Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions .gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -20,3 +20,5 @@ reference/di/*.wav
!reference/di/README.md
# Marshall drawings: kept locally, not redistributed (DECISIONS D-021)
reference/schematics/
reference/speaker/*.jpg
reference/speaker/*.png # third-party impedance plots: local only (as the drawings, D-021)
13 changes: 13 additions & 0 deletions docs/DECISIONS.md
Original file line number Diff line number Diff line change
Expand Up @@ -1054,3 +1054,16 @@ Consequences: DC unchanged. Every SPICE reference containing V1a is re-rendered;
Also found: the standalone blocks' default supply voltages (Preamp, StageV1aV1b, StageV2aV2b, StageV2aV2bToneStack,
PhaseInverter, PowerStage) were still the pre-D-058 values (bpA 389.35, bpX 402.34, ct 488.71, nScr 485.65 V); updated to
CIRCUIT.md §5 (391.92, 405.00, 491.35, 488.88 V) with the block DC fixtures (the full amp sets its supplies itself).

## D-063 Speaker load: D-059 kept after comparison with measured 4x12 impedance curves
Date: 2026-10-18 · Status: accepted (owner: "keep D-059 and record the comparison")
Context: the owner supplied a third-party plot of measured in-cabinet 4x12 impedance curves (local only,
`reference/speaker/impedance-curves-412.jpg`, git-ignored as D-021): Celestion T1511 (G12M "Blackback" 55 Hz, early 1970s),
T1281 (G12M family, 1960s–70s) and G12K-85; none is the JCM800-era G12-65 / G12M-70 / G12T-75. Read from the plot (by eye):
resonance ≈ 44 Ω at ≈ 108 Hz (T1511), ≈ 76 Ω at ≈ 104 Hz (T1281), ≈ 101 Ω at ≈ 118 Hz (G12K-85); minimum ≈ 13–15 Ω at
300–500 Hz; ≈ 34–35 Ω at 5 kHz and ≈ 45–50 Ω at 10 kHz for all three. D-059: 50 Ω at 100 Hz, 14 Ω at 400 Hz, 35 Ω at 5 kHz,
48 Ω at 10 kHz.
Decision: keep the D-059 load. Above ≈ 200 Hz it matches all three measured curves; its resonance lies inside their range
(the curves differ mainly in resonance height and frequency).
Consequences: none for the model; the D-059 parameters are now backed by measured curves of comparable cabinets, though not of
the 2203's era. An era-matched curve (G12-65, G12M-70, G12T-75) would still be the better reference if one becomes available.
4 changes: 3 additions & 1 deletion docs/PROGRESS.md
Original file line number Diff line number Diff line change
Expand Up @@ -562,7 +562,9 @@ Notes:
- [~] Step 2: device equations
- [~] Step 3: dense linear algebra / SIMD
- [ ] Step 4: linear parts at a lower rate (analysis first)
- [ ] Step 5: re-verification, M6 report
- [x] Step 5: re-verification, M6 report (`harness/reports/M6/20261018_M6_report.md`, 2026-10-18; for owner review)
- 2026-10-18: owner-supplied measured 4x12 impedance curves (T1511, T1281, G12K-85) compared with D-059: match above ≈ 200 Hz,
resonance within their range; D-059 kept (D-063).
Plan (2026-10-14; SPEC §7 M6 "performance work", WDF_DESIGN §4; goal: real-time headroom at the 8x default
without loosening any tolerance):

Expand Down
101 changes: 101 additions & 0 deletions harness/reports/M6/20261018_M6_report.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,101 @@
# M6 report — performance, real-time configuration and model refinements

Status: for owner review (2026-10-18). Scope set by the owner on 2026-10-14: performance work (the other SPEC §7 M6
items stay open); widened by owner decisions during M6 to the accuracy target (D-051), the real-time configuration
(D-054 – D-060) and reference-model refinements from listening (D-058, D-059, D-061, D-062).

- macOS 15.7, Apple clang 17, Release · JUCE 8.0.15 · pluginval 1.0.4 · ngspice 47 · main at 0b5f06e (PR #4)
- Details per step: `docs/PROGRESS.md` (M6), DECISIONS D-049 … D-062, the reports cited below.

