common/lwlibav_video.c: fix incorrect seek for some mkv files

The issue is that somehow newer mkvmerge generates way less cue points
than ffmpeg and older versions of mkvmerge. This makes av_seek_frame
much less reliable. For the specific test case mentioned in #14, rap
for frame 32 is frame 24, however, the best av_seek_frame can do is
just seek to the very beginning of the video track.

lsmas does not depend on fully reliable av_seek_frame, however, in this
particular cases, because the mkv is using PTS as seeking method, lsmas
will not attempt to call correct_current_frame_number to correct for the
mis-seek, which means it will assume the current position (frame 0) as
actually the rap (frame 24), then it follows that it will return frame
8 as frame 32.

To fix this, we must call correct_current_frame_number even for SEEK_PTS_BASED
formats. This fixes much of the issues, except for the first few frames.

Then there is another issue lurking in correct_current_frame_number:
when it finds that the current packet has a dts of AV_NOPTS_VALUE, it
will always abort. Unfortunately, as frame 0 typically has such a dts,
it will not correct the current frame number when the actual current
is frame 0. We fix this by disregarding pkt->dts == AV_NOPTS_VALUE
when the seek mode is not just SEEK_DTS_BASED (note that mkv has a
seek mode of SEEK_DTS_BASED | SEEK_POS_CORRECTION, which means that
we can reliably use the pkt->pos field to determine the current frame.)

Fixes #14.

Signed-off-by: akarin <[email protected]>

common/lwlibav_video.c

@@ -432,7 +432,14 @@ static uint32_t correct_current_frame_number
     order_converter_t  *oc   = vdhp->order_converter;
     video_frame_info_t *info = vdhp->frame_list;
     uint32_t p = oc ? oc[i].decoding_to_presentation : i;
-    if( pkt->dts == AV_NOPTS_VALUE || MATCH_DTS( p ) || MATCH_POS( p ) )
+    // It is possible that the first frame has dts == AV_NOPTS_VALUE and we happen to seek to frame 0,
+    // even when rap is strictly > 0. This happens with recent mkvmerge versions where the #cuepoints
+    // are dramatically cut down and thus av_seek_frame is much less accurate and might give us a (much)
+    // earlier frame.
+    // Therefore, we should not just give up when dts == AV_NOPTS_VALUE, and we have to also check some
+    // others fields, especially when lw_seek_flags is more than just SEEK_DTS_BASED.
+    if( (pkt->dts == AV_NOPTS_VALUE && ((vdhp->lw_seek_flags & ~SEEK_DTS_BASED) == 0))
+            || MATCH_DTS( p ) || MATCH_POS( p ) )
         return i;
     if( pkt->dts > info[p].dts )
     {
@@ -487,7 +494,7 @@ static int decode_video_picture
         return ret;
     /* Correct the current picture number in order to match DTS since libavformat might have sought wrong position. */
     uint32_t correction_distance = 0;
-    if( picture_number == rap_number && (vdhp->lw_seek_flags & SEEK_DTS_BASED) )
+    if( picture_number == rap_number && (vdhp->lw_seek_flags & (SEEK_DTS_BASED | SEEK_PTS_BASED)) )
     {
         picture_number = correct_current_frame_number( vdhp, pkt, picture_number, goal );
         if( picture_number == 0