Add animation support for WebP images

Source/core/platform/image-decoders/webp/WEBPImageDecoder.cpp

 /*
  * Copyright (C) 2010 Google Inc. All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions
  * are met:
  *
  * 1.  Redistributions of source code must retain the above copyright
  *     notice, this list of conditions and the following disclaimer.
  * 2.  Redistributions in binary form must reproduce the above copyright
@@ skipping 15 matching lines @@
  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
 #include "config.h"
 #include "core/platform/image-decoders/webp/WEBPImageDecoder.h"
 
 #include "core/platform/PlatformInstrumentation.h"
 
 #ifdef QCMS_WEBP_COLOR_CORRECTION
 #include "qcms.h"
-#include "webp/demux.h"
-#else
-#undef ICCP_FLAG
-#define ICCP_FLAG 0
 #endif
 
-// Backward emulation for earlier versions than 0.1.99.
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+#include "RuntimeEnabledFeatures.h"
+#include "webp/format_constants.h"
+#endif
+
 #if (WEBP_DECODER_ABI_VERSION < 0x0163)
+// Backward emulation for versions earlier than 0.1.99.
 #define MODE_rgbA MODE_RGBA
 #define MODE_bgrA MODE_BGRA
+#define ALPHA_FLAG 0
+#elif (WEBP_DECODER_ABI_VERSION <= 0x0200)
+// Backward emulation for versions earlier than 0.3.0.
+#define ALPHA_FLAG 0x000010
 #endif
 
 #if CPU(BIG_ENDIAN) || CPU(MIDDLE_ENDIAN)
 inline WEBP_CSP_MODE outputMode(bool hasAlpha) { return hasAlpha ? MODE_rgbA : MODE_RGBA; }
 #elif SK_B32_SHIFT
 inline WEBP_CSP_MODE outputMode(bool hasAlpha) { return hasAlpha ? MODE_rgbA : MODE_RGBA; }
 #else // LITTLE_ENDIAN, output BGRA pixels.
 inline WEBP_CSP_MODE outputMode(bool hasAlpha) { return hasAlpha ? MODE_bgrA : MODE_BGRA; }
 #endif
 
 namespace WebCore {
 
 WEBPImageDecoder::WEBPImageDecoder(ImageSource::AlphaOption alphaOption,
                                    ImageSource::GammaAndColorProfileOption gammaAndColorProfileOption)
     : ImageDecoder(alphaOption, gammaAndColorProfileOption)
     , m_decoder(0)
-    , m_hasAlpha(false)
     , m_formatFlags(0)
+    , m_frameBackgroundHasAlpha(false)
 #ifdef QCMS_WEBP_COLOR_CORRECTION
     , m_haveReadProfile(false)
     , m_transform(0)
+#endif
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+    , m_demux(0)
+    , m_demuxState(WEBP_DEMUX_PARSING_HEADER)
+    , m_haveAlreadyParsedThisData(false)
+    , m_haveReadAnimationParameters(false)
+    , m_repetitionCount(cAnimationLoopOnce)
     , m_decodedHeight(0)
 #endif
 {
-    WebPInitDecBuffer(&m_decoderBuffer);
 }
 
 WEBPImageDecoder::~WEBPImageDecoder()
 {
     clear();
 }
 
 void WEBPImageDecoder::clear()
 {
 #ifdef QCMS_WEBP_COLOR_CORRECTION
     if (m_transform)
         qcms_transform_release(m_transform);
     m_transform = 0;
 #endif
-    WebPFreeDecBuffer(&m_decoderBuffer);
-    if (m_decoder)
-        WebPIDelete(m_decoder);
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+    WebPDemuxDelete(m_demux);
+    m_demux = 0;
+#endif
+    clearDecoder();
+}
+
+void WEBPImageDecoder::clearDecoder()
+{
+    WebPIDelete(m_decoder);
     m_decoder = 0;
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+    m_decodedHeight = 0;
+    m_frameBackgroundHasAlpha = false;
+#endif
 }
 
 bool WEBPImageDecoder::isSizeAvailable()
 {
-    if (!ImageDecoder::isSizeAvailable())
-         decode(true);
-
+    if (!ImageDecoder::isSizeAvailable()) {
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+        updateDemuxer();
+#else
+        decode(reinterpret_cast<const uint8_t*>(m_data->data()), m_data->size(), true, 0);
+#endif
+    }
     return ImageDecoder::isSizeAvailable();
 }
 