## Outcome in one paragraph

The plugin now runs the circuit model at 2x oversampling (96 kHz inside at 48 kHz) with the exact preamp → power-amp
coupling, the power amp at full rate and a Newton tolerance of 1e-4 V; it meets the −20 dB ESR target against SPICE
on every preset and knob test, including HOT. The original M6 goal (real-time headroom at 8x with the old
tolerances) was replaced by the owner: the tolerances became −20 dB (literature range, D-051) and the rate study
showed 2x is indistinguishable by ear from the 16x reference. The reference model gained a fitted EL34, a
representative speaker load, a physical plate-current clamp and C2. Against a NAM capture of a real 2203 at its
published settings the plugin sounded "pretty accurate" to the owner. CPU at high Master (≈ 47 % of one core) stays
above the 20 % working budget.

## Plugin configuration (defaults)

| Setting | Value | Decision |
|---|---|---|
| Oversampling | 2x (2x / 4x / 8x / 16x selectable) | D-054 |
| Power Amp Rate | Full (Half selectable) | D-060, confirmed after a listening test |
| Preamp → power amp junction | exact J4 (joint solve) | X1, D-053 |
| Newton | maxIter 32 (both roots), residual tolerance 1e-4 V, adaptive half-steps on | D-052, D-055, D-056 |
| Knob changes | matrix updates in double precision every 128 samples while a knob moves | D-057 |
| Latency at 48 kHz | 49 samples FIR (≈ 1.0 ms); IIR lower | oversampler only |

## Verification (final, after D-062; reports dated 2026-10-18)

| Check | Result | Status |
|---|---|---|
| M4 full acceptance at 768 kHz (`M4/20261018_full_acceptance.md`) | 107 PASS; CRUNCH / DRIVE DI −60.2 … −62.1 dB | PASS |
| Full acceptance at 384 kHz, exact J4 (`…_fs384000_j4exact.md`) | 107 PASS | PASS |
| HOT (Master, Presence 1.0) at 384 kHz (`…_j4exact_hot.md`) | 20 PASS / 1 FAIL: DI −40.3 … −41.5 dB; FAIL = 2 rescued first-solve Newton failures on one clip | 1 FAIL (Newton count) |
| Plugin model check, 2x Full, 1e-4 V (`20261018_x2_rates_d062_full2x.md`) | vs SPICE worst CRUNCH / DRIVE / HOT −45.3 / −48.5 / −24.0 dB; unseen settings vs 16x worst −20.8 dB | PASS (≤ −20 dB) |
| Block tests (M2 / M3 / M4, 2026-10-18) | b12 105 / 106 / 106, b1234 119, b3 87, b34 94, b56 106, b5open 88 PASS, 0 FAIL; b6 96 PASS / 1 FAIL (known 5 kHz 20 V Newton median, D-043) | PASS except the known b6 item |
| Knob jumps vs SPICE | b12 −55.9, b1234 −39.1 … −55.1, b34 −53.1 … −67.6, b56 Presence −62.9 dB | PASS (≤ −20 dB) |
| Plugin vs ampwav null test, 2x Full (`20261018_plugin_null_2x_full.md`) | 26 / 26, worst −138.5 dB | PASS (≤ −60 dB) |
| pluginval strictness 5 (VST3, AU), auval | SUCCESS / SUCCESS / SUCCEEDED (`20261018_pluginval_*.log`, `20261018_auval.log`) | PASS |
| Automation at audio rate (96 kHz, Full, exact J4, 1e-4 V) | 56 runs, no NaN / Inf, 0 unrescued | PASS |

Notes: the b56 / b6 SWP-lo-0.1V references stop with "timestep too small" at RELTOL 1e-8 (recorded-input nodes;
independent of the tube model) and are rendered once more at RELTOL 1e-7 by `render_spice`, noted in their
reports (as the M3 cascade does). Earlier reports describe earlier models (Koren EL34, 16 Ω resistor, no C2).

