Some files not only use rounded timestamps, but they also do it
incorrectly. They may jitter between up to 4 specific frame durations.
In this case, I found a file that mostly used 41ms and 42ms, but also
had 40ms and 43ms outliers (often but not always following each other).
This breaks the assumption of the framerate estimation code that the
frame duration can deviate up to 1ms. If it jitters around 4 possible
frame durations, the maximum deviation is 3ms. Increase it accordingly.
The change might make playback of "true VFR" video via display-sync mode
worse, but it's not like it was particularly good in the first place.
Also, the check whether to usen the container FPS needs to be stricter.
In the worst case, num_dur is 1, which doesn't really indicate any
evidence that the framerate is correct. Only if there are "enough"
frames the deviation check will become meaningful. 16 is an arbitrary
value that has been designated "enough" by myself.
Also otuput the frame duration values for --dump-stats.
Instead of letting it keep decoding by trying to find a new frame,
"plug" the frame queue by not removing it. (Or actually, by putting
it back instead of discarding it.)
Matters for seamless looping (following commits), and possibly some
other corner cases.
The added function vf_unread_output_frame() is a bit of a sin, but still
reasonable, since its implementation is trivial.
The --image-display-duration option controls how long an image is
displayed. It's also possible to display the image forever (until manual
user interaction stops playback).
With this, the core drops the old method to "drain" video (i.e. waiting
for the last frame duration on end of playback). Instead, we reuse
MPContext.time_frame. The old mechanism was disabled for non-images
anyway.
Fixes#3425.
Change the last parameter from a bool to an int, which is supposed to
take bit-flags. The at this point only flag is MPSEEK_FLAG_DELAY, which
replaces the previous bool parameter. The old false parameter becomes 0,
the old true parameter becomes MPSEEK_FLAG_DELAY.
Since the old "immediate" parameter is now essentially inverted, two
coalesced immediate and delayed seeks end up as delayed instead of
immediate. This change doesn't matter, since there are no relative
immediate seeks anyway.
In display-sync mode, the very first video frame is idiotically fully
timed, even though audio has not been synced yet at this point, and the
video frame is more like a "preview" frame. But since it's fully timed,
an underflow is detected if audio takes longer than the display time of
the frame (we send the second frame only after audio is done).
The timing code will try to compensate for the determined desync, but it
really shouldn't. So explicitly discard the timing info in this specific
case. On the other hand, if the first frame still hasn't finished
display, we can pretend everything is ok.
This is a hack - ideally, we either would send a frame without timing
info (and then send it again or so when playback starts properly), or we
would add real pause support to the VO, and pause it during syncing.
Otherwise it behaves dumb. (Although you could argue it shouldn't try to
guess whether speed changes work, but instead simply disable DS if they
don't work.)
--deinterlace=auto is the default, and has the obscure semantics that
deinterlacing is disabled, unless the user has manually inserted a
deinterlacing filter.
While in software decoding this doesn't matter, and we will happily
insert 2 yadif filters (if the user has already added one), or not
remove the yadif filter (if deinterlacing is disabled, but the user has
added the filter manually), this is different with hardware deinterlacer
filters. These support VFCTRL_SET_DEINTERLACE for toggling deinterlacing
filtering at runtime. It exists mainly for legacy reasons, and possibly
because it makes switching deinterlacing modes more efficient. It might
also gives us an entry-point for VO deinterlacing, maybe. For whatever
reasons this mechanism exists, we still support and use it.
This commit fixes that video.c always used VFCTRL_SET_DEINTERLACE to
disable deinterlacing, even if --deinterlace=auto was set. Fix this by
checking the value of the option directly.
Commit 771a8bf5 added code to avoid unnecessary vf_reconfig() calls for
unrelated reasons, but forget to consider that it has to be called at
least once if the input format changes. As a consequence it got "stuck"
due to not being able to decode more frames.
vo_frame can have more than 1 frame - the extra frames are future
references, which are sometimes useful for filtering (vo_opengl
interpolation). There's no harm in reducing the number of frames sent to
the VO requested amount of future frames, so do that.
