Essentially, this lets video.c decide whether to consider a video track
cover art, instead of having the decoder wrapper use the lower level
sh_stream flag.
Some pain because of the dumb threading shit. Moving the code further
down to make some of it part of the lock should not change behavior,
although it could (framedrop nonsense).
This commit should not change actual behavior, and is only preparation
for the following commit.
It's relevant in some obscure corner cases (EDL file that has a segment
without audio). Didn't test what's actually going on (is ad_lavc.c
behaving wrong? is libavcodec behaving wrong or in an unexpected way? is
lavc_process wrong?) and just patched it over with some bullshit, so the
fix might be too complicated, and could be reworked at some later point.
This sure is a real data flow fuckmess.
After calling the main filter's destroy callback, all child filters are
destroyed. But one of them still tried to access the cache_lock mutex
(which is destroyed in said destroy callback). This actually caused a
crash on Android with _FORTIFY_SOURCE.
Fix this by destroying the child filters first.
As an unfortunate disaster, min/max values use the type double, which
causes tons of issues with int64_t types. Anyway, OPT_BYTE_SIZE is often
used as maximum for size_t quantities, which can have a size different
from (u)int64_t.
OPT_BYTE_SIZE still uses in64_t, because in theory, you could use it for
file sizes. (demux.c would for example be capable of caching more than
2GB on 32 bit platforms if a file cache is used. Though for some reason
the accounting code still uses size_t, so that use case is broken. But
still insist that it _could_ be used this way.)
There were various inconsistent attempts to set m_option.max to a value
such that the size_t/int64_t upper limit is not exceeded. Due to the
double max field, this didn't really work correctly. Try to fix this
with the M_MAX_MEM_BYTES constant. It's a good approximation, because on
32 bit it should allow 2GB (untested, also would probably exhaust
address space in practice but whatever), and something "high enough" in
64 bit.
For some reason, clang 11 still warns. But I think this might be a clang
bug, or I'm crazy. The result is correct anyway.
Change all OPT_* macros such that they don't define the entire m_option
initializer, and instead expand only to a part of it, which sets certain
fields. This requires changing almost every option declaration, because
they all use these macros. A declaration now always starts with
{"name", ...
followed by designated initializers only (possibly wrapped in macros).
The OPT_* macros now initialize the .offset and .type fields only,
sometimes also .priv and others.
I think this change makes the option macros less tricky. The old code
had to stuff everything into macro arguments (and attempted to allow
setting arbitrary fields by letting the user pass designated
initializers in the vararg parts). Some of this was made messy due to
C99 and C11 not allowing 0-sized varargs with ',' removal. It's also
possible that this change is pointless, other than cosmetic preferences.
Not too happy about some things. For example, the OPT_CHOICE()
indentation I applied looks a bit ugly.
Much of this change was done with regex search&replace, but some places
required manual editing. In particular, code in "obscure" areas (which I
didn't include in compilation) might be broken now.
In wayland_common.c the author of some option declarations confused the
flags parameter with the default value (though the default value was
also properly set below). I fixed this with this change.
Before this commit, option declarations used M_OPT_MIN/M_OPT_MAX (and
some other identifiers based on these) to signal whether an option had
min/max values. Remove these flags, and make it use a range implicitly
on the condition if min<max is true.
This requires care in all cases when only M_OPT_MIN or M_OPT_MAX were
set (instead of both). Generally, the commit replaces all these
instances with using DBL_MAX/DBL_MIN for the "unset" part of the range.
This also happens to fix some cases where you could pass over-large
values to integer options, which were silently truncated, but now cause
an error.
This commit has some higher potential for regressions.
The mp_filter_run() invocation blocks as long as the demuxer provides
packets and the queue can be filled. That means it may block quite a
long time of the decoder queue size is large (since we use libavcodec in
a blocking manner; it regrettably does not have an async. API).
This made the main thread freeze in certain situations, because it has
to wait on the decoder thread.
