AOs can now call ao_underrun_event() (in any context) if an underrun has
happened. It will print a message.
This will be used in the following commits. But for now, audio.c only
clears the underrun bit, so that subsequent underruns still print the
warning message.
Since the underrun flag will be used in fragile ways by the playback
state machine, there is the "reports_underruns" field that signals
strong support for underrun reporting. (Otherwise, underrun events will
not be used by it.)
Until recently, ao_lavc and vo_lavc started encoding whenever the core
happened to send them data. Since audio and video are not initialized at
the same time, and the muxer was not necessarily opened when the first
encoder started to produce data, the resulting packets were put into a
queue. As soon as the muxer was opened, the queue was flushed.
Change this to make the core wait with sending data until all encoders
are initialized. This has the advantage that we don't need to queue up
the packets.
stdatomic.h defines no atomic_float typedef. We can't just use _Atomic
unconditionally, because we support compilers without C11 atomics. So
just create a custom atomic_float typedef in the wrapper, which uses
_Atomic in the C11 code path.
This does what af_volume used to do. Since we couldn't relicense it,
just rewrite it. Since we don't have a new filter mechanism yet, and the
libavfilter is too inconvenient, do applying the volume gain in ao.c
directly. This is done before handling the audio data to the driver.
Since push.c runs a separate thread, and pull.c is called asynchronously
from the audio driver's thread, the volume value needs to be
synchronized. There's no existing central mutex, so do some shit with
atomics. Since there's no atomic_float type predefined (which is at
least needed when using the legacy wrapper), do some nonsense about
reinterpret casting the float value to an int for the purpose of atomic
access. Not sure if using memcpy() is undefined behavior, but for now I
don't care.
The advantage of not using a filter is lower complexity (no filter auto
insertion), and lower latency (gain processing is done after our
internal audio buffer of at least 200ms).
Disavdantages include inability to use native volume control _before_
other filters with custom filter chains, and the need to add new
processing for each new sample type.
Since this doesn't reuse any of the old GPL code, nor does indirectly
rely on it, volume and replaygain handling now works in LGPL mode.
How to process the gain is inspired by libavfilter's af_volume (LGPL).
In particular, we use exactly the same rounding, and we quantize
processing for integer sample types by 256 steps. Some of libavfilter's
copyright may or may not apply, but I think not, and it's the same
license anyway.
I plan to remove the S24 sample formats in mpv. It seems like we should
still support this _somehow_ in AOs though. So the idea is to convert
the data to more obscure representations (that would not be useful for
filtering etc. anyway) within the AO.
This commit adds helper to enable this. ao_convert_fmt is meant to
provide mechanisms for this, rather than a generic audio format
description (as the latter leads only to overly generic misery). The
conversion also supports only cases which we think will be needed at
all.
The main advantage of this approach is that we get S24 out of sight,
and that we could support other crazy formats (like S20). The main
disadvantage is that usually S32 will be selected (if both S32 and S24
are available), and there's no user control to force S24. That doesn't
really matter though, and at worst makes testing harder or will lead
to unpleasant arguments with audiophiles (they'd be wrong anyway).
ao_convert_fmt.pad_lsb is ignored, although if we ever find a case in
which playing S32 with data in the LSBs breaks when playing it as padded
24 bit format. (For example, WAVEFORMATEXTENSIBLE recommends setting the
unused bits to 0 if wValidBitsPerSample implies LSB padding.)
Before this change, AOs could have internal alignment, and play() would
not consume the trailing data if the size passed to it is not aligned.
Change this to require AOs to report their alignment (via period_size),
and make sure to always send aligned data.
The buffer reported by get_space() now always has to be correct and
reliable. If play() does not consume all data provided (which is bounded
by get_space()), an error is printed.
This is preparation for potential further AO changes.
I casually checked alsa/lavc/null/pcm, the other AOs might or might not
work.
All contributors of the current code have agreed. ao.c requires a
"driver" entry for each audio output - we assume that if someone who
didn't agree to LGPL added a line, it's fine for ao.c to be LGPL
anyway. If the affected audio output is not disabled at compilation
time, the resulting binary will be GPL anyway, and ootherwise the
code is not included.