Doesn't actually reduce the number of concurrently in use frames in
practice.
Some filters support VFCTRL_SET_DEINTERLACE. This affects most hardware
deinterlace filters. They can be inserted by the user manually, or auto-
inserted by vf.c itself as conversion filter (vf_d3d11vpp). In these
cases, we shouldn't insert or remove filters outselves, and instead
VFCTRL_SET_DEINTERLACE should be invoked to switch the mode.
This wasn't done correctly in the recently refactored code and could
have broken with --deinterlace. (The refactor only considered switching
via property in this case.) Fix it by making it a proper part of the
filter_reconfig() function, and making set_deinterlacing() (which is
called by the property handler) merely call filter_reconfig() in all
cases to do the real work.
We can even avoid rebuilding the filter chain - though only if no other
auto-filters are inserted. It probably also provides a slightly cleaner
way to implement functionality in the VO while still inserting video
filter fallbacks correctly if required.
The test scenario at hand was hardware decoding a file with d3d11 and
with deinterlacing enabled. The file switches to a non-hardware
dedocdeable format mid-stream. This failed, because it tried to call
vf_reconfig() with the old filters inserted, with was fatal due to
vf_d3d11vpp accepting only hardware input formats.
Fix this by always strictly removing all auto-inserted filters
(including the deinterlacing one), and reconfiguring only after that.
Note that this change is good for other situations too, because we
generally don't want to use a hardware deinterlacer for software
decoding by default. They're not necessarily optimal, and VAAPI VPP even
has incomprehensible deinterlacer bugs specifically with software frames
not coming from a hardware decoder.
Instead of using the "vf" command code (which changes filters at runtime
on user input), use the general filter-insertion code. The latter was
added later, and is more suitable for automatically inserted filters.
The old code failed in particular when using watch-later saving, which
stored the filter list in the resume config file. If a user changed the
hardware decoding mode via command line, the stored filter chain was out
of date and could cause failure due to not working with hardware or
software decoding mode. Storing the deinterlace filter in the filter
list was unavoidable, because it was part of the user state. (The new
code only edits the actually instantiated filters.)
The main change is with video/hwdec.h. mp_hwdec_info is made opaque (and
renamed to mp_hwdec_devices). Its accessors are mainly thread-safe (or
documented where not), which makes the whole thing saner and cleaner. In
particular, thread-safety rules become less subtle and more obvious.
The new internal API makes it easier to support multiple OpenGL interop
backends. (Although this is not done yet, and it's not clear whether it
ever will.)
This also removes all the API-specific fields from mp_hwdec_ctx and
replaces them with a "ctx" field. For d3d in particular, we drop the
mp_d3d_ctx struct completely, and pass the interfaces directly.
Remove the emulation checks from vaapi.c and vdpau.c; they are
pointless, and the checks that matter are done on the VO layer.
The d3d hardware decoders might slightly change behavior: dxva2-copy
will not use the VO device anymore if the VO supports proper interop.
This pretty much assumes that any in such cases the VO will not use any
form of exclusive mode, which makes using the VO device in copy mode
unnecessary.
This is a big refactor. Some things may be untested and could be broken.
Another crappy fix for timestamp reset issues. This time, we try to fix
files which have very weird but legitimate frame durations, such as
cdgraphics. It can have many short frames, but once in a while there are
potentially very long frames.
Fixes#3027.
This would get stuck in reconfiguring the filter chain forever, because
params was mutated ("params.rotate = 0;"). This was used as input for
vf_reconfig(), but the filter chain input must always be equivalent to
the decoder output, or filter chain reconfiguration will be triggered.
The line of code to reset the rotation is from a time when this used to
work differently.
Also remove the unnecessary try_filter() parameter.
Unfortunately I see no better solution.
The refresh seek is skipped if the amount of buffered audio is not
overly huge.
Unfortunately softvol af_volume insertion still can cause this issue,
because it's outside of the normal dynamic filter chain changing code.
Move the video refresh call to reinit_video_filters() to make it more
uniform along with the audio code.
