Fix Android black frames from MSE config changes.

media/filters/gpu_video_decoder.cc

Index: media/filters/gpu_video_decoder.cc
diff --git a/media/filters/gpu_video_decoder.cc b/media/filters/gpu_video_decoder.cc
index 260bfaabb3e6777e42f6d3a823589101190ae587..e3a3a14245e9548c61d84be33104d3a57f015289 100644
--- a/media/filters/gpu_video_decoder.cc
+++ b/media/filters/gpu_video_decoder.cc
@@ -476,7 +476,18 @@ void GpuVideoDecoder::PictureReady(const media::Picture& picture) {
     NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE);
     return;
   }
-  const PictureBuffer& pb = it->second;
+  PictureBuffer& pb = it->second;
+  if (picture.size_changed()) {
+    // Update the PictureBuffer size to match that of the Picture. Some VDA's
+    // (e.g. Android) will handle resolution changes internally without
+    // requesting new PictureBuffers. Sending a Picture of unmatched size is
+    // the signal that we should update the size of our PictureBuffer.
+    DCHECK(pb.size() != picture.visible_rect().size());
+    DVLOG(3) << __FUNCTION__ << " Updating size of PictureBuffer[" << pb.id()
+             << "] from:" << pb.size().ToString()
+             << " to:" << picture.visible_rect().size().ToString();
+    pb.set_size(picture.visible_rect().size());

[jbauman, 2016/03/07 23:53:50]

Hmm, should it be using the coded size instead? I'm not sure if that actually
matters, though.

[chcunningham, 2016/03/08 03:19:36]

I had the same concern, but Frank convinced me this should be ok. Coded size is
definitely distinct from visible size, but for AVDA this is always the same. See
Frank's reply earlier: "i agree that the flag is very AVDA-specific..."

+  }
 
   // Update frame's timestamp.
   base::TimeDelta timestamp;