MPC renders 128-frame blocks: 2902 µs per block, per plugin instance. A plugin passes at p99 ≤ 15% and max ≤ 50% of the block (warns up to 35% / 80%), the rule PolyForce uses.
SubForce is one voice, so its cost depends little on the patch: Init (one saw) and the heaviest case (everything on, both busses in Hi range) differ by about 1.5% of the block; only noise, the feedback loop, Analog and Hi-range busses add work, and only while they are up.
make bench-device on the Force (Cortex-A17 at 1.8 GHz, MPC OS 3.9, MPC running; 2026-10-05),
percent of the 2902 µs block, the profile-guided build:
| Case | Phase 0 avg / p99 | 0.0.1 avg / p99 |
|---|---|---|
| idle (no note) | 0.38% / 0.62% | 0.17% / 0.36% |
| Init, one held note | 3.64% / 4.09% | 2.79% / 3.31% |
| heavy patch, Duo (noise and feedback on) | 3.70% / 4.19% | 3.55% / 4.07% |
| heavy, retrig + glide every 50 ms | 3.75% / 4.28% | 3.60% / 4.15% |
0.0.5 (2026-10-07, Analog up in every case: the default 25%, the heavy patches 100%): idle 0.21%, Init 2.85% / 3.21%, heavy 3.71% / 4.10%, both busses in Hi range 4.33% / 4.68%, EG Time LFO and Hi FM 4.06% / 4.45% (avg / p99).
0.0.3 (2026-10-07, the same device and build): idle 0.17% / 0.24%, Init 2.65% / 3.04%, heavy 3.49% / 4.06%, retrig + glide 3.55% / 4.00% — the existing cases unchanged — and two new ones: the heavy patch with both busses in Hi range (on pitch, cutoff, wave and volume, every sample) 4.10% / 4.64%, and with an LFO on EG Time (new envelope coefficients every control step) and a Hi-range FM bus 3.87% / 4.48%.
The heavy patch now runs the original's feedback loop (the mixer's output back into it), which is serial by nature; Init leaves noise and feedback off and pays for neither.
The profiling build puts about 73 µs per block in the voice (Init) and 8–17 µs in the control steps (it reads the clock between stages, so it reads higher than the plain build).
make bench (x86) only proves the bench works; the Force is far slower per sample.
- One voice. Everything PolyForce spends on voice management and vectorising across voices isn't needed.
- No divisions waiting on each other. VFP's divider takes ~18 cycles and accepts one every
~14 (measured on the device); the Phase 0 voice queued 14 of them per sample. Now:
- The ladder (
dsp/ladder.h, 162 → 103 ns a tick): the four stages' nonlinear gains and their 1 / (1 + f t) are four NEON lanes with reciprocal estimates refined by Newton-Raphson; the loop is solved as an affine chain (each stage's output a + b · y3) and closed with one division; inputs, outputs and integrators are one vector step each. - The oscillators take each stretch's length instead of dividing for it; the two divisions left (at a wrap, at a pulse edge) are reciprocals.
- The cutoff coefficients (
exp2andtan) are worked out four at a time for each control run, ahead of the audio. - Measured on this core: a NEON q-register op issues every 2 cycles (a 64-bit datapath), so the serial parts stay in VFP; an all-NEON ladder (a prefix scan) was slower.
- The ladder (
- Idle costs next to nothing: when the amp EG has finished and the output died away, the engine stops rendering, and the control grid only keeps time (glides, busses, drift) until a note wakes it; block sizes still never change the sound.
- Analog costs a random number per cycle (jitter), a bowed ramp, and one low-pass and one high-pass on the oscillators' mix (the same as one per oscillator, a third of the work): +0.2% of a block on the Force; at 0 it is skipped.
- Hi range only where it is on: a Hi-range bus is worked out every sample (its phase, its value, four pitch multipliers at a time with NEON), and only on what needs it at audio rate: pitch, cutoff, wave and volume; everything else, and every other bus, at the control rate.
- Control rate: modulation and glide every 8 samples, gliding in between; no libm calls
(
log2Fast,exp2Fast, the cutoff note and the fixed coefficients worked out ahead), multiplications instead of divisions. - Pay for what is on: the noise filters (pink, dark, the 30 Hz high-pass) and the feedback loop only run while the noise or feedback level is up.
- Short polynomials instead of libm for
exp2,log2,tanandtanh(x)/x; a polyphase IIR decimator (8 multiplies per output sample) instead of a long FIR. - Profile-guided device build, trained under
qemu-armon a spread of patches and the factory presets.
From the test suite (make test prints these):
| What | Measured |
|---|---|
| Worst alias, any wave shape, C4..C7 | −53 dB against the strongest partial (the narrowest pulse the worst) |
| Pulse width swept by a bus | every edge corrected: the largest step between two samples is 1.5 (an uncorrected edge is 2) |
| Worst alias, hard sync | −65 dB |
| Decimator | passband flat to 20 kHz within 0.001 dB, stopband from 24.2 kHz at −85 dB |
| Self-oscillation | past 70% of the knob, as the original's (65%: none; 76%: rms 0.07); tracks the cutoff 15–50 cents flat (110 Hz–7 kHz), as a ladder does |
| Slopes | 5.4 / 10.9 / 16.3 / 21.7 dB per octave between 880 Hz and 1.76 kHz with a 400 Hz cutoff (6/12/18/24 nominal, reached further up) |
| Bass loss | 65% resonance (just under the edge): the passband drops by 13 dB, the ladder's 1 / (1 + r) |
| Multidrive | 0 → 100%: +4 dB louder, much denser; at 50% a triangle's 2nd harmonic at −20 dB (asymmetric, tube-like), none clean |
| Feedback | 0 → 100%: +6 dB, the 9th harmonic +9 dB against the fundamental; no motorboating (an octave under the note: −101 dB) |
| Noise colour | white, pink (−10 dB a decade, Paul Kellet's filter), dark: within 0.3 dB of each other through the open filter |
- 4x oversampling. 2x with polyBLEP oscillators already puts aliasing far below the signal; 4x would double the voice's cost for the ladder's own nonlinear harmonics, which the halfband removes above 24 kHz anyway. The device bench leaves room (3.6% of a block with everything on); a quality switch can come if listening shows a need.
- Oscillator 2 as a bus source (FM that tracks the keys): Hi range covers audio-rate modulation up to 1 kHz, as the original's LFOs do; oscillator 2 as a source would need its output per 2x sample on the pitch of oscillator 1. Not yet.
- Compensating the bass loss. The ladder's thinning with resonance is kept, as on the original; Multidrive and the mixer make up for it.
- Hand-written assembly for the ladder. GCC moves some vector lanes through core registers; a hand-scheduled ladder might save another ~30 ns a tick (~8% of the voice). Not worth giving up the one C++ source the x86 tests also run.