-ImageFrame* WEBPImageDecoder::frameBufferAtIndex(size_t index)
-{
-    if (index)
+size_t WEBPImageDecoder::frameCount()
+{
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+    if (!updateDemuxer())
         return 0;
-
+#else
     if (m_frameBufferCache.isEmpty()) {
         m_frameBufferCache.resize(1);
         m_frameBufferCache[0].setPremultiplyAlpha(m_premultiplyAlpha);
     }
-
-    ImageFrame& frame = m_frameBufferCache[0];
-    if (frame.status() != ImageFrame::FrameComplete) {
-        PlatformInstrumentation::willDecodeImage("WEBP");
-        decode(false);
-        PlatformInstrumentation::didDecodeImage();
-    }
+#endif
+    return m_frameBufferCache.size();
+}
+
+ImageFrame* WEBPImageDecoder::frameBufferAtIndex(size_t index)
+{
+    if (index >= frameCount())
+        return 0;
+
+    ImageFrame& frame = m_frameBufferCache[index];
+    if (frame.status() == ImageFrame::FrameComplete)
+        return &frame;
+
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+    if (RuntimeEnabledFeatures::animatedWebPEnabled()) {
+        Vector<size_t> framesToDecode;
+        size_t frameToDecode = index;
+        do {
+            framesToDecode.append(frameToDecode);
+            frameToDecode = m_frameBufferCache[frameToDecode].requiredPreviousFrameIndex();
+        } while (frameToDecode != notFound && m_frameBufferCache[frameToDecode].status() != ImageFrame::FrameComplete);
+
+        ASSERT(m_demux);
+        for (size_t i = framesToDecode.size(); i > 0; --i) {
+            size_t frameIndex = framesToDecode[i - 1];
+            WebPIterator webpFrame;
+            if (!WebPDemuxGetFrame(m_demux, frameIndex + 1, &webpFrame))
+                return 0;
+            if ((m_formatFlags & ANIMATION_FLAG) && !initFrameBuffer(webpFrame, frameIndex)) {
+                WebPDemuxReleaseIterator(&webpFrame);
+                return 0;
+            }
+            PlatformInstrumentation::willDecodeImage("WEBP");
+            decode(webpFrame.fragment.bytes, webpFrame.fragment.size, false, frameIndex);
+            PlatformInstrumentation::didDecodeImage();
+            WebPDemuxReleaseIterator(&webpFrame);
+
+            // We need more data to continue decoding.
+            if (m_frameBufferCache[frameIndex].status() != ImageFrame::FrameComplete)
+                break;
+        }
+
+        // It is also a fatal error if all data is received and we have decoded all
+        // frames available but the file is truncated.
+        if (index >= m_frameBufferCache.size() - 1 && isAllDataReceived() && m_demux && m_demuxState != WEBP_DEMUX_DONE)
+            setFailed();
+
+        return &frame;
+    }
+#endif
+
+    ASSERT(!index);
+    PlatformInstrumentation::willDecodeImage("WEBP");
+    decode(reinterpret_cast<const uint8_t*>(m_data->data()), m_data->size(), false, index);
+    PlatformInstrumentation::didDecodeImage();
     return &frame;
 }
 
