It is possible that the callback caused a reinit task to be queued while at the same time audiounit_stream_destroy got called.
We need to abort early as otherwise both stm->input_unit and stm->output_unit would have been cleared.
After calling the resampler, we would reduce the number of frames in the input buffer correctly by the number of frames used, but would always fully clear the input buffer after.
Resulting with a input_buffer and its frame counters to be potentially out of sync
In duplex mode, audiounit_input_callback and audiounit_output_callback are always called on the same thread. input_linear_buffer is only ever accessed serially.
Additionally, the use of atomic variables was incorrect as they should only ever be modified when input_linear_buffer was modified.
So we remove the lock and the related atomics.
remove input_buffer_frames member as its value is always latency_frames
* Use DEVICES_PROPERTY_ADDRESS in audiounit_get_devices_of_type
* Use exactly same AudioObjectPropertyAddress in audiounit_add_device_listener() and audiounit_remove_device_listener()
We have duplicated AudioObjectPropertyAddress variables in audiounit_add_device_listener() and audiounit_remove_device_listener(). This two functions use same values to add and remove listeners on same system events. Therefore, we should use the exactly same value for both instead of creating same values in two functions, in case the local variables are changed carelessly.
* Use global AudioObjectPropertyAddress settings for both audiounit_add_listener() and audiounit_remove_listener()
* audiounit_get_default_device_datasource
* Simplify audiounit_add/remove_listener() by introducing property_listener
audiounit_add_listener() and audiounit_remove_listener() should use exactly same variable to register and unregister event listener. Originally, these two functions are set by different local variables with same values: audiounit_add_listener(X, Y, Z, ...) and audiounit_remove_listener(X', Y', Z', ...), where X = X', Y = Y',...etc. There are too many function parameters so it's not easy to check if we have same parameters for audiounit_add_listener() and audiounit_remove_listener().
property_listener is introduced to simplify the parameter issues by using audiounit_add/remove_listener(property_listener L). A property_listener object L is created to register a event listener, and then be used to unregister the listener later. It'll be safer if audiounit_add_listener() and audiounit_remove_listener() use the exactly same object as their parameters. In addition, The readability is better since there is a "listener" be added/removed when we call audiounit_add/remove_listener().
* Use DEFAULT_INPUT/OUTPUT_DEVICE_PROPERTY_ADDRESS in audiounit_get_default_device_id()
* Make property_listener pointer be const
* Remove stm->XXX_listners if it's not nullptr no matter the input/output_unit exists or not.
stm->XXX_listener will be set based on existence of the input/output_unit. Therefore, if it's not nullptr, the input/output_unit must exist. Removing listeners or not only depends on existeneces of the listeners themselves.
* Add a test for cubeb_stream_get_current_device() and cubeb_stream_device_destroy()
* Reduce duplicates in audiounit_stream_get_current_device()
The audiounit_stream_get_current_device() will set the default device's name for input and output by the following steps:
1. Get default device's data, whose type is uint32
2. Convert the uint32 data into a string
3. Set the string into cubeb_device's name
Hence we can split it into functions to do the above steps:
i. audiounit_get_default_device_data(...)
Do the step 1
ii. allocate_and_convert_uint32_into_string(...)
Do the step 2 and 3. It will allocate memory for a string and put the converted string right there. (If we don't allocate and then set the values into the memory at the same time, we need to find a way to check the memory's boundary.)
They are called by a function named audiounit_get_default_device_name(...), with cubeb_device_type parameter to set the device name of input or output. And audiounit_get_default_device_name(...) will be used in audiounit_stream_get_current_device().
* Replace allocate_and_convert_uint32_into_string(...) by convert_uint32_into_string(...)
1. Rename allocate_and_convert_uint32_into_string(...) to convert_uint32_into_string(...)
2. Return a unique_ptr<char[]> from convert_uint32_into_string(...) pointing to an allocated memory
* Hard-coding the convertion from uint32 into string
This change is to make sure the conversion only ever needs to handle 4 bytes.
* Add more information to log when calling audiounit_stream_get_current_device(...)
We add a log with error message when we cannot get datasource data from the devices. However, we don't return an error code in this case since it's quite common when we try getting that data from USB devices. We will convert the datasource data into a string. If there is no data, the string is empty. Instead of logging when we cannot get the datasource data, it's better to log that the converted name is empty. It'll give more meaning (Users are more likely confused about what the empty datasource means.).
