32_t* const histogram_argb =
+          (uint32_t*)WebPSafeMalloc((uint64_t)histogram_image_xysize,
+                                    sizeof(*histogram_argb));
+      int max_index = 0;
+      uint32_t i;
+      if (histogram_argb == NULL) goto Error;
+      for (i = 0; i < histogram_image_xysize; ++i) {
+        const int symbol_index = histogram_symbols[i] & 0xffff;
+        histogram_argb[i] = 0xff000000 | (symbol_index << 8);
+        if (symbol_index >= max_index) {
+          max_index = symbol_index + 1;
+        }
+      }
+      histogram_image_size = max_index;
+
+      VP8LWriteBits(bw, 3, histogram_bits - 2);
+      ok = EncodeImageNoHuffman(bw, histogram_argb,
+                                VP8LSubSampleSize(width, histogram_bits),
+                                VP8LSubSampleSize(height, histogram_bits),
+                                quality);
+      free(histogram_argb);
+      if (!ok) goto Error;
+    }
+  }
+
+  // Store Huffman codes.
+  {
+    int i;
+    for (i = 0; i < 5 * histogram_image_size; ++i) {
+      HuffmanTreeCode* const codes = &huffman_codes[i];
+      if (!StoreHuffmanCode(bw, codes)) goto Error;
+      ClearHuffmanTreeIfOnlyOneSymbol(codes);
+    }
+  }
+
+  // Store actual literals.
+  StoreImageToBitMask(bw, width, histogram_bits, &refs,
+                      histogram_symbols, huffman_codes);
+  ok = 1;
+
+ Error:
+  free(histogram_image);
+
+  VP8LClearBackwardRefs(&refs);
+  if (huffman_codes != NULL) {
+    free(huffman_codes->codes);
+    free(huffman_codes);
+  }
+  free(histogram_symbols);
+  return ok;
+}
+
+// -----------------------------------------------------------------------------
+// Transforms
+
+// Check if it would be a good idea to subtract green from red and blue. We
+// only impact entropy in red/blue components, don't bother to look at others.
+static int EvalAndApplySubtractGreen(VP8LEncoder* const enc,
+                                     int width, int height,
+                                     VP8LBitWriter* const bw) {
+  if (!enc->use_palette_) {
+    int i;
+    const uint32_t* const argb = enc->argb_;
+    double bit_cost_before, bit_cost_after;
+    VP8LHistogram* const histo = (VP8LHistogram*)malloc(sizeof(*histo));
+    if (histo == NULL) return 0;
+
+    VP8LHistogramInit(histo, 1);
+    for (i = 0; i < width * height; ++i) {
+      const uint32_t c = argb[i];
+      ++histo->red_[(c >> 16) & 0xff];
+      ++histo->blue_[(c >> 0) & 0xff];
+    }
+    bit_cost_before = VP8LHistogramEstimateBits(histo);
+
+    VP8LHistogramInit(histo, 1);
+    for (i = 0; i < width * height; ++i) {
+      const uint32_t c = argb[i];
+      const int green = (c >> 8) & 0xff;
+      ++histo->red_[((c >> 16) - green) & 0xff];
+      ++histo->blue_[((c >> 0) - green) & 0xff];
+    }
+    bit_cost_after = VP8LHistogramEstimateBits(histo);
+    free(histo);
+
+    // Check if subtracting green yields low entropy.
+    enc->use_subtract_green_ = (bit_cost_after < bit_cost_before);
+    if (enc->use_subtract_green_) {
+      VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
+      VP8LWriteBits(bw, 2, SUBTRACT_GREEN);
+      VP8LSubtractGreenFromBlueAndRed(enc->argb_, width * height);
+    }
+  }
+  return 1;
+}
+
+static int ApplyPredictFilter(const VP8LEncoder* const enc,
+                              int width, int height, int quality,
+                              VP8LBitWriter* const bw) {
+  const int pred_bits = enc->transform_bits_;
+  const int transform_width = VP8LSubSampleSize(width, pred_bits);
+  const int transform_height = VP8LSubSampleSize(height, pred_bits);
+
+  VP8LResidualImage(width, height, pred_bits, enc->argb_, enc->argb_scratch_,
+                    enc->transform_data_);
+  VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
+  VP8LWriteBits(bw, 2, PREDICTOR_TRANSFORM);
+  assert(pred_bits >= 2);
+  VP8LWriteBits(bw, 3, pred_bits - 2);
+  if (!EncodeImageNoHuffman(bw, enc->transform_data_,
+                            transform_width, transform_height, quality)) {
+    return 0;
+  }
+  return 1;
+}
+
+static int ApplyCrossColorFilter(const VP8LEncoder* const enc,
+                                 int width, int height, int quality,
+                                 VP8LBitWriter* const bw) {
+  const int ccolor_transform_bits = enc->transform_bits_;
+  const int transform_width = VP8LSubSampleSize(width, ccolor_transform_bits);
+  const int transform_height = VP8LSubSampleSize(height, ccolor_transform_bits);
+  const int step = (quality < 25) ? 32 : (quality > 50) ? 8 : 16;
+
+  VP8LColorSpaceTransform(width, height, ccolor_transform_bits, step,
+                          enc->argb_, enc->transform_data_);
+  VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
+  VP8LWriteBits(bw, 2, CROSS_COLOR_TRANSFORM);
+  assert(ccolor_transform_bits >= 2);
+  VP8LWriteBits(bw, 3, ccolor_transform_bits - 2);
+  if (!EncodeImageNoHuffman(bw, enc->transform_data_,
+                            transform_width, transform_height, quality)) {
+    return 0;
+  }
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+
+static WebPEncodingError WriteRiffHeader(const WebPPicture* const pic,
+                                         size_t riff_size, size_t vp8l_size) {
+  uint8_t riff[RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE + VP8L_SIGNATURE_SIZE] = {
+    'R', 'I', 'F', 'F', 0, 0, 0, 0, 'W', 'E', 'B', 'P',
+    'V', 'P', '8', 'L', 0, 0, 0, 0, VP8L_MAGIC_BYTE,
+  };
+  PutLE32(riff + TAG_SIZE, (uint32_t)riff_size);
+  PutLE32(riff + RIFF_HEADER_SIZE + TAG_SIZE, (uint32_t)vp8l_size);
+  if (!pic->writer(riff, sizeof(riff), pic)) {
+    return VP8_ENC_ERROR_BAD_WRITE;
+  }
+  return VP8_ENC_OK;
+}
+
+static int WriteImageSize(const WebPPicture* const pic,
+                          VP8LBitWriter* const bw) {
+  const int width = pic->width - 1;
+  const int height = pic->height - 1;
+  assert(width < WEBP_MAX_DIMENSION && height < WEBP_MAX_DIMENSION);
+
+  VP8LWriteBits(bw, VP8L_IMAGE_SIZE_BITS, width);
+  VP8LWriteBits(bw, VP8L_IMAGE_SIZE_BITS, height);
+  return !bw->error_;
+}
+
+static int WriteRealAlphaAndVersion(VP8LBitWriter* const bw, int has_alpha) {
+  VP8LWriteBits(bw, 1, has_alpha);
+  VP8LWriteBits(bw, VP8L_VERSION_BITS, VP8L_VERSION);
+  return !bw->error_;
+}
+
+static WebPEncodingError WriteImage(const WebPPicture* const pic,
+                                    VP8LBitWriter* const bw,
+                                    size_t* const coded_size) {
+  WebPEncodingError err = VP8_ENC_OK;
+  const uint8_t* const webpll_data = VP8LBitWriterFinish(bw);
+  const size_t webpll_size = VP8LBitWriterNumBytes(bw);
+  const size_t vp8l_size = VP8L_SIGNATURE_SIZE + webpll_size;
+  const size_t pad = vp8l_size & 1;
+  const size_t riff_size = TAG_SIZE + CHUNK_HEADER_SIZE + vp8l_size + pad;
+
+  err = WriteRiffHeader(pic, riff_size, vp8l_size);
+  if (err != VP8_ENC_OK) goto Error;
+
+  if (!pic->writer(webpll_data, webpll_size, pic)) {
+    err = VP8_ENC_ERROR_BAD_WRITE;
+    goto Error;
+  }
+
+  if (pad) {
+    const uint8_t pad_byte[1] = { 0 };
+    if (!pic->writer(pad_byte, 1, pic)) {
+      err = VP8_ENC_ERROR_BAD_WRITE;
+      goto Error;
+    }
+  }
+  *coded_size = CHUNK_HEADER_SIZE + riff_size;
+  return VP8_ENC_OK;
+
+ Error:
+  return err;
+}
+
+// -----------------------------------------------------------------------------
+
+// Allocates the memory for argb (W x H) buffer, 2 rows of context for
+// prediction and transform data.
+static WebPEncodingError AllocateTransformBuffer(VP8LEncoder* const enc,
+                                                 int width, int height) {
+  WebPEncodingError err = VP8_ENC_OK;
+  const int tile_size = 1 << enc->transform_bits_;
+  const uint64_t image_size = width * height;
+  const uint64_t argb_scratch_size = tile_size * width + width;
+  const uint64_t transform_data_size =
+      (uint64_t)VP8LSubSampleSize(width, enc->transform_bits_) *
+      (uint64_t)VP8LSubSampleSize(height, enc->transform_bits_);
+  const uint64_t total_size =
+      image_size + argb_scratch_size + transform_data_size;
+  uint32_t* mem = (uint32_t*)WebPSafeMalloc(total_size, sizeof(*mem));
+  if (mem == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+  enc->argb_ = mem;
+  mem += image_size;
+  enc->argb_scratch_ = mem;
+  mem += argb_scratch_size;
+  enc->transform_data_ = mem;
+  enc->current_width_ = width;
+
+ Error:
+  return err;
+}
+
+static void ApplyPalette(uint32_t* src, uint32_t* dst,
+                         uint32_t src_stride, uint32_t dst_stride,
+                         const uint32_t* palette, int palette_size,
+                         int width, int height, int xbits, uint8_t* row) {
+  int i, x, y;
+  int use_LUT = 1;
+  for (i = 0; i < palette_size; ++i) {
+    if ((palette[i] & 0xffff00ffu) != 0) {
+      use_LUT = 0;
+      break;
+    }
+  }
+
+  if (use_LUT) {
+    uint8_t inv_palette[MAX_PALETTE_SIZE] = { 0 };
+    for (i = 0; i < palette_size; ++i) {
+      const int color = (palette[i] >> 8) & 0xff;
+      inv_palette[color] = i;
+    }
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const int color = (src[x] >> 8) & 0xff;
+        row[x] = inv_palette[color];
+      }
+      VP8LBundleColorMap(row, width, xbits, dst);
+      src += src_stride;
+      dst += dst_stride;
+    }
+  } else {
+    // Use 1 pixel cache for ARGB pixels.
+    uint32_t last_pix = palette[0];
+    int last_idx = 0;
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; ++x) {
+        const uint32_t pix = src[x];
+        if (pix != last_pix) {
+          for (i = 0; i < palette_size; ++i) {
+            if (pix == palette[i]) {
+              last_idx = i;
+              last_pix = pix;
+              break;
+            }
+          }
+        }
+        row[x] = last_idx;
+      }
+      VP8LBundleColorMap(row, width, xbits, dst);
+      src += src_stride;
+      dst += dst_stride;
+    }
+  }
+}
+
+// Note: Expects "enc->palette_" to be set properly.
+// Also, "enc->palette_" will be modified after this call and should not be used
+// later.
+static WebPEncodingError EncodePalette(VP8LBitWriter* const bw,
+                                       VP8LEncoder* const enc, int quality) {
+  WebPEncodingError err = VP8_ENC_OK;
+  int i;
+  const WebPPicture* const pic = enc->pic_;
+  uint32_t* src = pic->argb;
+  uint32_t* dst;
+  const int width = pic->width;
+  const int height = pic->height;
+  uint32_t* const palette = enc->palette_;
+  const int palette_size = enc->palette_size_;
+  uint8_t* row = NULL;
+  int xbits;
+
+  // Replace each input pixel by corresponding palette index.
+  // This is done line by line.
+  if (palette_size <= 4) {
+    xbits = (palette_size <= 2) ? 3 : 2;
+  } else {
+    xbits = (palette_size <= 16) ? 1 : 0;
+  }
+
+  err = AllocateTransformBuffer(enc, VP8LSubSampleSize(width, xbits), height);
+  if (err != VP8_ENC_OK) goto Error;
+  dst = enc->argb_;
+
+  row = (uint8_t*)WebPSafeMalloc((uint64_t)width, sizeof(*row));
+  if (row == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
+
+  ApplyPalette(src, dst, pic->argb_stride, enc->current_width_,
+               palette, palette_size, width, height, xbits, row);
+
+  // Save palette to bitstream.
+  VP8LWriteBits(bw, 1, TRANSFORM_PRESENT);
+  VP8LWriteBits(bw, 2, COLOR_INDEXING_TRANSFORM);
+  assert(palette_size >= 1);
+  VP8LWriteBits(bw, 8, palette_size - 1);
+  for (i = palette_size - 1; i >= 1; --i) {
+    palette[i] = VP8LSubPixels(palette[i], palette[i - 1]);
+  }
+  if (!EncodeImageNoHuffman(bw, palette, palette_size, 1, quality)) {
+    err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
+    goto Error;
+  }
+
+ Error:
+  free(row);
+  return err;
+}
+
+// -----------------------------------------------------------------------------
+
+static int GetHistoBits(int method, int use_palette, int width, int height) {
+  const uint64_t hist_size = sizeof(VP8LHistogram);
+  // Make tile size a function of encoding method (Range: 0 to 6).
+  int histo_bits = (use_palette ? 9 : 7) - method;
+  while (1) {
+    const uint64_t huff_image_size = VP8LSubSampleSize(width, histo_bits) *
+                                     VP8LSubSampleSize(height, histo_bits) *
+                                     hist_size;
+    if (huff_image_size <= MAX_HUFF_IMAGE_SIZE) break;
+    ++histo_bits;
+  }
+  return (histo_bits < MIN_HUFFMAN_BITS) ? MIN_HUFFMAN_BITS :
+         (histo_bits > MAX_HUFFMAN_BITS) ? MAX_HUFFMAN_BITS : histo_bits;
+}
+
+static void FinishEncParams(VP8LEncoder* const enc) {
+  const WebPConfig* const config = enc->config_;
+  const WebPPicture* const pic = enc->pic_;
+  const int method = config->method;
+  const float quality = config->quality;
+  const int use_palette = enc->use_palette_;
+  enc->transform_bits_ = (method < 4) ? 5 : (method > 4) ? 3 : 4;
+  enc->histo_bits_ = GetHistoBits(method, use_palette, pic->width, pic->height);
+  enc->cache_bits_ = (quality <= 25.f) ? 0 : 7;
+}
+
+// -----------------------------------------------------------------------------
+// VP8LEncoder
+
+static VP8LEncoder* VP8LEncoderNew(const WebPConfig* const config,
+                                   const WebPPicture* const picture) {
+  VP8LEncoder* const enc = (VP8LEncoder*)calloc(1, sizeof(*enc));
+  if (enc == NULL) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return NULL;
+  }
+  enc->config_ = config;
+  enc->pic_ = picture;
+
+  VP8LDspInit();
+
+  return enc;
+}
+
+static void VP8LEncoderDelete(VP8LEncoder* enc) {
+  free(enc->argb_);
+  free(enc);
+}
+
+// -----------------------------------------------------------------------------
+// Main call
+
+WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
+                                   const WebPPicture* const picture,
+                                   VP8LBitWriter* const bw) {
+  WebPEncodingError err = VP8_ENC_OK;
+  const int quality = (int)config->quality;
+  const int width = picture->width;
+  const int height = picture->height;
+  VP8LEncoder* const enc = VP8LEncoderNew(config, picture);
+  const size_t byte_position = VP8LBitWriterNumBytes(bw);
+
+  if (enc == NULL) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  // ---------------------------------------------------------------------------
+  // Analyze image (entropy, num_palettes etc)
+
+  if (!VP8LEncAnalyze(enc, config->image_hint)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  FinishEncParams(enc);
+
+  if (enc->use_palette_) {
+    err = EncodePalette(bw, enc, quality);
+    if (err != VP8_ENC_OK) goto Error;
+    // Color cache is disabled for palette.
+    enc->cache_bits_ = 0;
+  }
+
+  // In case image is not packed.
+  if (enc->argb_ == NULL) {
+    int y;
+    err = AllocateTransformBuffer(enc, width, height);
+    if (err != VP8_ENC_OK) goto Error;
+    for (y = 0; y < height; ++y) {
+      memcpy(enc->argb_ + y * width,
+             picture->argb + y * picture->argb_stride,
+             width * sizeof(*enc->argb_));
+    }
+    enc->current_width_ = width;
+  }
+
+  // ---------------------------------------------------------------------------
+  // Apply transforms and write transform data.
+
+  if (!EvalAndApplySubtractGreen(enc, enc->current_width_, height, bw)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  if (enc->use_predict_) {
+    if (!ApplyPredictFilter(enc, enc->current_width_, height, quality, bw)) {
+      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
+      goto Error;
+    }
+  }
+
+  if (enc->use_cross_color_) {
+    if (!ApplyCrossColorFilter(enc, enc->current_width_, height, quality, bw)) {
+      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
+      goto Error;
+    }
+  }
+
+  VP8LWriteBits(bw, 1, !TRANSFORM_PRESENT);  // No more transforms.
+
+  // ---------------------------------------------------------------------------
+  // Estimate the color cache size.
+
+  if (enc->cache_bits_ > 0) {
+    if (!VP8LCalculateEstimateForCacheSize(enc->argb_, enc->current_width_,
+                                           height, &enc->cache_bits_)) {
+      err = VP8_ENC_ERROR_INVALID_CONFIGURATION;
+      goto Error;
+    }
+  }
+
+  // ---------------------------------------------------------------------------
+  // Encode and write the transformed image.
+
+  if (!EncodeImageInternal(bw, enc->argb_, enc->current_width_, height,
+                           quality, enc->cache_bits_, enc->histo_bits_)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  if (picture->stats != NULL) {
+    WebPAuxStats* const stats = picture->stats;
+    stats->lossless_features = 0;
+    if (enc->use_predict_) stats->lossless_features |= 1;
+    if (enc->use_cross_color_) stats->lossless_features |= 2;
+    if (enc->use_subtract_green_) stats->lossless_features |= 4;
+    if (enc->use_palette_) stats->lossless_features |= 8;
+    stats->histogram_bits = enc->histo_bits_;
+    stats->transform_bits = enc->transform_bits_;
+    stats->cache_bits = enc->cache_bits_;
+    stats->palette_size = enc->palette_size_;
+    stats->lossless_size = (int)(VP8LBitWriterNumBytes(bw) - byte_position);
+  }
+
+ Error:
+  VP8LEncoderDelete(enc);
+  return err;
+}
+
+int VP8LEncodeImage(const WebPConfig* const config,
+                    const WebPPicture* const picture) {
+  int width, height;
+  int has_alpha;
+  size_t coded_size;
+  int percent = 0;
+  WebPEncodingError err = VP8_ENC_OK;
+  VP8LBitWriter bw;
+
+  if (picture == NULL) return 0;
+
+  if (config == NULL || picture->argb == NULL) {
+    err = VP8_ENC_ERROR_NULL_PARAMETER;
+    WebPEncodingSetError(picture, err);
+    return 0;
+  }
+
+  width = picture->width;
+  height = picture->height;
+  if (!VP8LBitWriterInit(&bw, (width * height) >> 1)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  if (!WebPReportProgress(picture, 1, &percent)) {
+ UserAbort:
+    err = VP8_ENC_ERROR_USER_ABORT;
+    goto Error;
+  }
+  // Reset stats (for pure lossless coding)
+  if (picture->stats != NULL) {
+    WebPAuxStats* const stats = picture->stats;
+    memset(stats, 0, sizeof(*stats));
+    stats->PSNR[0] = 99.f;
+    stats->PSNR[1] = 99.f;
+    stats->PSNR[2] = 99.f;
+    stats->PSNR[3] = 99.f;
+    stats->PSNR[4] = 99.f;
+  }
+
+  // Write image size.
+  if (!WriteImageSize(picture, &bw)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  has_alpha = WebPPictureHasTransparency(picture);
+  // Write the non-trivial Alpha flag and lossless version.
+  if (!WriteRealAlphaAndVersion(&bw, has_alpha)) {
+    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+    goto Error;
+  }
+
+  if (!WebPReportProgress(picture, 5, &percent)) goto UserAbort;
+
+  // Encode main image stream.
+  err = VP8LEncodeStream(config, picture, &bw);
+  if (err != VP8_ENC_OK) goto Error;
+
+  // TODO(skal): have a fine-grained progress report in VP8LEncodeStream().
+  if (!WebPReportProgress(picture, 90, &percent)) goto UserAbort;
+
+  // Finish the RIFF chunk.
+  err = WriteImage(picture, &bw, &coded_size);
+  if (err != VP8_ENC_OK) goto Error;
+
+  if (!WebPReportProgress(picture, 100, &percent)) goto UserAbort;
+
+  // Save size.
+  if (picture->stats != NULL) {
+    picture->stats->coded_size += (int)coded_size;
+    picture->stats->lossless_size = (int)coded_size;
+  }
+
+  if (picture->extra_info != NULL) {
+    const int mb_w = (width + 15) >> 4;
+    const int mb_h = (height + 15) >> 4;
+    memset(picture->extra_info, 0, mb_w * mb_h * sizeof(*picture->extra_info));
+  }
+
+ Error:
+  if (bw.error_) err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+  VP8LBitWriterDestroy(&bw);
+  if (err != VP8_ENC_OK) {
+    WebPEncodingSetError(picture, err);
+    return 0;
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/enc/vp8li.h b/drivers/webp/enc/vp8li.h
new file mode 100644
index 0000000000..96d6faed64
--- /dev/null
+++ b/drivers/webp/enc/vp8li.h
@@ -0,0 +1,70 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Lossless encoder: internal header.
+//
+// Author: Vikas Arora (vikaas.arora@gmail.com)
+
+#ifndef WEBP_ENC_VP8LI_H_
+#define WEBP_ENC_VP8LI_H_
+
+#include "./histogram.h"
+#include "../utils/bit_writer.h"
+#include "../webp/encode.h"
+#include "../webp/format_constants.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  const WebPConfig* config_;    // user configuration and parameters
+  const WebPPicture* pic_;      // input picture.
+
+  uint32_t* argb_;              // Transformed argb image data.
+  uint32_t* argb_scratch_;      // Scratch memory for argb rows
+                                // (used for prediction).
+  uint32_t* transform_data_;    // Scratch memory for transform data.
+  int       current_width_;     // Corresponds to packed image width.
+
+  // Encoding parameters derived from quality parameter.
+  int histo_bits_;
+  int transform_bits_;
+  int cache_bits_;        // If equal to 0, don't use color cache.
+
+  // Encoding parameters derived from image characteristics.
+  int use_cross_color_;
+  int use_subtract_green_;
+  int use_predict_;
+  int use_palette_;
+  int palette_size_;
+  uint32_t palette_[MAX_PALETTE_SIZE];
+} VP8LEncoder;
+
+//------------------------------------------------------------------------------
+// internal functions. Not public.
+
+// Encodes the picture.
+// Returns 0 if config or picture is NULL or picture doesn't have valid argb
+// input.
+int VP8LEncodeImage(const WebPConfig* const config,
+                    const WebPPicture* const picture);
+
+// Encodes the main image stream using the supplied bit writer.
+WebPEncodingError VP8LEncodeStream(const WebPConfig* const config,
+                                   const WebPPicture* const picture,
+                                   VP8LBitWriter* const bw);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_ENC_VP8LI_H_ */
diff --git a/drivers/webp/enc/webpenc.c b/drivers/webp/enc/webpenc.c
new file mode 100644
index 0000000000..207cce6beb
--- /dev/null
+++ b/drivers/webp/enc/webpenc.c
@@ -0,0 +1,404 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// WebP encoder: main entry point
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+
+#include "./vp8enci.h"
+#include "./vp8li.h"
+#include "../utils/utils.h"
+
+// #define PRINT_MEMORY_INFO
+
+#ifdef PRINT_MEMORY_INFO
+#include <stdio.h>
+#endif
+
+//------------------------------------------------------------------------------
+
+int WebPGetEncoderVersion(void) {
+  return (ENC_MAJ_VERSION << 16) | (ENC_MIN_VERSION << 8) | ENC_REV_VERSION;
+}
+
+//------------------------------------------------------------------------------
+// WebPPicture
+//------------------------------------------------------------------------------
+
+static int DummyWriter(const uint8_t* data, size_t data_size,
+                       const WebPPicture* const picture) {
+  // The following are to prevent 'unused variable' error message.
+  (void)data;
+  (void)data_size;
+  (void)picture;
+  return 1;
+}
+
+int WebPPictureInitInternal(WebPPicture* picture, int version) {
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_ENCODER_ABI_VERSION)) {
+    return 0;   // caller/system version mismatch!
+  }
+  if (picture != NULL) {
+    memset(picture, 0, sizeof(*picture));
+    picture->writer = DummyWriter;
+    WebPEncodingSetError(picture, VP8_ENC_OK);
+  }
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+// VP8Encoder
+//------------------------------------------------------------------------------
+
+static void ResetSegmentHeader(VP8Encoder* const enc) {
+  VP8SegmentHeader* const hdr = &enc->segment_hdr_;
+  hdr->num_segments_ = enc->config_->segments;
+  hdr->update_map_  = (hdr->num_segments_ > 1);
+  hdr->size_ = 0;
+}
+
+static void ResetFilterHeader(VP8Encoder* const enc) {
+  VP8FilterHeader* const hdr = &enc->filter_hdr_;
+  hdr->simple_ = 1;
+  hdr->level_ = 0;
+  hdr->sharpness_ = 0;
+  hdr->i4x4_lf_delta_ = 0;
+}
+
+static void ResetBoundaryPredictions(VP8Encoder* const enc) {
+  // init boundary values once for all
+  // Note: actually, initializing the preds_[] is only needed for intra4.