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+
+void WEBPImageDecoder::setData(SharedBuffer* data, bool allDataReceived)
+{
+    if (failed())
+        return;
+
+    ImageDecoder::setData(data, allDataReceived);
+
+    if (m_demuxState != WEBP_DEMUX_DONE)
+        m_haveAlreadyParsedThisData = false;
+}
+
+int WEBPImageDecoder::repetitionCount() const
+{
+    return failed() ? cAnimationLoopOnce : m_repetitionCount;
+}
+
+bool WEBPImageDecoder::frameIsCompleteAtIndex(size_t index) const
+{
+    if (!RuntimeEnabledFeatures::animatedWebPEnabled())
+        return ImageDecoder::frameIsCompleteAtIndex(index);
+    if (!m_demux || m_demuxState <= WEBP_DEMUX_PARSING_HEADER)
+        return false;
+    if (!(m_formatFlags & ANIMATION_FLAG))
+        return ImageDecoder::frameIsCompleteAtIndex(index);
+    bool frameIsLoadedAtIndex = index < m_frameBufferCache.size();
+    return frameIsLoadedAtIndex;
+}
+
+float WEBPImageDecoder::frameDurationAtIndex(size_t index) const
+{
+    return index < m_frameBufferCache.size() ? m_frameBufferCache[index].duration() : 0;
+}
+
+bool WEBPImageDecoder::updateDemuxer()
+{
+    if (m_haveAlreadyParsedThisData)
+        return true;
+
+    m_haveAlreadyParsedThisData = true;
+
+    if (m_data->size() < RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE)
+        return false; // Wait for headers so that WebPDemuxPartial doesn't return null.
+
+    WebPDemuxDelete(m_demux);
+    WebPData inputData = { reinterpret_cast<const uint8_t*>(m_data->data()), m_data->size() };
+    m_demux = WebPDemuxPartial(&inputData, &m_demuxState);
+    if (!m_demux)
+        return setFailed();
+
+    if (m_demuxState <= WEBP_DEMUX_PARSING_HEADER)
+        return false; // Not enough data for parsing canvas width/height yet.
+
+    bool hasAnimation = (m_formatFlags & ANIMATION_FLAG);
+    if (!ImageDecoder::isSizeAvailable()) {
+        m_formatFlags = WebPDemuxGetI(m_demux, WEBP_FF_FORMAT_FLAGS);
+        hasAnimation = (m_formatFlags & ANIMATION_FLAG);
+        if (hasAnimation && !RuntimeEnabledFeatures::animatedWebPEnabled())
+            return setFailed();
+        if (!setSize(WebPDemuxGetI(m_demux, WEBP_FF_CANVAS_WIDTH), WebPDemuxGetI(m_demux, WEBP_FF_CANVAS_HEIGHT)))
+            return setFailed();
+    }
+    ASSERT(ImageDecoder::isSizeAvailable());
+    const size_t newFrameCount = WebPDemuxGetI(m_demux, WEBP_FF_FRAME_COUNT);
+    if (hasAnimation && !m_haveReadAnimationParameters && newFrameCount) {
+        // As we have parsed at least one frame (even if partially),
+        // we must already have parsed the animation properties.
+        // This is because ANIM chunk always precedes ANMF chunks.
+        m_repetitionCount = WebPDemuxGetI(m_demux, WEBP_FF_LOOP_COUNT);
+        ASSERT(m_repetitionCount == (m_repetitionCount & 0xffff)); // Loop count is always <= 16 bits.
+        if (!m_repetitionCount)
+            m_repetitionCount = cAnimationLoopInfinite;
+        m_haveReadAnimationParameters = true;
+    }
+    const size_t oldFrameCount = m_frameBufferCache.size();
+    if (newFrameCount > oldFrameCount) {
+        m_frameBufferCache.resize(newFrameCount);
+        for (size_t i = oldFrameCount; i < newFrameCount; ++i) {
+            m_frameBufferCache[i].setPremultiplyAlpha(m_premultiplyAlpha);
+            if (!hasAnimation) {
+                ASSERT(!i);
+                m_frameBufferCache[i].setRequiredPreviousFrameIndex(notFound);
+                continue;
+            }
+            WebPIterator animatedFrame;
+            WebPDemuxGetFrame(m_demux, i + 1, &animatedFrame);
+            ASSERT(animatedFrame.complete == 1);
+            m_frameBufferCache[i].setDuration(animatedFrame.duration);
+            m_frameBufferCache[i].setDisposalMethod(animatedFrame.dispose_method == WEBP_MUX_DISPOSE_BACKGROUND ? ImageFrame::DisposeOverwriteBgcolor : ImageFrame::DisposeKeep);
+            WebPDemuxReleaseIterator(&animatedFrame);
+            m_frameBufferCache[i].setRequiredPreviousFrameIndex(findRequiredPreviousFrame(i));
+        }
+    }
+    return true;
+}
+
+bool WEBPImageDecoder::initFrameBuffer(const WebPIterator& frame, size_t frameIndex)
+{
+    ImageFrame& buffer = m_frameBufferCache[frameIndex];
+    if (buffer.status() != ImageFrame::FrameEmpty) // Already initialized.
+        return true;
+
+    // Initialize the frame rect in our buffer.
+    IntRect frameRect(frame.x_offset, frame.y_offset, frame.width, frame.height);
+
+    // Make sure the frameRect doesn't extend outside the buffer.
+    if (frameRect.maxX() > size().width())
+        frameRect.setWidth(size().width() - frame.x_offset);
+    if (frameRect.maxY() > size().height())
+        frameRect.setHeight(size().height() - frame.y_offset);
+    buffer.setOriginalFrameRect(frameRect);
+
+    const size_t requiredPreviousFrameIndex = buffer.requiredPreviousFrameIndex();
+    if (requiredPreviousFrameIndex == notFound) {
+        // This frame doesn't rely on any previous data.
+        if (!buffer.setSize(size().width(), size().height()))
+            return setFailed();
+        m_frameBackgroundHasAlpha = !frameRect.contains(IntRect(IntPoint(), size()));
+    } else {
+        const ImageFrame& prevBuffer = m_frameBufferCache[requiredPreviousFrameIndex];
+        ASSERT(prevBuffer.status() == ImageFrame::FrameComplete);
+
+        // Preserve the last frame as the starting state for this frame.
+        if (!buffer.copyBitmapData(prevBuffer))
+            return setFailed();
+
+        if (prevBuffer.disposalMethod() == ImageFrame::DisposeOverwriteBgcolor) {
+            // We want to clear the previous frame to transparent, without
+            // affecting pixels in the image outside of the frame.
+            const IntRect& prevRect = prevBuffer.originalFrameRect();
+            ASSERT(!prevRect.contains(IntRect(IntPoint(), size())));
+            for (int y = prevRect.y(); y < prevRect.maxY(); ++y) {
+                for (int x = prevRect.x(); x < prevRect.maxX(); ++x)
+                    buffer.setRGBA(x, y, 0, 0, 0, 0);

