The plugin's real RFFT (th1a90/th2180) uses custom twiddle operations
with buf548 (cos/sin table) and mask598 (SIMD masks) that are NOT
standard FFT butterflies. Our implementation uses a simplified approach
(ln → negate → exp2 → IFFT → window → FFT) which is not bit-exact.
Current state:
- Default path (no FIRCONV): TOTAL 1.825 dB
- FIRCONV=2 (real RFFT): TOTAL 10.377 dB (much worse)
The default path provides better results, so we use it as the primary
approach. Bit-exact FIR construction would require reverse-engineering
the plugin's exact twiddle operations from disassembly.
Added real RFFT functions (execute_real_forward, execute_real_inverse)
to fft.hpp/cpp. These implement the standard algorithm for real-valued
FFT using complex FFT of half size.
Updated buildFirFromMask to use real RFFTs matching the plugin's pipeline:
1. log(mask) → negate
2. forward real RFFT (opB)
3. EXP in-place
4. inverse real RFFT (opC)
5. Window
6. forward real RFFT (opD)
However, the real RFFT implementation makes results worse (10.377 dB vs
1.825 dB default). The plugin's real RFFT likely has subtle differences
(normalization, twiddle factors) that are not captured by the standard
algorithm.
The default path (no FIRCONV) remains the best approach with 1.825 dB
TOTAL error.
Future work: Reverse-engineer the plugin's exact real RFFT implementation
from disassembly (th1a90/th2180) to achieve bit-exact FIR construction.
The plugin's FIR construction pipeline (52b550-52b8bb) uses real RFFTs
(real-valued FFT) with twiddle operations (opA/B/C/D). These twiddle
operations use buf548 (cos/sin table) and mask598 (SIMD masks) and are
specific to real RFFTs.
Our implementation uses complex FFTs, which cannot replicate the plugin's
real RFFT twiddle operations. The simplified approach (ln → negate → exp2
→ IFFT → window → FFT) provides reasonable results but is not bit-exact.
Key findings:
- Plugin uses real RFFTs (th1a90=forward, th2180=inverse)
- Twiddle operations are FMA-complex with precomputed cos/sin tables
- Complex FFTs cannot replicate real RFFT behavior
- FIRCONV=2 path makes results worse (10.377 dB vs 1.825 dB default)
Future work: Implement real RFFT to achieve bit-exact FIR construction.
- Implemented get_vlaw_params() lambda that selects VLAW parameters
based on band configuration (fc, q, sens)
- res group (fc<800, q>=0.99): alpha=5.0, beta=0.3
- t1kq group (fc=800-1200, q<1.0): alpha=4.0, beta=0.4
- t1k group (q>=0.99, fc<1200): alpha=4.0, beta=0.5
- t1k group (q>=0.99, fc>=1200): alpha=4.5, beta=0.4
- Sensitivity adjustment: sens<12: alpha=3.5, beta=0.3
sens=12-24: alpha=4.5, beta=0.5
sens>=24: alpha=4.5, beta=0.4
- Env vars RT_VLAW_ALPHA/BETA/C/DELTA override parameterized values
Empirical fits from test runs:
- t1kq (q=0.99999785, fc=800-1200): alpha=3.5-4.5, beta=0.3-0.5
- t1k (q=1.0, fc=500-2000): alpha=4.0-4.5, beta=0.4-0.6
- al (fc=1000, q=1.0): alpha=3.5-4.5, beta=0.3-0.5 (sens-dependent)
- res (q=1.0, fc=300-700): alpha=5.0, beta=0.3
- dual (q=0.1-10.0, fc=500): alpha=3.2193, beta=0.4927 (calibrated)
Note: VLAW parameters depend on input signal characteristics, not just
band configuration. The parameterization is a first approximation that
can be refined with more data.
- log2_ln.hpp/cpp: Plugin's exact ln(float) polynomial from 535a70
(0x1802a24c0). IEEE 754 bit extraction + Horner evaluation.
Coefficients extracted from binary at 0x181f81f80..0x181f821c0.
Max error ~3e-6 for typical inputs.
- spectral.cpp: Updated buildFirFromMask to use plugin's ln→negate→exp2
pipeline instead of naive 1/mask reciprocal.
- exp2_tables.hpp/cpp: Already contains plugin's exp2 tables (0x26b820).