See --lavfi-complex option.
This is still quite rough. There's no support for dynamic configuration
of any kind. There are probably corner cases where playback might freeze
or burn 100% CPU (due to dataflow problems when interaction with
libavfilter).
Future possible plans might include:
- freely switch tracks by providing some sort of default track graph
label
- automatically enabling audio visualization
- automatically mix audio or stack video when multiple tracks are
selected at once (similar to how multiple sub tracks can be selected)
Will be helpful for the coming filter support. I planned on merging
audio/video decoding, but this will have to wait a bit longer, so only
remove the duplicate status codes.
These changes don't make too much sense without context, but are
preparation for later. Then the audio_src/video_src fields will be
actually be NULL under circumstances.
Regression caused by commit 3b95dd47. Also see commit 4c25b000. We can
either use video_next_pts and add "delay", or we just use video_pts. Any
other combination breaks. The reason why the assumption that delay==0 at
this point was wrong exactly because after displaying the first video
frame (usually done before audio resync) a new frame might be "added"
immediately, resulting in a new video_next_pts and "delay", which will
still amount to video_pts.
Fixes#2770. (The reason why display-sync was blamed in this issue is
because enabling display-sync in the options forces a prefetch by 2
instead of 1 frames for seeks/playback restart, which triggers the
issue, even if display-sync is not actually enabled. In this case,
display-sync is never enabled because the frames have a unusually high
frame duration. This is also what exposed the initial desync issue.)
Even though the timing logic is correct, it tends to mess with looping
videos and such in unappreciated ways.
It also has to be admitted that most file formats seem not to properly
define the duration of the last video frame (or libavformat does not
export it in a useful way), so whether or not we should use the demuxer
reported framerate for the last frame is questionable. (Still, why would
you essentially just discard the last frame?)
The timing logic is kept, but disabled for video with "normal" FPS
values. In particular, we want to keep it for displaying images, which
implicitly set the frame duration to 1 second by reporting 1 FPS. It's
also good for slide shows with mf://.
Fixes#2745.
vo_chain_uninit() isn't supposed to care much about the decoder
(although decoders and outputs still go strictly together, so there is
not much of an actual difference now).
Also unset track.d_video correctly.
Remove a stale declaration from dec_video.h as well.
Similar to the video path. dec_audio.c now handles decoding only. It
also looks very similar to dec_video.c, and actually contains some of
the rewritten code from it. (A further goal might be unifying the
decoders, I guess.)
High potential for regressions.
Eventually we want the VO be driven by a A->V filter, so a decoder
doesn't even have to exist. Some features definitely require a decoder
though (like reporting the decoder in use, hardware decoding, etc.), so
for each thing which accessed d_video, it has to be redecided if and how
it can access decoder state.
At least the "framedrop" property slightly changes semantics: you can
now always set this property, even if no video is active.
Some untested changes in this commit, but our bio-based distributed
test suite has to take care of this.
This moves some code related to decoding from video.c to dec_video.c,
and also removes some accesses to dec_video.c from the filtering code.
dec_video.ch is starting to make sense, and simply returns video frames
from a demuxer stream. The API exposed is also somewhat intended to be
easily changeable to move decoding to a separate thread, if we ever want
this (due to libavcodec already being threaded, I don't see much of a
reason, but it might still be helpful).
Lots of noise to remove the vfilter/vo fields from dec_video.
From now on, video filtering and output will still be done together,
summarized under struct vo_chain.
There is the question where exactly the vf_chain should go in such a
decoupled architecture. The end goal is being able to place a "complex"
filter between video decoders and output (which will culminate in
natural integration of A->V filters for natural integration of
libavfilter audio visualizations). The vf_chain is still useful for
"final" processing, such as format conversions and deinterlacing. Also,
there's only 1 VO and 1 --vf option. So having 1 vf_chain for a VO seems
ideal, since otherwise there would be no natural way to handle all these
existing options and mechanisms.
There is still some work required to truly decouple decoding.
