Repository navigation
Fix audiomixer phase inversion and use DSP on Cortex-M33 - #11617
Merged
Merged
Conversation
On cores with DSP instructions, a level of 1.0 (1 << 15) shifted left by 16 overflowed into the sign bit, so the mixed sample came out negated: mono output was inverted and the right channel of a centered stereo voice was out of phase with the left. Shift by 15 instead. A centered pan now scales both channels by 1.0 instead of the left one by 32767/32768.
The SIMD helpers were limited to ARMv7E-M, so the Cortex-M33 in the RP2350 fell back to a float divide per sample. Check __ARM_FEATURE_DSP instead, and mix with integers where there is no DSP, which matches the DSP result bit for bit.
bdbarnett
added a commit
to PyDevices/micropython-pydevices
that referenced
this pull request
Oct 22, 2026
…, rebase the ESP-IDF components patch (#173) alpha.2 fixes the audiomixer level-1.0 rounding upstream (adafruit/circuitpython#11617: it does the multiply in integers, shifted by 15, so a level of 1.0 is exact), so patch 0007 is removed; the series skips its number. The espressif port moved to ESP-IDF 6.1 and rewrote its component list and Makefile, so 0006 (the ESP-IDF components a user C module asks for) is rebased onto it. The other patches apply unchanged.
bdbarnett
added a commit
to PyDevices/audiodsp
that referenced
this pull request
Oct 22, 2026
* Take CircuitPython 11.0.0-alpha.2's integer mixer and inline synthio_sat16 audiomixer (11a329885, fc57c0f36; adafruit/circuitpython#11617): a voice is scaled as (ai * mul) >> 15 where it was ai * (mul / 32767.0f) truncated toward zero, and a centred pan is 32768 - panning where it was 32767 - panning. Taken in upstream's words in src/audiomixer/Mixer.c, in the CPython extension's scaling kernel and in src/cpython/audiomixer.py, whose float32 helpers go. Every level below 1.0 moves by up to one LSB, and a small negative sample now rounds down (-1 at level 0.5 stays -1). A voice at level 1.0 is a wire without help, so the unity branch this port added in audiodsp#95 is deleted, and a stereo voice at pan 0 no longer has its left channel one step below the right. The CPython target's block-level pass-through at unity is keyed on Q15 32768 alone now: 32767 no longer multiplies by exactly 1. synthio (cbbde1208; adafruit/circuitpython#11621): synthio_sat16 is inline in src/synthio/__init__.h as upstream has it, with __ssat on cores that define __ARM_FEATURE_SAT. The rounding is the same arithmetic without a branch, so no sample changes. mixer_q15_level_probe.py replaces mixer_level_precision_probe.py, whose material was the float32-versus-float64 flips that no longer exist. It renders the rails and the values either side of zero at four levels and five pans, mono and stereo, and verify_dsp.py holds CircuitPython, MicroPython and CPython to the same bytes. test_cpython_audiomixer_level.py and test_cpython_mixer_scale.py are rewritten around the integer arithmetic. * Move the oracle to CircuitPython 11.0.0-alpha.2 and re-capture two digests CIRCUITPYTHON_ORACLE names d53b2277c, and tests/test_voice_ceiling_consistency.py pins the sha256 of the 64-voice binary built from it (ee0cc93c...). Two stored digests moved with the integer mixer, re-captured from the new oracle's own output, which CPython and desktop MicroPython reproduce: - streaming_component, streaming half: f3ffad8e... to 33b27a2f.... The mixer voices in streaming_component_probe.py are below unity. - synthtools_acceptance: b6e3b06f... to 757d944a.... The lead, bend and bass lines moved; the other five lines are unchanged. Every other stored digest held and reproduces on the new oracle: midi_component, effects_component, biquad_component and mixdown_knee. The oracle labels in effects_component and streaming_component follow the pin. * Record the alpha.2 resync and retire "a voice at level 1.0 is a wire" upstream-sync.md gets the new "last resync" entry (the alpha.1 one becomes the one before it), upstream-diff.md says what upstream's integer mixer changed and which digests moved, and the audiodsp#95 section becomes a record of a difference upstream now shares, as the alpha.1 resync did for its five. * Print the Q15 probe's levels by name A float32 MicroPython build prints 100 / 127 with fewer digits than CPython, so the probe's label line differed where the samples agreed.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
On boards with DSP instructions (SAMD51, nRF52840, i.MX RT, STM32F4), a voice at level 1.0 came out negated. The level is 1 << 15, and the code shifted it left by 16, which overflowed into the sign bit. Mono output was inverted. With centered stereo the right channel was out of phase with the left, so it cancelled out when the two channels were summed to mono.
Speed: The DSP code ran only on ARMv7E-M. The RP2350 has the same DSP instructions, so it now uses them too. Chips without DSP, like the RP2040, now mix with integers instead of float. The output is the same on every chip.
Time to mix 4 voices, 1024 samples:
Flash is 1.1 KB smaller on RP2040 and 1.8 KB smaller on RP2350.
Testing: I read the mixer output with audiocore.get_buffer() on all four boards.
The code was written with AI help.