* Rename audiounit_get_default_device_data() to audiounit_get_default_device_datasource()
We should use an explicit name for this function since the same device will return different datasources. For example, the default input device on macbook pro will return "imic" if it uses the default internal microphone or "emic" when it uses an external microphone plugged in audio jack.
TODO: it's better to rename audiounit_stream_get_current_device to audiounit_stream_get_current_device_source. The reason is same as above.
This allows to output what people typically expect when playing mono audio: sound coming from both left and right channels.
To force this conversion for happening on mac, we tag that layout are unknown as soon as a channel type is unknown
The method kAudioUnitProperty_AudioChannelLayout used to retrieve the channel layout wasn't introduced until 10.12. So we use kAudioDevicePropertyPreferredChannelLayout instead should it fails.
Fixes#448
This define was added as part of commit d2c45250 and is mingw-specific (only mingw uses __MSVCRT_VERSION__). I don't know why it was added, it shouldn't be needed.
Recently mingw-w64 added support for UCRT-based toolchains. In this case __MSVCRT_VERSION__ is set to 0x1400 and should not really be changed. UCRT-based builds support a lot of stdio.h function by inline wrappers. Those can't be disabled just for one file as they are not exported by ucrtbase.dll.
I found the problem while working on porting Firefox to clang+mingw-w64 toolchain that uses UCRT by default.
Apply volume in software as do other backends. This is necessary
because sndio volume may be controlled externally and there's no
volume getter in libcubeb to notify the caller about volume
changes.
The presentation latency of a stream is roughly kAudioDevicePropertyLatency +
kAudioStreamPropertyLatency + kAudioUnitProperty_Latency. Calculate this at
device setup time and apply the correction when calculating the current stream
position for stream_get_position.
This addresses the thread local COM initialization lifetime problem
discussed in issue #416 by moving the responsibility for COM
initialization from within cubeb to the caller.
* Add QUAD and QUAD_LFE layouts.
* Remove dual mono layout.
It makes no sense to have a case for those as the data structure
used (a bitmask) do not allow to represent this channel layout (a
channel can only be present once). As such it was a non-functional
layout
* Fix up cubeb_pulse compilation using C++ keyword.
* Remove the concept of preferred layout.
Channel layout is derived by the content being played. The concept of
preferred layout is meaningless. Either we have a layout defined, or
we don't. There's no in-between.
So we remove it.
* Remove CHANNEL_MONO concept.
* Add cubeb_sample_size convenience method.
* Rework cubeb_mixer.
This completely replace the existing remixer which had serious limitations:
1- Had no memory bound checks
2- Could only downmix 5.1 and 7.1 to stereo.
This mixer allows to convert from any sane layout to any other and work directly on interleaved samples.
This cubeb_mixer doesn't have an API compatible with the previous one.
This commit is non-fonctional, and was split for ease of review.
* Fix remixing on mac, windows and pulse backend.
* Make cubeb_mixer creation infallible.
Rather than ignore nonsensical layouts, we attempt to play it according to the stream channels count instead. The audio data will be played as-is, dropping the extra channels or inserting silence where needed.
* User proper sample size when calculating offsets.
Should the user data be of a different type to what the AudioUnit output is set to, we would have written outside the end of our allocated buffer.
* Fix input mixing and clarify frames vs samples terminology
* If a layout is unknown or invalid, always treat it as plain stereo or mono.
- For the silent loopback test we check that we don't get any sound above an
epsilon. The value of this epsilon has been increased to accommodate a wider
range of systems.
- The delay on all tests has been increased to allow for more time to capture
looped samples. The original 150ms delay was sufficient on the machine I tested
with, however, I've seen the tests fail due to not having enough looped input.
Some devices (namely, Bose QC35, mark 1 and 2), expose a single channel mapped
to the right for some reason, and this confuses our channel mapping code.
Alex Chronopoulos
Matthew Gregan
Jean-Yves Avenard
Paul Adenot
orbea
Chun-Min Chang
Andreas Pehrson
Jacek Caban
Alexandre Ratchov
Dan Glastonbury
Jean-Yves Avenard
Jan Malakhovski
Bryce Van Dyk