The previous commit broke these things, and fixing them is separate in
this commit in order to reduce the volume of changes.
Move the image queue from the VO to the playback core. The image queue
is a remnant of the old way how vdpau was implemented, and increasingly
became more and more an artifact. In the end, it did only one thing:
computing the duration of the current frame. This was done by taking the
PTS difference between the current and the future frame. We keep this,
but by moving it out of the VO, we don't have to special-case format
changes anymore. This simplifies the code a lot.
Since we need the queue to compute the duration only, a queue size
larger than 2 makes no sense, and we can hardcode that.
Also change how the last frame is handled. The last frame is a bit of a
problem, because video timing works by showing one frame after another,
which makes it a special case. Make the VO provide a function to notify
us when the frame is done, instead. The frame duration is used for that.
This is not perfect. For example, changing playback speed during the
last frame doesn't update the end time. Pausing will not stop the clock
that times the last frame. But I don't think this matters for such a
corner case.
The VO is run inside its own thread. It also does most of video timing.
The playloop hands the image data and a realtime timestamp to the VO,
and the VO does the rest.
In particular, this allows the playloop to do other things, instead of
blocking for video redraw. But if anything accesses the VO during video
timing, it will block.
This also fixes vo_sdl.c event handling; but that is only a side-effect,
since reimplementing the broken way would require more effort.
Also drop --softsleep. In theory, this option helps if the kernel's
sleeping mechanism is too inaccurate for video timing. In practice, I
haven't ever encountered a situation where it helps, and it just burns
CPU cycles. On the other hand it's probably actively harmful, because
it prevents the libavcodec decoder threads from doing real work.
Side note:
Originally, I intended that multiple frames can be queued to the VO. But
this is not done, due to problems with OSD and other certain features.
OSD in particular is simply designed in a way that it can be neither
timed nor copied, so you do have to render it into the video frame
before you can draw the next frame. (Subtitles have no such restriction.
sd_lavc was even updated to fix this.) It seems the right solution to
queuing multiple VO frames is rendering on VO-backed framebuffers, like
vo_vdpau.c does. This requires VO driver support, and is out of scope
of this commit.
As consequence, the VO has a queue size of 1. The existing video queue
is just needed to compute frame duration, and will be moved out in the
next commit.
Handle --term-playing-msg at a better place.
Move MPV_EVENT_TICK hack into a separate function. Also add some words
to the client API that you shouldn't use it. (But better leave breaking
it for later.)
Handle --frames and frame_step differently. Remove the mess from the
playloop, and do it after frame display. Give up on the weird semantics
for audio-only mode (they didn't make sense anyway), and adjust the
manpage accordingly.
If seeks take very long, it's better not to freeze up the display.
(This doesn't handle the case when decoding video frames is extremely
slow; just if hr-seek is used, or the demuxer is threaded and blocks on
network I/O.)
Achieve this by polling. Will be used by the OSC. Basically a bad hack -
but the point is that the mpv core itself is in the best position to
improve this later.
Basically move the code from playloop.c to video.c. The new function
write_video() now contains the code that was part of run_playloop().
There are no functional changes, except handling "new_frame_shown"
slightly differently. This is done so that we don't need new a new
MPContext field or a return value for write_video() to signal this
condition. Instead, it's handled indirectly.
This also reduces some code duplication with other parts of the code.
The changfe is mostly cosmetic, although there are also some subtle
changes in behavior. At least one change is that the big desync message
is now printed after every seek.
Regression since commit 261506e3. Internally speaking, playback was
often not properly terminated, and the main part of handle_keep_open()
was just executed once, instead of any time the user tries to seek. This
means playback_pts was not set, and the "current time" was determined by
the seek target PTS.
So fix this aspect of video EOF handling, and also remove the now
unnecessary eof_reached field.
The pause check before calling pause_player() is a lazy workaround for
a strange event feedback loop that happens on EOF with --keep-open.
If an imprecise seek is issues while a precise seek is ongoing,
don't wait up to 300ms (herustistic which usually improves user
experience), but instead let it cancel the seek.
Improves responsiveness of the OSC after the previous commit.
Note that we don't do this on "default-precise" seeks, because we
don't know if they're going to be precise or not.
Seeking in .ts files (and some other formats) is too unreliable, so
there's a separate code path for this case. But it breaks hr-seek.
