Skip to content

PARITY.md

Latest commit

 

History

History
58 lines (42 loc) · 3.16 KB

PARITY.md

File metadata and controls

58 lines (42 loc) · 3.16 KB
Raw

Parity posture

Where abgen's reproduction of the Unity asset-bundle-converter's output is byte-exact, where it is deterministic-but-configurable, and where it drops to render/visual equivalence.

Byte parity: the deterministic profile

The deterministic profile's output is bit-identical across rebuilds and compile targets, held by two gates:

  • Native reproducibility. With the deterministic profile selected, 32 representative bundles reproduce bit-for-bit against a frozen baseline. Release binaries are additionally built twice from a clean tree in CI and required bit-identical before publishing.
  • Cross-target parity. The native converter (with ABGEN_JPEG_GLB_9C=1) and the WebAssembly build agree on 56/56 measured bundle hashes. The wasm self-gate independently compares 26 windows/mac/webgl bundle hashes across 8 format fixtures (JPEG, crunched DXT5 normals, Draco, gamma, KHR_texture_transform, generated tangents, multi-material, a two-parcel scene with a baked LOD) — all bit-identical, wasm vs native.

Every byte-path transcendental is routed through a single pure-Rust math implementation on both targets, so the two cannot drift.

The native default trades bytes for fidelity on GLB-embedded JPEG

Native default JPEG decode uses turbojpeg: measurably closer to the upstream oracle but not bit-reproducible across platforms (SIMD IDCT isn't portable). ABGEN_JPEG_GLB_9C=1 switches to a vendored libjpeg9c integer-ISLOW path that is bit-reproducible — the deterministic profile the byte gates above are defined on.

Scoped to a small, bounded slice of output:

  • It affects only GLB-embedded JPEGs — about 10 of 205 measured bundles. Standalone .jpg texture bundles are byte-identical under either decoder.
  • On those ~10 bundles the deterministic path differs from upstream by a worst-case mean channel delta of ~3/255 (turbojpeg: ~0.03/255); roughly half are ~equal either way.

The gap is sub-perceptual: deterministic is the right default for reproducibility, turbojpeg for closest-to-upstream fidelity — hence the env switch, not a hard-coded choice.

Remaining texture delta vs upstream is encoder block choice

Aside from JPEG decode, the largest residual per-texel differences are all BC7-vs-BC7 at identical dimensions and mip counts: our encoder and upstream's pick different blocks/partitions for the same input. Not closable byte-wise — two compliant BC7 encoders may disagree — and it lands in the render-noise class, not a change in what a texture depicts.

Wearables: render parity, not byte parity

Wearable bundles can no longer be byte-gated against upstream by anyone: upstream purged the pre-2026 wearable asset-bundle payloads, and the remaining manifests have no payloads behind them. Validated by render/visual comparison instead.

The visual gate

The byte gates can't speak to encoder-block noise or wearables. The check for both is a Unity render gate: convert with abgen, render the result and the upstream bundle in a Unity host, compare images (screenshots + diff heatmaps) into pass/fail verdicts. The compare pipeline drives this stage; the README documents pointing it at a Unity Editor.