+  int i;
+  uint8_t* const top = enc->preds_ - enc->preds_w_;
+  uint8_t* const left = enc->preds_ - 1;
+  for (i = -1; i < 4 * enc->mb_w_; ++i) {
+    top[i] = B_DC_PRED;
+  }
+  for (i = 0; i < 4 * enc->mb_h_; ++i) {
+    left[i * enc->preds_w_] = B_DC_PRED;
+  }
+  enc->nz_[-1] = 0;   // constant
+}
+
+// Mapping from config->method_ to coding tools used.
+//-------------------+---+---+---+---+---+---+---+
+//   Method          | 0 | 1 | 2 | 3 |(4)| 5 | 6 |
+//-------------------+---+---+---+---+---+---+---+
+// fast probe        | x |   |   | x |   |   |   |
+//-------------------+---+---+---+---+---+---+---+
+// dynamic proba     | ~ | x | x | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// fast mode analysis|   |   |   |   | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// basic rd-opt      |   |   |   | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// disto-score i4/16 |   |   | x |   |   |   |   |
+//-------------------+---+---+---+---+---+---+---+
+// rd-opt i4/16      |   |   | ~ | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// token buffer (opt)|   |   |   | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// Trellis           |   |   |   |   |   | x |Ful|
+//-------------------+---+---+---+---+---+---+---+
+// full-SNS          |   |   |   |   | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+
+static void MapConfigToTools(VP8Encoder* const enc) {
+  const WebPConfig* const config = enc->config_;
+  const int method = config->method;
+  const int limit = 100 - config->partition_limit;
+  enc->method_ = method;
+  enc->rd_opt_level_ = (method >= 6) ? RD_OPT_TRELLIS_ALL
+                     : (method >= 5) ? RD_OPT_TRELLIS
+                     : (method >= 3) ? RD_OPT_BASIC
+                     : RD_OPT_NONE;
+  enc->max_i4_header_bits_ =
+      256 * 16 * 16 *                 // upper bound: up to 16bit per 4x4 block
+      (limit * limit) / (100 * 100);  // ... modulated with a quadratic curve.
+
+  enc->thread_level_ = config->thread_level;
+
+  enc->do_search_ = (config->target_size > 0 || config->target_PSNR > 0);
+  if (!config->low_memory) {
+#if !defined(DISABLE_TOKEN_BUFFER)
+    enc->use_tokens_ = (enc->rd_opt_level_ >= RD_OPT_BASIC);  // need rd stats
+#endif
+    if (enc->use_tokens_) {
+      enc->num_parts_ = 1;   // doesn't work with multi-partition
+    }
+  }
+}
+
+// Memory scaling with dimensions:
+//  memory (bytes) ~= 2.25 * w + 0.0625 * w * h
+//
+// Typical memory footprint (768x510 picture)
+// Memory used:
+//              encoder: 33919
+//          block cache: 2880
+//                 info: 3072
+//                preds: 24897
+//          top samples: 1623
+//             non-zero: 196
+//             lf-stats: 2048
+//                total: 68635
+// Transient object sizes:
+//       VP8EncIterator: 352
+//         VP8ModeScore: 912
+//       VP8SegmentInfo: 532
+//             VP8Proba: 31032
+//              LFStats: 2048
+// Picture size (yuv): 589824
+
+static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
+                                  WebPPicture* const picture) {
+  const int use_filter =
+      (config->filter_strength > 0) || (config->autofilter > 0);
+  const int mb_w = (picture->width + 15) >> 4;
+  const int mb_h = (picture->height + 15) >> 4;
+  const int preds_w = 4 * mb_w + 1;
+  const int preds_h = 4 * mb_h + 1;
+  const size_t preds_size = preds_w * preds_h * sizeof(uint8_t);
+  const int top_stride = mb_w * 16;
+  const size_t nz_size = (mb_w + 1) * sizeof(uint32_t) + ALIGN_CST;
+  const size_t info_size = mb_w * mb_h * sizeof(VP8MBInfo);
+  const size_t samples_size = 2 * top_stride * sizeof(uint8_t)  // top-luma/u/v
+                            + ALIGN_CST;                        // align all
+  const size_t lf_stats_size =
+      config->autofilter ? sizeof(LFStats) + ALIGN_CST : 0;
+  VP8Encoder* enc;
+  uint8_t* mem;
+  const uint64_t size = (uint64_t)sizeof(VP8Encoder)   // main struct
+                      + ALIGN_CST                      // cache alignment
+                      + info_size                      // modes info
+                      + preds_size                     // prediction modes
+                      + samples_size                   // top/left samples
+                      + nz_size                        // coeff context bits
+                      + lf_stats_size;                 // autofilter stats
+
+#ifdef PRINT_MEMORY_INFO
+  printf("===================================\n");
+  printf("Memory used:\n"
+         "             encoder: %ld\n"
+         "                info: %ld\n"
+         "               preds: %ld\n"
+         "         top samples: %ld\n"
+         "            non-zero: %ld\n"
+         "            lf-stats: %ld\n"
+         "               total: %ld\n",
+         sizeof(VP8Encoder) + ALIGN_CST, info_size,
+         preds_size, samples_size, nz_size, lf_stats_size, size);
+  printf("Transient object sizes:\n"
+         "      VP8EncIterator: %ld\n"
+         "        VP8ModeScore: %ld\n"
+         "      VP8SegmentInfo: %ld\n"
+         "            VP8Proba: %ld\n"
+         "             LFStats: %ld\n",
+         sizeof(VP8EncIterator), sizeof(VP8ModeScore),
+         sizeof(VP8SegmentInfo), sizeof(VP8Proba),
+         sizeof(LFStats));
+  printf("Picture size (yuv): %ld\n",
+         mb_w * mb_h * 384 * sizeof(uint8_t));
+  printf("===================================\n");
+#endif
+  mem = (uint8_t*)WebPSafeMalloc(size, sizeof(*mem));
+  if (mem == NULL) {
+    WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
+    return NULL;
+  }
+  enc = (VP8Encoder*)mem;
+  mem = (uint8_t*)DO_ALIGN(mem + sizeof(*enc));
+  memset(enc, 0, sizeof(*enc));
+  enc->num_parts_ = 1 << config->partitions;
+  enc->mb_w_ = mb_w;
+  enc->mb_h_ = mb_h;
+  enc->preds_w_ = preds_w;
+  enc->mb_info_ = (VP8MBInfo*)mem;
+  mem += info_size;
+  enc->preds_ = ((uint8_t*)mem) + 1 + enc->preds_w_;
+  mem += preds_w * preds_h * sizeof(uint8_t);
+  enc->nz_ = 1 + (uint32_t*)DO_ALIGN(mem);
+  mem += nz_size;
+  enc->lf_stats_ = lf_stats_size ? (LFStats*)DO_ALIGN(mem) : NULL;
+  mem += lf_stats_size;
+
+  // top samples (all 16-aligned)
+  mem = (uint8_t*)DO_ALIGN(mem);
+  enc->y_top_ = (uint8_t*)mem;
+  enc->uv_top_ = enc->y_top_ + top_stride;
+  mem += 2 * top_stride;
+  assert(mem <= (uint8_t*)enc + size);
+
+  enc->config_ = config;
+  enc->profile_ = use_filter ? ((config->filter_type == 1) ? 0 : 1) : 2;
+  enc->pic_ = picture;
+  enc->percent_ = 0;
+
+  MapConfigToTools(enc);
+  VP8EncDspInit();
+  VP8DefaultProbas(enc);
+  ResetSegmentHeader(enc);
+  ResetFilterHeader(enc);
+  ResetBoundaryPredictions(enc);
+
+  VP8EncInitAlpha(enc);
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+  VP8EncInitLayer(enc);
+#endif
+
+  VP8TBufferInit(&enc->tokens_);
+  return enc;
+}
+
+static int DeleteVP8Encoder(VP8Encoder* enc) {
+  int ok = 1;
+  if (enc != NULL) {
+    ok = VP8EncDeleteAlpha(enc);
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    VP8EncDeleteLayer(enc);
+#endif
+    VP8TBufferClear(&enc->tokens_);
+    free(enc);
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+
+static double GetPSNR(uint64_t err, uint64_t size) {
+  return (err > 0 && size > 0) ? 10. * log10(255. * 255. * size / err) : 99.;
+}
+
+static void FinalizePSNR(const VP8Encoder* const enc) {
+  WebPAuxStats* stats = enc->pic_->stats;
+  const uint64_t size = enc->sse_count_;
+  const uint64_t* const sse = enc->sse_;
+  stats->PSNR[0] = (float)GetPSNR(sse[0], size);
+  stats->PSNR[1] = (float)GetPSNR(sse[1], size / 4);
+  stats->PSNR[2] = (float)GetPSNR(sse[2], size / 4);
+  stats->PSNR[3] = (float)GetPSNR(sse[0] + sse[1] + sse[2], size * 3 / 2);
+  stats->PSNR[4] = (float)GetPSNR(sse[3], size);
+}
+
+static void StoreStats(VP8Encoder* const enc) {
+  WebPAuxStats* const stats = enc->pic_->stats;
+  if (stats != NULL) {
+    int i, s;
+    for (i = 0; i < NUM_MB_SEGMENTS; ++i) {
+      stats->segment_level[i] = enc->dqm_[i].fstrength_;
+      stats->segment_quant[i] = enc->dqm_[i].quant_;
+      for (s = 0; s <= 2; ++s) {
+        stats->residual_bytes[s][i] = enc->residual_bytes_[s][i];
+      }
+    }
+    FinalizePSNR(enc);
+    stats->coded_size = enc->coded_size_;
+    for (i = 0; i < 3; ++i) {
+      stats->block_count[i] = enc->block_count_[i];
+    }
+  }
+  WebPReportProgress(enc->pic_, 100, &enc->percent_);  // done!
+}
+
+int WebPEncodingSetError(const WebPPicture* const pic,
+                         WebPEncodingError error) {
+  assert((int)error < VP8_ENC_ERROR_LAST);
+  assert((int)error >= VP8_ENC_OK);
+  ((WebPPicture*)pic)->error_code = error;
+  return 0;
+}
+
+int WebPReportProgress(const WebPPicture* const pic,
+                       int percent, int* const percent_store) {
+  if (percent_store != NULL && percent != *percent_store) {
+    *percent_store = percent;
+    if (pic->progress_hook && !pic->progress_hook(percent, pic)) {
+      // user abort requested
+      WebPEncodingSetError(pic, VP8_ENC_ERROR_USER_ABORT);
+      return 0;
+    }
+  }
+  return 1;  // ok
+}
+//------------------------------------------------------------------------------
+
+int WebPEncode(const WebPConfig* config, WebPPicture* pic) {
+  int ok = 0;
+
+  if (pic == NULL)
+    return 0;
+  WebPEncodingSetError(pic, VP8_ENC_OK);  // all ok so far
+  if (config == NULL)  // bad params
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_NULL_PARAMETER);
+  if (!WebPValidateConfig(config))
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_INVALID_CONFIGURATION);
+  if (pic->width <= 0 || pic->height <= 0)
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_BAD_DIMENSION);
+  if (pic->width > WEBP_MAX_DIMENSION || pic->height > WEBP_MAX_DIMENSION)
+    return WebPEncodingSetError(pic, VP8_ENC_ERROR_BAD_DIMENSION);
+
+  if (pic->stats != NULL) memset(pic->stats, 0, sizeof(*pic->stats));
+
+  if (!config->lossless) {
+    VP8Encoder* enc = NULL;
+    if (pic->y == NULL || pic->u == NULL || pic->v == NULL) {
+      // Make sure we have YUVA samples.
+      float dithering = 0.f;
+      if (config->preprocessing & 2) {
+        const float x = config->quality / 100.f;
+        const float x2 = x * x;
+        // slowly decreasing from max dithering at low quality (q->0)
+        // to 0.5 dithering amplitude at high quality (q->100)
+        dithering = 1.0f + (0.5f - 1.0f) * x2 * x2;
+      }
+      if (!WebPPictureARGBToYUVADithered(pic, WEBP_YUV420, dithering)) {
+        return 0;
+      }
+    }
+
+    enc = InitVP8Encoder(config, pic);
+    if (enc == NULL) return 0;  // pic->error is already set.
+    // Note: each of the tasks below account for 20% in the progress report.
+    ok = VP8EncAnalyze(enc);
+
+    // Analysis is done, proceed to actual coding.
+    ok = ok && VP8EncStartAlpha(enc);   // possibly done in parallel
+    if (!enc->use_tokens_) {
+      ok = ok && VP8EncLoop(enc);
+    } else {
+      ok = ok && VP8EncTokenLoop(enc);
+    }
+    ok = ok && VP8EncFinishAlpha(enc);
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    ok = ok && VP8EncFinishLayer(enc);
+#endif
+
+    ok = ok && VP8EncWrite(enc);
+    StoreStats(enc);
+    if (!ok) {
+      VP8EncFreeBitWriters(enc);
+    }
+    ok &= DeleteVP8Encoder(enc);  // must always be called, even if !ok
+  } else {
+    // Make sure we have ARGB samples.
+    if (pic->argb == NULL && !WebPPictureYUVAToARGB(pic)) {
+      return 0;
+    }
+
+    ok = VP8LEncodeImage(config, pic);  // Sets pic->error in case of problem.
+  }
+
+  return ok;
+}
+
diff --git a/drivers/webp/encode.h b/drivers/webp/encode.h
new file mode 100644
index 0000000000..7a428b4e6e
--- /dev/null
+++ b/drivers/webp/encode.h
@@ -0,0 +1,498 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//   WebP encoder: main interface
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_WEBP_ENCODE_H_
+#define WEBP_WEBP_ENCODE_H_
+
+#include "./types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define WEBP_ENCODER_ABI_VERSION 0x0202    // 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.
+WEBP_EXTERN(int) WebPGetEncoderVersion(void);
+
+//------------------------------------------------------------------------------
+// One-stop-shop call! No questions asked:
+
+// 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)'.
+// 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).
+WEBP_EXTERN(size_t) WebPEncodeRGB(const uint8_t* rgb,
+                                  int width, int height, int stride,
+                                  float quality_factor, uint8_t** output);
+WEBP_EXTERN(size_t) WebPEncodeBGR(const uint8_t* bgr,
+                                  int width, int height, int stride,
+                                  float quality_factor, uint8_t** output);
+WEBP_EXTERN(size_t) WebPEncodeRGBA(const uint8_t* rgba,
+                                   int width, int height, int stride,
+                                   float quality_factor, uint8_t** output);
+WEBP_EXTERN(size_t) WebPEncodeBGRA(const uint8_t* bgra,
+                                   int width, int height, int stride,
+                                   float quality_factor, uint8_t** output);
+
+// These functions are the equivalent of the above, but compressing in a
+// lossless manner. Files are usually larger than lossy format, but will
+// not suffer any compression loss.
+WEBP_EXTERN(size_t) WebPEncodeLosslessRGB(const uint8_t* rgb,
+                                          int width, int height, int stride,
+                                          uint8_t** output);
+WEBP_EXTERN(size_t) WebPEncodeLosslessBGR(const uint8_t* bgr,
+                                          int width, int height, int stride,
+                                          uint8_t** output);
+WEBP_EXTERN(size_t) WebPEncodeLosslessRGBA(const uint8_t* rgba,
+                                           int width, int height, int stride,
+                                           uint8_t** output);
+WEBP_EXTERN(size_t) WebPEncodeLosslessBGRA(const uint8_t* bgra,
+                                           int width, int height, int stride,
+                                           uint8_t** output);
+
+//------------------------------------------------------------------------------
+// Coding parameters
+
+// Image characteristics hint for the underlying encoder.
+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
+  WEBP_HINT_GRAPH,        // Discrete tone image (graph, map-tile etc).
+  WEBP_HINT_LAST
+} WebPImageHint;
+
+// 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)
+
+  WebPImageHint image_hint;  // Hint for image type (lossless only for now).
+
+  // Parameters related to lossy compression only:
+  int target_size;        // if non-zero, set the desired target size in bytes.
+                          // Takes precedence over the 'compression' parameter.
+  float target_PSNR;      // if non-zero, specifies the minimal distortion to
+                          // try to achieve. Takes precedence over target_size.
+  int segments;           // maximum number of segments to use, in [1..4]
+  int sns_strength;       // Spatial Noise Shaping. 0=off, 100=maximum.
+  int filter_strength;    // range: [0 = off .. 100 = strongest]
+  int filter_sharpness;   // range: [0 = off .. 7 = least sharp]
+  int filter_type;        // filtering type: 0 = simple, 1 = strong (only used
+                          // if filter_strength > 0 or autofilter > 0)
+  int autofilter;         // Auto adjust filter's strength [0 = off, 1 = on]
+  int alpha_compression;  // Algorithm for encoding the alpha plane (0 = none,
+                          // 1 = compressed with WebP lossless). Default is 1.
+  int alpha_filtering;    // Predictive filtering method for alpha plane.
+                          //  0: none, 1: fast, 2: best. Default if 1.
+  int alpha_quality;      // Between 0 (smallest size) and 100 (lossless).
+                          // Default is 100.
+  int pass;               // number of entropy-analysis passes (in [1..10]).
+
+  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, 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).
+  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).
+
+  uint32_t pad[5];        // padding for later use
+};
+
+// Enumerate some predefined settings for WebPConfig, depending on the type
+// of source picture. These presets are used when calling WebPConfigPreset().
+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
+  WEBP_PRESET_DRAWING,      // hand or line drawing, with high-contrast details
+  WEBP_PRESET_ICON,         // small-sized colorful images
+  WEBP_PRESET_TEXT          // text-like
+} WebPPreset;
+
+// Internal, version-checked, entry point
+WEBP_EXTERN(int) WebPConfigInitInternal(WebPConfig*, WebPPreset, float, int);
+
+// Should always be called, to initialize a fresh WebPConfig structure before
+// modification. Returns false in case of version mismatch. WebPConfigInit()
+// must have succeeded before using the 'config' object.
+// Note that the default values are lossless=0 and quality=75.
+static WEBP_INLINE int WebPConfigInit(WebPConfig* config) {
+  return WebPConfigInitInternal(config, WEBP_PRESET_DEFAULT, 75.f,
+                                WEBP_ENCODER_ABI_VERSION);
+}
+
+// This function will initialize the configuration according to a predefined
+// set of parameters (referred to by 'preset') and a given quality factor.
+// This function can be called as a replacement to WebPConfigInit(). Will
+// return false in case of error.
+static WEBP_INLINE int WebPConfigPreset(WebPConfig* config,
+                                        WebPPreset preset, float quality) {
+  return WebPConfigInitInternal(config, preset, quality,
+                                WEBP_ENCODER_ABI_VERSION);
+}
+
+// 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
+// Structure for storing auxiliary statistics (mostly for lossy encoding).
+
+struct WebPAuxStats {
+  int coded_size;         // final size
+
+  float PSNR[5];          // peak-signal-to-noise ratio for Y/U/V/All/Alpha
+  int block_count[3];     // number of intra4/intra16/skipped macroblocks
+  int header_bytes[2];    // approximate number of bytes spent for header
+                          // and mode-partition #0
+  int residual_bytes[3][4];  // approximate number of bytes spent for
+                             // DC/AC/uv coefficients for each (0..3) segments.
+  int segment_size[4];    // number of macroblocks in each segments
+  int segment_quant[4];   // quantizer values for each segments
+  int segment_level[4];   // filtering strength for each segments [0..63]
+
+  int alpha_data_size;    // size of the transparency data
+  int layer_data_size;    // size of the enhancement layer data
+
+  // lossless encoder statistics
+  uint32_t lossless_features;  // bit0:predictor bit1:cross-color transform
+                               // bit2:subtract-green bit3:color indexing
+  int histogram_bits;          // number of precision bits of histogram
+  int transform_bits;          // precision bits for transform
+  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
+
+  uint32_t pad[4];        // 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
+// reference (and so one can make use of picture->custom_ptr).
+typedef int (*WebPWriterFunction)(const uint8_t* data, size_t data_size,
+                                  const WebPPicture* picture);
+
+// WebPMemoryWrite: a special WebPWriterFunction that writes to memory using
+// the following WebPMemoryWriter object (to be set as a custom_ptr).
+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
+};
+
+// The following must be called first before any use.
+WEBP_EXTERN(void) WebPMemoryWriterInit(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 using the call 'free(writer.mem)'.
+WEBP_EXTERN(int) WebPMemoryWrite(const uint8_t* data, size_t data_size,
+                                 const WebPPicture* picture);
+
+// Progress hook, called from time to time to report progress. It can return
+// false to request an abort of the encoding process, or true otherwise if
+// everything is OK.
+typedef int (*WebPProgressHook)(int percent, const WebPPicture* picture);
+
+// 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_CSP_ALPHA_BIT = 4   // bit that is set if alpha is present
+} WebPEncCSP;
+
+// Encoding error conditions.
+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
+  VP8_ENC_ERROR_NULL_PARAMETER,           // a pointer parameter is NULL
+  VP8_ENC_ERROR_INVALID_CONFIGURATION,    // configuration is invalid
+  VP8_ENC_ERROR_BAD_DIMENSION,            // picture has invalid width/height
+  VP8_ENC_ERROR_PARTITION0_OVERFLOW,      // partition is bigger than 512k
+  VP8_ENC_ERROR_PARTITION_OVERFLOW,       // partition is bigger than 16M
+  VP8_ENC_ERROR_BAD_WRITE,                // error while flushing bytes
+  VP8_ENC_ERROR_FILE_TOO_BIG,             // file is bigger than 4G
+  VP8_ENC_ERROR_USER_ABORT,               // abort request by user
+  VP8_ENC_ERROR_LAST                      // list terminator. always last.
+} WebPEncodingError;
+
+// maximum width/height allowed (inclusive), in pixels
+#define WEBP_MAX_DIMENSION 16383
+
+// Main exchange structure (input samples, output bytes, statistics)
+struct WebPPicture {
+  //   INPUT
+  //////////////
+  // Main flag for encoder selecting between ARGB or YUV input.
+  // It is recommended to use ARGB input (*argb, argb_stride) for lossless
+  // compression, and YUV input (*y, *u, *v, etc.) for lossy compression
+  // since these are the respective native colorspace for these formats.
+  int use_argb;
+
+  // YUV input (mostly used for input to lossy compression)
+  WebPEncCSP colorspace;     // colorspace: should be YUV420 for now (=Y'CbCr).
+  int width, height;         // dimensions (less or equal to WEBP_MAX_DIMENSION)
+  uint8_t *y, *u, *v;        // pointers to luma/chroma planes.
+  int y_stride, uv_stride;   // luma/chroma strides.
+  uint8_t* a;                // pointer to the alpha plane
+  int a_stride;              // stride of the alpha plane
+  uint32_t pad1[2];          // padding for later use
+
+  // ARGB input (mostly used for input to lossless compression)
+  uint32_t* argb;            // Pointer to argb (32 bit) plane.
+  int argb_stride;           // This is stride in pixels units, not bytes.
+  uint32_t pad2[3];          // padding for later use
+
+  //   OUTPUT
+  ///////////////
+  // Byte-emission hook, to store compressed bytes as they are ready.
+  WebPWriterFunction writer;  // can be NULL
+  void* custom_ptr;           // can be used by the writer.
+
+  // map for extra information (only for lossy compression mode)
+  int extra_info_type;    // 1: intra type, 2: segment, 3: quant
+                          // 4: intra-16 prediction mode,
+                          // 5: chroma prediction mode,
+                          // 6: bit cost, 7: distortion
+  uint8_t* extra_info;    // if not NULL, points to an array of size
+                          // ((width + 15) / 16) * ((height + 15) / 16) that
+                          // will be filled with a macroblock map, depending
+                          // on extra_info_type.
+
+  //   STATS AND REPORTS
+  ///////////////////////////
+  // Pointer to side statistics (updated only if not NULL)
+  WebPAuxStats* stats;
+
+  // Error code for the latest error encountered during encoding
+  WebPEncodingError error_code;
+
+  // If not NULL, report progress during encoding.
+  WebPProgressHook progress_hook;
+
+  void* user_data;        // this field is free to be set to any value and
+                          // used during callbacks (like progress-report e.g.).
+
+  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
+
+  // 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
+};
+
+// Internal, version-checked, entry point
+WEBP_EXTERN(int) WebPPictureInitInternal(WebPPicture*, int);
+
+// Should always be called, to initialize the structure. Returns false in case
+// of version mismatch. WebPPictureInit() must have succeeded before using the
+// 'picture' object.
+// Note that, by default, use_argb is false and colorspace is WEBP_YUV420.
+static WEBP_INLINE int WebPPictureInit(WebPPicture* picture) {
+  return WebPPictureInitInternal(picture, WEBP_ENCODER_ABI_VERSION);
+}
+
+//------------------------------------------------------------------------------
+// WebPPicture utils
+
+// Convenience allocation / deallocation based on picture->width/height:
+// Allocate y/u/v buffers as per colorspace/width/height specification.
+// Note! This function will free the previous buffer if needed.
+// Returns false in case of memory error.
+WEBP_EXTERN(int) WebPPictureAlloc(WebPPicture* picture);
+
+// Release the memory allocated by WebPPictureAlloc() or WebPPictureImport*().
+// Note that this function does _not_ free the memory used by the 'picture'
+// object itself.
+// Besides memory (which is reclaimed) all other fields of 'picture' are
+// 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). 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, SSIM or LSIM distortion metric between two pictures.
+// Result is in dB, stores in result[] in the Y/U/V/Alpha/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* 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.
+// Returns false in case of memory allocation error, or if the rectangle is
+// outside of the source picture.
+// The rectangle for the view is defined by the top-left corner pixel
+// coordinates (left, top) as well as its width and height. This rectangle
+// must be fully be comprised inside the 'src' source picture. If the source
+// picture uses the YUV420 colorspace, the top and left coordinates will be
+// snapped to even values.
+WEBP_EXTERN(int) WebPPictureCrop(WebPPicture* picture,
+                                 int left, int top, int width, int height);
+
+// Extracts a view from 'src' picture into 'dst'. The rectangle for the view
+// is defined by the top-left corner pixel coordinates (left, top) as well
+// as its width and height. This rectangle must be fully be comprised inside
+// the 'src' source picture. If the source picture uses the YUV420 colorspace,
+// 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). 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,
+                                 WebPPicture* dst);
+
+// Returns true if the 'picture' is actually a view and therefore does
+// not own the memory for pixels.
+WEBP_EXTERN(int) WebPPictureIsView(const WebPPicture* picture);
+
+// Rescale a picture to new dimension width x height.
