Conversation
audio_out is a format factory; PaceOutput.write returns immediately and exposes space(); I2S polls on service. FACTORY_ROLES bind those callables instead of constructing on getattr.
pcm_out/pcm_in return raw PCM devices; audio_out returns an AudioOut sample player. That holds on hosts and on boards alike, which it did not before: audiodev.auto.audio_out handed back a PCMOutput while board_peripherals .audio_out handed back an AudioOut, so an app that could reach either wrote getattr(pcm, 'transport', pcm) to find out which it held. Both spellings in audiodev.auto are renamed rather than aliased -- an alias would keep two names for one thing, which is the quirk being removed. There is deliberately no audio_in. Output has a player layer; input does not, and audioif has no input-side node to give it one. A name implying otherwise is the same vagueness that let audio_out mean two things. pwm_tone's factory becomes tone_out for the same reason: it returns a ToneOutput, so calling it pcm_out would have relocated the lie rather than fixed it. Boards now declare an AudioCapability and audiodev owns negotiate(), instead of each board hand-writing _require_format() with its own error text. The capability splits channels (what the wire opens) from native_channels (what reaches a transducer): the P4's ES8311 clocks two slots into one speaker, so 44.1k stereo opens I2S.STEREO and nothing is mixed down. I2SWire publishes the pins so a consumer that drives the peripheral itself -- usbif's C pump -- can stop reaching into a board module's private names. The audiomath import PR #31 added to audiodev/__init__.py is gone; the remix implementation is injected instead, defaulting to the portable one, with audiodev.accel as the opt-in seam for the C path. A test now asserts that layering by walking the AST, because nothing caught it the first time. boarddev resolves host-vs-board, which audiodev cannot do without importing board configs and losing its standalone-ness. It resolves to board_peripherals, never board_config -- importing that on the P4 starts MIPI DSI and stalls the VM. Suite: 434 tests, green.
Brad asked that P4 changes stay on a branch until he has heard them. This is that branch. Everything measurable without ears is measured; what is left is named below. Builds on PR #31, keeping its measured work (the pace factory, the POLL ioctl that arms I2S non-blocking, the es8311 clock-tree change) and replacing its hand-written _require_format with the shared capability and negotiate(). That resolves a contradiction PR #31 carried: _MAX_CHANNELS = 1 sat beside an audio_out that accepted 2 channels, while audio_in raised on the same request -- one constant asked to mean two different things. Now AUDIO_OUT declares channels=(1,2) with native_channels=1 (the ES8311 clocks two slots into one speaker, so stereo needs no mixdown) and AUDIO_IN declares channels=(1,), and each does what it says. New role audio_power(enable, volume=): brings the analog path up with NO I2S stream opened. usbif's C pump opens I2S itself from AUDIO_OUT.wire, and two owners of one I2S channel is a silent failure, not an error -- but the codec still has to be powered or every byte moves inaudibly. A defect the tests caught: pcm_in was ignoring negotiate()'s verdict, so a caller asking for stereo capture on a mono ADC got a mono stream and no error. There is no input-side adapter, so it raises now. Fixed on all five I2S boards, since all five had it. Measured on the board (non-destructively, under /tst with sys.path ahead of /lib, so Brad's spremote setup is untouched): PHASE_A 24000 mono measured 23995 err -1 permille PHASE_A 44100 stereo measured 44137 err 0 permille plus the capability and wire readable without opening, and pcm_in/pcm_out refusing stereo capture, 32-bit and 3 channels with readable errors. NOT VERIFIED: that it sounds right. The P4's output and input share I2S(0) and one set of clocks under AudioSession(duplex=False), so it cannot run the acoustic gate that the T-Embed can. Drain timing proves the ESP32 clocks at the rate asked; it cannot prove the ES8311's internal dividers agree, and a wrong REG06 clocks correctly and sounds like garbage. Suite: 444 tests, green, stable across four runs.
pcm_out/pcm_in/audio_out, boards declaring an AudioCapability with audiodev.negotiate() holding the policy, and channels (what the wire opens) split from native_channels (what reaches a transducer). Verified on three boards across three ESP32 families -- T-Embed acoustically via its own microphones, and the QT Py ESP32 Pico with an Adafruit Audio BFF by ear. Rates 8 kHz to 48 kHz, mono and stereo, all within 1 permille. Supersedes PR #31, keeping its measured work.
Held on a branch until Brad heard it, which he has: two passes of the same arpeggio at 24 kHz mono and 44.1 kHz stereo, same pitch and clean. That was the one thing no measurement here could establish -- drain timing proves the ESP32 clocks at the rate asked but cannot prove the ES8311's own dividers agree, and 44.1 kHz stereo is a rate this board had never run at. Also resolves the contradiction PR #31 carried, where _MAX_CHANNELS = 1 sat beside an audio_out accepting 2 channels while audio_in raised on the same request.
|
Superseded and closed in favour of what landed on What was kept
What changed, and why
The VerifiedThree boards, three ESP32 families. ESP32-P4 by ear (24 kHz mono and 44.1 kHz stereo, same pitch, clean); T-Embed acoustically through its own microphones via Two gaps recorded rather than papered over: #34 (nothing here has a stereo transducer, so channel order is untested) and PyDevices/usbif#12 (the C pump hard-codes 512fs MCLK and cannot exceed 32 kHz). Full write-up: |
1 participant
Summary
board_config.audio_out/board_peripherals.audio_outis a format factory (FACTORY_ROLESbind the callable, not a constructed object). P4 Connect opens I2S at the app rate, including stereo slots.PaceOutput.write()returns immediately, copies so the caller may reuse a buffer, caps the stash withqueue_ms, and exposesspace()so a VM pump can gate on DMA room instead ofsleep_ms.service()/write(); WASAPIwin_audio.audio_outtakes anAudioFormat. ES8311 advertiseschannels = 1.wifi.connectsyncs NTP after STA is up (Connect TLS).Test plan
SDL_VIDEODRIVER=dummy SDL_AUDIODRIVER=dummy python3 -m unittest discover -s tests -p 'test_audiodev.py'(43 tests)python3 -m unittest discover -s tests -p 'test_boarddev.py'python3 -m unittest tests.test_es8311 tests.test_esp32_p4_audioaudio_out(AudioFormat(44100, 2, 16), latency="low")then Connect drain;latency="low"must not shrink the output ibufboard_peripherals, notboard_config