From 98342fb53eae4d32440fc69c279f2ca0fef785b5 Mon Sep 17 00:00:00 2001
From: Carl Hetherington <cth@carlh.net>
Date: Sun, 10 Nov 2019 22:59:39 +0100
Subject: [PATCH] Don't trust video timestamps from FFmpegDecoder.

---
 src/lib/video_decoder.cc | 44 ++++++++++++++++++++++++++++++++++++----
 src/lib/video_decoder.h  |  2 +-
 2 files changed, 41 insertions(+), 5 deletions(-)

diff --git a/src/lib/video_decoder.cc b/src/lib/video_decoder.cc
index 7a3a3e19b..19a994197 100644
--- a/src/lib/video_decoder.cc
+++ b/src/lib/video_decoder.cc
@@ -54,12 +54,44 @@ VideoDecoder::VideoDecoder (Decoder* parent, shared_ptr<const Content> c)
  *  and so on.
  */
 void
-VideoDecoder::emit (shared_ptr<const Film> film, shared_ptr<const ImageProxy> image, Frame frame)
+VideoDecoder::emit (shared_ptr<const Film> film, shared_ptr<const ImageProxy> image, Frame decoder_frame)
 {
 	if (ignore ()) {
 		return;
 	}
 
+	double const afr = _content->active_video_frame_rate(film);
+
+	Frame frame;
+	if (!_position) {
+		/* This is the first data we have received since initialisation or seek.  Set
+		   the position based on the frame that was given.  After this first time
+		   we just cound frames, since (as with audio) it seems that ContentTimes
+		   are unreliable from FFmpegDecoder.  They are much better than audio times
+		   but still we get the occasional one which is duplicated.  In this case
+		   ffmpeg seems to carry on regardless, processing the video frame as normal.
+		   If we drop the frame with the duplicated timestamp we obviously lose sync.
+		*/
+		_position = ContentTime::from_frames (decoder_frame, afr);
+		if (_content->video->frame_type() == VIDEO_FRAME_TYPE_3D_ALTERNATE) {
+			frame = decoder_frame / 2;
+			_last_emitted_eyes = EYES_RIGHT;
+		} else {
+			frame = decoder_frame;
+		}
+	} else {
+		if (_content->video->frame_type() == VIDEO_FRAME_TYPE_3D_ALTERNATE) {
+			DCPOMATIC_ASSERT (_last_emitted_eyes);
+			if (_last_emitted_eyes.get() == EYES_RIGHT) {
+				frame = _position->frames_round(afr) + 1;
+			} else {
+				frame = _position->frames_round(afr);
+			}
+		} else {
+			frame = _position->frames_round(afr) + 1;
+		}
+	}
+
 	switch (_content->video->frame_type ()) {
 	case VIDEO_FRAME_TYPE_2D:
 		Data (ContentVideo (image, frame, EYES_BOTH, PART_WHOLE));
@@ -91,9 +123,13 @@ VideoDecoder::emit (shared_ptr<const Film> film, shared_ptr<const ImageProxy> im
 		break;
 	}
 	case VIDEO_FRAME_TYPE_3D_ALTERNATE:
-		Data (ContentVideo (image, frame / 2, (frame % 2) ? EYES_RIGHT : EYES_LEFT, PART_WHOLE));
-		frame /= 2;
+	{
+		DCPOMATIC_ASSERT (_last_emitted_eyes);
+		Eyes const eyes = _last_emitted_eyes.get() == EYES_LEFT ? EYES_RIGHT : EYES_LEFT;
+		Data (ContentVideo (image, frame, eyes, PART_WHOLE));
+		_last_emitted_eyes = eyes;
 		break;
+	}
 	case VIDEO_FRAME_TYPE_3D_LEFT_RIGHT:
 		Data (ContentVideo (image, frame, EYES_LEFT, PART_LEFT_HALF));
 		Data (ContentVideo (image, frame, EYES_RIGHT, PART_RIGHT_HALF));
@@ -112,7 +148,7 @@ VideoDecoder::emit (shared_ptr<const Film> film, shared_ptr<const ImageProxy> im
 		DCPOMATIC_ASSERT (false);
 	}
 
-	_position = ContentTime::from_frames (frame, _content->active_video_frame_rate(film));
+	_position = ContentTime::from_frames (frame, afr);
 }
 
 void
diff --git a/src/lib/video_decoder.h b/src/lib/video_decoder.h
index 34ce8f288..954b8f365 100644
--- a/src/lib/video_decoder.h
+++ b/src/lib/video_decoder.h
@@ -63,7 +63,7 @@ public:
 
 private:
 	boost::shared_ptr<const Content> _content;
-	/** Frame of last thing to be emitted */
+	/** Frame of last thing to be emitted; only used for 3D */
 	boost::optional<Frame> _last_emitted_frame;
 	boost::optional<Eyes> _last_emitted_eyes;
 	boost::optional<dcpomatic::ContentTime> _position;
-- 
2.30.2