Module lmodel3 

Learned probabilities for TCG3 code-blocks, in 32-bit integer arithmetic end to end, so a WGSL kernel (no i64) reproduces every probability.

• Features: the TCG3-visible context (Feats::raw: block-local causal neighbours, coarser-level parent, residual magnitudes) plus the class one-hot, as signed Q8 log-magnitudes (integer lg_q8).
• Network: M1b’s MLP (learned::Net, ReLU, Q8 weights, Q16 biases), evaluated with i32 accumulators: weights clamped to ±2047 (±8.0), activations to [0, 2^14] (hidden) / ±2^14 (outputs), biases to ±2^24, so no partial sum can overflow (28·2^11·2^14 + 2^24 < 2^31).
• Output: a 2-component discretized logistic mixture (π logit, μ₁, μ₂, log₂ s₁, log₂ s₂, all Q8). CDF in Q16 from two integer tables (σ in Q16 over |z| ≤ 16, 2^{f/256} in Q16), built on the CPU with basic IEEE operations only and uploaded to the GPU as data.
• Binarization (as M1b, without the APM): zero flag, sign, exp-Golomb bucket by unary tail ratios, mantissa by integer bisection; each decision’s P(1) is an integer ratio of Q16 CDF masses.

Structs

AdCfg

Adaptive-layer settings (stream header).

Adapt

Per-lane adaptive state over the MLP output: an adaptive probability map (secondary estimation) per decision type, interpolated over AB stretch-domain bins, count-based power-of-two rate. Integer only; reset per block.

Mix3

Mixture parameters.

Net32

A network prepared for 32-bit inference.

Constants

AB

APM bins per context (step 1.0 in the logistic domain, ±8).

AC

APM contexts.

NFL

Features actually used (the rest of the NF inputs are zero).

ZM

|z| range of the σ table (Q8).

Functions

ad_init

A lane’s fresh adaptive state as the kernels hold it: (p << 8) | count per bin (count 0), the layout of the adaptive residual graph’s adt.

code_learned

Code residual r under mx on lane.

code_learned_ad

code_learned through the lane’s adaptive layer.

feat_vec3

Integer feature vector of a coefficient.

ratio16

P(1)·2^16 from masses (num ≤ den ≤ 2^16), clamped for rANS.

sig16

σ(z), z in Q8 → Q16.

stretch_tab

Inverse of sig16 on a p >> 4 grid: the smallest Q8 z with sig16(z) ≥ p (p = 16·i + 8), so stretch(p) ∈ [−ZM, ZM].

tables

σ table (Q16, k/256, k = 0..=ZM) and 2^{f/256} table (Q16).

zval

d · 2^{−ls/256} (both Q8; result Q8) saturated to ±(ZM+1), with ls clamped to [−16·256, 24·256]: t = ⌊d·m/2^16⌉ with the Q16 mantissa m = 2^{(256−f)/256}, then an arithmetic or saturating shift.