Other than I suspected (I wrote that code, but that doesn't mean I know
how the hell it works), this did not freeze during seeking: seek resets
flushed the queue, which also prevented the decoder thread from adding
more frames to it, thus stopping decoding and responding to the main
thread in time. But it does fix the issue that exiting the player waited
for the decoder to finish filling the queue when stopping playback.
(This happened because it called mp_decoder_wrapper_set_play_dir()
before any resets. Related to the somewhat messy way play_dir is
generally set. But it affects all "synchronous" decoder wrapper API
calls.)
This uses pretty weird mechanisms in filter.h and dispatch.h. The
resulting durr hurr interactions are probably hard to follow, and this
whole thing is a sin. On the other hand, this is a _very_ user visible
issue, and I'm happy that it can be fixed in such an unintrusive way.
p->log has a prefix set that gives some context and distinguishes audio
and video decoders. The "public" wrapper filter didn't use it, which is
a regression since commit a3823ce0e0.
This was forgotten.
Hardware decoding typically breaks immediately, because many hw decoding
APIs require allocating all surfaces in advance (and/or libavcodec was
not made flexible enough to add new surfaces later). If the queue is
large enough, it will run out of surfaces, fail decoding, and fall back
to software decoding. We consider this the user's fault.
--hwdec-extra-frames can be used to avoid this, if you have enough GPU
memory, and the needed number of frames is lower than the arbitrary
mpv-set maximum limit of that option.
Instead of having f_decoder_wrapper create its own copy of the entire
mpv option tree, create a struct local to that file and move all used
options to there.
movie_aspect is used by the "video-aspect" deprecated property code. I
think it's probably better not to remove the property yet, but
fortunately it's easy to work around without needing special handling
for this option or so.
correct_pts is used to prevent use of hr-seek in playloop.c. Ignore
that, if you use --no-correct-pts you're asking for trouble anyway. This
is the only behavior change.
See manpage additions. This has been a topic in MPlayer/mplayer2/mpv
since forever. But since libavcodec multi-threaded decoding was added,
I've always considered this pointless. libavcodec requires you to
"preload" it with packets, and then you can pretty much avoid blocking
on it, if decoding is fast enough.
But in some cases, a decoupled decoder thread _might_ help. Users have
for example come up with cases where decoding video in a separate
process and piping it as raw video to mpv helped. (Or my memory is
false, and it was about vapoursynth filtering, who knows.) So let's just
see whether this helps with anything.
Note that this would have been _much_ easier if libavcodec had an
asynchronous (or rather, non-blocking) API. It could probably have
easily gained that with a small change to its multi-threading code and a
small extension to its API, but I guess not.
Unfortunately, this uglifies f_decoder_wrapper quite a lot. Part of this
is due to annoying corner cases like legacy frame dropping and hardware
decoder state. These could probably be prettified later on.
There is also a change in playloop.c: this is because there is a need to
coordinate playback resets between demuxer thread, decoder thread, and
playback logic. I think this SEEK_BLOCK idea worked out reasonably well.
There are still a number of problems. For example, if the demuxer cache
is full, the decoder thread will simply block hard until the output
queue is full, which interferes with seeking. Could also be improved
later. Hardware decoding will probably die in a fire, because it will
run out of surfaces quickly. We could reduce the queue to size 1...
maybe later. We could update the queue options at runtime easily, but
currently I'm not going to bother.
I could only have put the lavc wrapper itself on a separate thread. But
there is some annoying interaction with EDL and backward playback shit,
and also you would have had to loop demuxer packets through the
playloop, so this sounded less annoying.
The food my mother made for us today was delicious.
Because audio uses the same code, also for audio (even if completely
pointless).
