MPC renders 128-frame blocks: 2902 µs per block, per plugin instance. A plugin passes at p99 ≤ 15% of the block (warns up to 35%). EffectForce's ceiling is every module on at its heaviest settings with every matrix slot busy: what the worst preset could reach.
make bench-device copies the .so and efbench (tools/bench.cpp) to the Force, runs them pinned
to core 1 while MPC keeps running, and deletes them. Each case feeds noise in place through the
plugin's own entry points and times every block with the thread's CPU clock.
2026-10-06, the Force, the ten modules and the performance layer, the profile-guided build (GCC 13, local; glibc 2.38), percent of the 2902 µs block:
| Case | avg | p99 | max |
|---|---|---|---|
| Everything off | 0.27 | 0.58 | 0.87 |
| Drive / Filter / EQ / Comp alone | 1.37 / 0.76 / 0.68 / 0.58 | 1.79 / 1.08 / 1.08 / 0.89 | 2.13 / 1.42 / 1.31 / 1.20 |
| Chorus / Phaser / Pulse alone | 0.89 / 0.86 / 0.42 | 1.30 / 1.21 / 0.77 | 1.66 / 1.79 / 1.00 |
| Grain / Delay / Reverb alone | 2.09 / 0.94 / 2.42 | 2.73 / 1.42 / 3.20 | 2.94 / 1.68 / 3.60 |
| Everything on, heaviest, 8 mod slots | 11.97 | 14.90 | 16.37 |
| Looper rolling, fader sweeping (scene 2: four sends, a dozen locks) | 6.01 | 7.70 | 8.27 |
| That, everything on heaviest | 13.40 | 16.94 (WARN) | 18.98 |
| A scene holding 14 locks still (six modules on) | 4.50 | 5.78 | 6.64 |
| The same scene moving all 14 every chunk (a 1-beat Loop, 7 on a log curve) | 5.43 | 7.08 | 8.30 |
| That, everything on heaviest | 13.32 | 16.56 (WARN) | 18.63 |
Everything on at its heaviest passes with nothing to spare (p99 14.9 of 15; the first eight modules had 11.3, below), and the performance layer on top of it goes past (16.6-16.9): a warning, not a failure, and the worst a patch can reach, not what Perform presets do (6-8). A scene move costs what any moving parameter costs, mostly the modules working out their coefficients again: 14 locks moving every chunk add 0.9 points on average and 1.3 at p99 to the same scene held still. An effect of the FX library moves one to three over bars. Run to run, the device's p99 moves by about 0.5.
2026-10-05, the Force (Cortex-A17), a local build (GCC 13, not profile-guided), percent of the 2902 µs block. "Alone" includes the plugin's own overhead (the first row).
| Case | avg | p99 | max | Design target (the module's share) |
|---|---|---|---|---|
| Everything off | 0.33 | 0.62 | 0.89 | |
| Drive alone (Tube, 24 dB, 80% mix: both halfband pairs) | 1.44 | 1.85 | 2.19 | ≤ 1.5 |
| Filter alone (LP 24, resonant, drive, spread) | 0.72 | 1.07 | 1.25 | ≤ 0.8 |
| EQ alone (every band) | 0.71 | 1.01 | 1.34 | ≤ 0.8 |
| Comp alone (compressor, sidechain cut) | 0.66 | 1.00 | 1.27 | ≤ 0.8 |
| Chorus alone (Ensemble) | 0.95 | 1.30 | 1.53 | ≤ 0.8 |
| Phaser alone (Phaser 8) | 0.94 | 1.32 | 1.51 | ≤ 0.8 |
| Delay alone (ping-pong, wow, drive, ducking) | 0.98 | 1.40 | 1.73 | ≤ 0.8 |
| Reverb alone (Hall, 6 s, full modulation) | 2.58 | 3.30 | 4.46 | ≤ 4 |
| Everything on, heaviest (Drive Fold, OTT, Phaser 12, Space reverb), 8 mod slots | 9.10 | 11.31 | 13.32 | p99 ≤ 15 |
Every case passes. Net of the overhead, every module is within (or, for Chorus and Delay, about
0.1-0.3 point above) its share of the design budget; the total, which is what MPC sees, has 3.7
points to spare. These are the first eight modules: Pulse, Grain and the shimmer came after, and
their device run is still to do (Roadmap); until then, the instruction counts below.
The release build is profile-guided (as PolyForce's and SubForce's: make arm-plugin with qemu-arm,
and CI), worth about 2% here: the modules are hand-tuned already.
Without the device, ARM instructions per 128-frame block are the measure: the bench's cases
(tools/bench.cpp) through the plain .so (GCC 13, the device flags) under a plugin-enabled
qemu-arm, per processReplacing call. On the first eight modules' everything-on case, 328k
instructions were 9.1% of the block on the Force, Hall alone 98k net for 2.25%: about 36k-44k
instructions to a point of the block. They miss what the Force adds (cache misses, VFP divisions and
flag transfers that stall), which the device's p99 over its average (1.24) shows.
2026-10-05, the ten modules, thousands of instructions per block:
| Case | At the hand-off (avg / p99) | Now (avg / p99) |
|---|---|---|
| Everything off (the plugin's own work) | 17.8 / – | 9.8 / – |
| Pulse alone (Gate) | 25.1 / 25.3 | 17.1 / 17.3 |
| Grain alone (Cloud, densest, +12, feedback) | 97.2 / 108.1 | 72.1 / 79.8 |
| Reverb alone (Space, full modulation, shimmer 0.6) | 140.7 / 284.6 | 120.0 / 146.1 |
| Everything on, heaviest, 8 mod slots | 446.5 / 592.7 | 402.5 / 431.3 |
Everything on is about 9-11% of the block on average by that measure. At the hand-off its p99 was the shimmer's splice search (a correlation of a few thousand multiply-adds in one sample, every 80 ms: one block in 28, a p99 of over 16% before the device's own jitter); now the search sums sixteen lags at a time in vector lanes, and Grain, the rack's per-sample overhead and the control paths are leaner too (the commits of 2026-10-05 say what and by how much). Most of the changes leave the output bit for bit as it was; four move it by float rounding, the compiler fusing multiply-adds differently: the shimmer's shifter reads and its path after the network loop (-130 dB), Grain's read segments aligned to time since reset (-100 dB) and the modulation of log-curve targets by the fast exp2 (1e-7 of the value).
- A module that is off costs nothing: the rack skips it entirely.
- Control rate: parameters, modulation and coefficients are worked out once per 32-sample chunk; every module glides across the chunk.
- NEON where the work is parallel: left and right as two lanes (Filter, EQ, Comp, Delay), the halfband filters of both sides in one vector (Drive), four chorus voices per load, the reverb's eight lines as two vectors, two of a grain's samples (left and right each) per vector, the shimmer's splice search sixteen lags at a time.
- No VFP compares per sample where NEON lanes do: VFP has no min or max, and moving its flags to the core stalls; limiters, envelope followers and gates keep their state in a NEON lane.
- No libm per sample: SubForce's fast math (
dsp/fastmath.h) and per-chunk coefficients; what needs libm (a tail length, a pitch's speed, a loop gain) is worked out again only when what it depends on changes. - Settled is free: a gain or mix that isn't moving runs as a constant, four samples at a time (and a unity gain not at all).
- Cheap resets: a module switched back on doesn't clear megabytes of delay line on the audio thread; the Delay and the Reverb count what they have written since and read older samples as silence.
- Cheap getters: MPC polls names and texts hundreds of times a second; they read caches.