+// Now 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);
+
+// Colorspace conversion function to import RGB samples.
+// Previous buffer will be free'd, if any.
+// *rgb buffer should have a size of at least height * rgb_stride.
+// Returns false in case of memory error.
+WEBP_EXTERN(int) WebPPictureImportRGB(
+    WebPPicture* picture, const uint8_t* rgb, int rgb_stride);
+// Same, but for RGBA buffer.
+WEBP_EXTERN(int) WebPPictureImportRGBA(
+    WebPPicture* picture, const uint8_t* rgba, int rgba_stride);
+// Same, but for RGBA buffer. Imports the RGB direct from the 32-bit format
+// input buffer ignoring the alpha channel. Avoids needing to copy the data
+// to a temporary 24-bit RGB buffer to import the RGB only.
+WEBP_EXTERN(int) WebPPictureImportRGBX(
+    WebPPicture* picture, const uint8_t* rgbx, int rgbx_stride);
+
+// Variants of the above, but taking BGR(A|X) input.
+WEBP_EXTERN(int) WebPPictureImportBGR(
+    WebPPicture* picture, const uint8_t* bgr, int bgr_stride);
+WEBP_EXTERN(int) WebPPictureImportBGRA(
+    WebPPicture* picture, const uint8_t* bgra, int bgra_stride);
+WEBP_EXTERN(int) WebPPictureImportBGRX(
+    WebPPicture* picture, const uint8_t* bgrx, int bgrx_stride);
+
+// Converts picture->argb data to the YUVA format specified by 'colorspace'.
+// 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);
+
+// 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);
+
+// Converts picture->yuv to picture->argb and sets picture->use_argb to true.
+// The input format must be YUV_420 or YUV_420A.
+// Note that the use of this method is discouraged if one has access to the
+// raw ARGB samples, since using YUV420 is comparatively lossy. Also, the
+// conversion from YUV420 to ARGB incurs a small loss too.
+// 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).
+WEBP_EXTERN(void) WebPCleanupTransparentArea(WebPPicture* picture);
+
+// Scan the picture 'picture' for the presence of non fully opaque alpha values.
+// Returns true in such case. Otherwise returns false (indicating that the
+// 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
+
+// Main encoding call, after config and picture have been initialized.
+// 'picture' must be less than 16384x16384 in dimension (cf WEBP_MAX_DIMENSION),
+// and the 'config' object must be a valid one.
+// Returns false in case of error, true otherwise.
+// In case of error, picture->error_code is updated accordingly.
+// 'picture' can hold the source samples in both YUV(A) or ARGB input, depending
+// on the value of 'picture->use_argb'. It is highly recommended to use
+// the former for lossy encoding, and the latter for lossless encoding
+// (when config.lossless is true). Automatic conversion from one format to
+// another is provided but they both incur some loss.
+WEBP_EXTERN(int) WebPEncode(const WebPConfig* config, WebPPicture* picture);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_WEBP_ENCODE_H_ */
diff --git a/drivers/webp/format_constants.h b/drivers/webp/format_constants.h
new file mode 100644
index 0000000000..4c04b50c6a
--- /dev/null
+++ b/drivers/webp/format_constants.h
@@ -0,0 +1,88 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//  Internal header for constants related to WebP file format.
+//
+// Author: Urvang (urvang@google.com)
+
+#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) ((uint32_t)(a) | (b) << 8 | (c) << 16 | (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
+#define VP8_MAX_PARTITION_SIZE  (1 << 24)   // max size for token partition
+#define VP8_FRAME_HEADER_SIZE 10  // Size of the frame header within VP8 data.
+
+// VP8L related constants.
+#define VP8L_SIGNATURE_SIZE          1      // VP8L signature size.
+#define VP8L_MAGIC_BYTE              0x2f   // VP8L signature byte.
+#define VP8L_IMAGE_SIZE_BITS         14     // Number of bits used to store
+                                            // width and height.
+#define VP8L_VERSION_BITS            3      // 3 bits reserved for version.
+#define VP8L_VERSION                 0      // version 0
+#define VP8L_FRAME_HEADER_SIZE       5      // Size of the VP8L frame header.
+
+#define MAX_PALETTE_SIZE             256
+#define MAX_CACHE_BITS               11
+#define HUFFMAN_CODES_PER_META_CODE  5
+#define ARGB_BLACK                   0xff000000
+
+#define DEFAULT_CODE_LENGTH          8
+#define MAX_ALLOWED_CODE_LENGTH      15
+
+#define NUM_LITERAL_CODES            256
+#define NUM_LENGTH_CODES             24
+#define NUM_DISTANCE_CODES           40
+#define CODE_LENGTH_CODES            19
+
+#define MIN_HUFFMAN_BITS             2  // min number of Huffman bits
+#define MAX_HUFFMAN_BITS             9  // max number of Huffman bits
+
+#define TRANSFORM_PRESENT            1  // The bit to be written when next data
+                                        // to be read is a transform.
+#define NUM_TRANSFORMS               4  // Maximum number of allowed transform
+                                        // in a bitstream.
+typedef enum {
+  PREDICTOR_TRANSFORM      = 0,
+  CROSS_COLOR_TRANSFORM    = 1,
+  SUBTRACT_GREEN           = 2,
+  COLOR_INDEXING_TRANSFORM = 3
+} VP8LImageTransformType;
+
+// Alpha related constants.
+#define ALPHA_HEADER_LEN            1
+#define ALPHA_NO_COMPRESSION        0
+#define ALPHA_LOSSLESS_COMPRESSION  1
+#define ALPHA_PREPROCESSED_LEVELS   1
+
+// Mux related constants.
+#define TAG_SIZE           4     // Size of a chunk tag (e.g. "VP8L").
+#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 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 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.
+#define MAX_CHUNK_PAYLOAD (~0U - CHUNK_HEADER_SIZE - 1)
+
+#endif  /* WEBP_WEBP_FORMAT_CONSTANTS_H_ */
diff --git a/drivers/webp/image_loader_webp.cpp b/drivers/webp/image_loader_webp.cpp
new file mode 100644
index 0000000000..9d8a616556
--- /dev/null
+++ b/drivers/webp/image_loader_webp.cpp
@@ -0,0 +1,165 @@
+/*************************************************/
+/*  image_loader_webp.cpp                        */
+/*************************************************/
+/*            This file is part of:              */
+/*                GODOT ENGINE                   */
+/*************************************************/
+/*       Source code within this file is:        */
+/*  (c) 2007-2010 Juan Linietsky, Ariel Manzur   */
+/*             All Rights Reserved.              */
+/*************************************************/
+
+#include "image_loader_webp.h"
+
+#include "print_string.h"
+#include "os/os.h"
+#include "drivers/webp/decode.h"
+#include "drivers/webp/encode.h"
+#include "io/marshalls.h"
+#include <stdlib.h>
+
+static DVector<uint8_t> _webp_lossy_pack(const Image& p_image,float p_quality) {
+
+	ERR_FAIL_COND_V(p_image.empty(),DVector<uint8_t>());
+
+	Image img=p_image;
+	if (img.detect_alpha())
+		img.convert(Image::FORMAT_RGBA);
+	else
+		img.convert(Image::FORMAT_RGB);
+
+	Size2 s(img.get_width(),img.get_height());
+	DVector<uint8_t> data = img.get_data();
+	DVector<uint8_t>::Read r = data.read();
+
+	uint8_t *dst_buff=NULL;
+	size_t dst_size=0;
+	if (img.get_format()==Image::FORMAT_RGB) {
+
+		dst_size = WebPEncodeRGB(r.ptr(),s.width,s.height,3*s.width,CLAMP(p_quality*100.0,0,100.0),&dst_buff);
+	} else {
+		dst_size = WebPEncodeRGBA(r.ptr(),s.width,s.height,4*s.width,CLAMP(p_quality*100.0,0,100.0),&dst_buff);
+	}
+
+	ERR_FAIL_COND_V(dst_size==0,DVector<uint8_t>());
+	DVector<uint8_t> dst;
+	dst.resize(4+dst_size);
+	DVector<uint8_t>::Write w = dst.write();
+	w[0]='W';
+	w[1]='E';
+	w[2]='B';
+	w[3]='P';
+	copymem(&w[4],dst_buff,dst_size);
+	free(dst_buff);
+	w=DVector<uint8_t>::Write();
+	return dst;
+}
+
+static Image _webp_lossy_unpack(const DVector<uint8_t>& p_buffer) {
+
+	int size = p_buffer.size()-4;
+	ERR_FAIL_COND_V(size<=0,Image());
+	DVector<uint8_t>::Read r = p_buffer.read();
+
+	ERR_FAIL_COND_V(r[0]!='W' || r[1]!='E' || r[2]!='B' || r[3]!='P',Image());
+	WebPBitstreamFeatures features;
+	if (WebPGetFeatures(&r[4],size,&features)!=VP8_STATUS_OK) {
+		ERR_EXPLAIN("Error unpacking WEBP image:");
+		ERR_FAIL_V(Image());
+	}
+
+	//print_line("width: "+itos(features.width));
+	//print_line("height: "+itos(features.height));
+	//print_line("alpha: "+itos(features.has_alpha));
+
+	DVector<uint8_t> dst_image;
+	int datasize = features.width*features.height*(features.has_alpha?4:3);
+	dst_image.resize(datasize);
+
+	DVector<uint8_t>::Write dst_w = dst_image.write();
+
+	bool errdec=false;
+	if (features.has_alpha)	 {
+		errdec = WebPDecodeRGBAInto(&r[4],size,dst_w.ptr(),datasize,4*features.width)==NULL;
+	} else {
+		errdec = WebPDecodeRGBInto(&r[4],size,dst_w.ptr(),datasize,3*features.width)==NULL;
+
+	}
+
+	//ERR_EXPLAIN("Error decoding webp! - "+p_file);
+	ERR_FAIL_COND_V(errdec,Image());
+
+	dst_w = DVector<uint8_t>::Write();
+
+	return Image(features.width,features.height,0,features.has_alpha?Image::FORMAT_RGBA:Image::FORMAT_RGB,dst_image);
+
+}
+
+
+Error ImageLoaderWEBP::load_image(Image *p_image,FileAccess *f) {
+
+
+	uint32_t size = f->get_len();
+	DVector<uint8_t> src_image;
+	src_image.resize(size);
+
+	WebPBitstreamFeatures features;
+
+	DVector<uint8_t>::Write src_w = src_image.write();
+	f->get_buffer(src_w.ptr(),size);
+	ERR_FAIL_COND_V(f->eof_reached(), ERR_FILE_EOF);
+
+	if (WebPGetFeatures(src_w.ptr(),size,&features)!=VP8_STATUS_OK) {
+		f->close();
+		//ERR_EXPLAIN("Error decoding WEBP image: "+p_file);
+		ERR_FAIL_V(ERR_FILE_CORRUPT);
+	}
+
+	print_line("width: "+itos(features.width));
+	print_line("height: "+itos(features.height));
+	print_line("alpha: "+itos(features.has_alpha));
+
+	src_w = DVector<uint8_t>::Write();
+
+	DVector<uint8_t> dst_image;
+	int datasize = features.width*features.height*(features.has_alpha?4:3);
+	dst_image.resize(datasize);
+
+	DVector<uint8_t>::Read src_r = src_image.read();
+	DVector<uint8_t>::Write dst_w = dst_image.write();
+
+
+	bool errdec=false;
+	if (features.has_alpha)	 {
+		errdec = WebPDecodeRGBAInto(src_r.ptr(),size,dst_w.ptr(),datasize,4*features.width)==NULL;
+	} else {
+		errdec = WebPDecodeRGBInto(src_r.ptr(),size,dst_w.ptr(),datasize,3*features.width)==NULL;
+
+	}
+
+	//ERR_EXPLAIN("Error decoding webp! - "+p_file);
+	ERR_FAIL_COND_V(errdec,ERR_FILE_CORRUPT);
+
+	src_r = DVector<uint8_t>::Read();
+	dst_w = DVector<uint8_t>::Write();
+
+	*p_image = Image(features.width,features.height,0,features.has_alpha?Image::FORMAT_RGBA:Image::FORMAT_RGB,dst_image);
+
+
+	return OK;
+
+}
+
+void ImageLoaderWEBP::get_recognized_extensions(List<String> *p_extensions) const {
+	
+	p_extensions->push_back("webp");
+}
+
+
+ImageLoaderWEBP::ImageLoaderWEBP() {
+
+	Image::lossy_packer=_webp_lossy_pack;
+	Image::lossy_unpacker=_webp_lossy_unpack;
+}
+
+
diff --git a/drivers/webp/image_loader_webp.h b/drivers/webp/image_loader_webp.h
new file mode 100644
index 0000000000..8fc188cc9c
--- /dev/null
+++ b/drivers/webp/image_loader_webp.h
@@ -0,0 +1,32 @@
+/*************************************************/
+/*  image_loader_webp.h                          */
+/*************************************************/
+/*            This file is part of:              */
+/*                GODOT ENGINE                   */
+/*************************************************/
+/*       Source code within this file is:        */
+/*  (c) 2007-2010 Juan Linietsky, Ariel Manzur   */
+/*             All Rights Reserved.              */
+/*************************************************/
+
+#ifndef IMAGE_LOADER_WEBP_H
+#define IMAGE_LOADER_WEBP_H
+
+#include "io/image_loader.h"
+
+/**
+	@author Juan Linietsky <reduzio@gmail.com>
+*/
+class ImageLoaderWEBP : public ImageFormatLoader {
+
+
+public:
+
+	virtual Error load_image(Image *p_image,FileAccess *f);
+	virtual void get_recognized_extensions(List<String> *p_extensions) const;	
+	ImageLoaderWEBP();
+};
+
+
+
+#endif
diff --git a/drivers/webp/mux.h b/drivers/webp/mux.h
new file mode 100644
index 0000000000..eb57f51c64
--- /dev/null
+++ b/drivers/webp/mux.h
@@ -0,0 +1,377 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//  RIFF container manipulation 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, 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);
+*/
+
+#ifndef WEBP_WEBP_MUX_H_
+#define WEBP_WEBP_MUX_H_
+
+#include "./mux_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define WEBP_MUX_ABI_VERSION 0x0101        // 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 WebPMuxError WebPMuxError;
+// typedef enum WebPChunkId WebPChunkId;
+typedef struct WebPMux WebPMux;   // main opaque object.
+typedef struct WebPMuxFrameInfo WebPMuxFrameInfo;
+typedef struct WebPMuxAnimParams WebPMuxAnimParams;
+
+// Error codes
+typedef enum WebPMuxError {
+  WEBP_MUX_OK                 =  1,
+  WEBP_MUX_NOT_FOUND          =  0,
+  WEBP_MUX_INVALID_ARGUMENT   = -1,
+  WEBP_MUX_BAD_DATA           = -2,
+  WEBP_MUX_MEMORY_ERROR       = -3,
+  WEBP_MUX_NOT_ENOUGH_DATA    = -4
+} WebPMuxError;
+
+// IDs for different types of chunks.
+typedef enum WebPChunkId {
+  WEBP_CHUNK_VP8X,     // VP8X
+  WEBP_CHUNK_ICCP,     // ICCP
+  WEBP_CHUNK_ANIM,     // ANIM
+  WEBP_CHUNK_ANMF,     // ANMF
+  WEBP_CHUNK_FRGM,     // FRGM
+  WEBP_CHUNK_ALPHA,    // ALPH
+  WEBP_CHUNK_IMAGE,    // VP8/VP8L
+  WEBP_CHUNK_EXIF,     // EXIF
+  WEBP_CHUNK_XMP,      // XMP
+  WEBP_CHUNK_UNKNOWN,  // Other chunks.
+  WEBP_CHUNK_NIL
+} WebPChunkId;
+
+//------------------------------------------------------------------------------
+
+// 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
+
+// Internal, version-checked, entry point
+WEBP_EXTERN(WebPMux*) WebPNewInternal(int);
+
+// Creates an empty mux object.
+// Returns:
+//   A pointer to the newly created empty mux object.
+static WEBP_INLINE WebPMux* WebPMuxNew(void) {
+  return WebPNewInternal(WEBP_MUX_ABI_VERSION);
+}
+
+// Deletes the mux object.
+// Parameters:
+//   mux - (in/out) object to be deleted
+WEBP_EXTERN(void) WebPMuxDelete(WebPMux* mux);
+
+//------------------------------------------------------------------------------
+// Mux creation.
+
+// Internal, version-checked, entry point
+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 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.
+static WEBP_INLINE WebPMux* WebPMuxCreate(const WebPData* bitstream,
+                                          int copy_data) {
+  return WebPMuxCreateInternal(bitstream, copy_data, WEBP_MUX_ABI_VERSION);
+}
+
+//------------------------------------------------------------------------------
+// Non-image chunks.
+
+// 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().
+
+// 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 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, 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) WebPMuxSetChunk(
+    WebPMux* mux, const char fourcc[4], const WebPData* chunk_data,
+    int copy_data);
+
+// 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 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 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) WebPMuxGetChunk(
+    const WebPMux* mux, const char fourcc[4], WebPData* chunk_data);
+
+// Deletes the chunk with the given 'fourcc' from the mux object.
+// Parameters:
+//   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 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) WebPMuxDeleteChunk(
+    WebPMux* mux, const char fourcc[4]);
+
+//------------------------------------------------------------------------------
+// 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 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 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);
+
+// 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 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 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 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
+//   frame - (out) data of the returned frame
+// Returns:
+//   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, WebPMuxFrameInfo* frame);
+
+// 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_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);
+
+//------------------------------------------------------------------------------
+// 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 ANIM chunk is to be set/added
+//   params - (in) animation parameters.
+// Returns:
+//   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) WebPMuxSetAnimationParams(
+    WebPMux* mux, const WebPMuxAnimParams* params);
+
+// Gets the animation parameters from the mux object.
+// Parameters:
+//   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 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) WebPMuxGetAnimationParams(
+    const WebPMux* mux, WebPMuxAnimParams* params);
+
+//------------------------------------------------------------------------------
+// Misc Utilities.
+
+// 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) 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, 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) 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_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 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 mux, or num_elements is NULL.
+//   WEBP_MUX_OK - on success.
+WEBP_EXTERN(WebPMuxError) WebPMuxNumChunks(const WebPMux* mux,
+                                           WebPChunkId id, int* num_elements);
+
+// Assembles all chunks in WebP RIFF format and returns in 'assembled_data'.
+// This function also validates the mux object.
+// 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().
+// 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 mux or assembled_data is NULL.
+//   WEBP_MUX_MEMORY_ERROR - on memory allocation error.
+//   WEBP_MUX_OK - on success.
+WEBP_EXTERN(WebPMuxError) WebPMuxAssemble(WebPMux* mux,
+                                          WebPData* assembled_data);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_WEBP_MUX_H_ */
diff --git a/drivers/webp/mux/muxedit.c b/drivers/webp/mux/muxedit.c
new file mode 100644
index 0000000000..25770b3546
--- /dev/null
+++ b/drivers/webp/mux/muxedit.c
@@ -0,0 +1,652 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Set and delete APIs for mux.
+//
+// Authors: Urvang (urvang@google.com)
+//          Vikas (vikasa@google.com)
+
+#include <assert.h>
+#include "./muxi.h"
+#include "../utils/utils.h"
+
+//------------------------------------------------------------------------------
+// Life of a mux object.
+
+static void MuxInit(WebPMux* const mux) {
+  if (mux == NULL) return;
+  memset(mux, 0, sizeof(*mux));
+}
+
+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);
+    return mux;
+  }
+}
+
+// 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;
+  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);
+}
+
+//------------------------------------------------------------------------------
+// Helper method(s).
+
+// Handy MACRO, makes MuxSet() very symmetric to MuxGet().
+#define SWITCH_ID_LIST(INDEX, LIST)                                            \
+  if (idx == (INDEX)) {                                                        \
+    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, 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_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
+
+// 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;
+
+  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_frgm_bytes = (uint8_t*)malloc(frame_frgm_size);
+  if (frame_frgm_bytes == NULL) return WEBP_MUX_MEMORY_ERROR;
+
+  PutLE24(frame_frgm_bytes + 0, info->x_offset / 2);
+  PutLE24(frame_frgm_bytes + 3, info->y_offset / 2);
+
+  if (is_frame) {
+    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_frgm->bytes = frame_frgm_bytes;
+  frame_frgm->size = frame_frgm_size;
+  return WEBP_MUX_OK;
+}
+
+// Outputs image data given a bitstream. The bitstream can either be a
+// single-image WebP file or raw VP8/VP8L data.
+// Also outputs 'is_lossless' to be true if the given bitstream is lossless.
+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)) {
+    // It is NOT webp file data. Return input data as is.
+    *image = *bitstream;
+  } else {
+    // It is webp file data. Extract image data from it.
+    const WebPMuxImage* wpi;
+    WebPMux* const mux = WebPMuxCreate(bitstream, 0);
+    if (mux == NULL) return WEBP_MUX_BAD_DATA;
+    wpi = mux->images_;
+    assert(wpi != NULL && wpi->img_ != NULL);
+    *image = wpi->img_->data_;
+    if (wpi->alpha_ != NULL) {
+      *alpha = wpi->alpha_->data_;
+    }
+    WebPMuxDelete(mux);
+  }
+  *is_lossless = VP8LCheckSignature(image->bytes, image->size);
+  return WEBP_MUX_OK;
+}
+
+static WebPMuxError DeleteChunks(WebPChunk** chunk_list, uint32_t tag) {
+  WebPMuxError err = WEBP_MUX_NOT_FOUND;
+  assert(chunk_list);
+  while (*chunk_list) {
+    WebPChunk* const chunk = *chunk_list;
+    if (chunk->tag_ == tag) {
+      *chunk_list = ChunkDelete(chunk);
+      err = WEBP_MUX_OK;
+    } else {
+      chunk_list = &chunk->next_;
+    }
+  }
+  return err;
+}
+
+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;
+  return DeleteChunks(MuxGetChunkListFromId(mux, id), tag);
+}
+
+//------------------------------------------------------------------------------
+// Set API(s).
+
+WebPMuxError WebPMuxSetChunk(WebPMux* mux, const char fourcc[4],
+                             const WebPData* chunk_data, int copy_data) {
+  uint32_t tag;
+  WebPMuxError err;
+  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);
+
+  // Delete existing chunk(s) with the same 'fourcc'.
+  err = MuxDeleteAllNamedData(mux, tag);
+  if (err != WEBP_MUX_OK && err != WEBP_MUX_NOT_FOUND) return err;
+
+  // Add the given chunk.
+  return MuxSet(mux, tag, 1, chunk_data, copy_data);
+}
+
+// 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, data, copy_data, tag);
+  if (err != WEBP_MUX_OK) goto Err;
+  err = ChunkSetNth(&chunk, chunk_list, 1);
+  if (err != WEBP_MUX_OK) goto Err;
+  return WEBP_MUX_OK;
+ Err:
+  ChunkRelease(&chunk);
+  return err;
+}
+
+// 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;
+  }
+  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 WebPMuxSetImage(WebPMux* mux, const WebPData* bitstream,
+                             int copy_data) {
+  WebPMuxImage wpi;
+  WebPMuxError err;
+
+  // Sanity checks.
+  if (mux == NULL || bitstream == NULL || bitstream->bytes == NULL ||
+      bitstream->size > MAX_CHUNK_PAYLOAD) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+
+  if (mux->images_ != NULL) {
+    // Only one 'simple image' can be added in mux. So, remove present images.
+    DeleteAllImages(&mux->images_);
+  }
+
+  MuxImageInit(&wpi);
+  err = SetAlphaAndImageChunks(bitstream, copy_data, &wpi);
+  if (err != WEBP_MUX_OK) goto Err;
+
+  // Add this WebPMuxImage to mux.
+  err = MuxImagePush(&wpi, &mux->images_);
+  if (err != WEBP_MUX_OK) goto Err;
+
+  // All is well.
+  return WEBP_MUX_OK;
+
+ Err:  // Something bad happened.
+  MuxImageRelease(&wpi);
+  return err;
+}
+
+WebPMuxError WebPMuxPushFrame(WebPMux* mux, const WebPMuxFrameInfo* frame,
+                              int copy_data) {
+  WebPMuxImage wpi;
+  WebPMuxError err;
+  int is_frame;
+  const WebPData* const bitstream = &frame->bitstream;
+
+  // Sanity checks.
+  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;
+  }
+#ifndef WEBP_EXPERIMENTAL_FEATURES
+  if (frame->id == WEBP_CHUNK_FRGM) {     // disabled for now.
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+#endif
+
+  if (bitstream->bytes == NULL || bitstream->size > MAX_CHUNK_PAYLOAD) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+
+  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.
+    }
+  }
+
+  MuxImageInit(&wpi);
+  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;
+    // 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;
+  }
+
+  // Add this WebPMuxImage to mux.
+  err = MuxImagePush(&wpi, &mux->images_);
+  if (err != WEBP_MUX_OK) goto Err;
+
+  // All is well.
+  return WEBP_MUX_OK;
+
+ Err:  // Something bad happened.
+  MuxImageRelease(&wpi);
+  return err;
+}
+
+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 || params == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+  if (params->loop_count < 0 || params->loop_count >= MAX_LOOP_COUNT) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+
+  // 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;
+
+  // Set the animation parameters.
+  PutLE32(data, params->bgcolor);
+  PutLE16(data + 4, params->loop_count);
+  return MuxSet(mux, kChunks[IDX_ANIM].tag, 1, &anim, 1);
+}
+
+//------------------------------------------------------------------------------
+// Delete API(s).
+
+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) {
+  if (mux == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+  return MuxImageDeleteNth(&mux->images_, nth);
+}
+
+//------------------------------------------------------------------------------
+// Assembly of the WebP RIFF file.
+
+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 ? 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;
+}
+
+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 frame_frgm_chunk = wpi->header_;
+  WebPMuxError err;
+  assert(wpi != NULL);
+  assert(frame_frgm_chunk != NULL);
+
+  // 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.
