Discontinuity detection/correction is left in the main thread, as it is
entangled with InputStream.next_dts and related variables, which may be
set by decoding code.
Fixes races e.g. in fate-ffmpeg-streamloop after
aae9de0cb2.
This will allow to move normal offset handling to demuxer thread, since
discontinuities currently have to be processed in the main thread, as
the code uses some decoder-produced values.
InputFile.ts_offset can change during transcoding, due to discontinuity
correction. This should not affect the streamcopy starting timestamp.
Cf. bf2590aed3
-stream_loop is currently handled by destroying the demuxer thread,
seeking, then recreating it anew. This is very messy and conflicts with
the future goal of moving each major ffmpeg component into its own
thread.
Handle -stream_loop directly in the demuxer thread. Looping requires the
demuxer to know the duration of the file, which takes into account the
duration of the last decoded audio frame (if any). Use a thread message
queue to communicate this information from the main thread to the
demuxer thread.
This avoids a potential race with the demuxer adding new streams. It is
also more efficient, since we no longer do inter-thread transfers of
packets that will be just discarded.
This undocumented feature runtime-enables dumping input packets. I can
think of no reasonable real-world use case that cannot also be
accomplished in a different way. Keeping this functionality would
interfere with the following commit moving it to the input thread (then
setting the variable would require locking or atomics, which would be
unnecessarily complicated for a feature that probably nobody uses).
There are currently three possible modes for an output stream:
1) The stream is produced by encoding output from some filtergraph. This
is true when ost->enc_ctx != NULL, or equivalently when
ost->encoding_needed != 0.
2) The stream is produced by copying some input stream's packets. This
is true when ost->enc_ctx == NULL && ost->source_index >= 0.
3) The stream is produced by attaching some file directly. This is true
when ost->enc_ctx == NULL && ost->source_index < 0.
OutputStream.stream_copy is currently used to identify case 2), and
sometimes to confusingly (or even incorrectly) identify case 1). Remove
it, replacing its usage with checking enc_ctx/source_index values.
The streamcopy initialization code briefly needs an AVCodecContext to
apply AVOptions to. Allocate a temporary codec context, do not use the
encoding one.
It retrieves the muxer's internal timestamp with under-defined
semantics. Continuing to use this value would also require
synchronization once the muxer is moved to a separate thread.
Replace the value with last_mux_dts.
This field means different things when the video is encoded (number of
frames emitted to the encoding sync queue/encoder by the video sync
code) or copied (number of packets sent to the muxer sync queue).
Print the value of packets_written instead, which means the same thing
in both cases. It is also more accurate, since packets may be dropped by
the sync queue or bitstream filters.
Same issues apply to it as to -shortest.
Changes the results of the following tests:
- matroska-flac-extradata-update
The test reencodes two input FLAC streams into three output FLAC
streams. The last output stream is limited to 8 frames. The current
code results in the first two output streams having 12 frames, after
this commit all three streams have 8 frames and are the same length.
This new result is better, since it is predictable.
- mkv-1242
The test streamcopies one video and one audio stream, video is limited
to 11 frames. The new result shortens the audio stream so that it is
not longer than the video.
The -shortest option (which finishes the output file at the time the
shortest stream ends) is currently implemented by faking the -t option
when an output stream ends. This approach is fragile, since it depends
on the frames/packets being processed in a specific order. E.g. there
are currently some situations in which the output file length will
depend unpredictably on unrelated factors like encoder delay. More
importantly, the present work aiming at splitting various ffmpeg
components into different threads will make this approach completely
unworkable, since the frames/packets will arrive in effectively random
order.
This commit introduces a "sync queue", which is essentially a collection
of FIFOs, one per stream. Frames/packets are submitted to these FIFOs
and are then released for further processing (encoding or muxing) when
it is ensured that the frame in question will not cause its stream to
get ahead of the other streams (the logic is similar to libavformat's
interleaving queue).
These sync queues are then used for encoding and/or muxing when the
-shortest option is specified.
A new option – -shortest_buf_duration – controls the maximum number of
queued packets, to avoid runaway memory usage.
This commit changes the results of the following tests:
- copy-shortest[12]: the last audio frame is now gone. This is
correct, since it actually outlasts the last video frame.
- shortest-sub: the video packets following the last subtitle packet are
now gone. This is also correct.
The following commits will add a new buffering stage after bitstream
filters, which should not be taken into account for choosing next
output.
OutputStream.last_mux_dts is also used by the muxing code to make up
missing DTS values - that field is now moved to the muxer-private
MuxStream object.
The current placement of this free is historical - it used to be
followed by avcodec_close(), since removed.
The proper place for freeing the stats is currently right before the
encoder context itself is freed.
Anton Khirnov
Andreas Rheinhardt