Crate gamut_dsp 

Shared digital signal processing routines for the gamut codecs.

Provides the AV1 1-D transform kernels:

• the lossless 4×4 Walsh–Hadamard pair (fwht4x4 / iwht4x4, AV1 §7.13.2.10),
• the discrete cosine transform pair (forward_dct / inverse_dct, AV1 §7.13.2.2–.3),
• the asymmetric discrete sine pair (forward_adst / inverse_adst, AV1 §7.13.2.4–.9 — DST-VII at size 4, DST-IV at 8/16), with the FLIPADST flip (flip_in_place), and
• the identity transforms (forward_identity / inverse_identity, AV1 §7.13.2.11–.15).

The 2-D assembly that selects per PlaneTxType and applies the per-pass normalization shifts (AV1 §7.13.3) is tracked in gamut-avif/STATUS.md.

Alongside the transforms, [mod@math] exposes the small AV1 §4.7 integer arithmetic primitives (round2, round2_signed, clip3) and the shared forward-quantize rounding (round_div_nearest) that the codec crates build on, and [mod@mulaw] adds µ-law companding/quantization (mu_compress / mu_expand / mu_quantize / mu_dequantize).

Functions

clip3

Clip3(low, high, x) (AV1 §4.7): clamp x to the inclusive range [low, high].

flip_in_place

Reverse the first len elements of t in place — the FLIPADST sample-order flip (AV1 §7.13.3 applies this around the shared inverse ADST for the FLIPADST transform variants).

forward_adst

In-place 1-D forward ADST of the length-2^n array t (n ∈ {2, 3, 4}) — the transpose of the inverse ADST matrix, built from the inverse’s impulse responses (see module docs).

forward_dct

In-place 1-D forward DCT-II of the length-2^n array t (2 ≤ n ≤ 6).

forward_identity

In-place 1-D forward identity transform of the length-2^n array t (2 ≤ n ≤ 5): the reciprocal per-element scaling that pairs with inverse_identity.

fwht4x4

Forward 4×4 Walsh–Hadamard transform: maps a row-major residual block to the lossless AV1 coefficients to be entropy-coded.

inverse_adst

In-place 1-D inverse ADST of the length-2^n array t (n ∈ {2, 3, 4}), per AV1 §7.13.2.9.

inverse_dct

In-place 1-D inverse DCT of the length-2^n array t (2 ≤ n ≤ 6), per AV1 §7.13.2.3.

inverse_identity

In-place 1-D inverse identity transform of the length-2^n array t (2 ≤ n ≤ 5), per AV1 §7.13.2.15.

iwht4x4

Lossless 4×4 reconstruct: dequantizes (* 4), applies the AV1 row (shift 2) and column (shift 0) inverse WHT, and returns the residual. Bit-exact with the AV1 decoder for an 8-bit base_q_idx == 0 block (AV1 §7.12.3, §7.13.3, §7.13.2.10).

mu_compress

µ-law compress a value in [-1, 1] to a companded value in [-1, 1]. value is clamped to [-1, 1]. Precondition: mu > 0.

mu_dequantize

Dequantize an integer index back to a value in [-1, 1] through µ-law — the inverse of mu_quantize. The never-written top code clamps down to 2^bits−2 for robustness.

mu_expand

µ-law expand a companded value in [-1, 1] back to [-1, 1] — the inverse of mu_compress. Precondition: mu > 0.

mu_quantize

Quantize a value in [-1, 1] through µ-law to an integer index in 0..=2^bits−2 (odd level count; the center index is exactly 0.0).

round2

Round2(x, n) (AV1 §4.7): rounding right shift, x for n == 0 (arithmetic >>, ties up).

round2_signed

Round2Signed(x, n) (AV1 §4.7): symmetric rounding right shift — rounds the magnitude and keeps the sign (ties away from zero), unlike round2 which ties toward +∞.

round_div_nearest

Rounds num / den to the nearest integer, ties away from zero — the encoder forward-quantize step shared by AV1 and VP8 (level ≈ coeff / q). Requires den > 0.