audio: apply mix and decay in fixed point, not floating point

[peterbay]

Code written by Claude Code, guided and corrected by @peterbay.

The problem

An audio effect reads its mix, decay or level once per block from a BlockInput, and those values do not change again until the next block. Four of the callbacks were nonetheless applying them to every sample in floating point.

On a part with a floating point unit that is a couple of instructions. On one without — RP2040, SAMD21, the RISC-V ESP32s — every one of them is an out-of-line call into a soft-float routine, and it happens inside a callback that has to keep up with the sample rate. nm -u on the built Mixer.o shows U __divsf3, which in an RP2040 firmware resolves to 0x40002274: the ROM routine, a real unpack, divide and round.

audiodelays/Flanger, audiofilters/Phaser and audiofreeverb/Freeverb already avoid this — they convert their parameters to fixed point once per block and keep the sample loop in integers. The five files this PR touches did not.

What this does

• Each block-constant parameter is converted to Q15 where it is already being read, and the per-sample arithmetic becomes an integer multiply and a shift. No behaviour is added or removed; the same values are applied to the same samples.

• The scale is 1 << 15, which is exactly unity. audiomixer had been dividing by 32767, and its level reaches mult16signed as a float in 0.0–1.0 scaled by 1 << 15, so unity became a gain of 32768/32767 — slightly above one. A sample of 32766 came back as 32767, and full-scale samples clipped. The portable branch also disagreed with the ARM branch it exists to mirror. Both are now exact.

• The effects that already use fixed point are left alone. They scale by 32767, which errs the other way — a nominal unity mix is 32767/32768 — and by less. Changing them is a separate question from removing the soft-float, so this PR does not.

What it costs and what it saves

Measured on a Waveshare RP2040-GEEK, two channels at 22050 Hz, 512-byte buffers, PWM output. A fixed Python loop is timed with the effect feeding the output and again with it stopped; the shortfall is the callback's share of the core. Three rounds each, spread within a round under half a point:

effect	before	after
audiodelays/Echo	57.5 %	18.1 %
audiodelays/MultiTapDelay	36.9 %	10.6 %
audiofilters/Filter	29.4 %	10.5 %
audiodelays/Chorus	15.3 %	9.4 %

audiomixer was measured separately, one 16-bit stereo voice at 44.1 kHz with I2S running throughout: 4.07 % of the core before and 1.81 % after, five rounds each, ranges 3.93–4.64 % and 1.25–2.31 %. Its compiled function went from 5394 to 4170 bytes, which matters on its own — it was the largest in shared-bindings and shared-module against a 16 kB instruction cache.

Soft-float calls remaining in each callback, counted in the built objects. What is left is the once-per-block setup, not per-sample work, which is why MultiTapDelay loses only three calls and still runs three and a half times lighter:

	before	after
Echo	36	14
Filter	16	6
MultiTapDelay	16	13
Chorus	12	11

Firmware size is 32 bytes smaller on that build. No new translatable strings.

Accuracy

Checked exhaustively on the host against the floating point it replaces.

For audiomixer, over all 2 147 549 184 (sample, multiplier) pairs the mixer can produce — multiplier 0–32768, sample -32768–32767: every product stays inside int32, the new result never differs from the old by more than one LSB, and the old code lands on a rail three times as often as the new one.

For the effects, over 65 601 536 (sample, coefficient) combinations: at most one LSB per term, two for a dry/wet sum, which is the difference between truncating toward zero and shifting — about −84 dB.

One thing noticed and not verified

audiomixer's ARMv7EM branch does lomul <<= 16 on an int32_t that reaches 32768 at level 1.0, which shifts into the sign bit. Modelling smulwb and ssat on the host makes that branch return -sample there. I have no M4 or M7 board with an audio setup to check it on, so I am only mentioning it — it is not addressed here and it is not what this PR changes.

