DigiFilter

Extra per-track filter modes for the Elektron Digitakt mk1 (OS 1.53).

A format-2 elekloader mod, id digifilter. It adds six filter types on top of the stock OFF / LP / HP / EQ:1..5:

TYPE	mode	what it is
8	BP	band pass (state-variable filter)
9	BP2	a wider band pass (about half the resonance)
10	COMB	feedback comb, delay from FREQ, feedback from RESO
11	TRASH	second comb at half the delay (a higher harmonic series)
12	OB4	the Oberheim OB-1's 4-pole lowpass ladder, self-oscillating at the top of RESO
13	OB2	the OB-1's 2-pole output: 12 dB/oct, resonating from the same 4-pole loop

Each mode runs per track and uses that track's own FREQ, RESO and filter envelope, exactly like the stock types. Stock TYPE values keep running the stock filter — nothing here reimplements a stock mode.

The FLTR page redraws its FREQ/RES response curve for the new types (a band-pass bell for BP/BP2, comb teeth for COMB, a trash can for TRASH) and shows a small fixed version of each shape in the TYPE box to the right of the graph. The screenshot shows the page for every type, top to bottom — LP (stock), BP, BP2, COMB and TRASH:

OB4 and OB2 are modelled on the OB-1's VCF: a ladder of four OTA stages with the resonance fed back from the fourth, saturating where the circuit does. The FLTR page draws their lowpass response; top to bottom, LP (stock), OB4 and OB2. How they were made and checked (SPICE, a model, the fixed-point kernel, digiemu, a unit) is in docs/ob1.md.

Their FREQ control now follows the measured reference curve, from 10 Hz to 21 kHz with cubic travel. Resonance starts oscillating between RESO 80 and 81. Existing OB4/OB2 Sounds therefore change at the same knob settings; their saved format is unchanged. See the calibration and its limits.

Digitakt mk1 only. Digitakt II's filter machines are out of scope.

Status

The audio path is done and validated (BP / BP2 / COMB / TRASH and OB4 / OB2), including the filter envelope, the comb's own controls (feedback on the RESO/GAIN knob, Harmonics/Damping on the free C/G encoders), and the FLTR page's FREQ/RES response curve and TYPE-box glyphs, which now redraw for the new modes (a band-pass bell for BP/BP2, comb teeth for COMB/TRASH, a lowpass for OB4/OB2). See RE_NOTES.md.

Stress-tested on hardware over 3 days: PASS (a Digitakt mk1). 1.1a was tested on hardware (Digitakt mk1) on both OS 1.53 and OS 1.54: PASS. 1.1b is emulator-verified (bit-exact for BP/BP2 and COMB/TRASH; the OB saturation table is bounded at ~-100 dBFS); hardware testing is pending.

Everything in this repo builds (BUILT), lints against core and combines with digihealth (OK: the mods combine).

Known issues

• BP (TYPE 8) output is now saturated to ±2^27 in the SVF (like the comb), so a high-Q peak on hot material clips cleanly instead of wrapping the 32-bit sample. It is still a hard clip at the top of the range; back off RESO/ENV depth or use BP2 for very hot samples. (The resonance range is also capped at Q ~5.)

Controls

• FREQ (E) sets the cutoff / comb pitch; RESO / GAIN (F) is the resonance, and for COMB/TRASH it is the comb feedback (0…127, up to 31/32 for a metallic, near-self-oscillating ring); the filter envelope (H + ADSR) modulates the cutoff exactly as the stock types do.

• For OB4/OB2, RESO / GAIN (F) is the resonance: it self-oscillates from about 121, so the filter rings at FREQ even with no input.

• Second FLTR page, while TYPE is COMB or TRASH, adds C = Harmonics (delay divider 1–4) and G = Damping (a one-pole on the feedback), each shown as a round dial like the stock SRR knob. They sit alongside the stock Base / Width / Env Delay / SRR knobs, which keep working and saving normally — nothing is replaced. The two values are mirrored into saved kit words (best-effort persistence, since every sound slot is otherwise taken); their defaults (0) leave the plain comb.