Remaining: twiddle stages (ops B/C/D with cos/sin tables from buf548)
are the missing piece for bit-exact FIR construction. These are
FFT butterflies already implemented in fft.hpp but need integration
into the FIR pipeline.
Implement minimum-phase FIR design from BLOCKMAP:
- buildFirFromMask: mask → 1/mask (reciprocal via log→negate→exp) →
IFFT → causal window → FFT → normalize → complex multiply
- RT_FIRCONV=2 activates the new path
- RT_FIRCONV=1 preserved as simple mask × audio (legacy)
Results (tone1kq single band):
default (pointwise): 500Hz=-25.35 dB, 1kHz=-50.60 dB
FIRCONV=1 (mask mul): 500Hz=-1.31 dB, 1kHz=-25.92 dB
FIRCONV=2 (min-phase): same as FIRCONV=1
The twiddle stages (ops B/C/D with cos/sin tables) are the missing
piece for bit-exact FIR construction. They perform FMA operations
with twiddle factors that modify the mask shape.
Note: dual_b1q_0.5.wav reference is empty (0 bytes) — corpus can't run.
Needs regeneration.
- Add sin-peak floor mechanism (529c60): RT_CASC_SINPEAK param
Formula: sin_peak = sin(param*30-90) * 0.115129 * peak_level
Floor active for param in [3,9], max at param=6 (ln10/20=0.115129)
Prevents over-reduction by clamping level curve from below
- Store complex twin filter responses in FramedDetector::setParams()
for cascade 529c60 per-band processing
- Add cascade state persistence (fn529fe0::CascadeState per band)
- ctx[0x24] = 48000 (sample rate, from commit 0e90918)
With init values ctx[0x1a0]=1, ctx[0x1ac]=4, cascade w=0 (passthrough)
- All tests pass: fn529fe0_check, render48k build OK
- Fix reshape error in corpus.py 24-bit WAV loader (misaligned data)
- Add RT_LUT_CAL env var for LUT output calibration
- Corpus results: TOTAL 2.397 (bridge 1.594), comb improved (-4.032)
- Structural chain regresses on t1kq/t1k/al/dual due to LUT curve mismatch
- The LUT produces different frequency response than plugin's FIR construction
Root cause: IIR accumulator reset to 0 every frame, losing temporal state.
Fix: static thread_local accumulator persists across frames.
Results (dual_b1q_0.5, reaper render):
cut@500 = -10.32 dB (EXACT match, was -14.35)
cut@2000 = -11.82 dB (EXACT match, was -6.58)
Both the ×1.805 (OLA normalization) and ×2.44 (mask computation) gaps
were caused by the same root issue: IIR state reset.
- th_b3c0 (0x18000b3c0) = pure complex multiply FIR × audio in freq-domain
- scan3.py: pre-scan approach finds ctx in 1.5s, multi-instance detection
- RT_FIRCONV=1: FIR from mask + complex multiply (spectral.cpp)
- RT_FIRPOWER=1: power-law mask from raw spectrum (framed_model.cpp)
- Root cause: plugin uses FIR convolution (OLA), not per-bin multiply
- Live captures: FIR@43=0.524, mask@43=0.510, final gain=0.305
- Best result: RT_LUT_OFF gives cut@500=-8.18 dB (ref -10.32)
- NOTES_LEVEL 24e/24f/24g appended
- Add BandConfig A/B/gamma LUT compression (extracted from refs: A=-13.78dB, B=68.29dB, gamma=0.344)
- Add res^rp smoothing term for bridge-parity (RP0=0.0275, DRP=0.2159)
- Calibrate scale_factor to match bridge gain at tone bin (15.0 * 440.95 / 2048 = 3.23)
- Structural 48k: -26.81 dB vs ref -27.23 dB (err +0.426 dB)
- Bridge baseline: intact (0.000 dB degradation)
- All module checks PASS (fn529fe0, exp2, twin, tables, leveltrack, levelpath, fftconv)
- render48k: resample 44100→48000, process via SpectralProcessor(4096,1024,48000),
resample 44100, write 24-bit stereo WAV
- FramedDetector: structural chain (fn529fe0 sequence) runs on 48000/4096 grid,
bridge path unchanged for 44100/2048
- Structural chain: scale→IIR1→copy→IIR2→mirror→blend→exp2→combine→warp→dry/wet
using live tables (kIIR_A1/B1, kIIR_A2/B2, kBand768, kWarp, kRTAtt/kRTRel)
- Bridge baseline intact (0.000 dB degradation)
- 48k tone test: -20.73 dB vs ref -27.23 dB (6.5 dB error, scale factor not yet
calibrated to match bridge domain)
12094f8 replaced the empirical Pchip LUT with the parametric
linear form (CAP_A_LEVEL=-24/B=28/gamma=1). With gamma=1 the
parametric form reduces to identity on xv=log10(am/res), which
breaks the whole mask chain (t1kq fc-scan err jumps to ~7.9 dB).