Instead of handling this on filter chain reinit, do it directly after
the decoder. This makes the code less entangled. In particular, this
gets rid of the really weird "override params" concept in the video
filter code.
The last_format/fixed_formats have some redundance with decoder_output,
but unfortunately the latter has a slightly different use.
Basically reimplement it. The old implementation was quite stupid, and
was probably done this way because video filtering and output used to be
way less decoupled. Now we can reimplement it in a very simple way: when
backstepping, seek to current time, but keep the last frame that was
supposed to be discarded when reaching the target time. When the seek
finishes, prepend the saved frame to the video frame queue.
A disadvantage is that the new implementation fails to skip over
timeline boundaries (ordered chapters etc.), but this never worked
properly anyway. It's possible that this will be fixed some time in the
future.
This is mainly a refactor. I'm hoping it will make some things easier
in the future due to cleanly separating codec metadata and stream
metadata.
Also, declare that the "codec" field can not be NULL anymore. demux.c
will set it to "" if it's NULL when added. This gets rid of a corner
case everything had to handle, but which rarely happened.
This is another attempt at making files with sparse video frames work
better.
The problem is that you generally can't know whether a jump in video
timestamps is just a (very) long video frame, or a timestamp reset. Due
to the existence of files with sparse video frames (new frame only every
few seconds or longer), every heuristic will be arbitrary (in general,
at least).
But we can use the fact that if video is continuous, audio should also
be continuous. Audio discontinuities can be easily detected, and if that
happens, reset some of the playback state.
The way the playback state is reset is rather radical (resets decoders
as well), but it's just better not to cause too much obscure stuff to
happen here. If the A/V sync code were to be rewritten, it should
probably strictly use PTS values (not this strange time_frame/delay
stuff), which would make it much easier to detect such situations and
to react to them.
Slightly change how it is decided when a new packet should be read.
Switch to demux_read_packet_async(), and let the player "wait properly"
until required subtitle packets arrive, instead of blocking everything.
Move distinguishing the cases of passive and active reading into the
demuxer, where it belongs.
MPlayer traditionally always used the display aspect ratio, e.g. 16:9,
while FFmpeg uses the sample (aka pixel) aspect ratio.
Both have a bunch of advantages and disadvantages. Actually, it seems
using sample aspect ratio is generally nicer. The main reason for the
change is making mpv closer to how FFmpeg works in order to make life
easier. It's also nice that everything uses integer fractions instead
of floats now (except --video-aspect option/property).
Note that there is at least 1 user-visible change: vf_dsize now does
not set the display size, only the display aspect ratio. This is
because the image_params d_w/d_h fields did not just set the display
aspect, but also the size (except in encoding mode).
Helps with files that have occasional broken timestamps. For larger
discontinuities, e.g. caused by actual timestamp resets, we still want
to realign audio.
(I guess in general, this should be removed and replaced by a more
general resync-on-desync logic, but not now.)
At least I hope so.
Deriving the duration from the pts was not really correct. It doesn't
include speed adjustments, and becomes completely wrong of the user e.g.
changes the playback speed by a huge amount. Pass through the accurate
duration value by adding a new vo_frame field.
The value for vsync_offset was not correct either. We don't need the
error for the next frame, but the error for the current one. This wasn't
noticed because it makes no difference in symmetric cases, like 24 fps
on 60 Hz.
I'm still not entirely confident in the correctness of this, but it sure
is an improvement.
Also, remove the MP_STATS() calls - they're not really useful to debug
anything anymore.
This was just converting back and forth between int64_t/microseconds and
double/seconds. Remove this stupidity. The pts/duration fields are still
in microseconds, but they have no meaning in the display-sync case (also
drop printing the pts field from opengl/video.c - it's always 0).
Instead of periodically trying to enable it again. There are two cases
that can happen:
1. A random discontinuity messed everything up,
2. Things are just broken and will desync all the time
Until now, it tried to deal with case 1 - but maybe this is really rare,
and we don't really need to care about it. On the other hand, case 2 is
kind of hard to diagnose if the user doesn't use the terminal.