Maybe hr-seek could actually be enabled in this case if we're careful
enough about timestamp resets, but for now nothing changes.
If the actual PTS is not known yet right after a seek, the "time-pos"
property will just return the seek target PTS. For this purpose, trigger
a change event to make the client API update the "time-pos" and related
properties. (MPV_EVENT_TICK triggers this update.)
Commit 261506e3 made constant seeking feel slower, because a subtle
change in the restart logic makes it now waste time showing another
video frame. The slowdown is about 20%.
(Background: the seek logic explicitly waits until a video frame is
displayed, because this makes it easier for the user to search for
something in the video. Without this logic, the display would freeze
until the user stops giving seek commands.)
Fix this by letting the seek logic issue another seek as soon as the
first video frame is displayed. This will prevent it from showing a
(useless, slow) second frame. Now it seems to be as fast as before the
change.
One side-effect is that the next seek happens after the first video
frame, but _before_ audio is restarted. Seeking is now silent. I guess
this is ok, so we don't do anything about it. Actually, I think whether
this happens is probably random; the seeking logic simply doesn't make
this explicit, so anything can happen.
This commit makes audio decoding non-blocking. If e.g. the network is
too slow the playloop will just go to sleep, instead of blocking until
enough data is available.
For video, this was already done with commit 7083f88c. For audio, it's
unfortunately much more complicated, because the audio decoder was used
in a blocking manner. Large changes are required to get around this.
The whole playback restart mechanism must be turned into a statemachine,
especially since it has close interactions with video restart. Lots of
video code is thus also changed.
(For the record, I don't think switching this code to threads would
make this conceptually easier: the code would still have to deal with
external input while blocked, so these in-between states do get visible
[and thus need to be handled] anyway. On the other hand, it certainly
should be possible to modularize this code a bit better.)
This will probably cause a bunch of regressions.
Move a condition somewhere else, which makes it conceptually simpler.
Also, the assignment to full_audio_buffers removed with this commit was
dead, and its value never used.
Fatal errors in the vidoe chain (such as failing to initialize the video
chain) disable video decoding. Restart the playloop, instead of just
continuing the current iteration.
The resulting behavior should be the same, but it gets rid of possible
corner cases.
Commit dc00b146, which disables polling by default, missed another
instance of polling: when the player pauses automatically on low cache.
This could lead to apparent freezes when playing network streams.
In my opinion this is not really necessary, since there's only a single
user of update_video(), but others reading this code would probably hate
me for using magic integer values instead of symbolic constants.
This should be a purely cosmetic commit; any changes in behavior are
bugs.
Instead of blocking on the demuxer when reading a packet, let packets be
read asynchronously. Basically, it polls whether a packet is available,
and if not, the playloop goes to sleep until the demuxer thread wakes it
up.
Note that the player will still block for I/O, because audio is still
read synchronously. It's much harder to do the same change for audio
(because of the design of the audio decoding path and especially
initialization), so audio will have to be done later.
Mouse cursor handling, --heartbeat-cmd, and OSD messages basically
relied on polling. For this reason, the playloop always used a small
timeout (not more than 500ms).
Fix these cases, and raise the timeout to 100 seconds. There is no
reason behind this number; for this specific purpose it's as close to
infinity as any other number.
On MS Windows, or if vo_sdl is used, the timeout remains very small.
In these cases the GUI code doesn't do proper event handling in the
first place, and fixing it requires much more effort.
getch2_poll() still does polling, because as far as I'm aware no event-
based way to detect this state change exists.
This adds a thread to the demuxer which reads packets asynchronously.
It will do so until a configurable minimum packet queue size is
reached. (See options.rst additions.)
For now, the thread is disabled by default. There are some corner cases
that have to be fixed, such as fixing cache behavior with webradios.
Note that most interaction with the demuxer is still blocking, so if
e.g. network dies, the player will still freeze. But this change will
make it possible to remove most causes for freezing.
Most of the new code in demux.c actually consists of weird caches to
compensate for thread-safety issues (with the previously single-threaded
design), or to avoid blocking by having to wait on the demuxer thread.
Most of the changes in the player are due to the fact that we must not
access the source stream directly. the demuxer thread already accesses
it, and the stream stuff is not thread-safe.
For timeline stuff (like ordered chapters), we enable the thread for the
current segment only. We also clear its packet queue on seek, so that
the remaining (unconsumed) readahead buffer doesn't waste memory.
