@Ronflaix no timers (those are used for grayscale!) but DMA yes.

Every time the DMA finishes sending bits, it triggers an interrupt which will then generate a new PCM buffer and also convert that into the PWM bitstream.

No mixing at all, just single stream, but you could totally mix if you wanted too.

@Ronflaix ehh, well, there's a lot of interrupts, but generally speaking, it's not too bad. the CPU is suuper fast, you have lots of time to do most things (as long as you don't do soft-float calculations :P).

If you were really strapped for CPU time, I guess the routines could also be optimized a bunch. The audio samples and the big PWM lookup table are all just stored in the code right now, which I'm guessing will load them into the CPU d-cache at every firing of the interrupt.

@Ronflaix
In other words, every few milliseconds your actual main loop will get interrupted, the cache will be overwritten with audio data and then it'll return execution.

That could (and maybe should?) be benchmarked and probably fixed :)

@manawyrm The CPU might be super fast but only its cache and the DSP's memory banks are as fast as that, sadly. (Still haven't opened the repo, sorry) I suppose you're not really using Azure and you're using that free extra cache for shenanigans, aren't you? :D

(I'm using Azure as my own "equivalent" failed to follow up on performance as I was starting to add text for *that* project and I lacked patience, time and tooling to investigate the innards of the CPU.)