Testing

Waveshare RP2040-GEEK, built with CIRCUITPY_AUDIODELAYS, CIRCUITPY_AUDIOFILTERS and CIRCUITPY_AUDIOFREEVERB enabled. Each effect is constructed, played through audiopwmio.PWMAudioOut into two unused pins, and torn down before the next, with the timing loop described above run three times on each side of the switch. audiomixer was tested on a Seeed XIAO nRF52840 Sense with an Adafruit Audio BFF over I2S.

[FoamyGuy]

Thanks for working on this.

There are some cases where rounding behavior has changed, and a few spots where an existing helper synthio_sat16() can be used.

[FoamyGuy]

This should use synthio_sat16(). The old code rounds toward zero, but the new code always goes toward -infinity. The helper should resolve it.

delay_word = synthio_sat16(delay_word * decay_scaled, 15) + sample_word;

[FoamyGuy]

Suggested change

	word = (int16_t)((echo_buffer[(j % echo_buf_len) + echo_buffer_offset] * decay_scaled) >> 15);
	word = synthio_sat16(echo_buffer[(j % echo_buf_len) + echo_buffer_offset] * decay_scaled, 15);

[FoamyGuy]

Suggested change

	word = (int16_t)((echo * decay_scaled) >> 15);
	word = synthio_sat16(echo * decay_scaled, 15);

[FoamyGuy]

Suggested change

	word = (int16_t)((echo * echo_scaled) >> 15);
	word = synthio_sat16(echo * echo_scaled, 15);

[FoamyGuy]

Suggested change

	word = ((echo * decay_scaled) >> 15) + sample_word;
	synthio_sat16(echo * decay_scaled, 15) + sample_word;

[FoamyGuy]

Suggested change

	int32_t dry_scaled = (int32_t)((MICROPY_FLOAT_CONST(1.0) - mix) * MICROPY_FLOAT_CONST(32768.0));
	int32_t dry_scaled = 32768 - wet_scaled;

Both wet and dry truncate toward zero. Could cause off by 1 error when they're summed.

[FoamyGuy]

This section has different behavior rounding always down instead of toward zero. It should use the new helper function.

Add here:

int32_t filtered_word = self->filter_buffer[k + SYNTHIO_MAX_DUR * buf_offset];

[FoamyGuy]

Suggested change

	word_buffer[i + j] = synthio_mix_down_sample(((sample_src[i + j] * dry_scaled) >> 15) + ((self->filter_buffer[k + SYNTHIO_MAX_DUR * buf_offset] * wet_scaled) >> 15), SYNTHIO_MIX_DOWN_SCALE(2));
	word_buffer[i + j] = synthio_mix_down_sample(synthio_shift_round_to_zero(sample_src[i + j] * dry_scaled + filtered_word * wet_scaled, 15), SYNTHIO_MIX_DOWN_SCALE(2));

[FoamyGuy]

Suggested change

	hword_buffer[i + j] = (int8_t)(((sample_hsrc[i + j] * dry_scaled) >> 15) + ((self->filter_buffer[k + SYNTHIO_MAX_DUR * buf_offset] * wet_scaled) >> 15));
	hword_buffer[i + j] = (int8_t)synthio_shift_round_to_zero(sample_hsrc[i + j] * dry_scaled + filtered_word * wet_scaled, 15);

[FoamyGuy]

Suggested change

	hword_buffer[i + j] = (uint8_t)((((int8_t)(((uint8_t)sample_hsrc[i + j]) ^ 0x80) * dry_scaled) >> 15) + ((self->filter_buffer[k + SYNTHIO_MAX_DUR * buf_offset] * wet_scaled) >> 15)) ^ 0x80;
	hword_buffer[i + j] = (uint8_t)synthio_shift_round_to_zero((int8_t)(((uint8_t)sample_hsrc[i + j]) ^ 0x80) * dry_scaled + filtered_word * wet_scaled, 15) ^ 0x80;