[Peter Kasting, 2013/06/26 00:18:17]

I suspect there's a faster way of doing this, though it may require calling Skia
functions directly.

[urvang (Google), 2013/06/26 22:12:47]

GIFDecoder
(https://chromium.googlesource.com/chromium/blink/+/master/Source/core/platfor...)
uses the same method for clearing pixels to transparent; so I'm guessing this is
the best way.

[Peter Kasting, 2013/06/26 22:58:23]

That's only because before we forked Blink, the GIFDecoder was using generic
non-Skia-specific ImageDecoder functions.  If we can do better now, we should.

[urvang (Google), 2013/06/27 01:27:37]

Let me ask this question: do you see any method that can do this better? I did't
see any in
https://code.google.com/p/chromium/codesearch#chromium/src/third_party/WebKit...

Perhaps we could create a method, which, given a frame rectangle, clears all
pixels in there to (0, 0, 0, 0) -- using memset().
But, I feel that should belong to its own patch, no?

[skal, 2013/06/27 07:22:27]

In the WebP decoder, we grab the pointer directly from ImageFrame and write
directly into memory:
(see WebPImageDecoder.cpp:193)

uint8_t* dst = reinterpret_cast<uint8_t*>(buffer.getAddr(prevRect.x(),
prevRect.y()));
memset(dst, 0, (prevRect.maxX() - prevRect.x()) *
sizeof(ImageFrame::PixelData));

would work, i think.

[Peter Kasting, 2013/06/27 17:40:18]

Right, that's what I'm saying: up until the fork, that file had to be
cross-platform and have no Skia references.  Whereas if we have a fast method to
clear a rectangle out of an image, we'll need to look in Skia itself, and then
we'll want to make that available to both the WebP and GIF decoders.
It would probably be an improvement.  What I was really hoping for in asking
this question was for someone to contact Mike Reed (or another Skia wizard) and
ask what the fastest method to clear a rectangle out of an image is.

[urvang (Google), 2013/06/27 22:53:12]

As discussed offline, I added a method to zero-fill a frame rectangle using the
memset.
Contacting Mike offline to find out if there is a better way to do this
(separate patch).

+            }
+        }
+
+        m_frameBackgroundHasAlpha = prevBuffer.hasAlpha() || (prevBuffer.disposalMethod() == ImageFrame::DisposeOverwriteBgcolor);
+    }
+
+    buffer.setStatus(ImageFrame::FramePartial);
+    // The buffer is transparent outside the decoded area while the image is loading.
+    // The correct value of 'hasAlpha' for the frame will be set when it is fully decoded.
+    buffer.setHasAlpha(true);
+    return true;
+}
+
+size_t WEBPImageDecoder::clearCacheExceptFrame(size_t clearExceptFrame)
+{
+    // Don't clear if there are no frames or only one frame.
+    if (m_frameBufferCache.size() <= 1)
+        return 0;
+
+    // If |clearExceptFrame| has status FrameComplete, we preserve that frame.
+    // Otherwise, we preserve a previous frame with status FrameComplete whose data is required to decode |clearExceptFrame, either in initFrameBuffer() or ApplyPostProcessing().

[Peter Kasting, 2013/06/26 00:18:17]

Nit: Missing pipe

[urvang (Google), 2013/06/26 22:12:47]

Done.

+    // All other frames can be cleared.
+    while ((clearExceptFrame < m_frameBufferCache.size()) && (m_frameBufferCache[clearExceptFrame].status() != ImageFrame::FrameComplete))

[Peter Kasting, 2013/06/26 00:18:17]

This function differs only fractionally from the superclass method.  Copy and
pasting this much code is error-prone.  Find a way to share most of the code.

[urvang (Google), 2013/06/26 22:12:47]

Done. Now, the logic of finding which frame to preserve is part of GIF/WebP
decoders. And base class just performs clearing (except the special case of 0 or
1 frames).

