H264: Improve FFmpeg decoder performance by using I420BufferPool.

webrtc/common_video/include/video_frame_buffer.h

Index: webrtc/common_video/include/video_frame_buffer.h
diff --git a/webrtc/common_video/include/video_frame_buffer.h b/webrtc/common_video/include/video_frame_buffer.h
index 710d2862f0b9a37c88bf72898c24c80d60d52648..4be8f4322db5c9b5eecc66fb6384eeabcbe2f0a6 100644
--- a/webrtc/common_video/include/video_frame_buffer.h
+++ b/webrtc/common_video/include/video_frame_buffer.h
@@ -73,6 +73,8 @@ class I420Buffer : public VideoFrameBuffer {
   // Non-const data access is only allowed if HasOneRef() is true to protect
   // against unexpected overwrites.
   uint8_t* MutableData(PlaneType type) override;
+  // The size of the |data| buffer. The YUV planes are continuous in the buffer.
+  int DataSize() const;

[stefan-webrtc, 2016/02/05 10:18:53]

I'm not sure I like this. Should we instead provide an InitializeData() method
which does the memset? We're now exposing information about how I420Buffer
stores its the buffer internally, which will be bad if we ever decide to change
it to three separate buffers.

[hbos, 2016/02/05 11:23:44]

Fair enough. If that changes, we will no longer be able to rely on this buffer
implementation for ffmpeg (or change so that ffmpeg does not rely on the buffer
being continuous), but at least that's DCHECKed.

   int stride(PlaneType type) const override;
   void* native_handle() const override;
   rtc::scoped_refptr<VideoFrameBuffer> NativeToI420Buffer() override;