+  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) {
+  WebPMuxImage* wpi = NULL;
+  assert(mux != NULL);
+  assert(width != NULL && height != NULL);
+
+  wpi = mux->images_;
+  assert(wpi != NULL);
+  assert(wpi->img_ != NULL);
+
+  if (wpi->next_ != NULL) {
+    int max_x = 0;
+    int max_y = 0;
+    int64_t image_area = 0;
+    // 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 = 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;
+      const int max_y_pos = y_offset + h;
+      if (err != WEBP_MUX_OK) return err;
+      assert(x_offset < MAX_POSITION_OFFSET);
+      assert(y_offset < MAX_POSITION_OFFSET);
+
+      if (max_x_pos > max_x) max_x = max_x_pos;
+      if (max_y_pos > max_y) max_y = max_y_pos;
+      image_area += w * h;
+    }
+    *width = max_x;
+    *height = 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, canvas dimensions are same as image dimensions.
+    *width = wpi->width_;
+    *height = wpi->height_;
+  }
+  return WEBP_MUX_OK;
+}
+
+// VP8X format:
+// Total Size : 10,
+// Flags  : 4 bytes,
+// Width  : 3 bytes,
+// Height : 3 bytes.
+static WebPMuxError CreateVP8XChunk(WebPMux* const mux) {
+  WebPMuxError err = WEBP_MUX_OK;
+  uint32_t flags = 0;
+  int width = 0;
+  int height = 0;
+  uint8_t data[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) {
+    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, 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) {
+    flags |= ICCP_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_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);
+  if (err != WEBP_MUX_OK) return err;
+
+  if (width <= 0 || height <= 0) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+  if (width > MAX_CANVAS_SIZE || height > MAX_CANVAS_SIZE) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+
+  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;
+  }
+
+  PutLE32(data + 0, flags);   // VP8X chunk flags.
+  PutLE24(data + 4, width - 1);   // canvas width.
+  PutLE24(data + 7, height - 1);  // canvas height.
+
+  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 single fragment or frame, 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.
+    if (frame_frag->header_ != NULL) {
+      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;
+  WebPMuxError err;
+
+  if (mux == NULL || assembled_data == NULL) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+
+  // 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 = 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);
+  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->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);
+    data = NULL;
+    size = 0;
+  }
+
+  // Finalize data.
+  assembled_data->bytes = data;
+  assembled_data->size = size;
+
+  return err;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/mux/muxi.h b/drivers/webp/mux/muxi.h
new file mode 100644
index 0000000000..277d5fba1b
--- /dev/null
+++ b/drivers/webp/mux/muxi.h
@@ -0,0 +1,230 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Internal header for mux library.
+//
+// Author: Urvang (urvang@google.com)
+
+#ifndef WEBP_MUX_MUXI_H_
+#define WEBP_MUX_MUXI_H_
+
+#include <stdlib.h>
+#include "../dec/vp8i.h"
+#include "../dec/vp8li.h"
+#include "../webp/mux.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// Defines and constants.
+
+#define MUX_MAJ_VERSION 0
+#define MUX_MIN_VERSION 2
+#define MUX_REV_VERSION 0
+
+// Chunk object.
+typedef struct WebPChunk WebPChunk;
+struct WebPChunk {
+  uint32_t        tag_;
+  int             owner_;  // True if *data_ memory is owned internally.
+                           // 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 ANMF/FRGM chunk, ALPH
+// chunk and VP8/VP8L chunk),
+typedef struct WebPMuxImage WebPMuxImage;
+struct WebPMuxImage {
+  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_;
+};
+
+// Main mux object. Stores data chunks.
+struct WebPMux {
+  WebPMuxImage*   images_;
+  WebPChunk*      iccp_;
+  WebPChunk*      exif_;
+  WebPChunk*      xmp_;
+  WebPChunk*      anim_;
+  WebPChunk*      vp8x_;
+
+  WebPChunk*  unknown_;
+};
+
+// CHUNK_INDEX enum: used for indexing within 'kChunks' (defined below) only.
+// Note: the reason for having two enums ('WebPChunkId' and 'CHUNK_INDEX') is to
+// allow two different chunks to have the same id (e.g. WebPChunkId
+// 'WEBP_CHUNK_IMAGE' can correspond to CHUNK_INDEX 'IDX_VP8' or 'IDX_VP8L').
+typedef enum {
+  IDX_VP8X = 0,
+  IDX_ICCP,
+  IDX_ANIM,
+  IDX_ANMF,
+  IDX_FRGM,
+  IDX_ALPHA,
+  IDX_VP8,
+  IDX_VP8L,
+  IDX_EXIF,
+  IDX_XMP,
+  IDX_UNKNOWN,
+
+  IDX_NIL,
+  IDX_LAST_CHUNK
+} CHUNK_INDEX;
+
+#define NIL_TAG 0x00000000u  // To signal void chunk.
+
+typedef struct {
+  uint32_t      tag;
+  WebPChunkId   id;
+  uint32_t      size;
+} ChunkInfo;
+
+extern const ChunkInfo kChunks[IDX_LAST_CHUNK];
+
+//------------------------------------------------------------------------------
+// Chunk object management.
+
+// Initialize.
+void ChunkInit(WebPChunk* const chunk);
+
+// 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_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);
+
+// Fill the chunk with the given data.
+WebPMuxError ChunkAssignData(WebPChunk* chunk, const WebPData* const data,
+                             int copy_data, uint32_t tag);
+
+// Sets 'chunk' at nth position in the 'chunk_list'.
+// nth = 0 has the special meaning "last of the 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_.
+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;
+  assert(data_size < MAX_CHUNK_PAYLOAD);
+  return SizeWithPadding(data_size);
+}
+
+// Total size of a list of chunks.
+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);
+
+//------------------------------------------------------------------------------
+// MuxImage object management.
+
+// Initialize.
+void MuxImageInit(WebPMuxImage* const wpi);
+
+// Releases image 'wpi' and returns wpi->next.
+WebPMuxImage* MuxImageRelease(WebPMuxImage* const wpi);
+
+// Delete image 'wpi' and return the next image in the list or NULL.
+// 'wpi' can be NULL.
+WebPMuxImage* MuxImageDelete(WebPMuxImage* const wpi);
+
+// 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_ANMF:
+    case WEBP_CHUNK_FRGM:
+    case WEBP_CHUNK_ALPHA:
+    case WEBP_CHUNK_IMAGE:  return 1;
+    default:        return 0;
+  }
+}
+
+// Pushes 'wpi' at the end of 'wpi_list'.
+WebPMuxError MuxImagePush(const WebPMuxImage* wpi, WebPMuxImage** wpi_list);
+
+// Delete nth image in the image list.
+WebPMuxError MuxImageDeleteNth(WebPMuxImage** wpi_list, uint32_t nth);
+
+// Get nth image in the image list.
+WebPMuxError MuxImageGetNth(const WebPMuxImage** wpi_list, uint32_t nth,
+                            WebPMuxImage** wpi);
+
+// Total size of the given image.
+size_t MuxImageDiskSize(const WebPMuxImage* const wpi);
+
+// Write out the given image into 'dst'.
+uint8_t* MuxImageEmit(const WebPMuxImage* const wpi, uint8_t* dst);
+
+//------------------------------------------------------------------------------
+// Helper methods for mux.
+
+// 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.
+WebPChunk** MuxGetChunkListFromId(const WebPMux* mux, WebPChunkId id);
+
+// Validates the given mux object.
+WebPMuxError MuxValidate(const WebPMux* const mux);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_MUX_MUXI_H_ */
diff --git a/drivers/webp/mux/muxinternal.c b/drivers/webp/mux/muxinternal.c
new file mode 100644
index 0000000000..3f992ce130
--- /dev/null
+++ b/drivers/webp/mux/muxinternal.c
@@ -0,0 +1,551 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Internal objects and utils for mux.
+//
+// Authors: Urvang (urvang@google.com)
+//          Vikas (vikasa@google.com)
+
+#include <assert.h>
+#include "./muxi.h"
+#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('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('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 }
+};
+
+//------------------------------------------------------------------------------
+
+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) {
+  assert(chunk);
+  memset(chunk, 0, sizeof(*chunk));
+  chunk->tag_ = NIL_TAG;
+}
+
+WebPChunk* ChunkRelease(WebPChunk* const chunk) {
+  WebPChunk* next;
+  if (chunk == NULL) return NULL;
+  if (chunk->owner_) {
+    WebPDataClear(&chunk->data_);
+  }
+  next = chunk->next_;
+  ChunkInit(chunk);
+  return next;
+}
+
+//------------------------------------------------------------------------------
+// Chunk misc methods.
+
+CHUNK_INDEX ChunkGetIndexFromTag(uint32_t tag) {
+  int i;
+  for (i = 0; kChunks[i].tag != NIL_TAG; ++i) {
+    if (tag == kChunks[i].tag) return (CHUNK_INDEX)i;
+  }
+  return IDX_UNKNOWN;
+}
+
+WebPChunkId ChunkGetIdFromTag(uint32_t tag) {
+  int i;
+  for (i = 0; kChunks[i].tag != NIL_TAG; ++i) {
+    if (tag == kChunks[i].tag) return kChunks[i].id;
+  }
+  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);
+}
+
+//------------------------------------------------------------------------------
+// Chunk search methods.
+
+// Returns next chunk in the chunk list with the given tag.
+static WebPChunk* ChunkSearchNextInList(WebPChunk* chunk, uint32_t tag) {
+  while (chunk != NULL && chunk->tag_ != tag) {
+    chunk = chunk->next_;
+  }
+  return chunk;
+}
+
+WebPChunk* ChunkSearchList(WebPChunk* first, uint32_t nth, uint32_t tag) {
+  uint32_t iter = nth;
+  first = ChunkSearchNextInList(first, tag);
+  if (first == NULL) return NULL;
+
+  while (--iter != 0) {
+    WebPChunk* next_chunk = ChunkSearchNextInList(first->next_, tag);
+    if (next_chunk == NULL) break;
+    first = next_chunk;
+  }
+  return ((nth > 0) && (iter > 0)) ? NULL : first;
+}
+
+// Outputs a pointer to 'prev_chunk->next_',
+//   where 'prev_chunk' is the pointer to the chunk at position (nth - 1).
+// 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 != NULL);
+  *location = chunk_list;
+
+  while (*chunk_list != NULL) {
+    WebPChunk* const cur_chunk = *chunk_list;
+    ++count;
+    if (count == nth) return 1;  // Found.
+    chunk_list = &cur_chunk->next_;
+    *location = chunk_list;
+  }
+
+  // *chunk_list is ok to be NULL if adding at last location.
+  return (nth == 0 || (count == nth - 1)) ? 1 : 0;
+}
+
+//------------------------------------------------------------------------------
+// Chunk writer methods.
+
+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_ANIM].tag) {
+    copy_data = 1;
+  }
+
+  ChunkRelease(chunk);
+
+  if (data != NULL) {
+    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(WebPChunk* chunk, WebPChunk** chunk_list,
+                         uint32_t nth) {
+  WebPChunk* new_chunk;
+
+  if (!ChunkSearchListToSet(chunk_list, nth, &chunk_list)) {
+    return WEBP_MUX_NOT_FOUND;
+  }
+
+  new_chunk = (WebPChunk*)malloc(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;
+}
+
+//------------------------------------------------------------------------------
+// Chunk deletion method(s).
+
+WebPChunk* ChunkDelete(WebPChunk* const chunk) {
+  WebPChunk* const next = ChunkRelease(chunk);
+  free(chunk);
+  return next;
+}
+
+void ChunkListDelete(WebPChunk** const chunk_list) {
+  while (*chunk_list != NULL) {
+    *chunk_list = ChunkDelete(*chunk_list);
+  }
+}
+
+//------------------------------------------------------------------------------
+// Chunk serialization methods.
+
+static uint8_t* ChunkEmit(const WebPChunk* const chunk, uint8_t* dst) {
+  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);
+  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 != NULL) {
+    dst = ChunkEmit(chunk_list, dst);
+    chunk_list = chunk_list->next_;
+  }
+  return dst;
+}
+
+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 size;
+}
+
+//------------------------------------------------------------------------------
+// Life of a MuxImage object.
+
+void MuxImageInit(WebPMuxImage* const wpi) {
+  assert(wpi);
+  memset(wpi, 0, sizeof(*wpi));
+}
+
+WebPMuxImage* MuxImageRelease(WebPMuxImage* const wpi) {
+  WebPMuxImage* next;
+  if (wpi == NULL) return NULL;
+  ChunkDelete(wpi->header_);
+  ChunkDelete(wpi->alpha_);
+  ChunkDelete(wpi->img_);
+  ChunkListDelete(&wpi->unknown_);
+
+  next = wpi->next_;
+  MuxImageInit(wpi);
+  return next;
+}
+
+//------------------------------------------------------------------------------
+// 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_) {
+    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;
+}
+
+// Outputs a pointer to 'prev_wpi->next_',
+//   where 'prev_wpi' is the pointer to the image at position (nth - 1).
+// Returns true if nth image was found.
+static int SearchImageToGetOrDelete(WebPMuxImage** wpi_list, uint32_t nth,
+                                    WebPMuxImage*** const location) {
+  uint32_t count = 0;
+  assert(wpi_list);
+  *location = wpi_list;
+
+  if (nth == 0) {
+    nth = MuxImageCount(*wpi_list, WEBP_CHUNK_NIL);
+    if (nth == 0) return 0;  // Not found.
+  }
+
+  while (*wpi_list != NULL) {
+    WebPMuxImage* const cur_wpi = *wpi_list;
+    ++count;
+    if (count == nth) return 1;  // Found.
+    wpi_list = &cur_wpi->next_;
+    *location = wpi_list;
+  }
+  return 0;  // Not found.
+}
+
+//------------------------------------------------------------------------------
+// MuxImage writer methods.
+
+WebPMuxError MuxImagePush(const WebPMuxImage* wpi, WebPMuxImage** wpi_list) {
+  WebPMuxImage* new_wpi;
+
+  while (*wpi_list != NULL) {
+    WebPMuxImage* const cur_wpi = *wpi_list;
+    if (cur_wpi->next_ == NULL) break;
+    wpi_list = &cur_wpi->next_;
+  }
+
+  new_wpi = (WebPMuxImage*)malloc(sizeof(*new_wpi));
+  if (new_wpi == NULL) return WEBP_MUX_MEMORY_ERROR;
+  *new_wpi = *wpi;
+  new_wpi->next_ = NULL;
+
+  if (*wpi_list != NULL) {
+    (*wpi_list)->next_ = new_wpi;
+  } else {
+    *wpi_list = new_wpi;
+  }
+  return WEBP_MUX_OK;
+}
+
+//------------------------------------------------------------------------------
+// MuxImage deletion methods.
+
+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);
+  return next;
+}
+
+WebPMuxError MuxImageDeleteNth(WebPMuxImage** wpi_list, uint32_t nth) {
+  assert(wpi_list);
+  if (!SearchImageToGetOrDelete(wpi_list, nth, &wpi_list)) {
+    return WEBP_MUX_NOT_FOUND;
+  }
+  *wpi_list = MuxImageDelete(*wpi_list);
+  return WEBP_MUX_OK;
+}
+
+//------------------------------------------------------------------------------
+// MuxImage reader methods.
+
+WebPMuxError MuxImageGetNth(const WebPMuxImage** wpi_list, uint32_t nth,
+                            WebPMuxImage** wpi) {
+  assert(wpi_list);
+  assert(wpi);
+  if (!SearchImageToGetOrDelete((WebPMuxImage**)wpi_list, nth,
+                                (WebPMuxImage***)&wpi_list)) {
+    return WEBP_MUX_NOT_FOUND;
+  }
+  *wpi = (WebPMuxImage*)*wpi_list;
+  return WEBP_MUX_OK;
+}
+
+//------------------------------------------------------------------------------
+// MuxImage serialization methods.
+
+// Size of an image.
+size_t MuxImageDiskSize(const WebPMuxImage* const wpi) {
+  size_t size = 0;
+  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;
+}
+
+// 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 dst + ChunkDiskSize(header);
+}
+
+uint8_t* MuxImageEmit(const WebPMuxImage* const wpi, uint8_t* dst) {
+  // Ordering of chunks to be emitted is strictly as follows:
+  // 1. ANMF/FRGM chunk (if present).
+  // 2. ALPH chunk (if present).
+  // 3. VP8/VP8L chunk.
+  assert(wpi);
+  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);
+  if (wpi->unknown_ != NULL) dst = ChunkListEmit(wpi->unknown_, dst);
+  return dst;
+}
+
+//------------------------------------------------------------------------------
+// Helper methods for mux.
+
+int MuxHasAlpha(const WebPMuxImage* images) {
+  while (images != NULL) {
+    if (images->has_alpha_) return 1;
+    images = images->next_;
+  }
+  return 0;
+}
+
+uint8_t* MuxEmitRiffHeader(uint8_t* const data, size_t size) {
+  PutLE32(data + 0, MKFOURCC('R', 'I', 'F', 'F'));
+  PutLE32(data + TAG_SIZE, (uint32_t)size - CHUNK_HEADER_SIZE);
+  assert(size == (uint32_t)size);
+  PutLE32(data + TAG_SIZE + CHUNK_SIZE_BYTES, MKFOURCC('W', 'E', 'B', 'P'));
+  return data + RIFF_HEADER_SIZE;
+}
+
+WebPChunk** MuxGetChunkListFromId(const WebPMux* mux, WebPChunkId id) {
+  assert(mux != NULL);
+  switch (id) {
+    case WEBP_CHUNK_VP8X:    return (WebPChunk**)&mux->vp8x_;
+    case WEBP_CHUNK_ICCP:    return (WebPChunk**)&mux->iccp_;
+    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_;
+  }
+}
+
+static int IsNotCompatible(int feature, int num_items) {
+  return (feature != 0) != (num_items > 0);
+}
+
+#define NO_FLAG 0
+
+// Test basic constraints:
+// retrieval, maximum number of chunks by index (use -1 to skip)
+// and feature incompatibility (use NO_FLAG to skip).
+// 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,
+                                  uint32_t vp8x_flags,
+                                  int max, int* num) {
+  const WebPMuxError err =
+      WebPMuxNumChunks(mux, kChunks[idx].id, num);
+  if (err != WEBP_MUX_OK) return err;
+  if (max > -1 && *num > max) return WEBP_MUX_INVALID_ARGUMENT;
+  if (feature != NO_FLAG && IsNotCompatible(vp8x_flags & feature, *num)) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+  return WEBP_MUX_OK;
+}
+
+WebPMuxError MuxValidate(const WebPMux* const mux) {
+  int num_iccp;
+  int num_exif;
+  int num_xmp;
+  int num_anim;
+  int num_frames;
+  int num_fragments;
+  int num_vp8x;
+  int num_images;
+  int num_alpha;
+  uint32_t flags;
+  WebPMuxError err;
+
+  // Verify mux is not NULL.
+  if (mux == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+
+  // Verify mux has at least one image.
+  if (mux->images_ == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+
+  err = WebPMuxGetFeatures(mux, &flags);
+  if (err != WEBP_MUX_OK) return err;
+
+  // At most one color profile chunk.
+  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_XMP, XMP_FLAG, flags, 1, &num_xmp);
+  if (err != WEBP_MUX_OK) return err;
+
+  // 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_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_anim == 0 || num_frames == 0)) {
+      return WEBP_MUX_INVALID_ARGUMENT;
+    }
+    if (!has_animation && (num_anim == 1 || num_frames > 0)) {
+      return WEBP_MUX_INVALID_ARGUMENT;
+    }
+  }
+
+  // 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
+  // mux->images_.
+  err = ValidateChunk(mux, IDX_VP8X, NO_FLAG, flags, 1, &num_vp8x);
+  if (err != WEBP_MUX_OK) return err;
+  err = ValidateChunk(mux, IDX_VP8, NO_FLAG, flags, -1, &num_images);
+  if (err != WEBP_MUX_OK) return err;
+  if (num_vp8x == 0 && num_images != 1) return WEBP_MUX_INVALID_ARGUMENT;
+
+  // ALPHA_FLAG & alpha chunk(s) are consistent.
+  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_fragments & num_images are consistent.
+  if (num_fragments > 0 && num_images != num_fragments) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+
+  return WEBP_MUX_OK;
+}
+
+#undef NO_FLAG
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/mux/muxread.c b/drivers/webp/mux/muxread.c
new file mode 100644
index 0000000000..6003a25b71
--- /dev/null
+++ b/drivers/webp/mux/muxread.c
@@ -0,0 +1,540 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Read APIs for mux.
+//
+// Authors: Urvang (urvang@google.com)
+//          Vikas (vikasa@google.com)
+
+#include <assert.h>
+#include "./muxi.h"
+#include "../utils/utils.h"
+
+//------------------------------------------------------------------------------
+// Helper method(s).
+
+// Handy MACRO.
+#define SWITCH_ID_LIST(INDEX, LIST)                                           \
+  if (idx == (INDEX)) {                                                       \
+    const WebPChunk* const chunk = ChunkSearchList((LIST), nth,               \
+                                                   kChunks[(INDEX)].tag);     \
+    if (chunk) {                                                              \
+      *data = chunk->data_;                                                   \
+      return WEBP_MUX_OK;                                                     \
+    } else {                                                                  \
+      return WEBP_MUX_NOT_FOUND;                                              \
+    }                                                                         \
+  }
+
+static WebPMuxError MuxGet(const WebPMux* const mux, CHUNK_INDEX idx,
+                           uint32_t nth, WebPData* const data) {
+  assert(mux != NULL);
+  assert(!IsWPI(kChunks[idx].id));
+  WebPDataInit(data);
+
+  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 WEBP_MUX_NOT_FOUND;
+}
+#undef SWITCH_ID_LIST
+
+// Fill the chunk with the given data (includes chunk header bytes), after some
+// verifications.
+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;
+  chunk_size = GetLE32(data + TAG_SIZE);
+
+  {
+    const size_t chunk_disk_size = SizeWithPadding(chunk_size);
+    if (chunk_disk_size > riff_size) return WEBP_MUX_BAD_DATA;
+    if (chunk_disk_size > data_size) return WEBP_MUX_NOT_ENOUGH_DATA;
+  }
+
+  // Data assignment.
+  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.
+
+WebPMux* WebPMuxCreateInternal(const WebPData* bitstream, int copy_data,
+                               int version) {
+  size_t riff_size;
+  uint32_t tag;
+  const uint8_t* end;
+  WebPMux* mux = NULL;
+  WebPMuxImage* wpi = NULL;
+  const uint8_t* data;
+  size_t size;
+  WebPChunk chunk;
+  ChunkInit(&chunk);
+
+  // Sanity checks.
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_MUX_ABI_VERSION)) {
+    return NULL;  // version mismatch
+  }
+  if (bitstream == NULL) return NULL;
+
+  data = bitstream->bytes;
+  size = bitstream->size;
+
+  if (data == NULL) return NULL;
+  if (size < RIFF_HEADER_SIZE) return NULL;
+  if (GetLE32(data + 0) != MKFOURCC('R', 'I', 'F', 'F') ||
+      GetLE32(data + CHUNK_HEADER_SIZE) != MKFOURCC('W', 'E', 'B', 'P')) {
+    return NULL;
+  }
+
+  mux = WebPMuxNew();
+  if (mux == NULL) return NULL;
+
+  if (size < RIFF_HEADER_SIZE + TAG_SIZE) goto Err;
+
+  tag = GetLE32(data + RIFF_HEADER_SIZE);
+  if (tag != kChunks[IDX_VP8].tag &&
+      tag != kChunks[IDX_VP8L].tag &&
+      tag != kChunks[IDX_VP8X].tag) {
+    goto Err;  // First chunk should be VP8, VP8L or VP8X.
+  }
+
+  riff_size = SizeWithPadding(GetLE32(data + TAG_SIZE));
+  if (riff_size > MAX_CHUNK_PAYLOAD || riff_size > size) {
+    goto Err;
+  } else {
+    if (riff_size < size) {  // Redundant data after last chunk.
+      size = riff_size;  // To make sure we don't read any data beyond mux_size.
+    }
+  }
+
+  end = data + size;
+  data += RIFF_HEADER_SIZE;
+  size -= RIFF_HEADER_SIZE;
+
+  wpi = (WebPMuxImage*)malloc(sizeof(*wpi));
+  if (wpi == NULL) goto Err;
+  MuxImageInit(wpi);
+
+  // Loop over chunks.
+  while (data != end) {
+    size_t data_size;
+    WebPChunkId id;
+    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_);
+    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.
+        break;
+      case WEBP_CHUNK_ANMF:
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+      case WEBP_CHUNK_FRGM:
+#endif
+        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;
+        break;
+    }
+    data += data_size;
+    size -= data_size;
+    ChunkInit(&chunk);
+  }
+
+  // Validate mux if complete.
+  if (MuxValidate(mux) != WEBP_MUX_OK) goto Err;
+
+  MuxImageDelete(wpi);
+  return mux;  // All OK;
+
+ Err:  // Something bad happened.
+  ChunkRelease(&chunk);
+  MuxImageDelete(wpi);
+  WebPMuxDelete(mux);
+  return NULL;
+}
+
+//------------------------------------------------------------------------------
+// Get API(s).
+
+// 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;
+  }
+}
+
+// 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.
+  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 {  // Single image case.
+    const WebPMuxImage* const wpi = mux->images_;
+    WebPMuxError err = ValidateForSingleImage(mux);
+    if (err != WEBP_MUX_OK) return err;
+    assert(wpi != NULL);
+    w = wpi->width_;
+    h = wpi->height_;
+    if (wpi->has_alpha_) f |= ALPHA_FLAG;
+  }
+  if (w * (uint64_t)h >= MAX_IMAGE_AREA) return WEBP_MUX_BAD_DATA;
+
+  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;
+  assert(width >= 1 && height >= 1);
+  assert(width <= MAX_CANVAS_SIZE && height <= MAX_CANVAS_SIZE);
+  assert(width * (uint64_t)height < MAX_IMAGE_AREA);
+  PutLE32(dst, MKFOURCC('V', 'P', '8', 'X'));
+  PutLE32(dst + TAG_SIZE, VP8X_CHUNK_SIZE);
+  PutLE32(dst + CHUNK_HEADER_SIZE, flags);
+  PutLE24(dst + CHUNK_HEADER_SIZE + 4, width - 1);
+  PutLE24(dst + CHUNK_HEADER_SIZE + 7, height - 1);
+  return dst + vp8x_size;
+}
+
+// Assemble a single image WebP bitstream from 'wpi'.