+        clearExceptFrame = m_frameBufferCache[clearExceptFrame].requiredPreviousFrameIndex();
+
+    size_t frameBytesCleared = 0;
+    for (size_t i = 0; i < m_frameBufferCache.size(); ++i) {
+        if (i != clearExceptFrame) {
+            frameBytesCleared += frameBytesAtIndex(i);
+            clearFrameBuffer(i);
+        }
+    }
+    return frameBytesCleared;
+}
+
+void WEBPImageDecoder::clearFrameBuffer(size_t frameIndex)
+{
+    if (m_demux && m_demuxState >= WEBP_DEMUX_PARSED_HEADER && m_frameBufferCache[frameIndex].status() == ImageFrame::FramePartial) {
+        // Clear the decoder state so that this partial frame can be decoded again when requested.
+        clearDecoder();
+    }
+    ImageDecoder::clearFrameBuffer(frameIndex);
+}
+
+#endif // WEBP_ICC_ANIMATION_SUPPORT
+
 #ifdef QCMS_WEBP_COLOR_CORRECTION
 
 void WEBPImageDecoder::createColorTransform(const char* data, size_t size)
 {
     if (m_transform)
         qcms_transform_release(m_transform);
     m_transform = 0;
 
     qcms_profile* deviceProfile = ImageDecoder::qcmsOutputDeviceProfile();
     if (!deviceProfile)
         return;
     qcms_profile* inputProfile = qcms_profile_from_memory(data, size);
     if (!inputProfile)
         return;
 
     // We currently only support color profiles for RGB profiled images.
     ASSERT(icSigRgbData == qcms_profile_get_color_space(inputProfile));
     // The input image pixels are RGBA format.
     qcms_data_type format = QCMS_DATA_RGBA_8;
     // FIXME: Don't force perceptual intent if the image profile contains an intent.
     m_transform = qcms_transform_create(inputProfile, format, deviceProfile, QCMS_DATA_RGBA_8, QCMS_INTENT_PERCEPTUAL);
 
     qcms_profile_release(inputProfile);
 }
 
-void WEBPImageDecoder::readColorProfile(const uint8_t* data, size_t size)
+void WEBPImageDecoder::readColorProfile()
 {
     WebPChunkIterator chunkIterator;
-    WebPData inputData = { data, size };
-    WebPDemuxState state;
-
-    WebPDemuxer* demuxer = WebPDemuxPartial(&inputData, &state);
-    if (!WebPDemuxGetChunk(demuxer, "ICCP", 1, &chunkIterator)) {
+    if (!WebPDemuxGetChunk(m_demux, "ICCP", 1, &chunkIterator)) {
         WebPDemuxReleaseChunkIterator(&chunkIterator);
-        WebPDemuxDelete(demuxer);
         return;
     }
 
     const char* profileData = reinterpret_cast<const char*>(chunkIterator.chunk.bytes);
     size_t profileSize = chunkIterator.chunk.size;
 
     // Only accept RGB color profiles from input class devices.
     bool ignoreProfile = false;
     if (profileSize < ImageDecoder::iccColorProfileHeaderLength)
         ignoreProfile = true;
     else if (!ImageDecoder::rgbColorProfile(profileData, profileSize))
         ignoreProfile = true;
     else if (!ImageDecoder::inputDeviceColorProfile(profileData, profileSize))
         ignoreProfile = true;
 
     if (!ignoreProfile)
         createColorTransform(profileData, profileSize);
 
     WebPDemuxReleaseChunkIterator(&chunkIterator);
-    WebPDemuxDelete(demuxer);
 }
 
-void WEBPImageDecoder::applyColorProfile(const uint8_t* data, size_t dataSize, ImageFrame& buffer)
+#endif // QCMS_WEBP_COLOR_CORRECTION
+
+#ifdef WEBP_ICC_ANIMATION_SUPPORT
+void WEBPImageDecoder::applyPostProcessing(size_t frameIndex)
 {
+    ImageFrame& buffer = m_frameBufferCache[frameIndex];
     int width;
     int decodedHeight;
     if (!WebPIDecGetRGB(m_decoder, &decodedHeight, &width, 0, 0))
         return; // See also https://bugs.webkit.org/show_bug.cgi?id=74062
     if (decodedHeight <= 0)
         return;
 