  

How it works

Three patch sites plus ev_draw / ev_tick subscriptions (addresses and the full reverse-engineering are in RE_NOTES.md):

1. widen the Filter Type range field from max 7 to max 13;
2. re-point the TYPE label formatter so 8–13 print BP / BP2 / COMB / TRASH / OB4 / OB2;
3. hook the per-voice filter call in the render and run our own DSP for TYPE 8–13, tail-calling the stock filter for everything else;
4. subscribe to core's ev_draw (digifilter_draw, order 40): after the page has drawn, when the shown page kind is FLTR (6) and the active track's TYPE is 8–13, mask the stock graph and TYPE box and paint the BP/BP2/COMB/TRASH/OB4/OB2 response curve plus the mode's small TYPE-box glyph. (This composes with the event chain; patching the page drawView vtable entry 0x401842a0 directly lost the curve in multi-mod builds.)
5. on each UI tick, when the active track's TYPE is COMB/TRASH, put the comb's extra controls on the second FLTR page's free C/G encoders: commandeer two unused Error descriptors as filter descriptors on spare sound slots (0x2f/0x30), labelled HARM/DAMP. The comb feedback is the RESO/GAIN knob on page 1 (0…127). The stock Base/Width/Env Delay/SRR knobs are left untouched; the stock layout and descriptors are restored for any other TYPE.
6. in the same ev_draw handler, on page kind 7 (second FLTR page) for COMB/TRASH, paint a round dial (circle + needle, like SRR's) in each of the two free C/G cells — the commandeered descriptors give the label and input but no value widget, since their slot is past the model's range.

Our DSP is integer-only (no FPU, no libgcc): a trapezoidal state-variable filter for BP/BP2, two feedback combs for COMB/TRASH, and a zero-delay-feedback OTA ladder for OB4/OB2. The cutoff and resonance are smoothed (a coefficient ramp across the block, and a fractional, smoothed comb delay) so modulating the filter does not step or click. The filter envelope is added to the cutoff exactly as the stock filter does.

Per voice per render block (measured in digiemu / tests/df_bench.py), TYPE 1 (stock) costs ~2400 instructions; BP/BP2 ~1070 with static knobs and ~1640 while the cutoff is swept (the coefficients are cached and only recomputed when the cutoff/resonance actually move, and the only division — the SVF's a1 — is a precomputed table, so no software division runs at all); COMB/TRASH ~1690; OB4/OB2 ~3200 at RESO 64 (~1390 with the feedback at zero). So BP/BP2 are now cheaper than the stock filter even while the cutoff moves, and every mode stays within the stock filter's own order. The OB-1 ladder computes its coefficients at the block's ends and ramps between them, and the comb's interpolation and feedback each use a single 32×32 multiply.

Building

The mod builds for OS 1.53 and OS 1.54 from one source tree: mod.json carries a "ports": {"1.54": …} block (the moved firmware addresses and the site bytes, generated with tools/port_os.py from the DigiFilter ecosystem) and the sources name each moved address through an F_…/.LF_… macro that resolves per OS. The 1.53 build is byte-for-byte the same as before the port.

Needs the m68k cross toolchain, an elekloader checkout, a stock OS .syx, and core. Set up the shell, then build (1.53 shown; use the 1.54 stock and a 1.54 core for the other):

export PATH="<m68k-elf bin>:$PATH"          # e.g. C:\SysGCC\m68k-elf\bin on Windows
export ELEKLOADER_CROSS=m68k-elf-
export PYTHONPATH="<elekloader checkout>"

python -m elekloader.sdk.build . --stock <Digitakt_OS1.53.syx>
python -m elekloader.lint  out/digifilter-1.1b.elemod --stock <Digitakt_OS1.53.syx> --with <core.elemod>
python -m elekloader.patch --stock <Digitakt_OS1.53.syx> --mod <core.elemod> --mod out/digifilter-1.1b.elemod --out custom.syx --version 2.0t --check

That must print BUILT, then OK: links as core …, then OK: the mods combine. Drop --check on the last command to write custom.syx. Building against the OS 1.54 stock selects the port automatically (elekloader 0.5+; it needs a stock file whose sha256 matches the release, so use the official Digitakt_OS1.54.syx).

core is elekloader's mods/core (built, or its release). Add --with <digihealth.elemod> / --mod <digihealth.elemod> if you use it — the third site is chosen so this mod does not overlap digihealth 1.0.

