Without a channel mask, the channels are treated as "raw" or unknown. Auto-
detection will only go up to stereo (assuming the first two channels are front-
left and front-right), while explicit requests will work as long as there are
enough channels (the user is responsible for ensuring a correct setup).
So that it's tried before sndio. Many Linux systems seem to have libsndio
installed as a hard dependency from SDL2, but not the server component to
actually make it work. And because there's no way to test if it's usable
without opening a device, the backend would get selected over ALSA and
subsequently fail to work. As a result, Linux users that don't use PipeWire or
PulseAudio would have to explicitly configure for ALSA.
* Upload Win32 and x64 artifacts to Github actions
Provides an alternative Windows binary download location that expires in 3 months (2 months longer than Appveyor)
* Add Windows check to skip failed Github actions artifact upload for Linux and macOS
The write vector represents the chunks queued for OpenSL, and the write pointer
only increments when OpenSL fills in more samples. So requeueing from the start
of the write vector is requeueing chunks that are already queued, which is
obviously wrong. It instead needs to queue chunks that become available as
they're read.
Because the ring buffer holds more elements than can be written to at a givem
time, the read vector does not represent the next writable chunks to queue.
Advancing the read pointer increases the number of writable elements, so
instead of queueing from the read vector or the start of the write vector,
queue from the end of the write vector given the number of chunks read.
There's apparently no way to get the containing class/struct type from a static
member function, and operator new/delete are implicitly static member functions
so the macros to define them need to be told the type to allocate for. This
ensures the type specified matches the containing type.
We want to resample before minimum phase reconstruction since that changes the
phase relationship of the sampled signal, introducing a slight bit of noise
from truncated sampling. It's not clear that the frequency domain resampling
method is accurate, so resampling prior to frequency analysis is an alternative
to ensure the resulting frequencies are given the proper phase for sampling.
This also cleans up some micro allocations in loops.
Currently it makes the pkg-config file unusable as prefix is specified twice.
FULL variables do the necessary parsing and automatically prepend the prefix if needed.
Newer versions of PipeWire may add things to public structures. For example,
pw_buffer::requested added in 0.3.49. Building against 0.3.49 or newer, but
then running with 0.3.48 could result in invalid accesses since the returned
pw_buffer objects are shorter than the definition says to expect, creating
undefined behavior. Even if explicit access to the additional fields is
protected by a runtime check, the language allows the compiler to assume a
pointer to a pw_buffer object contains a complete pw_buffer, allowing the
optimizer to access the field earlier than the check (with the check only
controlling if the value gets used).
Another example is pw_time, which had a few fields added in 0.3.50 along with a
function, pw_stream_get_time_n, that provides the size of the pw_time struct
the application is using (so the library knows what version of the struct it
has to fill in). If a later version adds a new field, running it with an older
version will either fail (due to the library getting a size larger than it
knows about) or silently leave the newer fields as garbage.
Building against 0.3.50 or newer will require that version or newer at runtime.
Building against a version before 0.3.50 will work with newer versions.
The backend is apparently using some stuff that wasn't in earlier headers, but
it's not clear what was introduced in which versions. 0.3.23 should work,
though it may need to be higher or it could go a bit lower.