Fixes: #6926
I may (optionally) move decoding to a separate thread in a future
change. It's a bit attractive to move the entire decoder wrapper to
there, so if the demuxer has a new packet, it doesn't have to wake up
the main thread, and can directly wake up the decoder. (Although that's
bullshit, since there's a queue in between, and libavcodec's
multi-threaded decoding plays cross-threads ping pong with packets
anyway. On the other hand, the main thread would still have to shuffle
the packets around, so whatever, just seems like better design.)
As preparation, there shouldn't be any mutable state exposed by the
wrapper. But there's still a large number of corner-caseish crap, so
just use setters/getters for them. This recorder thing will inherently
not work, so it'll have to be disabled if threads are used.
This is a bit painful, but probably still the right thing. Like
speculatively pulling teeth.
This is for easier use with the "delay_open" feature added in the
previous commit. The "null" codec is reported if the codec is unknown
(because the stream was not opened yet at time the tracks were added).
The rest of the timeline mechanism will set the correct codec at
runtime. But this means every time a delay-loaded track is selected, it
wants to initialize a decoder for the "null" codec.
Accept a "null" decoder. But since FFmpeg has no such codec, and out of
my own laziness, just let it fall back to "common" codecs that need no
other initialization data.
Just an implementation detail that can be cleaned up now. Internally,
m_config maintains a tree of m_sub_options structs, except for the root
it was not defined explicitly. GLOBAL_CONFIG was a hack to get access to
it anyway. Define it explicitly instead.
This wasn't done, probably regression from one of the last dozen of
times this special code path was touched. This meant coverart images
ignored the user-set aspect ratio completely, and some other things.
Shovel the code around to make the data flow slightly simpler (?). At
least there's only one send_packet function now. The old code had the
problem that send_packet() could be called even if there were queued
packets; due to sending the queued packets in the receive_frame
function, this should not happen anymore (the code checking for this
case in send_packet should normally never be called).
Untested with actual full stream hw decoders (none available here); I
created a test case by making hwaccel decoding fail.
lavc_process() calls the receive/send callbacks, which mirror
libavcodec's send/receive API. The receive API in particular can return
both a status code and a frame. Normally, libavcodec is pretty explicit
that it can't return both a frame and an error. But the receive callback
does more stuff in addition (vd_lavc does hardware decoding fallbacks
etc.). The previous commit shows an instance where this happened, and
where it leaked a frame in this case.
For robustness, check whether the frame is set first, i.e. trust it over
the status code. Before this, it checked for an EOF status first.
Hopefully is of no consequence otherwise. I made this change after
testing everything (can someone implement a test suite which tests this
exhaustively).
ad_lavc and vd_lavc use the lavc_process() helper to translate the
FFmpeg push/pull API to the internal filter API (which completely
mismatch, even though I'm responsible for both, just fucking kill me).
This interface was "slightly" too tight. It returned only a bool
indicating "progress", which was not enough to handle some cases (see
following commit).
While we're at it, move all state into a struct. This is only a single
bool, but we get the chance to add more if needed.
This fixes mpv falling asleep if decoding returns an error during
draining. If decoding fails when we already sent EOF, the state machine
stopped making progress. This left mpv just sitting around and doing
nothing.
A test case can be created with: echo $RANDOM >> image.png
This makes libavformat read a proper packet plus a packet of garbage.
libavcodec will decode a frame, and then return an error code. The
lavc_process() wrapper could not deal with this, because there was no
way to differentiate between "retry" and "send new packet". Normally, it
would send a new packet, so decoding would make progress anyway. If
there was "progress", we couldn't just retry, because it'd retry
forever.
This is made worse by the fact that it tries to decode at least two
frames before starting display, meaning it will "sit around and do
nothing" before the picture is displayed.
Change it so that on error return, "receiving" a frame is retried. This
will make it return the EOF, so everything works properly.
This is a high-risk change, because all these funny bullshit exceptions
for hardware decoding are in the way, and I didn't retest them. For
example, if hardware decoding is enabled, it keeps a list of packets,
that are fed into the decoder again if hardware decoding fails, and a
software fallback is performed. Another case of horrifying accidental
complexity.