+static WebPMuxError SynthesizeBitstream(const WebPMuxImage* const wpi,
+                                        WebPData* const bitstream) {
+  uint8_t* dst;
+
+  // Allocate data.
+  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 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);
+  if (data == NULL) return WEBP_MUX_MEMORY_ERROR;
+
+  // Main RIFF header.
+  dst = MuxEmitRiffHeader(data, size);
+
+  if (need_vp8x) {
+    dst = EmitVP8XChunk(dst, wpi->width_, wpi->height_, ALPHA_FLAG);  // VP8X.
+    dst = ChunkListEmit(wpi->alpha_, dst);       // ALPH.
+  }
+
+  // Bitstream.
+  dst = ChunkListEmit(wpi->img_, dst);
+  assert(dst == data + size);
+
+  // Output.
+  bitstream->bytes = data;
+  bitstream->size = size;
+  return WEBP_MUX_OK;
+}
+
+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;
+  }
+  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;
+  }
+}
+
+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);
+}
+
+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;
+#ifndef WEBP_EXPERIMENTAL_FEATURES
+  if (!is_frame) return WEBP_MUX_INVALID_ARGUMENT;
+#endif
+  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);
+}
+
+WebPMuxError WebPMuxGetFrame(
+    const WebPMux* mux, uint32_t nth, WebPMuxFrameInfo* frame) {
+  WebPMuxError err;
+  WebPMuxImage* wpi;
+
+  // Sanity checks.
+  if (mux == NULL || frame == NULL) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+
+  // Get the nth WebPMuxImage.
+  err = MuxImageGetNth((const WebPMuxImage**)&mux->images_, nth, &wpi);
+  if (err != WEBP_MUX_OK) return err;
+
+  // Get frame info.
+  if (wpi->header_ == NULL) {
+    return MuxGetImageInternal(wpi, frame);
+  } else {
+    return MuxGetFrameFragmentInternal(wpi, frame);
+  }
+}
+
+WebPMuxError WebPMuxGetAnimationParams(const WebPMux* mux,
+                                       WebPMuxAnimParams* params) {
+  WebPData anim;
+  WebPMuxError err;
+
+  if (mux == NULL || params == NULL) return WEBP_MUX_INVALID_ARGUMENT;
+
+  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);
+
+  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 (CHUNK_INDEX)i;
+  }
+  return IDX_NIL;
+}
+
+// Count number of chunks matching 'tag' in the 'chunk_list'.
+// If tag == NIL_TAG, any tag will be matched.
+static int CountChunks(const WebPChunk* const chunk_list, uint32_t tag) {
+  int count = 0;
+  const WebPChunk* current;
+  for (current = chunk_list; current != NULL; current = current->next_) {
+    if (tag == NIL_TAG || current->tag_ == tag) {
+      count++;  // Count chunks whose tags match.
+    }
+  }
+  return count;
+}
+
+WebPMuxError WebPMuxNumChunks(const WebPMux* mux,
+                              WebPChunkId id, int* num_elements) {
+  if (mux == NULL || num_elements == NULL) {
+    return WEBP_MUX_INVALID_ARGUMENT;
+  }
+
+  if (IsWPI(id)) {
+    *num_elements = MuxImageCount(mux->images_, id);
+  } else {
+    WebPChunk* const* chunk_list = MuxGetChunkListFromId(mux, id);
+    const CHUNK_INDEX idx = ChunkGetIndexFromId(id);
+    *num_elements = CountChunks(*chunk_list, kChunks[idx].tag);
+  }
+
+  return WEBP_MUX_OK;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/mux_types.h b/drivers/webp/mux_types.h
new file mode 100644
index 0000000000..c94043a3c0
--- /dev/null
+++ b/drivers/webp/mux_types.h
@@ -0,0 +1,97 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Data-types common to the mux and demux libraries.
+//
+// Author: Urvang (urvang@google.com)
+
+#ifndef WEBP_WEBP_MUX_TYPES_H_
+#define WEBP_WEBP_MUX_TYPES_H_
+
+#include <stdlib.h>  // free()
+#include <string.h>  // memset()
+#include "./types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Note: forward declaring enumerations is not allowed in (strict) C and C++,
+// the types are left here for reference.
+// typedef enum WebPFeatureFlags WebPFeatureFlags;
+// typedef enum WebPMuxAnimDispose WebPMuxAnimDispose;
+// typedef enum WebPMuxAnimBlend WebPMuxAnimBlend;
+typedef struct WebPData WebPData;
+
+// VP8X Feature Flags.
+typedef enum WebPFeatureFlags {
+  FRAGMENTS_FLAG  = 0x00000001,
+  ANIMATION_FLAG  = 0x00000002,
+  XMP_FLAG        = 0x00000004,
+  EXIF_FLAG       = 0x00000008,
+  ALPHA_FLAG      = 0x00000010,
+  ICCP_FLAG       = 0x00000020
+} WebPFeatureFlags;
+
+// Dispose method (animation only). Indicates how the area used by the current
+// frame is to be treated before rendering the next frame on the canvas.
+typedef enum WebPMuxAnimDispose {
+  WEBP_MUX_DISPOSE_NONE,       // Do not dispose.
+  WEBP_MUX_DISPOSE_BACKGROUND  // Dispose to background color.
+} WebPMuxAnimDispose;
+
+// Blend operation (animation only). Indicates how transparent pixels of the
+// current frame are blended with those of the previous canvas.
+typedef enum WebPMuxAnimBlend {
+  WEBP_MUX_BLEND,              // Blend.
+  WEBP_MUX_NO_BLEND            // Do not blend.
+} WebPMuxAnimBlend;
+
+// Data type used to describe 'raw' data, e.g., chunk data
+// (ICC profile, metadata) and WebP compressed image data.
+struct WebPData {
+  const uint8_t* bytes;
+  size_t size;
+};
+
+// Initializes the contents of the 'webp_data' object with default values.
+static WEBP_INLINE void WebPDataInit(WebPData* webp_data) {
+  if (webp_data != NULL) {
+    memset(webp_data, 0, sizeof(*webp_data));
+  }
+}
+
+// Clears the contents of the 'webp_data' object by calling free(). Does not
+// deallocate the object itself.
+static WEBP_INLINE void WebPDataClear(WebPData* webp_data) {
+  if (webp_data != NULL) {
+    free((void*)webp_data->bytes);
+    WebPDataInit(webp_data);
+  }
+}
+
+// Allocates necessary storage for 'dst' and copies the contents of 'src'.
+// Returns true on success.
+static WEBP_INLINE 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;
+  }
+  return 1;
+}
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_WEBP_MUX_TYPES_H_ */
diff --git a/drivers/webp/types.h b/drivers/webp/types.h
new file mode 100644
index 0000000000..568d1f263f
--- /dev/null
+++ b/drivers/webp/types.h
@@ -0,0 +1,47 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+//  Common types
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_WEBP_TYPES_H_
+#define WEBP_WEBP_TYPES_H_
+
+#include <stddef.h>  // for size_t
+
+#ifndef _MSC_VER
+#include <inttypes.h>
+#ifdef __STRICT_ANSI__
+#define WEBP_INLINE
+#else  /* __STRICT_ANSI__ */
+#define WEBP_INLINE inline
+#endif
+#else
+typedef signed   char int8_t;
+typedef unsigned char uint8_t;
+typedef signed   short int16_t;
+typedef unsigned short uint16_t;
+typedef signed   int int32_t;
+typedef unsigned int uint32_t;
+typedef unsigned long long int uint64_t;
+typedef long long int int64_t;
+#define WEBP_INLINE __forceinline
+#endif  /* _MSC_VER */
+
+#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
+#endif  /* WEBP_EXTERN */
+
+// Macro to check ABI compatibility (same major revision number)
+#define WEBP_ABI_IS_INCOMPATIBLE(a, b) (((a) >> 8) != ((b) >> 8))
+
+#endif  /* WEBP_WEBP_TYPES_H_ */
diff --git a/drivers/webp/utils/alpha_processing.c b/drivers/webp/utils/alpha_processing.c
new file mode 100644
index 0000000000..7362ff94a5
--- /dev/null
+++ b/drivers/webp/utils/alpha_processing.c
@@ -0,0 +1,196 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Utilities for processing transparent channel.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include "./alpha_processing.h"
+
+// Tables can be faster on some platform but incur some extra binary size (~2k).
+// #define USE_TABLES_FOR_ALPHA_MULT
+
+// -----------------------------------------------------------------------------
+
+#define MFIX 24    // 24bit fixed-point arithmetic
+#define HALF ((1u << MFIX) >> 1)
+#define KINV_255 ((1u << MFIX) / 255u)
+
+static uint32_t Mult(uint8_t x, uint32_t mult) {
+  const uint32_t v = (x * mult + HALF) >> MFIX;
+  assert(v <= 255);  // <- 24bit precision is enough to ensure that.
+  return v;
+}
+
+#ifdef USE_TABLES_FOR_ALPHA_MULT
+
+static const uint32_t kMultTables[2][256] = {
+  {    // (255u << MFIX) / alpha
+    0x00000000, 0xff000000, 0x7f800000, 0x55000000, 0x3fc00000, 0x33000000,
+    0x2a800000, 0x246db6db, 0x1fe00000, 0x1c555555, 0x19800000, 0x172e8ba2,
+    0x15400000, 0x139d89d8, 0x1236db6d, 0x11000000, 0x0ff00000, 0x0f000000,
+    0x0e2aaaaa, 0x0d6bca1a, 0x0cc00000, 0x0c249249, 0x0b9745d1, 0x0b1642c8,
+    0x0aa00000, 0x0a333333, 0x09cec4ec, 0x0971c71c, 0x091b6db6, 0x08cb08d3,
+    0x08800000, 0x0839ce73, 0x07f80000, 0x07ba2e8b, 0x07800000, 0x07492492,
+    0x07155555, 0x06e45306, 0x06b5e50d, 0x0689d89d, 0x06600000, 0x063831f3,
+    0x06124924, 0x05ee23b8, 0x05cba2e8, 0x05aaaaaa, 0x058b2164, 0x056cefa8,
+    0x05500000, 0x05343eb1, 0x05199999, 0x05000000, 0x04e76276, 0x04cfb2b7,
+    0x04b8e38e, 0x04a2e8ba, 0x048db6db, 0x0479435e, 0x04658469, 0x045270d0,
+    0x04400000, 0x042e29f7, 0x041ce739, 0x040c30c3, 0x03fc0000, 0x03ec4ec4,
+    0x03dd1745, 0x03ce540f, 0x03c00000, 0x03b21642, 0x03a49249, 0x03976fc6,
+    0x038aaaaa, 0x037e3f1f, 0x03722983, 0x03666666, 0x035af286, 0x034fcace,
+    0x0344ec4e, 0x033a5440, 0x03300000, 0x0325ed09, 0x031c18f9, 0x0312818a,
+    0x03092492, 0x03000000, 0x02f711dc, 0x02ee5846, 0x02e5d174, 0x02dd7baf,
+    0x02d55555, 0x02cd5cd5, 0x02c590b2, 0x02bdef7b, 0x02b677d4, 0x02af286b,
+    0x02a80000, 0x02a0fd5c, 0x029a1f58, 0x029364d9, 0x028ccccc, 0x0286562d,
+    0x02800000, 0x0279c952, 0x0273b13b, 0x026db6db, 0x0267d95b, 0x026217ec,
+    0x025c71c7, 0x0256e62a, 0x0251745d, 0x024c1bac, 0x0246db6d, 0x0241b2f9,
+    0x023ca1af, 0x0237a6f4, 0x0232c234, 0x022df2df, 0x02293868, 0x02249249,
+    0x02200000, 0x021b810e, 0x021714fb, 0x0212bb51, 0x020e739c, 0x020a3d70,
+    0x02061861, 0x02020408, 0x01fe0000, 0x01fa0be8, 0x01f62762, 0x01f25213,
+    0x01ee8ba2, 0x01ead3ba, 0x01e72a07, 0x01e38e38, 0x01e00000, 0x01dc7f10,
+    0x01d90b21, 0x01d5a3e9, 0x01d24924, 0x01cefa8d, 0x01cbb7e3, 0x01c880e5,
+    0x01c55555, 0x01c234f7, 0x01bf1f8f, 0x01bc14e5, 0x01b914c1, 0x01b61eed,
+    0x01b33333, 0x01b05160, 0x01ad7943, 0x01aaaaaa, 0x01a7e567, 0x01a5294a,
+    0x01a27627, 0x019fcbd2, 0x019d2a20, 0x019a90e7, 0x01980000, 0x01957741,
+    0x0192f684, 0x01907da4, 0x018e0c7c, 0x018ba2e8, 0x018940c5, 0x0186e5f0,
+    0x01849249, 0x018245ae, 0x01800000, 0x017dc11f, 0x017b88ee, 0x0179574e,
+    0x01772c23, 0x01750750, 0x0172e8ba, 0x0170d045, 0x016ebdd7, 0x016cb157,
+    0x016aaaaa, 0x0168a9b9, 0x0166ae6a, 0x0164b8a7, 0x0162c859, 0x0160dd67,
+    0x015ef7bd, 0x015d1745, 0x015b3bea, 0x01596596, 0x01579435, 0x0155c7b4,
+    0x01540000, 0x01523d03, 0x01507eae, 0x014ec4ec, 0x014d0fac, 0x014b5edc,
+    0x0149b26c, 0x01480a4a, 0x01466666, 0x0144c6af, 0x01432b16, 0x0141938b,
+    0x01400000, 0x013e7063, 0x013ce4a9, 0x013b5cc0, 0x0139d89d, 0x01385830,
+    0x0136db6d, 0x01356246, 0x0133ecad, 0x01327a97, 0x01310bf6, 0x012fa0be,
+    0x012e38e3, 0x012cd459, 0x012b7315, 0x012a150a, 0x0128ba2e, 0x01276276,
+    0x01260dd6, 0x0124bc44, 0x01236db6, 0x01222222, 0x0120d97c, 0x011f93bc,
+    0x011e50d7, 0x011d10c4, 0x011bd37a, 0x011a98ef, 0x0119611a, 0x01182bf2,
+    0x0116f96f, 0x0115c988, 0x01149c34, 0x0113716a, 0x01124924, 0x01112358,
+    0x01100000, 0x010edf12, 0x010dc087, 0x010ca458, 0x010b8a7d, 0x010a72f0,
+    0x01095da8, 0x01084a9f, 0x010739ce, 0x01062b2e, 0x01051eb8, 0x01041465,
+    0x01030c30, 0x01020612, 0x01010204, 0x01000000 },
+  {   // alpha * KINV_255
+    0x00000000, 0x00010101, 0x00020202, 0x00030303, 0x00040404, 0x00050505,
+    0x00060606, 0x00070707, 0x00080808, 0x00090909, 0x000a0a0a, 0x000b0b0b,
+    0x000c0c0c, 0x000d0d0d, 0x000e0e0e, 0x000f0f0f, 0x00101010, 0x00111111,
+    0x00121212, 0x00131313, 0x00141414, 0x00151515, 0x00161616, 0x00171717,
+    0x00181818, 0x00191919, 0x001a1a1a, 0x001b1b1b, 0x001c1c1c, 0x001d1d1d,
+    0x001e1e1e, 0x001f1f1f, 0x00202020, 0x00212121, 0x00222222, 0x00232323,
+    0x00242424, 0x00252525, 0x00262626, 0x00272727, 0x00282828, 0x00292929,
+    0x002a2a2a, 0x002b2b2b, 0x002c2c2c, 0x002d2d2d, 0x002e2e2e, 0x002f2f2f,
+    0x00303030, 0x00313131, 0x00323232, 0x00333333, 0x00343434, 0x00353535,
+    0x00363636, 0x00373737, 0x00383838, 0x00393939, 0x003a3a3a, 0x003b3b3b,
+    0x003c3c3c, 0x003d3d3d, 0x003e3e3e, 0x003f3f3f, 0x00404040, 0x00414141,
+    0x00424242, 0x00434343, 0x00444444, 0x00454545, 0x00464646, 0x00474747,
+    0x00484848, 0x00494949, 0x004a4a4a, 0x004b4b4b, 0x004c4c4c, 0x004d4d4d,
+    0x004e4e4e, 0x004f4f4f, 0x00505050, 0x00515151, 0x00525252, 0x00535353,
+    0x00545454, 0x00555555, 0x00565656, 0x00575757, 0x00585858, 0x00595959,
+    0x005a5a5a, 0x005b5b5b, 0x005c5c5c, 0x005d5d5d, 0x005e5e5e, 0x005f5f5f,
+    0x00606060, 0x00616161, 0x00626262, 0x00636363, 0x00646464, 0x00656565,
+    0x00666666, 0x00676767, 0x00686868, 0x00696969, 0x006a6a6a, 0x006b6b6b,
+    0x006c6c6c, 0x006d6d6d, 0x006e6e6e, 0x006f6f6f, 0x00707070, 0x00717171,
+    0x00727272, 0x00737373, 0x00747474, 0x00757575, 0x00767676, 0x00777777,
+    0x00787878, 0x00797979, 0x007a7a7a, 0x007b7b7b, 0x007c7c7c, 0x007d7d7d,
+    0x007e7e7e, 0x007f7f7f, 0x00808080, 0x00818181, 0x00828282, 0x00838383,
+    0x00848484, 0x00858585, 0x00868686, 0x00878787, 0x00888888, 0x00898989,
+    0x008a8a8a, 0x008b8b8b, 0x008c8c8c, 0x008d8d8d, 0x008e8e8e, 0x008f8f8f,
+    0x00909090, 0x00919191, 0x00929292, 0x00939393, 0x00949494, 0x00959595,
+    0x00969696, 0x00979797, 0x00989898, 0x00999999, 0x009a9a9a, 0x009b9b9b,
+    0x009c9c9c, 0x009d9d9d, 0x009e9e9e, 0x009f9f9f, 0x00a0a0a0, 0x00a1a1a1,
+    0x00a2a2a2, 0x00a3a3a3, 0x00a4a4a4, 0x00a5a5a5, 0x00a6a6a6, 0x00a7a7a7,
+    0x00a8a8a8, 0x00a9a9a9, 0x00aaaaaa, 0x00ababab, 0x00acacac, 0x00adadad,
+    0x00aeaeae, 0x00afafaf, 0x00b0b0b0, 0x00b1b1b1, 0x00b2b2b2, 0x00b3b3b3,
+    0x00b4b4b4, 0x00b5b5b5, 0x00b6b6b6, 0x00b7b7b7, 0x00b8b8b8, 0x00b9b9b9,
+    0x00bababa, 0x00bbbbbb, 0x00bcbcbc, 0x00bdbdbd, 0x00bebebe, 0x00bfbfbf,
+    0x00c0c0c0, 0x00c1c1c1, 0x00c2c2c2, 0x00c3c3c3, 0x00c4c4c4, 0x00c5c5c5,
+    0x00c6c6c6, 0x00c7c7c7, 0x00c8c8c8, 0x00c9c9c9, 0x00cacaca, 0x00cbcbcb,
+    0x00cccccc, 0x00cdcdcd, 0x00cecece, 0x00cfcfcf, 0x00d0d0d0, 0x00d1d1d1,
+    0x00d2d2d2, 0x00d3d3d3, 0x00d4d4d4, 0x00d5d5d5, 0x00d6d6d6, 0x00d7d7d7,
+    0x00d8d8d8, 0x00d9d9d9, 0x00dadada, 0x00dbdbdb, 0x00dcdcdc, 0x00dddddd,
+    0x00dedede, 0x00dfdfdf, 0x00e0e0e0, 0x00e1e1e1, 0x00e2e2e2, 0x00e3e3e3,
+    0x00e4e4e4, 0x00e5e5e5, 0x00e6e6e6, 0x00e7e7e7, 0x00e8e8e8, 0x00e9e9e9,
+    0x00eaeaea, 0x00ebebeb, 0x00ececec, 0x00ededed, 0x00eeeeee, 0x00efefef,
+    0x00f0f0f0, 0x00f1f1f1, 0x00f2f2f2, 0x00f3f3f3, 0x00f4f4f4, 0x00f5f5f5,
+    0x00f6f6f6, 0x00f7f7f7, 0x00f8f8f8, 0x00f9f9f9, 0x00fafafa, 0x00fbfbfb,
+    0x00fcfcfc, 0x00fdfdfd, 0x00fefefe, 0x00ffffff }
+};
+
+static WEBP_INLINE uint32_t GetScale(uint32_t a, int inverse) {
+  return kMultTables[!inverse][a];
+}
+
+#else
+
+static WEBP_INLINE uint32_t GetScale(uint32_t a, int inverse) {
+  return inverse ? (255u << MFIX) / a : a * KINV_255;
+}
+
+#endif    // USE_TABLES_FOR_ALPHA_MULT
+
+void WebPMultARGBRow(uint32_t* const ptr, int width, int inverse) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    const uint32_t argb = ptr[x];
+    if (argb < 0xff000000u) {      // alpha < 255
+      if (argb <= 0x00ffffffu) {   // alpha == 0
+        ptr[x] = 0;
+      } else {
+        const uint32_t alpha = (argb >> 24) & 0xff;
+        const uint32_t scale = GetScale(alpha, inverse);
+        uint32_t out = argb & 0xff000000u;
+        out |= Mult(argb >>  0, scale) <<  0;
+        out |= Mult(argb >>  8, scale) <<  8;
+        out |= Mult(argb >> 16, scale) << 16;
+        ptr[x] = out;
+      }
+    }
+  }
+}
+
+void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows,
+                      int inverse) {
+  int n;
+  for (n = 0; n < num_rows; ++n) {
+    WebPMultARGBRow((uint32_t*)ptr, width, inverse);
+    ptr += stride;
+  }
+}
+
+void WebPMultRow(uint8_t* const ptr, const uint8_t* const alpha,
+                 int width, int inverse) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    const uint32_t a = alpha[x];
+    if (a != 255) {
+      if (a == 0) {
+        ptr[x] = 0;
+      } else {
+        const uint32_t scale = GetScale(a, inverse);
+        ptr[x] = Mult(ptr[x], scale);
+      }
+    }
+  }
+}
+
+void WebPMultRows(uint8_t* ptr, int stride,
+                  const uint8_t* alpha, int alpha_stride,
+                  int width, int num_rows, int inverse) {
+  int n;
+  for (n = 0; n < num_rows; ++n) {
+    WebPMultRow(ptr, alpha, width, inverse);
+    ptr += stride;
+    alpha += alpha_stride;
+  }
+}
+
+#undef KINV_255
+#undef HALF
+#undef MFIX
+
diff --git a/drivers/webp/utils/alpha_processing.h b/drivers/webp/utils/alpha_processing.h
new file mode 100644
index 0000000000..80e1ae45dd
--- /dev/null
+++ b/drivers/webp/utils/alpha_processing.h
@@ -0,0 +1,46 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Utilities for processing transparent channel.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_UTILS_ALPHA_PROCESSING_H_
+#define WEBP_UTILS_ALPHA_PROCESSING_H_
+
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// 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.
+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.
+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);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif    // WEBP_UTILS_ALPHA_PROCESSING_H_
diff --git a/drivers/webp/utils/bit_reader.c b/drivers/webp/utils/bit_reader.c
new file mode 100644
index 0000000000..bfa4d7d2e2
--- /dev/null
+++ b/drivers/webp/utils/bit_reader.c
@@ -0,0 +1,207 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Boolean decoder
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./bit_reader.h"
+
+#ifndef USE_RIGHT_JUSTIFY
+#define MK(X) (((range_t)(X) << (BITS)) | (MASK))
+#else
+#define MK(X) ((range_t)(X))
+#endif
+
+//------------------------------------------------------------------------------
+// VP8BitReader
+
+void VP8InitBitReader(VP8BitReader* const br,
+                      const uint8_t* const start, const uint8_t* const end) {
+  assert(br != NULL);
+  assert(start != NULL);
+  assert(start <= end);
+  br->range_   = MK(255 - 1);
+  br->buf_     = start;
+  br->buf_end_ = end;
+  br->value_   = 0;
+  br->bits_    = -8;   // to load the very first 8bits
+  br->eof_     = 0;
+}
+
+const uint8_t kVP8Log2Range[128] = {
+     7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4,
+  3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  0
+};
+
+// range = (range << kVP8Log2Range[range]) + trailing 1's
+const range_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)
+};
+
+#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_) {
+#ifndef USE_RIGHT_JUSTIFY
+    br->value_ |= (bit_t)(*br->buf_++) << ((BITS) - 8 - br->bits_);
+#else
+    br->value_ = (bit_t)(*br->buf_++) | (br->value_ << 8);
+#endif
+    br->bits_ += 8;
+  } else if (!br->eof_) {
+#ifdef USE_RIGHT_JUSTIFY
+    // These are not strictly needed, but it makes the behaviour
+    // consistent for both USE_RIGHT_JUSTIFY and !USE_RIGHT_JUSTIFY.
+    br->value_ <<= 8;
+    br->bits_ += 8;
+#endif
+    br->eof_ = 1;
+  }
+}
+
+//------------------------------------------------------------------------------
+// Higher-level calls
+
+uint32_t VP8GetValue(VP8BitReader* const br, int bits) {
+  uint32_t v = 0;
+  while (bits-- > 0) {
+    v |= VP8GetBit(br, 0x80) << bits;
+  }
+  return v;
+}
+
+int32_t VP8GetSignedValue(VP8BitReader* const br, int bits) {
+  const int value = VP8GetValue(br, bits);
+  return VP8Get(br) ? -value : value;
+}
+
+//------------------------------------------------------------------------------
+// VP8LBitReader
+
+#define MAX_NUM_BIT_READ 25
+
+#define LBITS 64      // Number of bits prefetched.
+#define WBITS 32      // Minimum number of bytes needed after VP8LFillBitWindow.