-    if (!m_haveReadProfile) {
-        readColorProfile(data, dataSize);
-        m_haveReadProfile = true;
+    const IntRect& frameRect = buffer.originalFrameRect();
+    ASSERT_WITH_SECURITY_IMPLICATION(width == frameRect.width());
+    ASSERT_WITH_SECURITY_IMPLICATION(decodedHeight <= frameRect.height());
+    const int left = frameRect.x();
+    const int top = frameRect.y();
+
+#ifdef QCMS_WEBP_COLOR_CORRECTION
+    if ((m_formatFlags & ICCP_FLAG) && !ignoresGammaAndColorProfile()) {
+        if (!m_haveReadProfile) {
+            readColorProfile();
+            m_haveReadProfile = true;
+        }
+        for (int y = m_decodedHeight; y < decodedHeight; ++y) {
+            const int canvasY = top + y;
+            uint8_t* row = reinterpret_cast<uint8_t*>(buffer.getAddr(left, canvasY));
+            if (qcms_transform* transform = colorTransform())
+                qcms_transform_data_type(transform, row, row, width, QCMS_OUTPUT_RGBX);
+            uint8_t* pixel = row;
+            for (int x = 0; x < width; ++x, pixel += 4) {
+                const int canvasX = left + x;
+                buffer.setRGBA(canvasX, canvasY, pixel[0], pixel[1], pixel[2], pixel[3]);
+            }
+        }
     }
+#endif // QCMS_WEBP_COLOR_CORRECTION
 
-    ASSERT(width == size().width());
-    ASSERT(decodedHeight <= size().height());
-
-    for (int y = m_decodedHeight; y < decodedHeight; ++y) {
-        uint8_t* row = reinterpret_cast<uint8_t*>(buffer.getAddr(0, y));
-        if (qcms_transform* transform = colorTransform())
-            qcms_transform_data_type(transform, row, row, width, QCMS_OUTPUT_RGBX);
-        uint8_t* pixel = row;
-        for (int x = 0; x < width; ++x, pixel += 4)
-            buffer.setRGBA(x, y, pixel[0], pixel[1], pixel[2], pixel[3]);
+    // During the decoding of current frame, we may have set some pixels to be transparent (i.e. alpha < 255).
+    // However, the value of each of these pixels should have been determined by blending it against the value
+    // of that pixel in the previous frame. So, we correct these pixels based on disposal method of the previous
+    // frame and the previous frame buffer.

[Peter Kasting, 2013/06/26 00:18:17]

Argh.  It really sucks to do this as another pass after we've decoded.  Why
can't we integrate this directly into the decode pass?  That should be both
faster and easier.

This would also eliminate the need to override clearCacheExceptFrame().

[urvang (Google), 2013/06/26 22:12:47]

Sorry, I should have added you earlier to the review, when this question was
answered. To repeat, this is because of the difference between WebP and GIF
decode:

In case of GIF, here's what happens:
- It first decodes current frame into its own buffer.
- Then, copies pixel-by-pixel to the ImageFrame buffer -- either the previous
frame pixel or current frame pixel -- as per the 'alpha' value.

In case of WebP:
- It decodes directly into the ImageFrame buffer. (No extra memory is used).
[One reason for this is that Decode Row callback is very hard to implement in
case of WebP].
- Only when prev frame and current frame have some common area, only that common
area would need to be re-looked to see if we should restore pixels to those in
prev frame.

So, technically, WebP takes less memory (no extra decode buffer) and "should I
restore to previous pixel" check happens for smaller area of the frame.
I feel these benefits compensate for the complexity.

[Peter Kasting, 2013/06/26 22:58:23]

I would imagine the common case is that the previous frame and the current frame
are both full frames, so we're running the check over the whole frame every
frame.  That's certainly the common case with GIF.

I think we can do better than this.  If the WebP decode function can decode
directly to an ImageFrame, then also hand it a pointer to the previous
ImageFrame (or frame data for it or whatever), so that when it hits a
transparent pixel it can do the blend as part of the decode.

The GIF decoder could be improved to do this as well by having it decode
directly to an ImageFrame instead of using the HaveDecodedRow callback
implementation.

[urvang (Google), 2013/06/27 01:27:37]

Well, that would be desirable, but it is also non-trivial. Consider what all the
decoder will have to take care of (in addition to alpha-blending):
- Using disposal method to find out previous required frames.
- The frame rectangle of current and previous required frame(s).
- Using color profile (dependency on QCMS, not available for all platforms)
- With all these, he library would need to work on all platforms.

This is almost duplicating what should be the job of a renderer. It would also
blow up the size of the library and dependencies. And 

Given the complexities, I could think about having this support in libwebp in
future, but not for now.

[Peter Kasting, 2013/06/27 17:40:18]

Not actually necessary.  The decoder assumes the embedder passes in the correct
image data to composite this frame against.  It's the embedder's job to figure
out which frame this is, ensure it has the right data, and provide it to the
decoder.
The decoder already has to know the current frame rect in order to write to the
correct coordinates within the image.  It doesn't actually need to know the
previous frame rect, it just composites against the same coordinates in the
previous frame.
Ultimately, this seems like the sort of thing the decoder should take care of. 
However, for now, it seems like this could be implemented as a callout hook to
the embedder, basically "here's the raw row data, please do any necessary color
correction".  The decoder would then call this before compositing.  This should
be approximately the same speed as your current implementation, I think?

