Module wavcodec 

The wavelet codec: CDF 9/7 pyramid → deadzone quantizer → context model → adaptive binary range coding, with a SPERR-style outlier stream for pointwise bounds and an integer 5/3 lossless mode.

Lossy path

1. Forward multi-level CDF 9/7 (crate::wavelet); every coefficient is multiplied by its band’s synthesis-basis norm, so one uniform step Δ is equal-slope (MSE-matched) across bands.
2. Deadzone uniform quantizer q = sgn(c)·⌊|c|/Δ + ρ⌋ (ρ < ½ widens the zero bin); reconstruction at the per-(class, |q|∈{1,2,≥3}) centroid |ĉ| = (|q| − δ)Δ, the 30 δs transmitted in 1/512 units.
3. Coefficients coded by crate::coefcode (hand contexts) or a learned model (crate::learned).
4. Rate control: for a PSNR/RMSE target Δ is bisected on the measured error of the actual decoder output.

Max-abs mode

For a bound b the wavelet layer is coded at Δ = f·b for several f; the encoder decodes its own output and every cell with |x − x̂| > b gets a correction k = round((x − x̂)/2b) (then |x − (x̂ + 2bk)| ≤ b), coded as a context-modelled significance map plus |k|; cells a float cast still pushes over b are stored raw. The f with the smallest total wins and the bound is re-verified on the exact decoder output before returning.

Structs

Coeffs

Forward-transformed, band-weighted coefficients.

WavConfig

Wavelet-codec options.

WavStats

What the encoder did.

Enums

ModelChoice

Which coefficient model codes the stream.

Constants

FAST_K_MAXABS

See FAST_K_RMSE.

FAST_K_RMSE

Fast-mode step ratios Δ / target: RMSE targets (Δ = k·rmse) and max-abs bounds (Δ = k·b). Calibrated on the 19 seen fields.

MAGIC

Stream magic of the wavelet codec.

Functions

decode

Decode a lossy wavelet stream.

decode_i32_lossless

Decode a lossless stream from encode_i32_lossless.

encode_f32

Encode f32 samples (row-major, dims 1D–3D).

encode_f64

Encode f64 samples.

encode_i32_lossless

Lossless coding of integer samples via the reversible 5/3 transform.

fast_step

Step for an RMSE target: start from the formula Δ = k·t, then take log–log secant steps on the measured error (rmse ∝ Δ^γ), aiming 1% under the target, for at most checks trials. Returns the largest tried step that met the target (or the last one, halved, when none did) and the number of trials.

quantize

Deadzone quantization with centroid offsets.

reconstruct

Dequantize + inverse transform → reconstruction (f64, row-major).