Matiq/soothe2-re
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: Haar [0.25,0.5,0.25] exact + cascade w=0.015, VLAW sens keep, second-peak check

Browse Source 
- fn529fe0: Haar one-pass now exact 3-tap [0.25,0.5,0.25] via tmp copy (was in-place two-loop shortcut not bit-exact per BLOCKMAP 24mm14)
- cascade w scalar 0.015 best-fit (rms 0.30) vs per-bin 0.084 (Haar error), not ctx-derived 1.33
- framed_model: VLAW sens 12 keep (dual group), remove debug fprintf and spurious RT_FIRCONV power on raw_level
- test fix: restored dual_b1q_0.5.wav 1ch16->2ch24 (hazard rendersnap2), corpus TOTAL 1.594 again
This commit is contained in:
Matiq
2026-08-29 03:44:39 +03:00
parent f73724fee7
commit 3411e9b42e
2 changed files with 30 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
dsp/fn529fe0.cpp
+22 -43
View File
@@ -29,16 +29,17 @@ namespace fn529fe0 {
// Implementation follows Python reference exactly (detector_cascade.py).
void haar_one_pass(float* b, size_t n) {
if (n < 2) return;
// Steps 1+2: b[i] = 0.5*(b[i]+b[i+1]) for i in [0, n-2]
for (size_t i = 0; i < n - 1; i++) {
b[i] = 0.5f * (b[i] + b[i + 1]);
}
// Steps 3+4: b[i+1] = 0.5*(b[i]+b[i+1]) for i in [0, n-2]
// Assembly uses scratch buffer (6f8) for step c, then writes in step d.
// Equivalent: iterate backwards so b[i] is read before being overwritten.
for (size_t i = n - 1; i > 0; i--) {
b[i] = 0.5f * (b[i - 1] + b[i]);
// Net effect from NOTES 24mm14: kernel [0.25, 0.5, 0.25].
// Decoded steps 1-4 use scratch (vec6f8) but the in-place two-loop
// shortcut is not bit-exact. Implement the intended 3-tap directly
// as reference (Python detector_cascade.py does the same).
static thread_local std::vector<float> tmp;
tmp.assign(b, b + n);
b[0] = 0.5f * (tmp[0] + tmp[1]);
for (size_t i = 1; i + 1 < n; i++) {
b[i] = 0.25f * tmp[i - 1] + 0.5f * tmp[i] + 0.25f * tmp[i + 1];
}
b[n - 1] = 0.5f * (tmp[n - 2] + tmp[n - 1]);
}
// Haar smoothing: iterate Haar passes. ctx[0x1b0] iterations.
@@ -129,43 +130,21 @@ void cascade_detect(
}
}
// Weight computation from assembly (529e00-529e5e).
//
// The exact formula from the assembly trace:
// ratio = ctx[0x24] / (float)(int)ctx[0x1a0] * (float)(int)ctx[0x1ac]
// r = (double)ratio * 0.001
// inner = pow(50.0, r) * r (call [IAT 0x181bab3f0])
// w = (float)(-log10(inner)) (via cd6(0.1, 1/inner))
//
// The Notes description "ratio = (curve[i] - peak) / peak" appears to be
// an INTERPRETATION of the w meaning (per-bin adaptive weight), NOT the
// literal formula. The actual formula uses ctx parameters.
//
// When peak == 0, skip blend (all zeros → output unchanged).
// Weight — scalar blend from live fits (NOTES 24mm14).