The audio output code itself was inspired or partially copied from
libao in 7a2eec4b59 (thus why MPlayer's audio code is named libao2).
Just to be sure we got permission from Aaron Holtzman, Jack Moffitt, and
Stan Seibert, who according to libao's SVN history and README are the
initial author. (Something similar was done for libvo, although the
commit relicensing it forgot to mention it.)
242aa6ebd40: anders mostly disagreed with the LGPL relicensing, but we
got permission for this particular commit.
0ef8e555735: nick could not be reached, but the include statement was
removed again anyway.
879e05a7c17: iive agreed to LGPL v3+ only, but this line of code was
removed anyway, so ao_null.c can be LGPL v2.1+.
9dd8f241ac2: patch author could not be reached, but the corresponding
code (old slave mode interface) was completely removed later.
Long planned. Leads to some sanity.
There still are some rather gross things. Especially g_groups is ugly,
and a hack that can hopefully be removed. (There is a plan for it, but
whether it's implemented depends on how much energy is left.)
Currently, calling mp_input_wakeup() will wake up the core thread (also
called the playloop). This seems odd, but currently the core indeed
calls mp_input_wait() when it has nothing more to do. It's done this way
because MPlayer used input_ctx as central "mainloop".
This is probably going to change. Remove direct calls to this function,
and replace it with mp_wakeup_core() calls. ao and vo are changed to use
opaque callbacks and not use input_ctx for this purpose. Other code
already uses opaque callbacks, or has legitimate reasons to use
input_ctx directly (such as sending actual user input).
I decided that it's too much work to convert all the VO/AOs to the new
option system manually at once. So here's a shitty hack instead, which
achieves almost the same thing. (The only user-visible difference is
that e.g. --vo=name:help will list the sub-options normally, instead of
showing them as deprecation placeholders. Also, the sub-option parser
will verify each option normally, instead of deferring to the global
option parser.)
Another advantage is that once we drop the deprecated options,
converting the remaining things will be easier, because we obviously
don't need to add the compatibility hacks.
Using this mechanism is separate in the next commit to keep the diff
noise down.
Instead of requiring each VO or AO to manually add members to MPOpts and
the global option table, make it possible to register them automatically
via vo_driver/ao_driver.global_opts members. This avoids modifying
options.c/options.h every time, including having to duplicate the exact
ifdeffery used to enable a driver.
This commit adds an --audio-channel=auto-safe mode, and makes it the
default. This mode behaves like "auto" with most AOs, except with
ao_alsa. The intention is to allow multichannel output by default on
sane APIs. ALSA is not sane as in it's so low level that it will e.g.
configure any layout over HDMI, even if the connected A/V receiver does
not support it. The HDMI fuckup is of course not ALSA's fault, but other
audio APIs normally isolate applications from dealing with this and
require the user to globally configure the correct output layout.
This will help with other AOs too. ao_lavc (encoding) is changed to the
new semantics as well, because it used to force stereo (perhaps because
encoding mode is supposed to produce safe files for crap devices?).
Exclusive mode output on Windows might need to be adjusted accordingly,
as it grants the same kind of low level access as ALSA (requires more
research).
In addition to the things mentioned above, the --audio-channels option
is extended to accept a set of channel layouts. This is supposed to be
the correct way to configure mpv ALSA multichannel output. You need to
put a list of channel layouts that your A/V receiver supports.
Not very important for the command line player; but GUI applications
will want to know about this.
This only adds the internal API; support for specific audio outputs
comes later.
This reuses the ao struct as context for the hotplug event listener,
similar to how the "old" device listing API did. This is probably a bit
unclean and confusing. One argument got reusing it is that otherwise
rewriting parts of ao_pulse would be required (because the PulseAudio
API requires so damn much boilerplate). Another is that --ao-defaults is
applied to the hotplug dummy ao struct, which automatically applies such
defaults even to the hotplug context.