+#define LOG8_WBITS 4  // Number of bytes needed to store 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
+};
+
+void VP8LInitBitReader(VP8LBitReader* const br,
+                       const uint8_t* const start,
+                       size_t length) {
+  size_t i;
+  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_ |= ((vp8l_val_t)br->buf_[br->pos_]) << (8 * i);
+    ++br->pos_;
+  }
+}
+
+void VP8LBitReaderSetBuffer(VP8LBitReader* const br,
+                            const uint8_t* const buf, size_t len) {
+  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;
+}
+
+// If not at EOS, reload up to LBITS byte-by-byte
+static void ShiftBytes(VP8LBitReader* const br) {
+  while (br->bit_pos_ >= 8 && br->pos_ < br->len_) {
+    br->val_ >>= 8;
+    br->val_ |= ((vp8l_val_t)br->buf_[br->pos_]) << (LBITS - 8);
+    ++br->pos_;
+    br->bit_pos_ -= 8;
+  }
+}
+
+void VP8LFillBitWindow(VP8LBitReader* const br) {
+  if (br->bit_pos_ >= WBITS) {
+#if (defined(__x86_64__) || defined(_M_X64))
+    if (br->pos_ + sizeof(br->val_) < br->len_) {
+      br->val_ >>= WBITS;
+      br->bit_pos_ -= WBITS;
+      // The expression below needs a little-endian arch to work correctly.
+      // This gives a large speedup for decoding speed.
+      br->val_ |= *(const vp8l_val_t*)(br->buf_ + br->pos_) << (LBITS - WBITS);
+      br->pos_ += LOG8_WBITS;
+      return;
+    }
+#endif
+    ShiftBytes(br);       // Slow path.
+    if (br->pos_ == br->len_ && br->bit_pos_ >= LBITS) {
+      br->eos_ = 1;
+    }
+  }
+}
+
+uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits) {
+  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) {
+    const uint32_t val =
+        (uint32_t)(br->val_ >> br->bit_pos_) & kBitMask[n_bits];
+    const int new_bits = br->bit_pos_ + n_bits;
+    br->bit_pos_ = new_bits;
+    // If this read is going to cross the read buffer, set the eos flag.
+    if (br->pos_ == br->len_) {
+      if (new_bits >= LBITS) {
+        br->eos_ = 1;
+      }
+    }
+    ShiftBytes(br);
+    return val;
+  } else {
+    br->error_ = 1;
+    return 0;
+  }
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/utils/bit_reader.h b/drivers/webp/utils/bit_reader.h
new file mode 100644
index 0000000000..98df98a767
--- /dev/null
+++ b/drivers/webp/utils/bit_reader.h
@@ -0,0 +1,334 @@
+// Copyright 2010 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Boolean decoder
+//
+// Author: Skal (pascal.massimino@gmail.com)
+//         Vikas Arora (vikaas.arora@gmail.com)
+
+#ifndef WEBP_UTILS_BIT_READER_H_
+#define WEBP_UTILS_BIT_READER_H_
+
+#include <assert.h>
+#ifdef _MSC_VER
+#include <stdlib.h>  // _byteswap_ulong
+#endif
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// 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 a few of them
+// ahead (actually, we cache BITS of them ahead. See below). There's two
+// strategies regarding how to shift these looked-ahead fresh bits into the
+// 8bit window of value_: either we shift them in, while keeping the position of
+// the window fixed. Or we slide the window to the right while keeping the cache
+// bits at a fixed, right-justified, position.
+//
+//  Example, for BITS=16: here is the content of value_ for both strategies:
+//
+//          !USE_RIGHT_JUSTIFY            ||        USE_RIGHT_JUSTIFY
+//                                        ||
+//   <- 8b -><- 8b -><- BITS bits  ->     ||  <- 8b+3b -><- 8b -><- 13 bits ->
+//   [unused][value_][cached bits][0]     ||  [unused...][value_][cached bits]
+//  [........00vvvvvvBBBBBBBBBBBBB000]LSB || [...........00vvvvvvBBBBBBBBBBBBB]
+//                                        ||
+// After calling VP8Shift(), where we need to shift away two zeros:
+//  [........vvvvvvvvBBBBBBBBBBB00000]LSB || [.............vvvvvvvvBBBBBBBBBBB]
+//                                        ||
+// Just before we need to call VP8LoadNewBytes(), the situation is:
+//  [........vvvvvv000000000000000000]LSB || [..........................vvvvvv]
+//                                        ||
+// And just after calling VP8LoadNewBytes():
+//  [........vvvvvvvvBBBBBBBBBBBBBBBB]LSB || [........vvvvvvvvBBBBBBBBBBBBBBBB]
+//
+// -> we're back to eight active 'value_' bits (marked 'v') and BITS cached
+// bits (marked 'B')
+//
+// The right-justify strategy tends to use less shifts and is often faster.
+
+//------------------------------------------------------------------------------
+// BITS can be any multiple of 8 from 8 to 56 (inclusive).
+// Pick values that fit natural register size.
+
+#if !defined(WEBP_REFERENCE_IMPLEMENTATION)
+
+#define USE_RIGHT_JUSTIFY
+
+#if defined(__i386__) || defined(_M_IX86)      // x86 32bit
+#define BITS 16
+#elif defined(__x86_64__) || defined(_M_X64)   // x86 64bit
+#define BITS 56
+#elif defined(__arm__) || defined(_M_ARM)      // ARM
+#define BITS 24
+#else                      // reasonable default
+#define BITS 24
+#endif
+
+#else     // reference choices
+
+#define USE_RIGHT_JUSTIFY
+#define BITS 8
+
+#endif
+
+//------------------------------------------------------------------------------
+// Derived types and constants
+
+// bit_t = natural register type
+// lbit_t = natural type for memory I/O
+
+#if (BITS > 32)
+typedef uint64_t bit_t;
+typedef uint64_t lbit_t;
+#elif (BITS == 32)
+typedef uint64_t bit_t;
+typedef uint32_t lbit_t;
+#elif (BITS == 24)
+typedef uint32_t bit_t;
+typedef uint32_t lbit_t;
+#elif (BITS == 16)
+typedef uint32_t bit_t;
+typedef uint16_t lbit_t;
+#else
+typedef uint32_t bit_t;
+typedef uint8_t lbit_t;
+#endif
+
+#ifndef USE_RIGHT_JUSTIFY
+typedef bit_t range_t;     // type for storing range_
+#define MASK ((((bit_t)1) << (BITS)) - 1)
+#else
+typedef uint32_t range_t;  // range_ only uses 8bits here. No need for bit_t.
+#endif
+
+//------------------------------------------------------------------------------
+// Bitreader
+
+typedef struct VP8BitReader VP8BitReader;
+struct VP8BitReader {
+  const uint8_t* buf_;        // next byte to be read
+  const uint8_t* buf_end_;    // end of read buffer
+  int eof_;                   // true if input is exhausted
+
+  // boolean decoder
+  range_t range_;            // current range minus 1. In [127, 254] interval.
+  bit_t value_;              // current value
+  int bits_;                 // number of valid bits left
+};
+
+// Initialize the bit reader and the boolean decoder.
+void VP8InitBitReader(VP8BitReader* const br,
+                      const uint8_t* const start, const uint8_t* const end);
+
+// return the next value made of 'num_bits' bits
+uint32_t VP8GetValue(VP8BitReader* const br, int num_bits);
+static WEBP_INLINE uint32_t VP8Get(VP8BitReader* const br) {
+  return VP8GetValue(br, 1);
+}
+
+// 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 range_t kVP8NewRange[128];
+
+void VP8LoadFinalBytes(VP8BitReader* const br);    // special case for the tail
+
+static WEBP_INLINE void VP8LoadNewBytes(VP8BitReader* const br) {
+  assert(br != NULL && br->buf_ != NULL);
+  // 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;
+    const lbit_t in_bits = *(const lbit_t*)br->buf_;
+    br->buf_ += (BITS) >> 3;
+#if !defined(__BIG_ENDIAN__)
+#if (BITS > 32)
+// gcc 4.3 has builtin functions for swap32/swap64
+#if defined(__GNUC__) && \
+           (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+    bits = (bit_t)__builtin_bswap64(in_bits);
+#elif defined(_MSC_VER)
+    bits = (bit_t)_byteswap_uint64(in_bits);
+#elif defined(__x86_64__)
+    __asm__ volatile("bswapq %0" : "=r"(bits) : "0"(in_bits));
+#else  // generic code for swapping 64-bit values (suggested by bdb@)
+    bits = (bit_t)in_bits;
+    bits = ((bits & 0xffffffff00000000ull) >> 32) |
+           ((bits & 0x00000000ffffffffull) << 32);
+    bits = ((bits & 0xffff0000ffff0000ull) >> 16) |
+           ((bits & 0x0000ffff0000ffffull) << 16);
+    bits = ((bits & 0xff00ff00ff00ff00ull) >> 8) |
+           ((bits & 0x00ff00ff00ff00ffull) << 8);
+#endif
+    bits >>= 64 - BITS;
+#elif (BITS >= 24)
+#if defined(__i386__) || defined(__x86_64__)
+    {
+      lbit_t swapped_in_bits;
+      __asm__ volatile("bswap %k0" : "=r"(swapped_in_bits) : "0"(in_bits));
+      bits = (bit_t)swapped_in_bits;   // 24b/32b -> 32b/64b zero-extension
+    }
+#elif defined(_MSC_VER)
+    bits = (bit_t)_byteswap_ulong(in_bits);
+#else
+    bits = (bit_t)(in_bits >> 24) | ((in_bits >> 8) & 0xff00)
+         | ((in_bits << 8) & 0xff0000)  | (in_bits << 24);
+#endif  // x86
+    bits >>= (32 - BITS);
+#elif (BITS == 16)
+    // gcc will recognize a 'rorw $8, ...' here:
+    bits = (bit_t)(in_bits >> 8) | ((in_bits & 0xff) << 8);
+#else   // BITS == 8
+    bits = (bit_t)in_bits;
+#endif
+#else    // BIG_ENDIAN
+    bits = (bit_t)in_bits;
+    if (BITS != 8 * sizeof(bit_t)) bits >>= (8 * sizeof(bit_t) - BITS);
+#endif
+#ifndef USE_RIGHT_JUSTIFY
+    br->value_ |= bits << (-br->bits_);
+#else
+    br->value_ = bits | (br->value_ << (BITS));
+#endif
+    br->bits_ += (BITS);
+  } else {
+    VP8LoadFinalBytes(br);    // no need to be inlined
+  }
+}
+
+static WEBP_INLINE int VP8BitUpdate(VP8BitReader* const br, range_t split) {
+  if (br->bits_ < 0) {  // Make sure we have a least BITS bits in 'value_'
+    VP8LoadNewBytes(br);
+  }
+#ifndef USE_RIGHT_JUSTIFY
+  split |= (MASK);
+  if (br->value_ > split) {
+    br->range_ -= split + 1;
+    br->value_ -= split + 1;
+    return 1;
+  } else {
+    br->range_ = split;
+    return 0;
+  }
+#else
+  {
+    const int pos = br->bits_;
+    const range_t value = (range_t)(br->value_ >> pos);
+    if (value > split) {
+      br->range_ -= split + 1;
+      br->value_ -= (bit_t)(split + 1) << pos;
+      return 1;
+    } else {
+      br->range_ = split;
+      return 0;
+    }
+  }
+#endif
+}
+
+static WEBP_INLINE void VP8Shift(VP8BitReader* const br) {
+#ifndef USE_RIGHT_JUSTIFY
+  // range_ is in [0..127] interval here.
+  const bit_t idx = br->range_ >> (BITS);
+  const int shift = kVP8Log2Range[idx];
+  br->range_ = kVP8NewRange[idx];
+  br->value_ <<= shift;
+  br->bits_ -= shift;
+#else
+  const int shift = kVP8Log2Range[br->range_];
+  assert(br->range_ < (range_t)128);
+  br->range_ = kVP8NewRange[br->range_];
+  br->bits_ -= shift;
+#endif
+}
+
+static WEBP_INLINE int VP8GetBit(VP8BitReader* const br, int prob) {
+#ifndef USE_RIGHT_JUSTIFY
+  // It's important to avoid generating a 64bit x 64bit multiply here.
+  // We just need an 8b x 8b after all.
+  const range_t split =
+      (range_t)((uint32_t)(br->range_ >> (BITS)) * prob) << ((BITS) - 8);
+  const int bit = VP8BitUpdate(br, split);
+  if (br->range_ <= (((range_t)0x7e << (BITS)) | (MASK))) {
+    VP8Shift(br);
+  }
+  return bit;
+#else
+  const range_t split = (br->range_ * prob) >> 8;
+  const int bit = VP8BitUpdate(br, split);
+  if (br->range_ <= (range_t)0x7e) {
+    VP8Shift(br);
+  }
+  return bit;
+#endif
+}
+
+static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v) {
+  const range_t split = (br->range_ >> 1);
+  const int bit = VP8BitUpdate(br, split);
+  VP8Shift(br);
+  return bit ? -v : v;
+}
+
+// -----------------------------------------------------------------------------
+// Bitreader for lossless format
+
+typedef uint64_t vp8l_val_t;  // right now, this bit-reader can only use 64bit.
+
+typedef struct {
+  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_;        // bitstream is finished
+  int            error_;      // an error occurred (buffer overflow attempt...)
+} VP8LBitReader;
+
+void VP8LInitBitReader(VP8LBitReader* const br,
+                       const uint8_t* const start,
+                       size_t length);
+
+//  Sets a new data buffer.
+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.
+uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits);
+
+// 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_);
+}
+
+// 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;
+}
+
+// Advances the read buffer by 4 bytes to make room for reading next 32 bits.
+void VP8LFillBitWindow(VP8LBitReader* const br);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_BIT_READER_H_ */
diff --git a/drivers/webp/utils/bit_writer.c b/drivers/webp/utils/bit_writer.c
new file mode 100644
index 0000000000..29810a1749
--- /dev/null
+++ b/drivers/webp/utils/bit_writer.c
@@ -0,0 +1,282 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Bit writing and boolean coder
+//
+// Author: Skal (pascal.massimino@gmail.com)
+//         Vikas Arora (vikaas.arora@gmail.com)
+
+#include <assert.h>
+#include <string.h>   // for memcpy()
+#include <stdlib.h>
+#include "./bit_writer.h"
+
+//------------------------------------------------------------------------------
+// VP8BitWriter
+
+static int BitWriterResize(VP8BitWriter* const bw, size_t extra_size) {
+  uint8_t* new_buf;
+  size_t new_size;
+  const uint64_t needed_size_64b = (uint64_t)bw->pos_ + extra_size;
+  const size_t needed_size = (size_t)needed_size_64b;
+  if (needed_size_64b != needed_size) {
+    bw->error_ = 1;
+    return 0;
+  }
+  if (needed_size <= bw->max_pos_) return 1;
+  // If the following line wraps over 32bit, the test just after will catch it.
+  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);
+  if (new_buf == NULL) {
+    bw->error_ = 1;
+    return 0;
+  }
+  if (bw->pos_ > 0) {
+    assert(bw->buf_ != NULL);
+    memcpy(new_buf, bw->buf_, bw->pos_);
+  }
+  free(bw->buf_);
+  bw->buf_ = new_buf;
+  bw->max_pos_ = new_size;
+  return 1;
+}
+
+static void kFlush(VP8BitWriter* const bw) {
+  const int s = 8 + bw->nb_bits_;
+  const int32_t bits = bw->value_ >> s;
+  assert(bw->nb_bits_ >= 0);
+  bw->value_ -= bits << s;
+  bw->nb_bits_ -= 8;
+  if ((bits & 0xff) != 0xff) {
+    size_t pos = bw->pos_;
+    if (!BitWriterResize(bw, bw->run_ + 1)) {
+      return;
+    }
+    if (bits & 0x100) {  // overflow -> propagate carry over pending 0xff's
+      if (pos > 0) bw->buf_[pos - 1]++;
+    }
+    if (bw->run_ > 0) {
+      const int value = (bits & 0x100) ? 0x00 : 0xff;
+      for (; bw->run_ > 0; --bw->run_) bw->buf_[pos++] = value;
+    }
+    bw->buf_[pos++] = bits;
+    bw->pos_ = pos;
+  } else {
+    bw->run_++;   // delay writing of bytes 0xff, pending eventual carry.
+  }
+}
+
+//------------------------------------------------------------------------------
+// renormalization
+
+static const uint8_t kNorm[128] = {  // renorm_sizes[i] = 8 - log2(i)
+     7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4,
+  3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  0
+};
+
+// range = ((range + 1) << kVP8Log2Range[range]) - 1
+static const uint8_t kNewRange[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
+};
+
+int VP8PutBit(VP8BitWriter* const bw, int bit, int prob) {
+  const int split = (bw->range_ * prob) >> 8;
+  if (bit) {
+    bw->value_ += split + 1;
+    bw->range_ -= split + 1;
+  } else {
+    bw->range_ = split;
+  }
+  if (bw->range_ < 127) {   // emit 'shift' bits out and renormalize
+    const int shift = kNorm[bw->range_];
+    bw->range_ = kNewRange[bw->range_];
+    bw->value_ <<= shift;
+    bw->nb_bits_ += shift;
+    if (bw->nb_bits_ > 0) kFlush(bw);
+  }
+  return bit;
+}
+
+int VP8PutBitUniform(VP8BitWriter* const bw, int bit) {
+  const int split = bw->range_ >> 1;
+  if (bit) {
+    bw->value_ += split + 1;
+    bw->range_ -= split + 1;
+  } else {
+    bw->range_ = split;
+  }
+  if (bw->range_ < 127) {
+    bw->range_ = kNewRange[bw->range_];
+    bw->value_ <<= 1;
+    bw->nb_bits_ += 1;
+    if (bw->nb_bits_ > 0) kFlush(bw);
+  }
+  return bit;
+}
+
+void VP8PutValue(VP8BitWriter* const bw, int value, int nb_bits) {
+  int mask;
+  for (mask = 1 << (nb_bits - 1); mask; mask >>= 1)
+    VP8PutBitUniform(bw, value & mask);
+}
+
+void VP8PutSignedValue(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);
+  } else {
+    VP8PutValue(bw, value << 1, nb_bits + 1);
+  }
+}
+
+//------------------------------------------------------------------------------
+
+int VP8BitWriterInit(VP8BitWriter* const bw, size_t expected_size) {
+  bw->range_   = 255 - 1;
+  bw->value_   = 0;
+  bw->run_     = 0;
+  bw->nb_bits_ = -8;
+  bw->pos_     = 0;
+  bw->max_pos_ = 0;
+  bw->error_   = 0;
+  bw->buf_     = NULL;
+  return (expected_size > 0) ? BitWriterResize(bw, expected_size) : 1;
+}
+
+uint8_t* VP8BitWriterFinish(VP8BitWriter* const bw) {
+  VP8PutValue(bw, 0, 9 - bw->nb_bits_);
+  bw->nb_bits_ = 0;   // pad with zeroes
+  kFlush(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
+  if (!BitWriterResize(bw, size)) return 0;
+  memcpy(bw->buf_ + bw->pos_, data, size);
+  bw->pos_ += size;
+  return 1;
+}
+
+void VP8BitWriterWipeOut(VP8BitWriter* const bw) {
+  if (bw) {
+    free(bw->buf_);
+    memset(bw, 0, sizeof(*bw));
+  }
+}
+
+//------------------------------------------------------------------------------
+// VP8LBitWriter
+
+// 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 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 (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);
+  if (allocated_buf == NULL) {
+    bw->error_ = 1;
+    return 0;
+  }
+  memcpy(allocated_buf, bw->buf_, current_size);
+  free(bw->buf_);
+  bw->buf_ = allocated_buf;
+  bw->max_bytes_ = allocated_size;
+  memset(allocated_buf + current_size, 0, allocated_size - current_size);
+  return 1;
+}
+
+int VP8LBitWriterInit(VP8LBitWriter* const bw, size_t expected_size) {
+  memset(bw, 0, sizeof(*bw));
+  return VP8LBitWriterResize(bw, expected_size);
+}
+
+void VP8LBitWriterDestroy(VP8LBitWriter* const bw) {
+  if (bw != NULL) {
+    free(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;
+  }
+#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 arithmetic
+    *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;
+      }
+    }
+    assert(n_bits <= 25);
+    *p = bits;
+    bw->bit_pos_ += 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;
+    }
+  }
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/utils/bit_writer.h b/drivers/webp/utils/bit_writer.h
new file mode 100644
index 0000000000..89a9ead488
--- /dev/null
+++ b/drivers/webp/utils/bit_writer.h
@@ -0,0 +1,125 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Bit writing and boolean coder
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_UTILS_BIT_WRITER_H_
+#define WEBP_UTILS_BIT_WRITER_H_
+
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// Bit-writing
+
+typedef struct VP8BitWriter VP8BitWriter;
+struct VP8BitWriter {
+  int32_t  range_;      // range-1
+  int32_t  value_;
+  int      run_;        // number of outstanding bits
+  int      nb_bits_;    // number of pending bits
+  uint8_t* buf_;        // internal buffer. Re-allocated regularly. Not owned.
+  size_t   pos_;
+  size_t   max_pos_;
+  int      error_;      // true in case of error
+};
+
+// Initialize the object. Allocates some initial memory based on expected_size.
+int VP8BitWriterInit(VP8BitWriter* const bw, size_t expected_size);
+// Finalize the bitstream coding. Returns a pointer to the internal buffer.
+uint8_t* VP8BitWriterFinish(VP8BitWriter* const bw);
+// Release any pending memory and zeroes the object. Not a mandatory call.
+// Only useful in case of error, when the internal buffer hasn't been grabbed!
+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);
+
+// Appends some bytes to the internal buffer. Data is copied.
+int VP8BitWriterAppend(VP8BitWriter* const bw,
+                       const uint8_t* data, size_t size);
+
+// return approximate write position (in bits)
+static WEBP_INLINE uint64_t VP8BitWriterPos(const VP8BitWriter* const bw) {
+  return (uint64_t)(bw->pos_ + bw->run_) * 8 + 8 + bw->nb_bits_;
+}
+
+// Returns a pointer to the internal buffer.
+static WEBP_INLINE uint8_t* VP8BitWriterBuf(const VP8BitWriter* const bw) {
+  return bw->buf_;
+}
+// Returns the size of the internal buffer.
+static WEBP_INLINE size_t VP8BitWriterSize(const VP8BitWriter* const bw) {
+  return bw->pos_;
+}
+
+//------------------------------------------------------------------------------
+// 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_;
+
+  // 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
+  // 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_;
+}
+
+// Returns 0 in case of memory allocation error.
+int VP8LBitWriterInit(VP8LBitWriter* const bw, size_t expected_size);
+
+void VP8LBitWriterDestroy(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.
+//
+// VP8LBitWriter's error_ flag is set in case of  memory allocation error.
+void VP8LWriteBits(VP8LBitWriter* const bw, int n_bits, uint32_t bits);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_BIT_WRITER_H_ */
diff --git a/drivers/webp/utils/color_cache.c b/drivers/webp/utils/color_cache.c
new file mode 100644
index 0000000000..66a44647fd
--- /dev/null
+++ b/drivers/webp/utils/color_cache.c
@@ -0,0 +1,39 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Color Cache for WebP Lossless
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include "./color_cache.h"
+#include "../utils/utils.h"
+
+//------------------------------------------------------------------------------
+// VP8LColorCache.
+
+int VP8LColorCacheInit(VP8LColorCache* const cc, int hash_bits) {
+  const int hash_size = 1 << hash_bits;
+  assert(cc != NULL);
+  assert(hash_bits > 0);
+  cc->colors_ = (uint32_t*)WebPSafeCalloc((uint64_t)hash_size,
+                                          sizeof(*cc->colors_));
+  if (cc->colors_ == NULL) return 0;
+  cc->hash_shift_ = 32 - hash_bits;
+  return 1;
+}
+
+void VP8LColorCacheClear(VP8LColorCache* const cc) {
+  if (cc != NULL) {
+    free(cc->colors_);
+    cc->colors_ = NULL;
+  }
+}
+
diff --git a/drivers/webp/utils/color_cache.h b/drivers/webp/utils/color_cache.h
new file mode 100644
index 0000000000..0f824ed457
--- /dev/null
+++ b/drivers/webp/utils/color_cache.h
@@ -0,0 +1,70 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Color Cache for WebP Lossless
+//
+// Authors: Jyrki Alakuijala (jyrki@google.com)
+//          Urvang Joshi (urvang@google.com)
+
+#ifndef WEBP_UTILS_COLOR_CACHE_H_
+#define WEBP_UTILS_COLOR_CACHE_H_
+
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Main color cache struct.
+typedef struct {
+  uint32_t *colors_;  // color entries
+  int hash_shift_;    // Hash shift: 32 - 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_));
+  return cc->colors_[key];
+}
+
+static WEBP_INLINE void VP8LColorCacheInsert(const VP8LColorCache* const cc,
+                                             uint32_t argb) {
+  const uint32_t key = (kHashMul * argb) >> cc->hash_shift_;
+  cc->colors_[key] = argb;
+}
+
+static WEBP_INLINE int VP8LColorCacheGetIndex(const VP8LColorCache* const cc,
+                                              uint32_t argb) {
+  return (kHashMul * argb) >> cc->hash_shift_;
+}
+
+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;
+}
+
+//------------------------------------------------------------------------------
+
+// Initializes the color cache with 'hash_bits' bits for the keys.
+// Returns false in case of memory error.
+int VP8LColorCacheInit(VP8LColorCache* const color_cache, int hash_bits);
+
+// Delete the memory associated to color cache.
+void VP8LColorCacheClear(VP8LColorCache* const color_cache);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // WEBP_UTILS_COLOR_CACHE_H_
diff --git a/drivers/webp/utils/filters.c b/drivers/webp/utils/filters.c
new file mode 100644
index 0000000000..2d15bd0e4a
--- /dev/null
+++ b/drivers/webp/utils/filters.c
@@ -0,0 +1,266 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Spatial prediction using various filters
+//
+// Author: Urvang (urvang@google.com)
+
+#include "./filters.h"
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+
+//------------------------------------------------------------------------------
+// Helpful macro.
+
+# define SANITY_CHECK(in, out)                                                 \
+  assert(in != NULL);                                                          \
+  assert(out != NULL);                                                         \
+  assert(width > 0);                                                           \
+  assert(height > 0);                                                          \
+  assert(stride >= width);                                                     \
+  assert(row >= 0 && num_rows > 0 && row + num_rows <= height);                \
+  (void)height;  // Silence unused warning.
+
+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 stride,
+                                           int row, int num_rows,
+                                           int inverse, uint8_t* out) {
+  const uint8_t* preds;
+  const size_t start_offset = row * stride;
+  const int last_row = row + num_rows;
+  SANITY_CHECK(in, out);
+  in += start_offset;
+  out += start_offset;
+  preds = inverse ? out : in;
+
+  if (row == 0) {
+    // Leftmost pixel is the same as input for topmost scanline.