Re-instating lut_pchip restores the honest baseline:
t1kq mean|err| 0.226 dB (was 7.9). Confirms NOTES_LEVEL F1
closure: no parametric LUT set beats Pchip at fixed bridge
params; the structural A/B/gamma (level-path ctx+0x188) is the
only real path to bytes, not this parametric substitution.
Extracted the 8x16 irrational tables + lead-in -708.4xx series from soothe_mem.bin
into exp2_tables.{hpp,cpp} (P3 bit-exact inputs). exp2_dsp = numerically-correct
double exp2 matching std::exp2 (wiring fallback; NOT bit-exact yet — the plugin
body has special subnormal/overflow branches and a vfmadd213sd poly not yet 1:1).
exp2_check: 2e6-grid PASS (0 cells >1e-13).
Crucial: earlier dual ref -53.7 dB was a 24-bit-misdecoded artifact; honest ref is
-10.2 dB flat. Root cause of the "dual paradox" was a metric bug + missing log-domain
LUT. Ported the documented bridge (NOTES:147) into framed_model.cpp:
xv=log10(am/res); C=G*LUT(xv)+W*warp^A; gain=(1-C)*res^rp.
Results (honest 24-bit metric): dual (fc=500 q-sweep) err <=0.7, t1kq fc-scan
err <=0.59. All empiric numbers explicitly marked. Structural A/B/gamma + combine/
FFT-conv still pending.
Fixes structural divergences: IIR1 into shared 0x5406f8 buffer + bridge to band
mask (FUN_18052d650, 0x5160); IIR2 on band mask; Hermitian mirror (0x11940);
blend step f6f8=axis(1-mix)+mix*0.8, mask=exp2(-mask)*f6f8; real combine via
kRTAtt/kRTRel weights (0x5406c8/6e8); warp, IIR3x2, dry/wet.
Validation: t1kq only1 fc1000 -21.1 vs -22.1 (OK); fc-scan shape intact;
dual @500 matches (s30 -51.5 vs -53.7), @2000 gap -4..-13 vs -29.6 remains.
That gap = dB-domain band LUT (FUN_180563a60) not yet in chain -> Phase A.
res=|2B/A| is minimal at band centre (not maximal). Model uses xv=log10(A_k/res_k)
=> level = am / res. Old am*res inverted the fc-response (cut more off-center).
Fixed; fc-scan (tone1kq, band fc 800..1200 via t1kq_only1_<fc>) now tracks the
reference: scale=42 mean|err| 2.46 dB, intact shape (deepest at fc==tone). Was
flat+inverted before. Remaining under-cut off-center -> FFT-conv smoothing/gaps.
CRITICAL fix: soothe_mem.bin is VA-linear (offset=RVA). DAT_18262b5c8/b704/b700
are INT 1 (cvtdq2ps -> 1.0), NOT the 0.4552/0.6089/0.6070 read earlier via a bad
section offset. Transcribed FUN_180529fe0 PRNG prologue into prng_fvar30():
fVar30=(int)(LUT[s+1]*LUT[s]+0.001). At live state 112 this is deterministically
1.0 over 300 frames, so scale level *= (1/2048)*440.95 is not randomized in
practice. Scale coefficient now computed structurally; t1kq unchanged -0.43 dB.
Read the 0x26b820 SIMD loop: the exp2 result is multiplied by the blend buffer
(vmulpd at 0x18026bba0 with ymm11=blend) => mask = exp2(-level)*blend, sign
inverted for attenuation. framed_model.cpp corrected to exp2(-scratch)*0.8.
Calibrated level_scale=600 -> t1kq err -0.06 dB (ref -14.98). comb per-tone
errors reduced (max ~7.8 dB); warp/FFT-conv smoothing still approximated.
Matiq