[urvang (Google), 2013/06/27 22:53:12]

That could probably be made to work; some details will have to be worked out
though:

e.g. The code which outputs the pixels is deep inside the codebase -- which
would affect lossy & lossless, incremental and non-incremental decoding. Will
have to be written carefully and well-tested.
Actually, the decoder isn't aware of the frame rectangle. It just writes into
the output buffer given to it.
But, that shouldn't be a problem, as the renderer can always pass the
appropriate output buffer (pointer to the top-left pixel, width, height and
stride) for prev and current frame.

To summarize: It's a good suggestion and adding such an API could be considered
for the next major libwebp release.
Until that API comes up, we'll have to make do with the current approach.

As discussed offline, putting a FIXME to take this up later.

+    if ((m_formatFlags & ANIMATION_FLAG) && frameIndex) {
+        ImageFrame& prevBuffer = m_frameBufferCache[frameIndex - 1];
+        ImageFrame::FrameDisposalMethod prevMethod = prevBuffer.disposalMethod();
+        if (prevMethod == ImageFrame::DisposeKeep) { // Restore transparent pixels to pixels in previous canvas.
+            ASSERT(prevBuffer.status() == ImageFrame::FrameComplete);
+            for (int y = m_decodedHeight; y < decodedHeight; ++y) {
+                const int canvasY = top + y;
+                for (int x = 0; x < width; ++x) {
+                    const int canvasX = left + x;
+                    ImageFrame::PixelData& pixel = *buffer.getAddr(canvasX, canvasY);
+                    // FIXME: Use alpha-blending when alpha is between 0 and 255.
+                    // Alpha-blending is being implemented in: https://bugs.webkit.org/show_bug.cgi?id=17022
+                    if (!((pixel >> SK_A32_SHIFT) & 0xff)) {
+                        ImageFrame::PixelData prevPixel = *prevBuffer.getAddr(canvasX, canvasY);
+                        pixel = prevPixel;
+                    }
+                }
+            }
+        } else if (prevMethod == ImageFrame::DisposeOverwriteBgcolor && buffer.requiredPreviousFrameIndex() != notFound) {
+            // Note: if the requiredPreviousFrameIndex is |notFound|, there's nothing to do.
+            ASSERT(prevBuffer.status() == ImageFrame::FrameComplete);
+            const IntRect& prevRect = prevBuffer.originalFrameRect();
+            // We need to restore transparent pixels to as they were just after initFrame() call. That is:
+            //   * Transparent if it belongs to prevRect <-- This is a no-op.
+            //   * Pixel in the previous canvas otherwise <-- Need to restore.
+            for (int y = m_decodedHeight; y < decodedHeight; ++y) {
+                const int canvasY = top + y;
+                for (int x = 0; x < width; ++x) {
+                    const int canvasX = left + x;
+                    ImageFrame::PixelData& pixel = *buffer.getAddr(canvasX, canvasY);
+                    // FIXME: Use alpha-blending when alpha is between 0 and 255.
+                    if (!((pixel >> SK_A32_SHIFT) & 0xff) && !prevRect.contains(IntPoint(canvasX, canvasY))) {
+                        ImageFrame::PixelData prevPixel = *prevBuffer.getAddr(canvasX, canvasY);
+                        pixel = prevPixel;
+                    }
+                }
+            }
+        }
     }
 
     m_decodedHeight = decodedHeight;
 }
+#endif // WEBP_ICC_ANIMATION_SUPPORT
 
-#endif // QCMS_WEBP_COLOR_CORRECTION
-
-bool WEBPImageDecoder::decode(bool onlySize)
+bool WEBPImageDecoder::decode(const uint8_t* dataBytes, size_t dataSize, bool onlySize, size_t frameIndex)
 {
     if (failed())
         return false;
 