+    out[0] = in[0];
+    PredictLine(in + 1, preds, out + 1, width - 1, inverse);
+    row = 1;
+    preds += stride;
+    in += stride;
+    out += stride;
+  }
+
+  // Filter line-by-line.
+  while (row < last_row) {
+    // Leftmost pixel is predicted from above.
+    PredictLine(in, preds - stride, out, 1, inverse);
+    PredictLine(in + 1, preds, out + 1, width - 1, inverse);
+    ++row;
+    preds += stride;
+    in += stride;
+    out += stride;
+  }
+}
+
+static void HorizontalFilter(const uint8_t* data, int width, int height,
+                             int stride, uint8_t* filtered_data) {
+  DoHorizontalFilter(data, width, height, stride, 0, height, 0, filtered_data);
+}
+
+static void HorizontalUnfilter(int width, int height, int stride, int row,
+                               int num_rows, uint8_t* data) {
+  DoHorizontalFilter(data, width, height, stride, row, num_rows, 1, data);
+}
+
+//------------------------------------------------------------------------------
+// Vertical filter.
+
+static WEBP_INLINE void DoVerticalFilter(const uint8_t* in,
+                                         int width, int height, int stride,
+                                         int row, int num_rows,
+                                         int inverse, uint8_t* out) {
+  const uint8_t* preds;
+  const size_t start_offset = row * stride;
+  const int last_row = row + num_rows;
+  SANITY_CHECK(in, out);
+  in += start_offset;
+  out += start_offset;
+  preds = inverse ? out : in;
+
+  if (row == 0) {
+    // Very first top-left pixel is copied.
+    out[0] = in[0];
+    // Rest of top scan-line is left-predicted.
+    PredictLine(in + 1, preds, out + 1, width - 1, inverse);
+    row = 1;
+    in += stride;
+    out += stride;
+  } else {
+    // We are starting from in-between. Make sure 'preds' points to prev row.
+    preds -= stride;
+  }
+
+  // Filter line-by-line.
+  while (row < last_row) {
+    PredictLine(in, preds, out, width, inverse);
+    ++row;
+    preds += stride;
+    in += stride;
+    out += stride;
+  }
+}
+
+static void VerticalFilter(const uint8_t* data, int width, int height,
+                           int stride, uint8_t* filtered_data) {
+  DoVerticalFilter(data, width, height, stride, 0, height, 0, filtered_data);
+}
+
+static void VerticalUnfilter(int width, int height, int stride, int row,
+                             int num_rows, uint8_t* data) {
+  DoVerticalFilter(data, width, height, stride, row, num_rows, 1, data);
+}
+
+//------------------------------------------------------------------------------
+// Gradient filter.
+
+static WEBP_INLINE int GradientPredictor(uint8_t a, uint8_t b, uint8_t c) {
+  const int g = a + b - c;
+  return ((g & ~0xff) == 0) ? g : (g < 0) ? 0 : 255;  // clip to 8bit
+}
+
+static WEBP_INLINE void DoGradientFilter(const uint8_t* in,
+                                         int width, int height, int stride,
+                                         int row, int num_rows,
+                                         int inverse, uint8_t* out) {
+  const uint8_t* preds;
+  const size_t start_offset = row * stride;
+  const int last_row = row + num_rows;
+  SANITY_CHECK(in, out);
+  in += start_offset;
+  out += start_offset;
+  preds = inverse ? out : in;
+
+  // left prediction for top scan-line
+  if (row == 0) {
+    out[0] = in[0];
+    PredictLine(in + 1, preds, out + 1, width - 1, inverse);
+    row = 1;
+    preds += stride;
+    in += stride;
+    out += stride;
+  }
+
+  // Filter line-by-line.
+  while (row < last_row) {
+    int w;
+    // leftmost pixel: predict from above.
+    PredictLine(in, preds - stride, out, 1, inverse);
+    for (w = 1; w < width; ++w) {
+      const int pred = GradientPredictor(preds[w - 1],
+                                         preds[w - stride],
+                                         preds[w - stride - 1]);
+      out[w] = in[w] + (inverse ? pred : -pred);
+    }
+    ++row;
+    preds += stride;
+    in += stride;
+    out += stride;
+  }
+}
+
+static void GradientFilter(const uint8_t* data, int width, int height,
+                           int stride, uint8_t* filtered_data) {
+  DoGradientFilter(data, width, height, stride, 0, height, 0, filtered_data);
+}
+
+static void GradientUnfilter(int width, int height, int stride, int row,
+                             int num_rows, uint8_t* data) {
+  DoGradientFilter(data, width, height, stride, row, num_rows, 1, data);
+}
+
+#undef SANITY_CHECK
+
+// -----------------------------------------------------------------------------
+// Quick estimate of a potentially interesting filter mode to try.
+
+#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) {
+  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;
+    int mean = p[0];
+    for (i = 2; i < width - 1; i += 2) {
+      const int diff0 = SDIFF(p[i], mean);
+      const int diff1 = SDIFF(p[i], p[i - 1]);
+      const int diff2 = SDIFF(p[i], p[i - width]);
+      const int grad_pred =
+          GradientPredictor(p[i - 1], p[i - width], p[i - width - 1]);
+      const int diff3 = SDIFF(p[i], grad_pred);
+      bins[WEBP_FILTER_NONE][diff0] = 1;
+      bins[WEBP_FILTER_HORIZONTAL][diff1] = 1;
+      bins[WEBP_FILTER_VERTICAL][diff2] = 1;
+      bins[WEBP_FILTER_GRADIENT][diff3] = 1;
+      mean = (3 * mean + p[i] + 2) >> 2;
+    }
+  }
+  {
+    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;
+      for (i = 0; i < SMAX; ++i) {
+        if (bins[filter][i] > 0) {
+          score += i;
+        }
+      }
+      if (score < best_score) {
+        best_score = score;
+        best_filter = (WEBP_FILTER_TYPE)filter;
+      }
+    }
+    return best_filter;
+  }
+}
+
+#undef SMAX
+#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 WebPUnfilterFunc WebPUnfilters[WEBP_FILTER_LAST] = {
+  NULL,                // WEBP_FILTER_NONE
+  HorizontalUnfilter,  // WEBP_FILTER_HORIZONTAL
+  VerticalUnfilter,    // WEBP_FILTER_VERTICAL
+  GradientUnfilter     // WEBP_FILTER_GRADIENT
+};
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/utils/filters.h b/drivers/webp/utils/filters.h
new file mode 100644
index 0000000000..dde39cb5c4
--- /dev/null
+++ b/drivers/webp/utils/filters.h
@@ -0,0 +1,59 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Spatial prediction using various filters
+//
+// Author: Urvang (urvang@google.com)
+
+#ifndef WEBP_UTILS_FILTERS_H_
+#define WEBP_UTILS_FILTERS_H_
+
+#include "../webp/types.h"
+
+#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 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 const 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 const WebPUnfilterFunc WebPUnfilters[WEBP_FILTER_LAST];
+
+// Fast estimate of a potentially good filter.
+WEBP_FILTER_TYPE EstimateBestFilter(const uint8_t* data,
+                                    int width, int height, int stride);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_FILTERS_H_ */
diff --git a/drivers/webp/utils/huffman.c b/drivers/webp/utils/huffman.c
new file mode 100644
index 0000000000..8c5739f633
--- /dev/null
+++ b/drivers/webp/utils/huffman.c
@@ -0,0 +1,288 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Utilities for building and looking up Huffman trees.
+//
+// Author: Urvang Joshi (urvang@google.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./huffman.h"
+#include "../utils/utils.h"
+#include "../webp/format_constants.h"
+
+// Uncomment the following to use look-up table for ReverseBits()
+// (might be faster on some platform)
+// #define USE_LUT_REVERSE_BITS
+
+#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);
+}
+
+static int IsFull(const HuffmanTree* const tree) {
+  return (tree->num_nodes_ == tree->max_nodes_);
+}
+
+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);
+}
+
+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;
+  assert(tree->max_nodes_ < (1 << 16));   // limit for the lut_jump_ table
+  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;
+  memset(tree->lut_bits_, 255, sizeof(tree->lut_bits_));
+  memset(tree->lut_jump_, 0, sizeof(tree->lut_jump_));
+  return 1;
+}
+
+void HuffmanTreeRelease(HuffmanTree* const tree) {
+  if (tree != NULL) {
+    free(tree->root_);
+    tree->root_ = NULL;
+    tree->max_nodes_ = 0;
+    tree->num_nodes_ = 0;
+  }
+}
+
+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;
+
+  assert(code_lengths != NULL);
+  assert(code_lengths_size > 0);
+  assert(huff_codes != NULL);
+
+  // Calculate max code length.
+  for (symbol = 0; symbol < code_lengths_size; ++symbol) {
+    if (code_lengths[symbol] > max_code_length) {
+      max_code_length = 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;
+  }
+
+  // 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;
+}
+
+#ifndef USE_LUT_REVERSE_BITS
+
+static int ReverseBitsShort(int bits, int num_bits) {
+  int retval = 0;
+  int i;
+  assert(num_bits <= 8);   // Not a hard requirement, just for coherency.
+  for (i = 0; i < num_bits; ++i) {
+    retval <<= 1;
+    retval |= bits & 1;
+    bits >>= 1;
+  }
+  return retval;
+}
+
+#else
+
+static const uint8_t kReversedBits[16] = {  // Pre-reversed 4-bit values.
+  0x0, 0x8, 0x4, 0xc, 0x2, 0xa, 0x6, 0xe,
+  0x1, 0x9, 0x5, 0xd, 0x3, 0xb, 0x7, 0xf
+};
+
+static int ReverseBitsShort(int bits, int num_bits) {
+  const uint8_t v = (kReversedBits[bits & 0xf] << 4) | kReversedBits[bits >> 4];
+  assert(num_bits <= 8);
+  return v >> (8 - num_bits);
+}
+
+#endif
+
+static int TreeAddSymbol(HuffmanTree* const tree,
+                         int symbol, int code, int code_length) {
+  int step = HUFF_LUT_BITS;
+  int base_code;
+  HuffmanTreeNode* node = tree->root_;
+  const HuffmanTreeNode* const max_node = tree->root_ + tree->max_nodes_;
+  assert(symbol == (int16_t)symbol);
+  if (code_length <= HUFF_LUT_BITS) {
+    int i;
+    base_code = ReverseBitsShort(code, code_length);
+    for (i = 0; i < (1 << (HUFF_LUT_BITS - code_length)); ++i) {
+      const int idx = base_code | (i << code_length);
+      tree->lut_symbol_[idx] = (int16_t)symbol;
+      tree->lut_bits_[idx] = code_length;
+    }
+  } else {
+    base_code = ReverseBitsShort((code >> (code_length - HUFF_LUT_BITS)),
+                                 HUFF_LUT_BITS);
+  }
+  while (code_length-- > 0) {
+    if (node >= max_node) {
+      return 0;
+    }
+    if (NodeIsEmpty(node)) {
+      if (IsFull(tree)) return 0;    // error: too many symbols.
+      AssignChildren(tree, node);
+    } else if (!HuffmanTreeNodeIsNotLeaf(node)) {
+      return 0;  // leaf is already occupied.
+    }
+    node += node->children_ + ((code >> code_length) & 1);
+    if (--step == 0) {
+      tree->lut_jump_[base_code] = (int16_t)(node - tree->root_);
+    }
+  }
+  if (NodeIsEmpty(node)) {
+    node->children_ = 0;      // turn newly created node into a leaf.
+  } else if (HuffmanTreeNodeIsNotLeaf(node)) {
+    return 0;   // trying to assign a symbol to already used code.
+  }
+  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);
+
+  // Find out number of symbols and the root symbol.
+  for (symbol = 0; symbol < code_lengths_size; ++symbol) {
+    if (code_lengths[symbol] > 0) {
+      // Note: code length = 0 indicates non-existent symbol.
+      ++num_symbols;
+      root_symbol = 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;
+
+    if (!HuffmanCodeLengthsToCodes(code_lengths, code_lengths_size, codes)) {
+      goto End;
+    }
+
+    // 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;
+        }
+      }
+    }
+    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;
+      }
+    }
+  }
+  ok = 1;
+ End:
+  ok = ok && IsFull(tree);
+  if (!ok) HuffmanTreeRelease(tree);
+  return ok;
+}
+
diff --git a/drivers/webp/utils/huffman.h b/drivers/webp/utils/huffman.h
new file mode 100644
index 0000000000..e8afd27f24
--- /dev/null
+++ b/drivers/webp/utils/huffman.h
@@ -0,0 +1,87 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Utilities for building and looking up Huffman trees.
+//
+// Author: Urvang Joshi (urvang@google.com)
+
+#ifndef WEBP_UTILS_HUFFMAN_H_
+#define WEBP_UTILS_HUFFMAN_H_
+
+#include <assert.h>
+#include "../webp/types.h"
+
+#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;
+
+// Huffman Tree.
+#define HUFF_LUT_BITS 7
+#define HUFF_LUT (1U << HUFF_LUT_BITS)
+typedef struct HuffmanTree HuffmanTree;
+struct HuffmanTree {
+  // Fast lookup for short bit lengths.
+  uint8_t lut_bits_[HUFF_LUT];
+  int16_t lut_symbol_[HUFF_LUT];
+  int16_t lut_jump_[HUFF_LUT];
+  // Complete tree for lookups.
+  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 not a leaf of the Huffman tree.
+static WEBP_INLINE int HuffmanTreeNodeIsNotLeaf(
+    const HuffmanTreeNode* const node) {
+  return node->children_;
+}
+
+// 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;
+}
+
+// Releases the nodes of the Huffman tree.
+// Note: It does NOT free 'tree' itself.
+void HuffmanTreeRelease(HuffmanTree* const tree);
+
+// 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);
+
+// 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);
+
+// 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);
+
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // WEBP_UTILS_HUFFMAN_H_
diff --git a/drivers/webp/utils/huffman_encode.c b/drivers/webp/utils/huffman_encode.c
new file mode 100644
index 0000000000..9c5986738d
--- /dev/null
+++ b/drivers/webp/utils/huffman_encode.c
@@ -0,0 +1,440 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+//
+// Entropy encoding (Huffman) for webp lossless.
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./huffman_encode.h"
+#include "../utils/utils.h"
+#include "../webp/format_constants.h"
+
+// -----------------------------------------------------------------------------
+// Util function to optimize the symbol map for RLE coding
+
+// Heuristics for selecting the stride ranges to collapse.
+static int ValuesShouldBeCollapsedToStrideAverage(int a, int b) {
+  return abs(a - b) < 4;
+}
+
+// Change the population counts in a way that the consequent
+// Huffman tree compression, especially its RLE-part, give smaller output.
+static int OptimizeHuffmanForRle(int length, int* const counts) {
+  uint8_t* good_for_rle;
+  // 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.
+    }
+    if (counts[length - 1] != 0) {
+      // Now counts[0..length - 1] does not have trailing zeros.
+      break;
+    }
+  }
+  // 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];
+    int stride = 0;
+    for (i = 0; i < length + 1; ++i) {
+      if (i == length || counts[i] != symbol) {
+        if ((symbol == 0 && stride >= 5) ||
+            (symbol != 0 && stride >= 7)) {
+          int k;
+          for (k = 0; k < stride; ++k) {
+            good_for_rle[i - k - 1] = 1;
+          }
+        }
+        stride = 1;
+        if (i != length) {
+          symbol = counts[i];
+        }
+      } else {
+        ++stride;
+      }
+    }
+  }
+  // 3) Let's replace those population counts that lead to more rle codes.
+  {
+    int stride = 0;
+    int limit = counts[0];
+    int 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;
+          // The stride must end, collapse what we have, if we have enough (4).
+          int count = (sum + stride / 2) / stride;
+          if (count < 1) {
+            count = 1;
+          }
+          if (sum == 0) {
+            // Don't make an all zeros stride to be upgraded to ones.
+            count = 0;
+          }
+          for (k = 0; k < stride; ++k) {
+            // We don't want to change value at counts[i],
+            // that is already belonging to the next stride. Thus - 1.
+            counts[i - k - 1] = count;
+          }
+        }
+        stride = 0;
+        sum = 0;
+        if (i < length - 3) {
+          // All interesting strides have a count of at least 4,
+          // at least when non-zeros.
+          limit = (counts[i] + counts[i + 1] +
+                   counts[i + 2] + counts[i + 3] + 2) / 4;
+        } else if (i < length) {
+          limit = counts[i];
+        } else {
+          limit = 0;
+        }
+      }
+      ++stride;
+      if (i != length) {
+        sum += counts[i];
+        if (stride >= 4) {
+          limit = (sum + stride / 2) / stride;
+        }
+      }
+    }
+  }
+  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) {
+  const HuffmanTree* const t1 = (const HuffmanTree*)ptr1;
+  const HuffmanTree* const t2 = (const HuffmanTree*)ptr2;
+  if (t1->total_count_ > t2->total_count_) {
+    return -1;
+  } else if (t1->total_count_ < t2->total_count_) {
+    return 1;
+  } else {
+    assert(t1->value_ != t2->value_);
+    return (t1->value_ < t2->value_) ? -1 : 1;
+  }
+}
+
+static void SetBitDepths(const HuffmanTree* const tree,
+                         const HuffmanTree* const pool,
+                         uint8_t* const bit_depths, int level) {
+  if (tree->pool_index_left_ >= 0) {
+    SetBitDepths(&pool[tree->pool_index_left_], pool, bit_depths, level + 1);
+    SetBitDepths(&pool[tree->pool_index_right_], pool, bit_depths, level + 1);
+  } else {
+    bit_depths[tree->value_] = level;
+  }
+}
+
+// Create an optimal Huffman tree.
+//
+// (data,length): population counts.
+// tree_limit: maximum bit depth (inclusive) of the codes.
+// bit_depths[]: how many bits are used for the symbol.
+//
+// Returns 0 when an error has occurred.
+//
+// The catch here is that the tree cannot be arbitrarily deep
+//
+// count_limit is the value that is to be faked as the minimum value
+// and this minimum value is raised until the tree matches the
+// maximum length requirement.
+//
+// This algorithm is not of excellent performance for very long data blocks,
+// especially when population counts are longer than 2**tree_limit, but
+// 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;
+  HuffmanTree* tree_pool;
+  HuffmanTree* tree;
+  int tree_size_orig = 0;
+  int i;
+
+  for (i = 0; i < histogram_size; ++i) {
+    if (histogram[i] != 0) {
+      ++tree_size_orig;
+    }
+  }
+
+  if (tree_size_orig == 0) {   // pretty optimal already!
+    return 1;
+  }
+
+  // 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;
+  tree_pool = tree + tree_size_orig;
+
+  // For block sizes with less than 64k symbols we never need to do a
+  // second iteration of this loop.
+  // If we actually start running inside this loop a lot, we would perhaps
+  // be better off with the Katajainen algorithm.
+  assert(tree_size_orig <= (1 << (tree_depth_limit - 1)));
+  for (count_min = 1; ; count_min *= 2) {
+    int tree_size = tree_size_orig;
+    // We need to pack the Huffman tree in tree_depth_limit bits.
+    // So, we try by faking histogram entries to be at least 'count_min'.
+    int idx = 0;
+    int j;
+    for (j = 0; j < histogram_size; ++j) {
+      if (histogram[j] != 0) {
+        const int count =
+            (histogram[j] < count_min) ? count_min : histogram[j];
+        tree[idx].total_count_ = count;
+        tree[idx].value_ = j;
+        tree[idx].pool_index_left_ = -1;
+        tree[idx].pool_index_right_ = -1;
+        ++idx;
+      }
+    }
+
+    // Build the Huffman tree.
+    qsort(tree, tree_size, sizeof(*tree), CompareHuffmanTrees);
+
+    if (tree_size > 1) {  // Normal case.
+      int tree_pool_size = 0;
+      while (tree_size > 1) {  // Finish when we have only one root.
+        int 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_size -= 2;
+        {
+          // Search for the insertion point.
+          int k;
+          for (k = 0; k < tree_size; ++k) {
+            if (tree[k].total_count_ <= count) {
+              break;
+            }
+          }
+          memmove(tree + (k + 1), tree + k, (tree_size - k) * sizeof(*tree));
+          tree[k].total_count_ = count;
+          tree[k].value_ = -1;
+
+          tree[k].pool_index_left_ = tree_pool_size - 1;
+          tree[k].pool_index_right_ = tree_pool_size - 2;
+          tree_size = tree_size + 1;
+        }
+      }
+      SetBitDepths(&tree[0], tree_pool, bit_depths, 0);
+    } else if (tree_size == 1) {  // Trivial case: only one element.
+      bit_depths[tree[0].value_] = 1;
+    }
+
+    {
+      // Test if this Huffman tree satisfies our 'tree_depth_limit' criteria.
+      int max_depth = bit_depths[0];
+      for (j = 1; j < histogram_size; ++j) {
+        if (max_depth < bit_depths[j]) {
+          max_depth = bit_depths[j];
+        }
+      }
+      if (max_depth <= tree_depth_limit) {
+        break;
+      }
+    }
+  }
+  free(tree);
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+// Coding of the Huffman tree values
+
+static HuffmanTreeToken* CodeRepeatedValues(int repetitions,
+                                            HuffmanTreeToken* tokens,
+                                            int value, int prev_value) {
+  assert(value <= MAX_ALLOWED_CODE_LENGTH);
+  if (value != prev_value) {
+    tokens->code = value;
+    tokens->extra_bits = 0;
+    ++tokens;
+    --repetitions;
+  }
+  while (repetitions >= 1) {
+    if (repetitions < 3) {
+      int i;
+      for (i = 0; i < repetitions; ++i) {
+        tokens->code = value;
+        tokens->extra_bits = 0;
+        ++tokens;
+      }
+      break;
+    } else if (repetitions < 7) {
+      tokens->code = 16;
+      tokens->extra_bits = repetitions - 3;
+      ++tokens;
+      break;
+    } else {
+      tokens->code = 16;
+      tokens->extra_bits = 3;
+      ++tokens;
+      repetitions -= 6;
+    }
+  }
+  return tokens;
+}
+
+static HuffmanTreeToken* CodeRepeatedZeros(int repetitions,
+                                           HuffmanTreeToken* tokens) {
+  while (repetitions >= 1) {
+    if (repetitions < 3) {
+      int i;
+      for (i = 0; i < repetitions; ++i) {
+        tokens->code = 0;   // 0-value
+        tokens->extra_bits = 0;
+        ++tokens;
+      }
+      break;
+    } else if (repetitions < 11) {
+      tokens->code = 17;
+      tokens->extra_bits = repetitions - 3;
+      ++tokens;
+      break;
+    } else if (repetitions < 139) {
+      tokens->code = 18;
+      tokens->extra_bits = repetitions - 11;
+      ++tokens;
+      break;
+    } else {
+      tokens->code = 18;
+      tokens->extra_bits = 0x7f;  // 138 repeated 0s
+      ++tokens;
+      repetitions -= 138;
+    }
+  }
+  return tokens;
+}
+
+int VP8LCreateCompressedHuffmanTree(const HuffmanTreeCode* const tree,
+                                    HuffmanTreeToken* tokens, int max_tokens) {
+  HuffmanTreeToken* const starting_token = tokens;
+  HuffmanTreeToken* const ending_token = tokens + max_tokens;
+  const int depth_size = tree->num_symbols;
+  int prev_value = 8;  // 8 is the initial value for rle.
+  int i = 0;
+  assert(tokens != NULL);
+  while (i < depth_size) {
+    const int value = tree->code_lengths[i];
+    int k = i + 1;
+    int runs;
+    while (k < depth_size && tree->code_lengths[k] == value) ++k;
+    runs = k - i;
+    if (value == 0) {
+      tokens = CodeRepeatedZeros(runs, tokens);
+    } else {
+      tokens = CodeRepeatedValues(runs, tokens, value, prev_value);
+      prev_value = value;
+    }
+    i += runs;
+    assert(tokens <= ending_token);
+  }
+  (void)ending_token;    // suppress 'unused variable' warning
+  return (int)(tokens - starting_token);
+}
+
+// -----------------------------------------------------------------------------
+
+// Pre-reversed 4-bit values.
+static const uint8_t kReversedBits[16] = {
+  0x0, 0x8, 0x4, 0xc, 0x2, 0xa, 0x6, 0xe,
+  0x1, 0x9, 0x5, 0xd, 0x3, 0xb, 0x7, 0xf
+};
+
+static uint32_t ReverseBits(int num_bits, uint32_t bits) {
+  uint32_t retval = 0;
+  int i = 0;
+  while (i < num_bits) {
+    i += 4;
+    retval |= kReversedBits[bits & 0xf] << (MAX_ALLOWED_CODE_LENGTH + 1 - i);
+    bits >>= 4;
+  }
+  retval >>= (MAX_ALLOWED_CODE_LENGTH + 1 - num_bits);
+  return retval;
+}
+
+// Get the actual bit values for a tree of bit depths.
+static void ConvertBitDepthsToSymbols(HuffmanTreeCode* const tree) {
+  // 0 bit-depth means that the symbol does not exist.
+  int i;
+  int len;
+  uint32_t next_code[MAX_ALLOWED_CODE_LENGTH + 1];
+  int depth_count[MAX_ALLOWED_CODE_LENGTH + 1] = { 0 };
+
+  assert(tree != NULL);
+  len = tree->num_symbols;
+  for (i = 0; i < len; ++i) {
+    const int code_length = tree->code_lengths[i];
+    assert(code_length <= MAX_ALLOWED_CODE_LENGTH);
+    ++depth_count[code_length];
+  }
+  depth_count[0] = 0;  // ignore unused symbol
+  next_code[0] = 0;
+  {
+    uint32_t code = 0;
+    for (i = 1; i <= MAX_ALLOWED_CODE_LENGTH; ++i) {
+      code = (code + depth_count[i - 1]) << 1;
+      next_code[i] = code;
+    }
+  }
+  for (i = 0; i < len; ++i) {
+    const int code_length = tree->code_lengths[i];
+    tree->codes[i] = ReverseBits(code_length, next_code[code_length]++);
+  }
+}
+
+// -----------------------------------------------------------------------------
+// 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;
+  }
+  // Create the actual bit codes for the bit lengths.
