This is the one that splits the "left wing" into two for loops to
bubble out the conditional that decides if it should read from the
left padding or the input buffer.
I still believe the optimization is good, but the basic logic of it
was incorrect, and needs to be reexamined and fixed before going
back into revision control.
- Calculate `j * RESAMPLER_SAMPLES_PER_ZERO_CROSSING` once per loop
iteration since we use it multiple times.
- Do the left-wing loop in two sections: while `srcframe < 0` and then
the remaining calculations when `srcframe >= 0`. This bubbles a conditional
out of every iteration of a tight loop, giving us a boost. We could
_probably_ do this to the right-wing loop too, but it's less straightforward
there.
- The real win: Use floats instead of doubles. This almost doubles the speed
of the entire function on Intel CPUs, and for embedded things without
hardware-level support for doubles, the speedup is enormous. This in
theory might reduce audio quality, though, and I had to put a check in
place to avoid a division-by-zero that we avoided at higher precision, but
this is likely to be worth keeping for at least the Sony PSP and other
smaller platforms, if not everyone.
Instead of waiting until the entire buffer from the SDL callback is ready
to be accepted by PulseAudio, we use pa_stream_set_write_callback and
feed some portion of the buffer as callbacks come in asking for more.
This lets us remove the halving of the buffer size during device open,
and also (hopefully) solves several strange hangs that happen in unusual
circumstances.
Fixes#4387Fixes#2262
if SDL_EnumUnixAudioDevices() fails to find any devices,
set an error message on the exit path. Without this,
SDL_Init() could fail without any message available
in SDL_GetError().
- no need to keep the error in a static variable
- always print the error code
- reduce the required stack-size
- reduce the number of snprintf calls (and code size)
The io_list_check_add() and io_list_remove() functions are only ever called from within the Pipewire thread loop, so the locks are redundant. io_list_sort() is called from within a lock in the device detection function, so those additional locks are redundant as well.
Remove the hard upper limit of 8192 samples and instead use the buffer sizes provided by Pipewire to determine the size of the intermediate input buffer and whether double buffering is required for output streams. This allows for higher latency streams to potentially avoid double-buffering in the output case, and we can guarantee that the intermediate input buffer will always be large enough to handle whatever Pipewire may deliver.
As the buffer size calculations occur in a callback in the Pipewire processing thread itself, the stream readiness check has been modified to wait on two distinct flags set when the buffers have been configured and when the stream is ready and running.
The context and stream creation functions will destroy the passed properties object on failure, so no need to do it manually.
The pw_properties_free() function pointer is no longer needed, so it can be removed.
- drop unnecessary hascapture check
- call SDL_InvalidParamError and return -1 in case the index is out of range
- do not zfill SDL_AudioSpec
- adjust documentation to reflect the behavior
- reorganize the loop which checks for the right wave-format
- use the return value of UpdateAudioStream
- ensure SetError is called in SDL_NewAudioStream
- use SDL_bool if possible
- assume NULL/SDL_FALSE filled impl
- skip zfill of current_audio at the beginning of SDL_AudioInit (done before the init() calls)
Pipewire, as of 0.3.22, uses client config files to load modules instead of explicitly specifying them (PW_KEY_CONTEXT_PROFILE_MODULES is deprecated). Use the new method to load the realtime module to boost the audio thread priority.
