The size of all forward buffered packets is used to control maximum
buffering.
Until now, this size was incrementally adjusted, but had to be
recomputed on seeks within the cache. Doing this was actually pretty
expensive. It iterates over a linked list of separate memory allocations
(which are probably spread all over the heap due to the allocation
behavior), and the demux_packet_estimate_total_size() call touches a lot
of further memory locations. I guess this affects the cache rather
negatively. In an unscientific test, the recompute_buffers() function
(which contained this loop) was responsible for roughly half of the time
seeking took.
Replace this with a way that computes the buffered size between 2
packets in constant times. The demux_packet.cum_pos field contains the
summed sizes of all previous packets, so subtracting cum_pos between two
packets yields the size of all packets in between. We can do this
because we never remove packets from the middle of the queue. We only
add packets to the end, or remove packets at the beginning.
The tail_cum_pos field is needed because we don't store the end position
of a packet, so the last packet's position would be unknown. We could
recompute the "estimated" packet size, or store the estimated size in
the packet struct, but I just didn't like this.
This also removes the cached fw_bytes fields. It's slightly nicer to
just recompute them when needed. Maintaining them incrementally was
annoying. total_size stays though, since recomputing it isn't that cheap
(would need to loop over all ranges every time).
I'm always using uint64_t for sizes. This is certainly needed (a stream
could easily burn through more than 4GB of data, even if much less of
that is cached). The actual cached amount should always fit into size_t,
so it's casted to size_t for printfs (yes, I hate the way you specify
stdint.h types in printfs, the less I have to use that crap, the
better).
See manpage additions. This is a huge hack. You can bet there are shit
tons of bugs. It's literally forcing square pegs into round holes.
Hopefully, the manpage wall of text makes it clear enough that the whole
shit can easily crash and burn. (Although it shouldn't literally crash.
That would be a bug. It possibly _could_ start a fire by entering some
sort of endless loop, not a literal one, just something where it tries
to do work without making progress.)
(Some obvious bugs I simply ignored for this initial version, but
there's a number of potential bugs I can't even imagine. Normal playback
should remain completely unaffected, though.)
How this works is also described in the manpage. Basically, we demux in
reverse, then we decode in reverse, then we render in reverse.
The decoding part is the simplest: just reorder the decoder output. This
weirdly integrates with the timeline/ordered chapter code, which also
has special requirements on feeding the packets to the decoder in a
non-straightforward way (it doesn't conflict, although a bugmessmass
breaks correct slicing of segments, so EDL/ordered chapter playback is
broken in backward direction).
Backward demuxing is pretty involved. In theory, it could be much
easier: simply iterating the usual demuxer output backward. But this
just doesn't fit into our code, so there's a cthulhu nightmare of shit.
To be specific, each stream (audio, video) is reversed separately. At
least this means we can do backward playback within cached content (for
example, you could play backwards in a live stream; on that note, it
disables prefetching, which would lead to losing new live video, but
this could be avoided).
The fuckmess also meant that I didn't bother trying to support
subtitles. Subtitles are a problem because they're "sparse" streams.
They need to be "passively" demuxed: you don't try to read a subtitle
packet, you demux audio and video, and then look whether there was a
subtitle packet. This means to get subtitles for a time range, you need
to know that you demuxed video and audio over this range, which becomes
pretty messy when you demux audio and video backwards separately.
Backward display is the most weird (and potentially buggy) part. To
avoid that we need to touch a LOT of timing code, we negate all
timestamps. The basic idea is that due to the navigation, all
comparisons and subtractions of timestamps keep working, and you don't
need to touch every single of them to "reverse" them.
E.g.:
bool before = pts_a < pts_b;
would need to be:
bool before = forward
? pts_a < pts_b
: pts_a > pts_b;
or:
bool before = pts_a * dir < pts_b * dir;
or if you, as it's implemented now, just do this after decoding:
pts_a *= dir;
pts_b *= dir;
and then in the normal timing/renderer code:
bool before = pts_a < pts_b;
Consequently, we don't need many changes in the latter code. But some
assumptions inhererently true for forward playback may have been broken
anyway. What is mainly needed is fixing places where values are passed
between positive and negative "domains". For example, seeking and
timestamp user display always uses positive timestamps. The main mess is
that it's not obvious which domain a given variable should or does use.
Well, in my tests with a single file, it suddenly started to work when I
did this. I'm honestly surprised that it did, and that I didn't have to
change a single line in the timing code past decoder (just something
minor to make external/cached text subtitles display). I committed it
immediately while avoiding thinking about it. But there really likely
are subtle problems of all sorts.