-    const uint8_t* dataBytes = reinterpret_cast<const uint8_t*>(m_data->data());
-    const size_t dataSize = m_data->size();
-
     if (!ImageDecoder::isSizeAvailable()) {
         static const size_t imageHeaderSize = 30;
         if (dataSize < imageHeaderSize)
             return false;
         int width, height;
-#ifdef QCMS_WEBP_COLOR_CORRECTION
-        WebPData inputData = { dataBytes, dataSize };
-        WebPDemuxState state;
-        WebPDemuxer* demuxer = WebPDemuxPartial(&inputData, &state);
-        if (!demuxer)
-            return setFailed();
-
-        width = WebPDemuxGetI(demuxer, WEBP_FF_CANVAS_WIDTH);
-        height = WebPDemuxGetI(demuxer, WEBP_FF_CANVAS_HEIGHT);
-        m_formatFlags = WebPDemuxGetI(demuxer, WEBP_FF_FORMAT_FLAGS);
-        m_hasAlpha = !!(m_formatFlags & ALPHA_FLAG);
-
-        WebPDemuxDelete(demuxer);
-        if (state <= WEBP_DEMUX_PARSING_HEADER)
-            return false;
-#elif (WEBP_DECODER_ABI_VERSION >= 0x0163)
+#if (WEBP_DECODER_ABI_VERSION >= 0x0163)
         WebPBitstreamFeatures features;
         if (WebPGetFeatures(dataBytes, dataSize, &features) != VP8_STATUS_OK)
             return setFailed();
         width = features.width;
         height = features.height;
-        m_hasAlpha = features.has_alpha;
+        m_formatFlags = features.has_alpha ? ALPHA_FLAG : 0;
 #else
         // Earlier version won't be able to display WebP files with alpha.
         if (!WebPGetInfo(dataBytes, dataSize, &width, &height))
             return setFailed();
-        m_hasAlpha = false;
 #endif
         if (!setSize(width, height))
             return setFailed();
     }
 
     ASSERT(ImageDecoder::isSizeAvailable());
     if (onlySize)
         return true;
 
-    ASSERT(!m_frameBufferCache.isEmpty());
-    ImageFrame& buffer = m_frameBufferCache[0];
+    ASSERT(m_frameBufferCache.size() > frameIndex);
+    ImageFrame& buffer = m_frameBufferCache[frameIndex];
     ASSERT(buffer.status() != ImageFrame::FrameComplete);
 
     if (buffer.status() == ImageFrame::FrameEmpty) {
         if (!buffer.setSize(size().width(), size().height()))
             return setFailed();
         buffer.setStatus(ImageFrame::FramePartial);
-        buffer.setHasAlpha(m_hasAlpha);
+        // The buffer is transparent outside the decoded area while the image is loading.
+        // The correct value of 'hasAlpha' for the frame will be set when it is fully decoded.
+        buffer.setHasAlpha(true);
         buffer.setOriginalFrameRect(IntRect(IntPoint(), size()));
     }
 
+    const IntRect& frameRect = buffer.originalFrameRect();
     if (!m_decoder) {
-        WEBP_CSP_MODE mode = outputMode(m_hasAlpha);
+        WEBP_CSP_MODE mode = outputMode(m_formatFlags & ALPHA_FLAG);
         if (!m_premultiplyAlpha)
             mode = outputMode(false);
+#ifdef QCMS_WEBP_COLOR_CORRECTION
         if ((m_formatFlags & ICCP_FLAG) && !ignoresGammaAndColorProfile())
             mode = MODE_RGBA; // Decode to RGBA for input to libqcms.
+#endif
+        WebPInitDecBuffer(&m_decoderBuffer);
         m_decoderBuffer.colorspace = mode;
         m_decoderBuffer.u.RGBA.stride = size().width() * sizeof(ImageFrame::PixelData);
-        m_decoderBuffer.u.RGBA.size = m_decoderBuffer.u.RGBA.stride * size().height();
+        m_decoderBuffer.u.RGBA.size = m_decoderBuffer.u.RGBA.stride * frameRect.height();
         m_decoderBuffer.is_external_memory = 1;
         m_decoder = WebPINewDecoder(&m_decoderBuffer);
         if (!m_decoder)
             return setFailed();
     }
 
-    m_decoderBuffer.u.RGBA.rgba = reinterpret_cast<uint8_t*>(buffer.getAddr(0, 0));
+    m_decoderBuffer.u.RGBA.rgba = reinterpret_cast<uint8_t*>(buffer.getAddr(frameRect.x(), frameRect.y()));
 
     switch (WebPIUpdate(m_decoder, dataBytes, dataSize)) {
     case VP8_STATUS_OK:
-        if ((m_formatFlags & ICCP_FLAG) && !ignoresGammaAndColorProfile()) 
-            applyColorProfile(dataBytes, dataSize, buffer);
+        applyPostProcessing(frameIndex);
+        buffer.setHasAlpha((m_formatFlags & ALPHA_FLAG) || m_frameBackgroundHasAlpha);
         buffer.setStatus(ImageFrame::FrameComplete);
-        clear();
+        clearDecoder();
         return true;
     case VP8_STATUS_SUSPENDED:
-        if ((m_formatFlags & ICCP_FLAG) && !ignoresGammaAndColorProfile()) 
-            applyColorProfile(dataBytes, dataSize, buffer);
+        applyPostProcessing(frameIndex);
         return false;
     default:
-        clear();                         
+        clear();
         return setFailed();
     }
 }
 
 } // namespace WebCore