Notification works through the property observation mechanism in the
client API. The notification chain is a bit complicated: the AO notifies
the player, which in turn notifies the clients, which in turn will
actually retrieve the device list. (It still has the advantage that it's
slightly cleaner, since the AO stuff doesn't need to know about client
API issues.)
The weird handling of atomic flags in ao.c is because we still don't
require real atomics from the compiler. Otherwise we'd just use atomic
bitwise operations.
This is what you would expect. Before this commit, each
ao_request_reload() call would just queue a reload command, and then
recreate the AO for the number of times the function was called.
Instead of sending a command, introduce some sort of event retrieval
mechanism. At least for the reload case, use atomics, because we're too
lazy to setup an extra mutex.
The main need I see for this is with libmpv - it would be confusing if
some application showed up as "mpv" on whateverthehell PulseAudio uses
it for (generally it does show up on various PA GUI tools).
Since the internal AO driver API has no proper way to determine EOF, we
need to guess by querying get_delay. But some AOs (e.g. ao_pulse with
no-latency-hacks set) may never reach 0, maybe because they naively add
the latency to the buffer level. In this case our heuristic can break.
Fix by always using the delay to estimate the EOF time. It's not even
that important - it's mostly used to avoid blocking draining. So this
should be ok.
CC: @mpv-player/stable (maybe)
Remove the unnecessary indirection through ao fields.
Also fix the inverted result of AOCONTROL_HAS_TEMP_VOLUME. Hopefully the
change is equivalent. But actually, it looks like the old code did it
wrong.
With --gapless-audio=no, changing from one file to the next apparently
made it hang, until the player was woken up by unrelated events like
input. The reason was that the AO doesn't notify the player of EOF
properly. the played was querying ao_eof_reached(), and then just went
to sleep, without anything waking it up.
Make it event-based: the AO wakes up the playloop if the EOF state
changes.
We could have fixed this in a simpler way by synchronously draining the
AO in these cases. But I think proper event handling is preferable.
Fixes: #1069
CC: @mpv-player/stable (perhaps)
Logic for this was missing from pull.c. For push.c it was missing if the
driver didn't support it. But even if the driver supported it (such as
with ao_alsa), strange behavior was observed by users. See issue #933.
Always check explicitly whether the AO is in paused mode, and if so,
don't drain.
Possibly fixes#933.
CC: @mpv-player/stable
Until now, we've always calculated a timeout based on a heuristic when
to refill the audio buffers. Allow AOs to do it completely event-based
by providing wait and wakeup callbacks.
This also shuffles around the heuristic used for other AOs, and there is
a minor possibility that behavior slightly changes in real-world cases.
But in general it should be much more robust now.
ao_pulse.c now makes use of event-based waiting. It already did before,
but the code for time-based waiting was also involved. This commit also
removes one awkward artifact of the PulseAudio API out of the generic
code: the callback asking for more data can be reentrant, and thus
requires a separate lock for waiting (or a recursive mutex).
There were subtle and minor race conditions in the pull.c code, and AOs
using it (jack, portaudio, sdl, wasapi). Attempt to remove these.
There was at least a race condition in the ao_reset() implementation:
mp_ring_reset() was called concurrently to the audio callback. While the
ringbuffer uses atomics to allow concurrent access, the reset function
wasn't concurrency-safe (and can't easily be made to).
Fix this by stopping the audio callback before doing a reset. After
that, we can do anything without needing synchronization. The callback
is resumed when resuming playback at a later point.
Don't call driver->pause, and make driver->resume and driver->reset
start/stop the audio callback. In the initial state, the audio callback
must be disabled.
JackAudio of course is different. Maybe there is no way to suspend the
audio callback without "disconnecting" it (what jack_deactivate() would
do), so I'm not trying my luck, and implemented a really bad hack doing
active waiting until we get the audio callback into a state where it
won't interfere. Once the callback goes from AO_STATE_WAIT to NONE, we
can be sure that the callback doesn't access the ringbuffer or anything
else anymore. Since both sched_yield() and pthread_yield() apparently
are not always available, use mp_sleep_us(1) to avoid burning CPU during
active waiting.