As far as I'm aware, gstreamer also supports backward playback. When I
looked at this years ago, I couldn't find a way to actually try this,
and I didn't revisit it now. Back then I also read talk slides from the
person who implemented it, and I'm not sure if and which ideas I might
have taken from it. It's possible that the timestamp reversal is
inspired by it, but I didn't check. (I think it claimed that it could
avoid large changes by changing a sign?)
VapourSynth has some sort of reverse function, which provides a backward
view on a video. The function itself is trivial to implement, as
VapourSynth aims to provide random access to video by frame numbers (so
you just request decreasing frame numbers). From what I remember, it
wasn't exactly fluid, but it worked. It's implemented by creating an
index, and seeking to the target on demand, and a bunch of caching. mpv
could use it, but it would either require using VapourSynth as demuxer
and decoder for everything, or replacing the current file every time
something is supposed to be played backwards.
FFmpeg's libavfilter has reversal filters for audio and video. These
require buffering the entire media data of the file, and don't really
fit into mpv's architecture. It could be used by playing a libavfilter
graph that also demuxes, but that's like VapourSynth but worse.
This makes ICY title changes show up at approximately the correct time,
even if the demuxer buffer is huge. (It'll still be wrong if the stream
byte cache contains a meaningful amount of data.)
It should have the same effect for mid-stream metadata changes in e.g.
OGG (untested).
This is still somewhat fishy, but in parts due to ICY being fishy, and
FFmpeg's metadata change API being somewhat fishy. For example, what
happens if you seek? With FFmpeg AVFMT_EVENT_FLAG_METADATA_UPDATED and
AVSTREAM_EVENT_FLAG_METADATA_UPDATED we hope that FFmpeg will correctly
restore the correct metadata when the first packet is returned.
If you seke with ICY, we're out of luck, and some audio will be
associated with the wrong tag until we get a new title through ICY
metadata update at an essentially random point (it's mostly inherent to
ICY). Then the tags will switch back and forth, and this behavior will
stick with the data stored in the demuxer cache. Fortunately, this can
happen only if the HTTP stream is actually seekable, which it usually is
not for ICY things. Seeking doesn't even make sense with ICY, since you
can't know the exact metadata location. Basically ICY metsdata sucks.
Some complexity is due to a microoptimization: I didn't want additional
atomic accesses for each packet if no timed metadata is used. (It
probably doesn't matter at all.)
This directly reads individual mkv sub-packets (block laces) into a
dedicated AVBufferRefs, which can be directly used for creating packets
without a additional copy of the packet data. This also means we switch
parsing of block header fields and lacing metadata to read directly from
the stream, instead of a memory buffer.
This could have been much easier if libavcodec didn't require padding
the packet data with zero bytes. We could just have each packet
reference a slice of the block data. But as it is, the only way to get
padding without a copy is to read the laces into individually allocated
(and padded) memory block, which required a larger rewrite.
This probably makes recovering from broken mkv files slightly worse if
the transport is unseekable. We just read, and then check if we've
overread. But I think that shouldn't be a real concern.
No actual measureable performance change. Potential for some
regressions, as this is quite intrusive, and touches weird obscure shit
like mkv lacing. Still keeping it because I like how it removes some
redundant EBML parsing functions.
The main purpose of this commit is avoiding any hidden O(n^2) algorithms
in the code for pruning the demuxer cache, and for determining the
seekable boundaries of the cache. The old code could loop over the whole
packet queue on every packet pruned in certain corner cases.
There are two ways how to reach the goal:
1) commit a cardinal sin
2) do everything incrementally
The cardinal sin is adding an extra field to demux_packet, which caches
the determined seekable range for a keyframe range. demux_packet is a
rather general data structure and thus shouldn't have any fields that
are not inherent to its use, and are only needed as an implementation
detail of code using it. But what are you gonna do, sue me?
In the future, demux.c might have its own packet struct though. Then the
other existing cardinal sin (the "next" field, from MPlayer times) could
be removed as well.
This commit also changes slightly how the seek end is determined. There
is a note on the manpage in case anyone finds the new behavior
confusing. It's somewhat cleaner and might be needed for supporting
multiple ranges (although that's unclear).
The new_segment field was used to track the decoder data flow handler of
timeline boundaries, which are used for ordered chapters etc. (anything
that sets demuxer_desc.load_timeline). This broke seeking with the
demuxer cache enabled. The demuxer is expected to set the new_segment
field after every seek or segment boundary switch, so the cached packets
basically contained incorrect values for this, and the decoders were not
initialized correctly.