Files

file	contents
mod.json	the mod: id, version, sites, sources, requires core
filter.c	modes, coefficients, the comb DSP, the site entry point
filter_dsp.s	the per-voice state-variable filter (ColdFire EMAC)
filter_glue.s	the raw site entry points (labels + the per-voice dispatcher)
filter_ui.c	the ev_draw handler (digifilter_draw: response curve, TYPE glyph, page-2 dials) and the page-2 ev_tick re-map (digifilter_page2_sync)
filter_tables.h	g, k, frequency and comb-delay tables
ob1_vcf.c, ob1_vcf.h, ob1_tables.h, ob1_fbsat_table.h	OB4/OB2: the OB-1 ladder, its tables and the interpolated saturation table; its model, checks, SPICE circuit and digiemu tests are tests/ob1_*, tests/spice/ and tests/digiemu_filter_ob1_* (docs/ob1.md)
tests/filter_model.py	floating-point reference (svf_block, comb_gain)
tests/df_bench.py, tests/df_bench.c	per-voice instructions and (with digiemu) cycle bench for every mode
tests/comb_equivalence.py	COMB/TRASH output and delay-line state versus a Git baseline (ColdFire)
tests/ob1_table_accuracy.py	the interpolated fbsat table's error versus the exact polynomial (ColdFire)
tests/digiemu_filter_*.py	headless digiemu traces/tests (need a digiemu checkout)
tools/disas.py	disassembly helper (m68k-elf-objdump -m m68k:cfv4e)
tools/gen_ob1_fbsat.py	generates ob1_fbsat_table.h
tools/re_locate.py, tools/emu_filter.py	early image/unicorn helpers
RE_NOTES.md	the reverse-engineering findings (and the 1.1a optimization ledger, §13)
HANDOFF-OPTIMIZATION.md	the 1.1a performance pass: what changed, how measured and verified

Testing

The OB4/OB2 validation commands cover the native kernel, measured calibration targets and focused ColdFire/firmware simulator checks.

# floating-point model self-test (no toolchain needed)
python tests/filter_model.py

# per-voice cost bench (needs m68k-elf-gcc + digiemu's Unicorn; from a
# digiemu checkout it also reports MCF54418 cycles via emu.cftiming):
python tests/df_bench.py

# bit-exact COMB/TRASH output + delay-line state versus a Git baseline:
python tests/comb_equivalence.py

# bound the OB saturation table's error against the exact polynomial:
python tests/ob1_table_accuracy.py

# live tests in digiemu (need a digiemu source checkout + a set-up firmware)
python tests/digiemu_filter_dsp.py  --digiemu <checkout> --fw <folder> --type 8
python tests/digiemu_filter_mod.py  --digiemu <checkout> --fw <folder> --type 10
python tests/digiemu_filter_cutoff.py --digiemu <checkout> --fw <folder>
python tests/digiemu_filter_curve.py --digiemu <checkout> --fw <folder>
python tests/digiemu_filter_page2.py --digiemu <checkout> --fw <folder>

Releases

Prebuilt packages live in releases/ and on the GitHub Releases page. The current one, DigiFilter-1.1b.zip, carries both OS builds — elemods/digifilter-1.1b.elemod for OS 1.53 and elemods/digifilter-1.1b-os1.54.elemod for OS 1.54 — plus the sources, docs and the install/build README. releases/elemods/ also keeps the older elemods (1.1a, 1.1, 1.0k, 1.0j, 1.0i, 1.0h).

Credits

• Digital Alchemy DSP — the mod: the band-pass BP / BP2 (state-variable), COMB and TRASH modes, the per-voice DSP and the FLTR-page response curve.
• Simon Johansson — the OB4 and OB2 modes: the Oberheim OB-1's VCF as a zero-delay-feedback ladder and its 2-pole output, from his SPICE study of the circuit (see docs/ob1.md).

Licence

GPL-2.0-or-later. No Elektron firmware is included or committed.