The ao_jack.c change also removes a race condition: apparently we didn't
initialize _all_ ao fields before starting the audio callback.
In ao_wasapi.c, I'm not sure whether reset really waits for the audio
callback to return. Kovensky says it's not guaranteed, so disable the
reset callback - for now the behavior of ao_wasapi.c is like with
ao_jack.c, and active waiting is used to deal with the audio callback.
For some reason, the buffered_audio variable was used to "cache" the
ao_get_delay() result. But I can't really see any reason why this should
be done, and it just seems to complicate everything.
One reason might be that the value should be checked only if the AO
buffers have been recently filled (as otherwise the delay could go low
and trigger an accidental EOF condition), but this didn't work anyway,
since buffered_audio is set from ao_get_delay() anyway at a later point
if it was unset. And in both cases, the value is used _after_ filling
the audio buffers anyway.
Simplify it. Also, move the audio EOF condition to a separate function.
(Note that ao_eof_reached() probably could/should whether the last
ao_play() call had AOPLAY_FINAL_CHUNK set to avoid accidental EOF on
underflows, but for now let's keep the code equivalent.)
And also add a function ao_need_data(), which AO drivers can call if
their audio buffer runs low.
This change intends to make it easier for the playback thread: instead
of making the playback thread calculate a timeout at which the audio
buffer should be refilled, make the push.c audio thread wakeup the core
instead.
ao_need_data() is going to be used by ao_pulse, and we need to
workaround a stupid situation with pulseaudio causing a deadlock because
its callback still holds the internal pulseaudio lock.
For AOs that don't call ao_need_data(), the deadline is calculated by
the buffer fill status and latency, as before.
Since the addition of the AO feed thread, 200ms of latency (MIN_BUFFER)
was added to all push-based AOs. This is not so nice, because even AOs
with relatively small buffering (e.g. ao_alsa on my system with ~170ms
of buffer size), the additional latency becomes noticable when e.g.
toggling mute with softvol.
Fix this by trying to keep not only 200ms minimum buffer, but also 200ms
maximum buffer. In other words, never buffer beyond 200ms in total. Do
this by estimating the AO's buffer fill status using get_space and the
initially known AO buffer size (the get_space return value on
initialization, before any audio was played). We limit the maximum
amount of data written to the soft buffer so that soft buffer size and
audio buffer size equal to 200ms (MIN_BUFFER).
To avoid weird problems with weird AOs, we buffer beyond MIN_BUFFER if
the AO's get_space requests more data than that, and as long as the soft
buffer is large enough.
Note that this is just a hack to improve the latency. When the audio
chain gains the ability to refilter data, this won't be needed anymore,
and instead we can introduce some sort of buffer replacement function in
order to update data in the soft buffer.
Until now, this was always conflated with uninit. This was ugly, and
also many AOs emulated this manually (or just ignored it). Make draining
an explicit operation, so AOs which support it can provide it, and for
all others generic code will emulate it.
For ao_wasapi, we keep it simple and basically disable the internal
draining implementation (maybe it should be restored later).
Tested on Linux only.
This has 2 goals:
- Ensure that AOs have always enough data, even if the device buffers
are very small.
- Reduce complexity in some AOs, which do their own buffering.
One disadvantage is that performance is slightly reduced due to more
copying.
Implementation-wise, we don't change ao.c much, and instead "redirect"
the driver's callback to an API wrapper in push.c.
Additionally, we add code for dealing with AOs that have a pull API.
These AOs usually do their own buffering (jack, coreaudio, portaudio),
and adding a thread is basically a waste. The code in pull.c manages
a ringbuffer, and allows callback-based AOs to read data directly.
We want to move the AO to its own thread. There's no technical reason
for making the ao struct opaque to do this. But it helps us sleep at
night, because we can control access to shared state better.
Aman Gupta
wm4
Marcin Kurczewski
Stefano Pigozzi