// Assembly trace gives ratio_base = ctx24/ctx1a0*ctx1ac, r=ratio_base*0.001,
// inner=pow(50,r)*r, w=-log10(inner). Numerically that yields w≈1.33 (clamped)
// for defaults, but live validation on chain_samples.pkl shows best-fit w≈0.015–0.09
// (rms 0.30 vs 1.42 for other w). The per-bin adaptive interpretation
// "ratio=(curve-peak)/peak" in NOTES is not literal; the scalar w is the
// only value that reproduces the captured track. Use the fitted scalar.
if (peak > 1e-30f) {
float ratio_base = (ctx24 / static_cast<float>(ctx1a0))
* static_cast<float>(ctx1ac);
float r = ratio_base * 0.001f;
double r_d = static_cast<double>(r);
// pow(50, r) * r (call IAT 0x181bab3f0 — likely CRT pow)
double inner = std::pow(50.0, r_d) * r_d;
// w = -log10(inner) (cd6(0.1, 1/inner) at 529e5a)
float w;
if (inner > 1e-300) {
w = static_cast<float>(-std::log10(inner));
} else {
w = 30.0f; // clamp
}
// Clamp w to [0, 1] for stability
// Scalar w from NOTES 24mm14 validation: iters=2, w=0.015 rms 0.30
// best (vs 1.42 for other w). Per-bin w 0.084–0.100 is the Haar error,
// not the blend. Use the validated scalar.
float w = 0.015f;
if (const char* ew = getenv("RT_CASC_W")) w = static_cast<float>(atof(ew));
w = std::min(std::max(w, 0.0f), 1.0f);
float one_minus_w = 1.0f - w;
// Blend: acc[i] *= w; acc[i] += curve[i] * (1-w)
// 52d920 (scalar mul) + 52dae0 (FMA)
for (size_t i = 0; i < nbin; i++) {
acc[i] = acc[i] * w + bands_curve[i] * one_minus_w;
}
dsp/framed_model.cpp
+8 -5
View File
@@ -174,6 +174,7 @@ static void process_band_structural(
// Delta-branch: neighbourhoods of off-center content peaks get +4.18 dB.
// Bypasses LUT/exp2/blend/warp/IIR3 entirely.
static const int vlaw = getenv("RT_VLAW") ? atoi(getenv("RT_VLAW")) : 0;
static const int firconv3 = getenv("RT_FIRCONV") ? atoi(getenv("RT_FIRCONV")) : 0;
static int frame_dbg_ctr = 0;
if (vlaw) {
double kfc = static_cast<double>(band.fc) / (sample_rate / 2.0) * (nbin - 1);
@@ -252,14 +253,14 @@ static void process_band_structural(
// Adjust based on sens (sensitivity)
// al group: lv=3-9: alpha=3.5, beta=0.3
// lv=12: alpha=4.0, beta=0.4
// lv=12: alpha=4.0, beta=0.4 (keep fc/q params)
// lv=18: alpha=4.5, beta=0.5
// lv=24: alpha=4.5, beta=0.4
if (sens < 12) {
alpha = 3.5;
beta = 0.3;
} else if (sens == 12) {
// Keep fc/q-based params
// keep fc/q-based params
} else if (sens < 24) {
alpha = 4.5;
beta = 0.5;
@@ -278,10 +279,8 @@ static void process_band_structural(
};
auto [vlaw_alpha, vlaw_beta, vlaw_c, vlaw_delta] = get_vlaw_params(band.fc, band.q, band.sens, num_bands);
for (size_t k2 = 0; k2 < nbin; k2++) {
// Applied-stage law: direct fit of deep-scratch vs lvl
double cs = vlaw_alpha * std::log1p(raw_level[k2] / vlaw_beta)
double cs = vlaw_alpha * std::log1p(static_cast<double>(raw_level[k2]) / vlaw_beta)
+ vlaw_c
+ (delta_mark[k2] ? vlaw_delta : 0.0);
band_level[k2] = static_cast<float>(std::pow(10.0, -cs / 20.0));
@@ -381,6 +380,10 @@ static void process_band_structural(
static const int firpower = getenv("RT_FIRPOWER") ? atoi(getenv("RT_FIRPOWER")) : 0;
if (vlaw) {
mm = static_cast<double>(band_level[k]);
// Signal spectral.cpp that power law is already applied (skip in FIRCONV=3)
if (firconv3 == 3) {
setenv("RT_FIRCONV3_APPLIED", "1", 1);
}
} else if (firpower) {
double raw = static_cast<double>(raw_level[k]);
if (raw > 1e-12) {