Fix this by getting rid of the flag completely. Let the decoders instead
compare the segment information by content, which is hopefully enough.
(In theory, two segments with same information could perhaps appear in
broken-ish corner cases, or in an attempt to simulate looping, and such.
I preferred the simple solution over others, such as generating unique
and stable segment IDs.)
We still add a "segmented" field to make it explicit whether segments
are used, instead of doing something silly like testing arbitrary other
segment fields for validity.
Cached seeking with timeline stuff is still slightly broken even with
this commit: the seek logic is not aware of the overlap that segments
can have, and the timestamp clamping that needs to be performed in
theory to account for the fact that a packet might contain a frame that
is always clipped off by segment handling. This can be fixed later.
All contributors have agreed. In 3a43f13fce, someone who potentially
disagreed reverted a commit by someone else (restoring the original
state). This shouldn't matter for Copyright, and all of the affected
code was rewritten/removed anyway.
It's all explained in the DOCS changes. Although this option was always
kind of obscure and pointless. Until it is removed, the only reason for
setting it would be to raise the static default limit, so change its
default to INT_MAX so that it does nothing by default.
This uses a different method to piece segments together. The old
approach basically changes to a new file (with a new start offset) any
time a segment ends. This meant waiting for audio/video end on segment
end, and then changing to the new segment all at once. It had a very
weird impact on the playback core, and some things (like truly gapless
segment transitions, or frame backstepping) just didn't work.
The new approach adds the demux_timeline pseudo-demuxer, which presents
an uniform packet stream from the many segments. This is pretty similar
to how ordered chapters are implemented everywhere else. It also reminds
of the FFmpeg concat pseudo-demuxer.
The "pure" version of this approach doesn't work though. Segments can
actually have different codec configurations (different extradata), and
subtitles are most likely broken too. (Subtitles have multiple corner
cases which break the pure stream-concatenation approach completely.)
To counter this, we do two things:
- Reinit the decoder with each segment. We go as far as allowing
concatenating files with completely different codecs for the sake
of EDL (which also uses the timeline infrastructure). A "lighter"
approach would try to make use of decoder mechanism to update e.g.
the extradata, but that seems fragile.
- Clip decoded data to segment boundaries. This is equivalent to
normal playback core mechanisms like hr-seek, but now the playback
core doesn't need to care about these things.
These two mechanisms are equivalent to what happened in the old
implementation, except they don't happen in the playback core anymore.
In other words, the playback core is completely relieved from timeline
implementation details. (Which honestly is exactly what I'm trying to
do here. I don't think ordered chapter behavior deserves improvement,
even if it's bad - but I want to get it out from the playback core.)
There is code duplication between audio and video decoder common code.
This is awful and could be shareable - but this will happen later.
Note that the audio path has some code to clip audio frames for the
purpose of codec preroll/gapless handling, but it's not shared as
sharing it would cause more pain than it would help.
Makes it somewhat more uniform, and breaks up the awfully deep nesting.
This implicitly changes multiple small details, rather than only moving
code around. In particular, this computes the packet fields first and
parses them afterwards, which is needed for the next commit.
This mechanism was introduced for Opus, and allows correct skipping of
"preroll" data, as well as discarding trailing audio if the file's
length isn't a multiple of the audio frame size.
Not sure how to handle seeking. I don't understand the purpose of the
SeekPreRoll element.
This was tested with correctness_trimming_nobeeps.opus, remuxed to mka
with mkvmerge v7.2.0. It seems to be correct, although the reported file
duration is incorrect (maybe a mkvmerge issue).
This is a simplification, because it lets us use the AVPacket
functions, instead of handling the details manually.
It also allows the libavcodec rawvideo decoder to use reference
counting, so it doesn't have to memcpy() the full image data. The change
in av_common.c enables this.
This change is somewhat risky, because we rely on the following AVPacket
implementation details and assumptions:
- av_packet_ref() doesn't access the input padding, and just copies the
data. By the API, AVPacket is always padded, and we violate this. The
lavc implementation would have to go out of its way to make this a
real problem, though.
- We hope that the way we make the AVPacket refcountable in av_common.c
is actually supported API-usage. It's hard to tell whether it is.
Of course we still use our own "old" demux_packet struct, just so that
libav* API usage is somewhat isolated.
Having the DTS directly can be useful for restoring PTS values.
The avi file format doesn't actually store PTS values, just DTS. An
older hack explicitly exported the DTS as PTS (ignoring the [I assume]
genpts generated non-sense PTS), which is not necessary anymore due to
this change.
wm4
Marcin Kurczewski