+  ConvertBitDepthsToSymbols(tree);
+  return 1;
+}
diff --git a/drivers/webp/utils/huffman_encode.h b/drivers/webp/utils/huffman_encode.h
new file mode 100644
index 0000000000..ee51c68c9f
--- /dev/null
+++ b/drivers/webp/utils/huffman_encode.h
@@ -0,0 +1,49 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Author: Jyrki Alakuijala (jyrki@google.com)
+//
+// Entropy encoding (Huffman) for webp lossless
+
+#ifndef WEBP_UTILS_HUFFMAN_ENCODE_H_
+#define WEBP_UTILS_HUFFMAN_ENCODE_H_
+
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Struct for holding the tree header in coded form.
+typedef struct {
+  uint8_t code;         // value (0..15) or escape code (16,17,18)
+  uint8_t extra_bits;   // extra bits for escape codes
+} HuffmanTreeToken;
+
+// Struct to represent the tree codes (depth and bits array).
+typedef struct {
+  int       num_symbols;   // Number of symbols.
+  uint8_t*  code_lengths;  // Code lengths of the symbols.
+  uint16_t* codes;         // Symbol Codes.
+} HuffmanTreeCode;
+
+// 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);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // WEBP_UTILS_HUFFMAN_ENCODE_H_
diff --git a/drivers/webp/utils/quant_levels.c b/drivers/webp/utils/quant_levels.c
new file mode 100644
index 0000000000..d7c8aab922
--- /dev/null
+++ b/drivers/webp/utils/quant_levels.c
@@ -0,0 +1,140 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Quantize levels for specified number of quantization-levels ([2, 256]).
+// Min and max values are preserved (usual 0 and 255 for alpha plane).
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+
+#include "./quant_levels.h"
+
+#define NUM_SYMBOLS     256
+
+#define MAX_ITER  6             // Maximum number of convergence steps.
+#define ERROR_THRESHOLD 1e-4    // MSE stopping criterion.
+
+// -----------------------------------------------------------------------------
+// Quantize levels.
+
+int QuantizeLevels(uint8_t* const data, int width, int height,
+                   int num_levels, uint64_t* const sse) {
+  int freq[NUM_SYMBOLS] = { 0 };
+  int q_level[NUM_SYMBOLS] = { 0 };
+  double inv_q_level[NUM_SYMBOLS] = { 0 };
+  int min_s = 255, max_s = 0;
+  const size_t data_size = height * width;
+  int i, num_levels_in, iter;
+  double last_err = 1.e38, err = 0.;
+  const double err_threshold = ERROR_THRESHOLD * data_size;
+
+  if (data == NULL) {
+    return 0;
+  }
+
+  if (width <= 0 || height <= 0) {
+    return 0;
+  }
+
+  if (num_levels < 2 || num_levels > 256) {
+    return 0;
+  }
+
+  {
+    size_t n;
+    num_levels_in = 0;
+    for (n = 0; n < data_size; ++n) {
+      num_levels_in += (freq[data[n]] == 0);
+      if (min_s > data[n]) min_s = data[n];
+      if (max_s < data[n]) max_s = data[n];
+      ++freq[data[n]];
+    }
+  }
+
+  if (num_levels_in <= num_levels) goto End;  // nothing to do!
+
+  // Start with uniformly spread centroids.
+  for (i = 0; i < num_levels; ++i) {
+    inv_q_level[i] = min_s + (double)(max_s - min_s) * i / (num_levels - 1);
+  }
+
+  // Fixed values. Won't be changed.
+  q_level[min_s] = 0;
+  q_level[max_s] = num_levels - 1;
+  assert(inv_q_level[0] == min_s);
+  assert(inv_q_level[num_levels - 1] == max_s);
+
+  // k-Means iterations.
+  for (iter = 0; iter < MAX_ITER; ++iter) {
+    double q_sum[NUM_SYMBOLS] = { 0 };
+    double q_count[NUM_SYMBOLS] = { 0 };
+    int s, slot = 0;
+
+    // Assign classes to representatives.
+    for (s = min_s; s <= max_s; ++s) {
+      // Keep track of the nearest neighbour 'slot'
+      while (slot < num_levels - 1 &&
+             2 * s > inv_q_level[slot] + inv_q_level[slot + 1]) {
+        ++slot;
+      }
+      if (freq[s] > 0) {
+        q_sum[slot] += s * freq[s];
+        q_count[slot] += freq[s];
+      }
+      q_level[s] = slot;
+    }
+
+    // Assign new representatives to classes.
+    if (num_levels > 2) {
+      for (slot = 1; slot < num_levels - 1; ++slot) {
+        const double count = q_count[slot];
+        if (count > 0.) {
+          inv_q_level[slot] = q_sum[slot] / count;
+        }
+      }
+    }
+
+    // Compute convergence error.
+    err = 0.;
+    for (s = min_s; s <= max_s; ++s) {
+      const double error = s - inv_q_level[q_level[s]];
+      err += freq[s] * error * error;
+    }
+
+    // Check for convergence: we stop as soon as the error is no
+    // longer improving.
+    if (last_err - err < err_threshold) break;
+    last_err = err;
+  }
+
+  // Remap the alpha plane to quantized values.
+  {
+    // double->int rounding operation can be costly, so we do it
+    // once for all before remapping. We also perform the data[] -> slot
+    // mapping, while at it (avoid one indirection in the final loop).
+    uint8_t map[NUM_SYMBOLS];
+    int s;
+    size_t n;
+    for (s = min_s; s <= max_s; ++s) {
+      const int slot = q_level[s];
+      map[s] = (uint8_t)(inv_q_level[slot] + .5);
+    }
+    // Final pass.
+    for (n = 0; n < data_size; ++n) {
+      data[n] = map[data[n]];
+    }
+  }
+ End:
+  // Store sum of squared error if needed.
+  if (sse != NULL) *sse = (uint64_t)err;
+
+  return 1;
+}
+
diff --git a/drivers/webp/utils/quant_levels.h b/drivers/webp/utils/quant_levels.h
new file mode 100644
index 0000000000..1cb5a32cae
--- /dev/null
+++ b/drivers/webp/utils/quant_levels.h
@@ -0,0 +1,36 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Alpha plane quantization utility
+//
+// Author:  Vikas Arora (vikasa@google.com)
+
+#ifndef WEBP_UTILS_QUANT_LEVELS_H_
+#define WEBP_UTILS_QUANT_LEVELS_H_
+
+#include <stdlib.h>
+
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Replace the input 'data' of size 'width'x'height' with 'num-levels'
+// quantized values. If not NULL, 'sse' will contain the sum of squared error.
+// Valid range for 'num_levels' is [2, 256].
+// Returns false in case of error (data is NULL, or parameters are invalid).
+int QuantizeLevels(uint8_t* const data, int width, int height, int num_levels,
+                   uint64_t* const sse);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_QUANT_LEVELS_H_ */
diff --git a/drivers/webp/utils/quant_levels_dec.c b/drivers/webp/utils/quant_levels_dec.c
new file mode 100644
index 0000000000..8489705a2d
--- /dev/null
+++ b/drivers/webp/utils/quant_levels_dec.c
@@ -0,0 +1,24 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// TODO(skal): implement gradient smoothing.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./quant_levels_dec.h"
+
+int DequantizeLevels(uint8_t* const data, int width, int height,
+                     int row, int num_rows) {
+  if (data == NULL || width <= 0 || height <= 0 || row < 0 || num_rows < 0 ||
+      row + num_rows > height) {
+    return 0;
+  }
+  return 1;
+}
+
diff --git a/drivers/webp/utils/quant_levels_dec.h b/drivers/webp/utils/quant_levels_dec.h
new file mode 100644
index 0000000000..0288383aeb
--- /dev/null
+++ b/drivers/webp/utils/quant_levels_dec.h
@@ -0,0 +1,34 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Alpha plane de-quantization utility
+//
+// Author:  Vikas Arora (vikasa@google.com)
+
+#ifndef WEBP_UTILS_QUANT_LEVELS_DEC_H_
+#define WEBP_UTILS_QUANT_LEVELS_DEC_H_
+
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Apply post-processing to input 'data' of size 'width'x'height' assuming that
+// the source was quantized to a reduced number of levels. The post-processing
+// will be applied to 'num_rows' rows of 'data' starting from 'row'.
+// Returns false in case of error (data is NULL, invalid parameters, ...).
+int DequantizeLevels(uint8_t* const data, int width, int height,
+                     int row, int num_rows);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_QUANT_LEVELS_DEC_H_ */
diff --git a/drivers/webp/utils/random.c b/drivers/webp/utils/random.c
new file mode 100644
index 0000000000..24e96ad648
--- /dev/null
+++ b/drivers/webp/utils/random.c
@@ -0,0 +1,43 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Pseudo-random utilities
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <string.h>
+#include "./random.h"
+
+//------------------------------------------------------------------------------
+
+// 31b-range values
+static const uint32_t kRandomTable[VP8_RANDOM_TABLE_SIZE] = {
+  0x0de15230, 0x03b31886, 0x775faccb, 0x1c88626a, 0x68385c55, 0x14b3b828,
+  0x4a85fef8, 0x49ddb84b, 0x64fcf397, 0x5c550289, 0x4a290000, 0x0d7ec1da,
+  0x5940b7ab, 0x5492577d, 0x4e19ca72, 0x38d38c69, 0x0c01ee65, 0x32a1755f,
+  0x5437f652, 0x5abb2c32, 0x0faa57b1, 0x73f533e7, 0x685feeda, 0x7563cce2,
+  0x6e990e83, 0x4730a7ed, 0x4fc0d9c6, 0x496b153c, 0x4f1403fa, 0x541afb0c,
+  0x73990b32, 0x26d7cb1c, 0x6fcc3706, 0x2cbb77d8, 0x75762f2a, 0x6425ccdd,
+  0x24b35461, 0x0a7d8715, 0x220414a8, 0x141ebf67, 0x56b41583, 0x73e502e3,
+  0x44cab16f, 0x28264d42, 0x73baaefb, 0x0a50ebed, 0x1d6ab6fb, 0x0d3ad40b,
+  0x35db3b68, 0x2b081e83, 0x77ce6b95, 0x5181e5f0, 0x78853bbc, 0x009f9494,
+  0x27e5ed3c
+};
+
+void VP8InitRandom(VP8Random* const rg, float dithering) {
+  memcpy(rg->tab_, kRandomTable, sizeof(rg->tab_));
+  rg->index1_ = 0;
+  rg->index2_ = 31;
+  rg->amp_ = (dithering < 0.0) ? 0
+           : (dithering > 1.0) ? (1 << VP8_RANDOM_DITHER_FIX)
+           : (uint32_t)((1 << VP8_RANDOM_DITHER_FIX) * dithering);
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/utils/random.h b/drivers/webp/utils/random.h
new file mode 100644
index 0000000000..08a83e9674
--- /dev/null
+++ b/drivers/webp/utils/random.h
@@ -0,0 +1,62 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Pseudo-random utilities
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_UTILS_RANDOM_H_
+#define WEBP_UTILS_RANDOM_H_
+
+#include <assert.h>
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define VP8_RANDOM_DITHER_FIX 8   // fixed-point precision for dithering
+#define VP8_RANDOM_TABLE_SIZE 55
+
+typedef struct {
+  int index1_, index2_;
+  uint32_t tab_[VP8_RANDOM_TABLE_SIZE];
+  int amp_;
+} VP8Random;
+
+// Initializes random generator with an amplitude 'dithering' in range [0..1].
+void VP8InitRandom(VP8Random* const rg, float dithering);
+
+// Returns a centered pseudo-random number with 'num_bits' amplitude.
+// (uses D.Knuth's Difference-based random generator).
+// 'amp' is in VP8_RANDOM_DITHER_FIX fixed-point precision.
+static WEBP_INLINE int VP8RandomBits2(VP8Random* const rg, int num_bits,
+                                      int amp) {
+  int diff;
+  assert(num_bits + VP8_RANDOM_DITHER_FIX <= 31);
+  diff = rg->tab_[rg->index1_] - rg->tab_[rg->index2_];
+  if (diff < 0) diff += (1u << 31);
+  rg->tab_[rg->index1_] = diff;
+  if (++rg->index1_ == VP8_RANDOM_TABLE_SIZE) rg->index1_ = 0;
+  if (++rg->index2_ == VP8_RANDOM_TABLE_SIZE) rg->index2_ = 0;
+  diff = (diff << 1) >> (32 - num_bits);         // sign-extend, 0-center
+  diff = (diff * amp) >> VP8_RANDOM_DITHER_FIX;  // restrict range
+  diff += 1 << (num_bits - 1);                   // shift back to 0.5-center
+  return diff;
+}
+
+static WEBP_INLINE int VP8RandomBits(VP8Random* const rg, int num_bits) {
+  return VP8RandomBits2(rg, num_bits, rg->amp_);
+}
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_RANDOM_H_ */
diff --git a/drivers/webp/utils/rescaler.c b/drivers/webp/utils/rescaler.c
new file mode 100644
index 0000000000..7061246024
--- /dev/null
+++ b/drivers/webp/utils/rescaler.c
@@ -0,0 +1,152 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Rescaling functions
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include "./rescaler.h"
+
+//------------------------------------------------------------------------------
+
+#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) {
+  wrk->x_expand = (src_width < dst_width);
+  wrk->src_width = src_width;
+  wrk->src_height = src_height;
+  wrk->dst_width = dst_width;
+  wrk->dst_height = dst_height;
+  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_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);
+  wrk->irow = work;
+  wrk->frow = work + num_channels * dst_width;
+}
+
+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];
+  }
+}
+
+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;
+
+    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
+    }
+    wrk->y_accum += wrk->y_add;
+    wrk->dst += wrk->dst_stride;
+    return dst;
+  } else {
+    return NULL;
+  }
+}
+
+#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);
+    }
+    src += src_stride;
+    ++total_imported;
+    wrk->y_accum -= wrk->y_sub;
+  }
+  return total_imported;
+}
+
+int WebPRescalerExport(WebPRescaler* const rescaler) {
+  int total_exported = 0;
+  while (WebPRescalerHasPendingOutput(rescaler)) {
+    WebPRescalerExportRow(rescaler);
+    ++total_exported;
+  }
+  return total_exported;
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/utils/rescaler.h b/drivers/webp/utils/rescaler.h
new file mode 100644
index 0000000000..68e49cee55
--- /dev/null
+++ b/drivers/webp/utils/rescaler.h
@@ -0,0 +1,84 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Rescaling functions
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_UTILS_RESCALER_H_
+#define WEBP_UTILS_RESCALER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "../webp/types.h"
+
+// Structure used for on-the-fly rescaling
+typedef struct {
+  int x_expand;               // true if we're expanding in the x 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.
+  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 src_width, src_height;  // source dimensions
+  int dst_width, dst_height;  // destination dimensions
+  uint8_t* dst;
+  int dst_stride;
+  int32_t* irow, *frow;       // work buffer
+} WebPRescaler;
+
+// Initialize a rescaler given scratch area 'work' and dimensions of src & dst.
+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);
+
+// 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 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);
+
+// 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.
+static WEBP_INLINE
+int WebPRescalerHasPendingOutput(const WebPRescaler* const rescaler) {
+  return (rescaler->y_accum <= 0);
+}
+
+// Export one row from rescaler. Returns the pointer where output was written,
+// or NULL if no row was pending.
+uint8_t* WebPRescalerExportRow(WebPRescaler* const rescaler);
+
+// Export as many rows as possible. Return the numbers of rows written.
+int WebPRescalerExport(WebPRescaler* const rescaler);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_RESCALER_H_ */
diff --git a/drivers/webp/utils/thread.c b/drivers/webp/utils/thread.c
new file mode 100644
index 0000000000..a9e3fae8de
--- /dev/null
+++ b/drivers/webp/utils/thread.c
@@ -0,0 +1,241 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Multi-threaded worker
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <string.h>   // for memset()
+#include "./thread.h"
+
+#ifdef WEBP_USE_THREAD
+
+#if defined(_WIN32)
+
+//------------------------------------------------------------------------------
+// simplistic pthread emulation layer
+
+#include <process.h>
+
+// _beginthreadex requires __stdcall
+#define THREADFN unsigned int __stdcall
+#define THREAD_RETURN(val) (unsigned int)((DWORD_PTR)val)
+
+static int pthread_create(pthread_t* const thread, const void* attr,
+                          unsigned int (__stdcall *start)(void*), void* arg) {
+  (void)attr;
+  *thread = (pthread_t)_beginthreadex(NULL,   /* void *security */
+                                      0,      /* unsigned stack_size */
+                                      start,
+                                      arg,
+                                      0,      /* unsigned initflag */
+                                      NULL);  /* unsigned *thrdaddr */
+  if (*thread == NULL) return 1;
+  SetThreadPriority(*thread, THREAD_PRIORITY_ABOVE_NORMAL);
+  return 0;
+}
+
+static int pthread_join(pthread_t thread, void** value_ptr) {
+  (void)value_ptr;
+  return (WaitForSingleObject(thread, INFINITE) != WAIT_OBJECT_0 ||
+          CloseHandle(thread) == 0);
+}
+
+// Mutex
+static int pthread_mutex_init(pthread_mutex_t* const mutex, void* mutexattr) {
+  (void)mutexattr;
+  InitializeCriticalSection(mutex);
+  return 0;
+}
+
+static int pthread_mutex_lock(pthread_mutex_t* const mutex) {
+  EnterCriticalSection(mutex);
+  return 0;
+}
+
+static int pthread_mutex_unlock(pthread_mutex_t* const mutex) {
+  LeaveCriticalSection(mutex);
+  return 0;
+}
+
+static int pthread_mutex_destroy(pthread_mutex_t* const mutex) {
+  DeleteCriticalSection(mutex);
+  return 0;
+}
+
+// Condition
+static int pthread_cond_destroy(pthread_cond_t* const condition) {
+  int ok = 1;
+  ok &= (CloseHandle(condition->waiting_sem_) != 0);
+  ok &= (CloseHandle(condition->received_sem_) != 0);
+  ok &= (CloseHandle(condition->signal_event_) != 0);
+  return !ok;
+}
+
+static int pthread_cond_init(pthread_cond_t* const condition, void* cond_attr) {
+  (void)cond_attr;
+  condition->waiting_sem_ = CreateSemaphore(NULL, 0, 1, NULL);
+  condition->received_sem_ = CreateSemaphore(NULL, 0, 1, NULL);
+  condition->signal_event_ = CreateEvent(NULL, FALSE, FALSE, NULL);
+  if (condition->waiting_sem_ == NULL ||
+      condition->received_sem_ == NULL ||
+      condition->signal_event_ == NULL) {
+    pthread_cond_destroy(condition);
+    return 1;
+  }
+  return 0;
+}
+
+static int pthread_cond_signal(pthread_cond_t* const condition) {
+  int ok = 1;
+  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_);
+    // wait until the event is consumed so the signaler cannot consume
+    // the event via its own pthread_cond_wait.
+    ok &= (WaitForSingleObject(condition->received_sem_, INFINITE) !=
+           WAIT_OBJECT_0);
+  }
+  return !ok;
+}
+
+static int pthread_cond_wait(pthread_cond_t* const condition,
+                             pthread_mutex_t* const mutex) {
+  int ok;
+  // note that there is a consumer available so the signal isn't dropped in
+  // pthread_cond_signal
+  if (!ReleaseSemaphore(condition->waiting_sem_, 1, NULL))
+    return 1;
+  // now unlock the mutex so pthread_cond_signal may be issued
+  pthread_mutex_unlock(mutex);
+  ok = (WaitForSingleObject(condition->signal_event_, INFINITE) ==
+        WAIT_OBJECT_0);
+  ok &= ReleaseSemaphore(condition->received_sem_, 1, NULL);
+  pthread_mutex_lock(mutex);
+  return !ok;
+}
+
+#else  // !_WIN32
+# define THREADFN void*
+# define THREAD_RETURN(val) val
+#endif  // _WIN32
+
+//------------------------------------------------------------------------------
+
+static THREADFN ThreadLoop(void* ptr) {
+  WebPWorker* const worker = (WebPWorker*)ptr;
+  int done = 0;
+  while (!done) {
+    pthread_mutex_lock(&worker->mutex_);
+    while (worker->status_ == OK) {   // wait in idling mode
+      pthread_cond_wait(&worker->condition_, &worker->mutex_);
+    }
+    if (worker->status_ == WORK) {
+      WebPWorkerExecute(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_);
+  }
+  return THREAD_RETURN(NULL);    // Thread is finished
+}
+
+// main thread state control
+static void ChangeState(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_);
+  }
+  pthread_mutex_unlock(&worker->mutex_);
+}
+
+#endif  // WEBP_USE_THREAD
+
+//------------------------------------------------------------------------------
+
+void WebPWorkerInit(WebPWorker* const worker) {
+  memset(worker, 0, sizeof(*worker));
+  worker->status_ = NOT_OK;
+}
+
+int WebPWorkerSync(WebPWorker* const worker) {
+#ifdef WEBP_USE_THREAD
+  ChangeState(worker, OK);
+#endif
+  assert(worker->status_ <= OK);
+  return !worker->had_error;
+}
+
+int WebPWorkerReset(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)) {
+      return 0;
+    }
+    pthread_mutex_lock(&worker->mutex_);
+    ok = !pthread_create(&worker->thread_, NULL, ThreadLoop, worker);
+    if (ok) worker->status_ = OK;
+    pthread_mutex_unlock(&worker->mutex_);
+#else
+    worker->status_ = OK;
+#endif
+  } else if (worker->status_ > OK) {
+    ok = WebPWorkerSync(worker);
+  }
+  assert(!ok || (worker->status_ == OK));
+  return ok;
+}
+
+void WebPWorkerExecute(WebPWorker* const worker) {
+  if (worker->hook != NULL) {
+    worker->had_error |= !worker->hook(worker->data1, worker->data2);
+  }
+}
+
+void WebPWorkerLaunch(WebPWorker* const worker) {
+#ifdef WEBP_USE_THREAD
+  ChangeState(worker, WORK);
+#else
+  WebPWorkerExecute(worker);
+#endif
+}
+
+void WebPWorkerEnd(WebPWorker* const worker) {
+  if (worker->status_ >= OK) {
+#ifdef WEBP_USE_THREAD
+    ChangeState(worker, NOT_OK);
+    pthread_join(worker->thread_, NULL);
+    pthread_mutex_destroy(&worker->mutex_);
+    pthread_cond_destroy(&worker->condition_);
+#else
+    worker->status_ = NOT_OK;
+#endif
+  }
+  assert(worker->status_ == NOT_OK);
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/utils/thread.h b/drivers/webp/utils/thread.h
new file mode 100644
index 0000000000..aef33bdaf2
--- /dev/null
+++ b/drivers/webp/utils/thread.h
@@ -0,0 +1,97 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Multi-threaded worker
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_UTILS_THREAD_H_
+#define WEBP_UTILS_THREAD_H_
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef 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;
+
+#else
+
+#include <pthread.h>
+
+#endif    /* _WIN32 */
+#endif    /* WEBP_USE_THREAD */
+
+// State of the worker thread object
+typedef enum {
+  NOT_OK = 0,   // object is unusable
+  OK,           // ready to work
+  WORK          // busy finishing the current task
+} WebPWorkerStatus;
+
+// Function to be called by the worker thread. Takes two opaque pointers as
+// 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
+typedef struct {
+#ifdef WEBP_USE_THREAD
+  pthread_mutex_t mutex_;
+  pthread_cond_t  condition_;
+  pthread_t       thread_;
+#endif
+  WebPWorkerStatus status_;
+  WebPWorkerHook hook;    // hook to call
+  void* data1;            // first argument passed to 'hook'
+  void* data2;            // second argument passed to 'hook'
+  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 to 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);
+// 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 WebPWorkerSync(WebPWorker* const worker);
+// Triggers 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);
+// This function is similar to WebPWorkerLaunch() except that it calls the
+// hook directly instead of using a thread. Convenient to bypass the thread
+// mechanism while still using the WebPWorker structs. WebPWorkerSync() must
+// still be called afterward (for error reporting).
+void WebPWorkerExecute(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);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_THREAD_H_ */
diff --git a/drivers/webp/utils/utils.c b/drivers/webp/utils/utils.c
new file mode 100644
index 0000000000..5592538988
--- /dev/null
+++ b/drivers/webp/utils/utils.c
@@ -0,0 +1,42 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Misc. common utility functions
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <stdlib.h>
+#include "./utils.h"
+
+//------------------------------------------------------------------------------
+// Checked memory allocation
+
+// 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;
+  return 1;
+}
+
+void* WebPSafeMalloc(uint64_t nmemb, size_t size) {
+  if (!CheckSizeArgumentsOverflow(nmemb, size)) return NULL;
+  assert(nmemb * size > 0);
+  return malloc((size_t)(nmemb * size));
+}
+
+void* WebPSafeCalloc(uint64_t nmemb, size_t size) {
+  if (!CheckSizeArgumentsOverflow(nmemb, size)) return NULL;
+  assert(nmemb * size > 0);
+  return calloc((size_t)nmemb, size);
+}
+
+//------------------------------------------------------------------------------
+
diff --git a/drivers/webp/utils/utils.h b/drivers/webp/utils/utils.h
new file mode 100644
index 0000000000..8bdf0f03db
--- /dev/null
+++ b/drivers/webp/utils/utils.h
@@ -0,0 +1,83 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Misc. common utility functions
+//
+// Authors: Skal (pascal.massimino@gmail.com)
+//          Urvang (urvang@google.com)
+
+#ifndef WEBP_UTILS_UTILS_H_
+#define WEBP_UTILS_UTILS_H_
+
+#include <assert.h>
+
+#include "../webp/types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//------------------------------------------------------------------------------
+// Memory allocation
+
+// This is the maximum memory amount that libwebp will ever try to allocate.
+#define WEBP_MAX_ALLOCABLE_MEMORY (1ULL << 40)
+
+// size-checking safe malloc/calloc: verify that the requested size is not too
+// large, or return NULL. You don't need to call these for constructs like
+// malloc(sizeof(foo)), but only if there's picture-dependent size involved
+// 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);
+// 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);
+
+//------------------------------------------------------------------------------
+// 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 (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));
+}
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_UTILS_H_ */