Fixes: #6618
Some state wasn't reset when decoding was started without a seek reset
before it. The code used to rely on reset_decoder() resetting this
state, but since the commit referenced below, reset_decoder() does less
than reset().
Fix this by explicitly calling reset() on initialization.
Fixes: "f_decoder_wrapper: avoid full reset on timeline switch etc."
Before this commit, there was a single process_decoded_frame() function.
It handled various aspects of dealing with a newly decoded frame. Move
some of these to a separate process_output_frame() function.
This new function is called in the order the frames are returned to the
playback core. Some correct_audio_pts() (was process_audio_frame())
becomes slightly less awkward due to this, and the timestamp smoothing
can actually work in backward playback mode now (thus moving p->pts out
of reset_decoder()).
Behavior for normal playback also changes subtly. This shouldn't matter
in sane cases, but if you mix broken files, --no-correct-pts, and
timeline stuff, differences in behavior might be visible.
Timeline clipping (EDL/ordered chapters) works now, because it's done
before "transforming" the timestamps. Audio timestamp smoothing happens
after it, which is a behavior change, but should be more correct. This
still runs crazy_video_pts_stuff() before everything else. On the pther
hand, --no-correct-pts or missing timestamp processing is done last. But
these things didn't really work with timeline before.
Slightly cleaner. We don't need to awkwardly backup "some" state on
backwards playback. Due to not resetting last_format, normal timeline
switches don't unconditionally trigger recomputing of certain image
parameters. Also probably doesn't reset framedrop parameters, although I
don't care about that part.
This could lead to nonsense when backward playback is involved. Better
reduce the possible interactions. Besides, it's better to fully reset
things on seeks in general.
The only exception is has_broken_packet_pts, which enables hr-seek if
everything looks good. It's intended to trigger at the second hr-seek or
so if the file is normal, and to disable it if the file is broken. It
tries to avoid enabling the hr-seek logic before it can know about
whether things are "good", so resetting it on seeks would obviously
never enable it. Document it as explicit exception.
Some audio codecs will discard or cut the first frames when starting
decoding. While some of that works through well-defined mechanisms (like
initial padding), it's in general very codec/decoder specific, and not
really predictable. In addition, libavcodec isn't very good with
reporting "dropped" frames (and our internal interface reflects this).
It seems our only chance to handle this is through timestamps.
In theory, it would be best to discard frames that have timestamps
before the "resume" position. But since video has reordered timestamps,
we'd need to put some effort into finding this position. The first video
packet doesn't necessarily contain this timestamp. (In theory, we could
just do this in the demuxer with some trivial additional work, and set
it on the packet's kf_seek_pts field. Although this field is supposed to
contain just this value, the field is considered demuxer-internal, and I
didn't want to make matters worse by reusing it for the interface to the
decoder. With some more effort and buffering, we could calculate this
value within the decoder, but fuck that.)
The approach chosen in this commit is setting the timestamp to NOPTS.
This will break in some obscure situations, but backward playback is a
pretty obscure feature to begin with, so I considered this a reasonable
implementation choice.
Before passing a preroll packet to the decoder, its timestamps are set
to NOPTS. Frames that are returned from the decoder and have the NOPTS
timestamp are considered preroll and are discarded. This happens only
during "preroll" mode (preroll_discard==true), so it doesn't affect
normal forward playback. It's disabled on the first packet with a
timestamp, so it can tolerate some crap even in backward playback mode.
We don't check the dts fields out of laziness (decoded audio frames
don't even have this field).
I considered using an approach using the EDL clipping infrastructure (as
mentioned in the last two paragraphs in the commit message of commit
" demux_lavf: implement bad hack for backward playback of wav"). This
didn't work, and I blamed timestamp rounding within mpv for it. But the
problem was actually due to Matroska-rounded timestamps. Since the audio
frame size isn't exactly aligned to 1ms, there will be an overlap (or
gap) in the timestamps. This overlap is much smaller than 1ms, since
it's just the sub-millisecond remainder part of the audio frame size.