Keep in mind that insane subtitles (such as ASS typesetting muxed into
mkv files) will practically disable the readahead, because the total
queue size is considered when checking whether the minimum queue size
was reached.
demux_seek() actually doesn't return seek success. Instead, it fails if
the demuxer is flagged as unseekable (but this is checked explicitly at
the beginning of this function), or if the seek target PTS is
MP_NOPTS_VALUE (which can never happen).
This should be unneeded, and the packet position is already sufficient
for this case.
Accessing the stream position directly is going to be a problem when the
stream is accessed from another thread later.
Let the VOs draw the OSD on their own, instead of making OSD drawing a
separate VO driver call. Further, let it be the VOs responsibility to
request subtitles with the correct PTS. We also basically allow the VO
to request OSD/subtitles at any time.
OSX changes untested.
Stop using it in most places, and prefer STREAM_CTRL_GET_SIZE. The
advantage is that always the correct size will be used. There can be no
doubt anymore whether the end_pos value is outdated (as it happens often
with files that are being downloaded).
Some streams still use end_pos. They don't change size, and it's easier
to emulate STREAM_CTRL_GET_SIZE using end_pos, instead of adding a
STREAM_CTRL_GET_SIZE implementation to these streams.
Make sure int64_t is always used for STREAM_CTRL_GET_SIZE (it was
uint64_t before).
Remove the seek flags mess, and replace them with a seekable flag. Every
stream must set it consistently now, and an assertion in stream.c checks
this. Don't distinguish between streams that can only be forward or
backwards seeked, since we have no such stream types.
stream.start_pos was needed for optical media only, and (apparently) not
for very good reasons. Just get rid of it.
For stream_dvd, we don't need to do anything. Byte seeking was already
removed from it earlier.
For stream_cdda and stream_vcd, emulate the start_pos by offsetting the
stream pos as seen by the rest of mpv.
The bits in discnav.c and loadfile.c were for dealing with the code
seeking back to the start in demux.c. Handle this differently by
assuming the demuxer is always initialized with the stream at start
position, and instead seek back if initializing the demuxer fails.
Remove the --sb option, which worked by modifying stream.start_pos. If
someone really wants this option, it could be added back by creating a
"slice" stream (actually ffmpeg already has such a thing).
Cover art is treated like video, but is not really video. In one case,
the audio sync code was accidentally still active. Fixes cover art
playback with --ao=null. (This is due to ao_null's latency emulation.
Although it's not very clear whether that is actually correct...)
Some options change from percentages to number of kilobytes; there are
no cache options using percentages anymore.
Raise the default values. The cache is now 25000 kilobytes, although if
your connection is slow enough, the maximum is probably never reached.
(Although all the memory will still be used as seekback-cache.)
Remove the separate --audio-file-cache option, and use the cache default
settings for it.
Until recently, the VO was an unavoidable part of the seeking code path.
This was because vdpau deinterlacing could double the framerate, and hr-
seek and framestepping etc. all had to "see" the additional frames. But
we've removed the frame doubling from the vdpau VO and moved it into a
video filter (vf_vdpaupp), and there's no reason left why the VO should
participate in seeking.
Instead of queuing frames to the VO during seek and skipping them
afterwards, drop the frames early.
This actually might make seeking with vo_vdpau and software decoding
faster, although I haven't measured it.
Now we avoid calling update_video() twice on reconfig (once to check
whether there are still new frames, and again to actually do the
reconfig). Instead, we check whether there's still something going on
before calling update_video() at all, and depending on that
update_video() will be allowed to reconfig or not.
This will simplify some things later.
Also remove MSGL_SMODE and friends.
Note: The indent in options.rst was added to work around a bug in
ReportLab that causes the PDF manual build to fail.
This wasn't really fine, and could (perhaps) cause weird corner cases on
reinit or when the player was paused.
Before eb9d20, video_left was also set to true if vo->frame_loaded was
set, and this variable basically indicated whether the previous
update_video() call was successful. This was overlooked when changing
everything. Simply always call update_video(), it should be equivalent.
Change how the video decoding loop works. The structure should now be a
bit easier to follow. The interactions on format changes are (probably)
simpler. This also aligns the decoding loop with future planned changes,
such as moving various things to separate threads.
wm4
Tsukasa OMOTO
Martin Herkt