Seeking will reenable display-sync, so you can fix playback if case 1
happens, but still get predictable behavior in case 2.
Use the demux_set_ts_offset() added in the previous commit to base each
timeline segment to use timestamps according to its relative position
within the overall timeline. As a consequence we don't need to care
about these timestamps anymore, and everything becomes simpler.
(Another minor but delicious nugget of sanity.)
If the player sends a frame with duration==0 to the VO, it can trivially
underrun. Don't panic, but keep the correct time.
Also, returning the absolute time from vo_get_next_frame_start_time()
just to turn it into a float with relative time was silly. Rename it and
make it return what the caller needs.
80ms allowable desync was a bit too much. It'd allow for a range of
160ms, which everyone can notice. It might also be a bother to apply
compensation resampling speed for that long.
We always let audio slowly desync until a threshold is reached, and then
pushed it back by applying a maximum compensation speed. Refine what
comes afterwards: instead of playing with the nominal video speed, use
the actual required audio speed for keeping sync as measured by the A/V
difference. (The "actual" speed is the ideal speed with A/V differences
added.)
Although this works in theory, it's somewhat questionable how much this
works in practice. The ideal time value is actually not exact, but is
the time at which the frame is scheduled (could be compensated by using
the time_left calculations in handle_display_sync_frame()). It doesn't
account for speed changes or catastrophic discontinuities. It uses only
10 past frames.
As long as it's within the desync tolerance, do not change the audio
speed at all for resampling. This reduces speed changes which might be
caused by jittering timestamps and similar cases.
(While in theory you could just not care and change speed every single
frame, I'm afraid that such changes could possibly cause audio
artifacts. So better just avoid it in the first place.)
This is very "illustrative", unlike the video-speed-correction
property, and thus useful. It can also be used to observe scheduling
errors, which are not detected by the core. (These happen due to
rounding errors; possibly not evne our fault, but coming from
files with rounded timestamps and so on.)
Instead of looking at the current frame duration for the intended
speedup, look at all past frames, and find a good average speed. This
ties in with not wanting to average _all_ frame durations, which
doesn't make sense in VFR situations.
This is currently done in the most naive way possible, but already sort
of works for VFR which switches between frame durations that are
integer multiples of a base rate. Certainly more improvements could
be made, such as trying to adjust directly on FPS changes, instead of
averaging everything, but for now this is not needed at all.
Helps somewhat with muxer-rounded timestamps.
There is some danger that this introduces a timestamp drift. But since
they are averaged values (unlike as when using an incorrect container
framerate hint), any potential drift shouldn't be too brutal, or
compensate itself soon. So I won't bother yet with comparing the results
with the real timestamp, unless we run into actual problems.
Of course we still prefer potentially real timestamps over the
approximated ones. But unless the timestamps match the container FPS,
we can't know whether they are (no, checking whether the they have
microsecond components would be cheating). Perhaps in future, we could
let the demuxer export the timebase - if the timebase is not 1000 (or
divisible by it), we know that millisecond-rounded timestamps won't
happen.
Get rid of get_past_frame_durations(), which was a bit too messy. Add
a past_frames array, which contains the same information in a more
reasonable way. This also means that we can get the exact current and
past frame durations without going through awful stuff. (The main
problem is that vo_pts_history contains future frames as well, which is
needed for frame backstepping etc., but gets in the way here.)
Also disable the automatic disabling of display-sync if the frame
duration changes, and extend the frame durations allowed for display
sync. To allow arbitrarily high durations, vo.c needs to be changed
to pause and potentially redraw OSD while showing a single frame, so
they're still limited.
In an attempt to deal with VFR, calculate the overall speed using the
average FPS. The frame scheduling itself does not use the average FPS,
but the duration of the current frame. This does not work too well,
but provides a good base for further improvements.
Where this commit actually helps a lot is dealing with rounded
timestamps, e.g. if the container framerate is wrong or unknown, or
if the muxer wrote incorrectly rounded timestamps. While the rounding
errors apparently can't be get rid of completely in the general case,
this is still much better than e.g. disabling display-sync completely
just because some frame durations go out of bounds.
