Pappus new

pappus

scatter to the wind

Two euclidean granular swarmers + resonant filterbank + multitap delay + colour + scene morphing, with modular routing and modulation.

A very deep sound design tool and soundscape instrument.

K2 back                  K3 forward
K2 & K3                  change lane
long K2                  reset parameter & clear modulation
long K3                  lock / freeze / hold module
E1                       select parameter
E2 value                 E3 sub-value / fine tune

Grid compatible.

Designed by Michael Manning. Inspired by Torso S-4.

Heavy LLM usage disclaimer. 100% Claude code.

---

requirements

norns. A grid is optional — everything is reachable without one.

install

;install https://github.com/FoundSoundsMM/pappus

the chain

GRAINSWARM 1 \
              >  RESONATOR  >  DELAY  >  COLOUR  >  REVERB  >  out
GRAINSWARM 2 /

FULL and LITE

The script asks what it is running on and builds itself to fit. A factory norns is a Raspberry Pi 3 compute module; a shield is whatever is underneath it, which since 2019 is usually a Pi 4. The whole instrument measures comfortable on a Pi 4 and lands at ninety per cent plus, with audio dropouts, on a Pi 3 — so on a Pi 3 it builds a smaller instrument instead.

Both halves run the same test, off /proc/device-tree: lite_mode in pappus.lua and prLiteMode in lib/Engine_Pappus.sc. bcm2711 (Pi 4, 400, CM4) and bcm2712 (Pi 5, CM5) get FULL; any other Raspberry Pi gets LITE. A machine with no device tree at all — a desktop running the tests — gets FULL, because unknown is not the same as slow.

PARAMS > PERFORMANCE > ENGINE shows what it decided and lets you disagree. FULL is there for a Pi 3 owner who wants both granulators and will pay for them; LITE is there for a Pi 4 owner who is running out of headroom anyway. It takes effect on the next load and has to: the engine is one SynthDef, and a SynthDef's topology is fixed the moment it is compiled. The choice is written to mode.txt in the script's data folder, which both sides read at load.

PAPPUS_LITE=1 forces LITE for the test harness, on either side.

what LITE gives up

	FULL	LITE
granulators	2	1 — GRAINSWARM 2's two pages are gone, and its box on SIGNAL with them
RESONATOR MODAL	48 resonators, six partials a voice	24, the first three of each voice
RESONATOR STRING	two combs a voice	one — STRUCTURE's dispersion is MODAL-only
LOSS	a real short-time Fourier transform	two cascaded lowpasses — the bandwidth half of it, not the holes
loop files in audio/	one PlayBuf each, always running, twice over	one shared player per bank, pointed at the chosen file
REVERB diffusion	four allpasses a side	two
capture buffers	4 × 60 s	2 × 60 s — 23 MB back
screen / grid / mod fps, screen detail	25 / 25 / 60 / FULL	15 / 15 / 30 / LITE by default

That is 1706 UGens against 2780, and the ones that went are the expensive ones — thirty-two GrainBufs, twenty-four modulated resonators, the transform and eight always-running stereo buffer reads.

The n_* parameters, commands and psets all still exist on LITE. A pset or a snapshot written on a Pi 4 loads on a factory norns without complaint; the second granulator's settings are simply carried and not used, so the same file comes back whole on a Pi 4.

performance

PARAMS > PERFORMANCE.

These four are the hand controls, and on LITE they already start at their cheaper end. On a norns the screen and the encoders are served by the same thread, so a frame that takes too long does not merely drop a frame — it holds the encoder queue shut, and what you feel is a knob that does nothing and then jumps several steps at once. A shield with a Pi 4 in it has the headroom to draw these pages at twenty-five a second; a factory norns, on a Pi 3, does not always.

• screen fps — 25 (default), 20, 15 or 10. Nothing about the instrument changes: every animation advances by elapsed time, so at 15 the same motion is drawn less often rather than more slowly. Fifteen is still smooth for what is on these pages and is a forty per cent cut in everything the screen costs.
• grid fps — 25 (default), 15 or 10. Separate because the grid is a different bottleneck: not cairo, but up to 128 LED writes and a serial frame. The display already skips a refresh entirely when nothing on the grid moved; this caps how often it may send one when things are moving.
• screen detail — FULL (default) or LITE. A different economy from screen fps: fps is how often a page is drawn, this is how much of it there is to draw, and dropping to 15 while keeping full detail leaves every frame as expensive as it was. LITE takes the antialiasing off COLOUR's sheet and PAPPUS's filaments, samples the sheet at twenty points across instead of thirty, and draws the GRAINSWARM waveform in two-pixel columns rather than one — taking the louder slot of each pair, so a transient is still on the screen rather than averaged away. Cells, headers, values, numbers and the grid are untouched. On the two heaviest pages that is about a quarter of the geometry, before counting what the antialiasing was costing on its own.
• mod fps — 60 (default) or 30, and this one is not free. The modulators run on their own metro, faster than the display, because an LFO stepped at the screen's rate is visibly stepped; sixty a second is eight LFOs advanced, sixteen routings accumulated and every touched destination re-sent, on the same thread as the screen and the encoders. Halving it halves that. What you give up is the top of the RATE knob: the ceiling tracks the update rate at a fifth of it — past which a sine stops being a sine and becomes a staircase — so at 30 the fastest an LFO can run drops from 12 Hz to 6. Everything below is unchanged, the RATE cell shows the resulting Hz either way, and an LFO already set above the new ceiling is pulled down to it rather than left aliasing.

If you are seeing high CPU or hearing clicks, first check PARAMS > PERFORMANCE > ENGINE says LITE — and the other things worth knowing:

• audio/ is scanned at load and every .wav in it becomes a NOISE type and a RESONATOR GRAIN type. Every one of them is held in memory and played continuously for the whole session, whichever is selected — so a folder full of long loops costs both resident RAM and memory bandwidth all the time.
• Every capture buffer is sixty seconds, four of them, whatever BUFFR is set to. That is fixed at load, not per patch.
• LOSS is the single most expensive control in the engine — it is a real short-time Fourier transform — but it costs the same whether the knob is up or down.

changing the engine

lib/Engine_Pappus.sc is one SynthDef sitting close to a hard limit: scsynth allocates a fixed pool of 64 audio interconnect buffers, and a def needing more is rejected at load. On norns that failure is silent — the class loads, every engine command is accepted and goes nowhere, Lua raises nothing, screen and grid work perfectly, and there is no audio at all.

test/wirecount.py models that allocation statically and undercounts: it scored a def at 63 that the real server refused. So before shipping any engine change, load it into an actual scsynth:

python3 test/defload.py            # ...and --loops N to fake N files in audio/

It loads both graphs, FULL and LITE, because "the def loads" is two claims now and the one that is easy to forget is the one that only ever runs on somebody else's factory norns. --full or --lite for one of them. Every other harness in test/ takes H.LITE = True (or build(lite=True)) the same way, and the Lua tests take PAPPUS_LITE=1.

And because the engine and the script have to name the same machine — one deciding whether GRAINSWARM 2 makes a sound, the other whether it has pages — both are put to the same eleven real device-tree strings:

python3 test/litemode_test.py

It needs SuperCollider installed, runs offline in NRT, and takes about a minute.