audiodev.accel
Optional C accelerators for audiodev, from audiodsp when it is present.
This module exists to be a choice. audiodev/__init__.py and every
transport backend import no audiodsp at all – that is what lets a headless
PCM consumer (a Spotify Connect speaker, a USB audio pump) run on firmware
with no DSP package in it. But the pure-Python channel remix in
audiodev is a per-sample Python loop, which is fine for a 10 ms chunk on
a desktop and much too slow for 44.1 kHz stereo on an MCU.
How slow is slow. Measured on a QT Py ESP32 Pico (240 MHz), converting a 10 ms chunk of stereo to mono with the portable implementation:
16000 Hz 6.3 ms 0.63x realtime usable 24000 Hz 9.4 ms 0.94x realtime marginal, no headroom for an app 44100 Hz 17.3 ms 1.73x realtime cannot feed a live stream
End to end, pushing real audio through the whole chain on the same board, 16 kHz stereo-to-mono sustains realtime (1778 ms of wall clock per 2000 ms of audio) and 44.1 kHz does not (4363 ms, 2.2x over).
So the common claim that pcm_out needs no audiodsp in firmware holds only
while the format matches the wire – which is the usual case, because a board
whose wire takes two slots needs no remix at all. A board that must genuinely
mix down at a high rate needs the C implementation, and therefore needs
audiodsp built in. Offline or low-rate work is fine on the portable one.
So the fast path is offered here rather than reached for there. A caller that has audiodsp asks:
from audiodev.accel import best_remix
pcm = adapt_channels(pcm, fmt, remix=best_remix())
and a caller that does not, does not import this module. The dependency
arrow never inverts, and nothing silently degrades: best_remix() returns
None when no C implementation is available, which adapt_channels
reads as “use the portable one”.
audiodev.sample_out already requires audiodsp (it pulls
audiocore.get_buffer), so it may use this freely.