We need a frame duration even on start, because the number of vsyncs
the frame is shown is predetermined. (vo_opengl actually makes use of
this property in certain cases.)
This check disables the display-sync resample method. If the filters
convert PCM to AC3, we can still insert a filter to change speed. This
is because filters are inserted at the beginning of the filter chain.
Move it (in a cosmetic sense), and also move its invocation to below all
the video handling.
All other changes remain cosmetic, including moving the framedrop
calculation code, and getting rid of the video_speed_correction
variable.
update_av_diff() works on the timestamps, while time_left is in real
time. When playing at not-1 speed, these are very different, and cause
the A/V difference to jitter. Fix this by scaling the expected A/V
desync to the correct range.
This didn't show up with cases where the frame pattern has a cycle of 1
or 2 like it is the case with 24-on-24 fps, or 24-on-60 fps. It did show
up with 25-on-60 fps. (We don't slow down 25 fps video to 24 on default
settings.)
In this case, we must not add the timing error of the next frame to the
A/V difference estimation of the current frame. Use the previous timing
error instead.
This is another bug resulting from the confusion about whether we
calculate parameters for the currently playing frame, or the one we're
about to queue.
Commit a1315c76 broke this slightly. Frame drops got counted multiple
times, and also vo.c was actually trying to "render" the dropped frame
over and over again (normally not a problem, since frames are always
queued "tightly" in display-sync mode, but could have caused 100% CPU
usage in some rare corner cases).
Do not repeat already dropped frames, but still treat new frames with
num_vsyncs==0 as dropped frames. Also, strictly count dropped frames in
the VO. This means we don't count "soft" dropped frames anymore (frames
that are shown, but for fewer vsyncs than intended). This will be
adjusted in the next commit.
Bump it to 80, and 2 vsyncs. This is another measure against vsync
jitter. Admittedly this is a bit simplistic (and we should probably
estimate a stable estimated vsync phase instead), but for now this will
do.
Calculate the A/V difference directly in the display sync code, instead
of the awkward current way, which reuses the fields for audio sync.
We still set time_frame, because it makes falling back to audio sync
somewhat smoother.
When dropping or repeating frames, we essentially influence when the
frame after the next frame will be shown, not the next frame. This led
to dropping/repeating frames 2 times, because the A/V difference had a
delay of one frame. Compensate it with the expected value.
This is all kinds of stupid - update_avsync_after_frame() will multiply
this value with the speed at a later point, and we only update this
field for this function. (This should be refactored.)
Apparently this function caused weird problems to me. I have no idea
why. The usage of the function looks perfectly fine to me, and even
rounding issues can be excluded. In any case, getting rid of this solved
my problem, and makes the code actually more readable.
This adjustment is supposed to improve the audio speed calculation in
case of unexpected desync. The flipped sign made it actually worse,
although the total impact of this bug was very minor.
If video EOF happens during playback restart, and audio is syncing, and
the demuxer packet queue overflows (i.e. no new packets will be read),
then it could happen that the player accidentally enters sleeping, and
continues playing anything only after e.g. user input wakes it up.
This parameter has been unused for years (the last flag was removed in
commit d658b115). Get rid of it.
This affects the general VO API, as well as the vo_opengl backend API,
so it touches a lot of files.
The VOFLAGs are still used to control OpenGL context creation, so move
them to the OpenGL backend code.
The vf_format suboption is replaced with --video-output-levels (a global
option and property). In particular, the parameter is removed from
mp_image_params. The mechanism is moved to the "video equalizer", which
also handles common video output customization like brightness and
contrast controls.
The new code is slightly cleaner, and the top-level option is slightly
more user-friendly than as vf_format sub-option.
Caused by one of the --force-window commits. The unconditional
uninit_video_out() call (which normally should be idempotent) raised
sporadic MPV_EVENT_VIDEO_RECONFIG events. This is ok, except for the
fact that clients (like a Lua script or libmpv users) would cause the
event loop to run again after receiving it, triggering a feedback loop.
Fix it by sending the events really only on a change.