## CPU (plugin, 2x Full, FIR, `M6/20261018_plugin_cpu.md` and PROGRESS)

| Situation | % of one core |
|---|---|
| Default knobs, static | 27.1 |
| Every knob 10, static | 50.8 |
| Owner's NAM-matched settings (Preamp 10, Master 10, Presence 8) | ≈ 47 |
| All six knobs moving (LFO / jumps, host blocks 512 or 64) | 32.8 … 34.7 |
| Same at Power Amp Rate Half (for comparison) | default ≈ 19, NAM settings ≈ 31.5 |

Before M6 (M5): real-time factor 0.53 at 16x (≈ 190 % of one core); knob automation alone had raised the 2x
configuration to 85 % until D-057.

## Experiments and studies (ARCHITECTURE_STUDY §6)

| Study | Result |
|---|---|
| X1 exact J4 | HOT 220 Hz ESR −12.6 → −58.9 dB; the Thevenin J4 had been the HOT error source |
| X2 / A1 lower rates | accuracy holds to 2x; the limit was the preamp's Newton budget (D-052); exact J4 at half rate without a loop filter (D-053) |
| X3 / X4 neural (LSTM-32) | −16.5 … −17.1 dB (one setting), −3 … −7 dB (three knobs); 2 % CPU; sounded weaker with less gain; set aside |
| X6 block boundaries | supply couples Master into every preamp boundary; per-block networks set aside |
| Listening (owner) | 4x and half 2x identical to 16x; Full vs Half not distinguishable; plugin "pretty accurate" vs a 2203 NAM capture at Master 10 |
| EL34 fit, speaker load, HOT convergence | see the model changes below |

## Reference-model changes (owner-approved; CIRCUIT.md §5 / §6, reference/README.md updated)

| Decision | Change | Why |
|---|---|---|
| D-058 | EL34 Koren parameters fitted to 81 digitised Ei-RC curve points (median error 3.7 / 2.7 % vs 11.5 / 28.2 %) | owner listening ("dark", "lacks gain") led to an A5 review |
| D-059 | 16 Ω resistor → representative nominal 16 Ω 4x12 impedance (Re 13 Ω, 1.0 mH ‖ 45 Ω, 100 Hz resonance) | closed-loop output impedance ≈ 20 Ω: a real speaker load adds +3 … +5 dB at 100 Hz and in the treble |
| D-061 | EL34 plate voltage through a 5 V softplus: no plate current below the cathode | the reverse plate current in Koren's form stalled Newton during speaker-load flybacks at HOT |
| D-062 | C2 100 pF V1a plate–cathode (missing from the transcription, O8) | on the drawing |

Also fixed during M6: the standalone blocks' default supplies were still pre-D-058 (found by the D-062 block tests).

## Open items

- HOT acceptance: 1 FAIL (2 rescued first-solve Newton failures on one DI clip at 384 kHz).
- Plugin model check: 102 first-solve failures (3 unrescued) over 28 runs, almost all at extreme unseen knob settings.
- CPU above the 20 % working budget at high Master (≈ 47 %); optimisation candidates: the Newton solve and the device
maths (profile: LU ≈ 29 %, libm ≈ 26 %, Newton loop ≈ 34 % at HOT).
- Speaker load: kept after comparison with measured 4x12 curves (T1511, T1281, G12K-85; D-063); an era-matched G12-65 /
G12M-70 / G12T-75 curve is not available.
- 12AX7: the Sarti DAFx-07 parameter set matches the RCA data sheet at 100 V where Koren's does not; studied, small
effect on the sound, not adopted.
- SPEC §7 M6 items outside the performance scope stay open.

## Decisions during M6

D-049 (performance options), D-050 (PyTorch for X3), D-051 (−20 dB ESR), D-052 (maxIter 32), D-053 (exact J4 at half
rate), D-054 (plugin defaults 2x / exact J4), D-055 (tolerance 1e-4 V), D-056 (half-steps on), D-057 (matrix updates),
D-058 (EL34 fit), D-059 (speaker load), D-060 (Full default), D-061 (plate clamp), D-062 (C2), D-063 (speaker load kept after comparison with measured curves).
Loading