This makes the timestamps discontinuous and unreliable for the purpose
we wanted to use it. We can't just smooth the timestamps in the demuxer
either.
Shitty ancient hack that wastes my time all the time.
demux.c: always return the coverart packet as soon as possible, and
don't let the backward demux state machine possibly stop it.
f_decoder_wrapper.c: mess with some shit until it somehow starts to
work. I think the old code tried to let it cleverly fall through so the
packet was processed "normally"; just make it run the "usual" code
instead.
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 was always a legacy thing. Remove it by applying an orgy of
mp_get_config_group() calls, and sometimes m_config_cache_alloc() or
mp_read_option_raw().
win32 changes untested.
Also rename stereo3d to stereo_in. The only real change is that the
vo_gpu OSD code now uses the actual stereo 3D mode, instead of the
--video-steroe-mode value. (Why does this vo_gpu code even exist?)
This means vf_vapoursynth doesn't need a hack to work around the filter
code, and libavfilter filters now actually get the frame_rate field on
input pads set.
The libavfilter doxygen says the frame_rate field is only to be set if
the frame rate is known to be constant, and uses the word "must" (which
probably means they really mean it?) - but ffmpeg.c sets the field to
mere guesses anyway, and it looks like this normally won't lead to
problems.
At least the libavcodec MediaCodec wrapper sometimes seems to break the
libavcodec API, and does the following sequence:
send_packet() -> EAGAIN
receive_frame() -> EAGAIN
send_packet() -> OK
The libavcodec API never allows returning EAGAIN from both functions, so
we discarded packets in this case. Change it to retrying decoding, for
the sake of MediaCodec. Note that it could also happen due to internal
bugs in the vd_lavc.c hw fallback code, but if there are any remaining,
they should be fixed properly instead.
Requested.
The recent changes to player/audio.c moved PTS jump detection to after
audio filtering. This was mostly done for convenience, because dataflow
between decoder and filters was made "automatic", and jump detection
would have to be done as filter. Now move it back to after decoders,
again out of convenience. The future direction is to make the dataflow
between filters and AO automatic, so this is a bit in the way. Another
reason is that speed changes tend to cause jumps - these are legitimate,
but get annoying quickly.
Use the decoder wrapper that was introduced for video. This removes all
code duplication the old audio decoder wrapper had with the video code.
(The audio wrapper was copy pasted from the video one over a decade ago,
and has been kept in sync ever since by the power of copy&paste. Since
the original copy&paste was possibly done by someone who did not answer
to the LGPL relicensing, this should also remove all doubts about
whether any of this code is left, since we now completely remove any
code that could possibly have been based on it.)
There is some complication with spdif handling, and a minor behavior
change (it will restrict the list of codecs to spdif if spdif is to be
used), but there should not be any difference in practice.
Move dec_video.c to filters/f_decoder_wrapper.c. It essentially becomes
a source filter. vd.h mostly disappears, because mp_filter takes care of
the dataflow, but its remains are in struct mp_decoder_fns.
One goal is to simplify dataflow by letting the filter framework handle
it (or more accurately, using its conventions). One result is that the
decode calls disappear from video.c, because we simply connect the
decoder wrapper and the filter chain with mp_pin_connect().
Another goal is to eventually remove the code duplication between the
audio and video paths for this. This commit prepares for this by trying
to make f_decoder_wrapper.c extensible, so it can be used for audio as
well later.
Decoder framedropping changes a bit. It doesn't seem to be worse than
before, and it's an obscure feature, so I'm content with its new state.
Some special code that was apparently meant to avoid dropping too many
frames in a row is removed, though.
I'm not sure how the source code tree should be organized. For one,
video/decode/vd_lavc.c is the only file in its directory, which is a bit
annoying.
wm4