Description
For real, glitch-free low latency audio performance, the audio processing path needs to be lock-free. That, in turn, implies no allocations or deallocations. The music-synthesizer-for-android project (which was a testbed for high performance audio on Android) dealt with this by doing fixed-size ring buffers, and preallocating the state for all DSP. That meet the lock-free needs, but is limiting.
Thinking about this more and starting to prototype, I believe the answer is lock-free queues such as Treiber stacks (though a number of other lock-free queue data structures would also work). That way, a main thread allocates nodes at will, sends them over a lock-free queue to the realtime audio thread, which at that point basically has &mut
access to the nodes. For garbage collection, the realtime thread can send them back over a return channel. Control and status can be multiplexed over the same channels (which has the additional benefit of ordering guarantees between these and graph mutation).
This is something of an exploratory issue. Is this a good repo to develop such ideas? The architectural changes are likely to be far-reaching. It may be less work to just develop a new graph runner and modules from scratch than retrofitting the RustAudio stack. But I'd like to at least sound out working with existing projects.