open-standards/wai/encode-conformance/README.md
Transaction Science 19b8efbd99 An encode record states null when nothing was measured, and run keeps its record then
The brief's form: quality is the figures, or null when not measured, and record::body writes null
until with_quality sets them. A record with no such member reads as before and states no figures
either, so every earlier record stands; a quality of any other form is record_malformed. The
corpus case that refused null is now accepted with pictures held and no verdict, and an empty
object takes its place among the malformed spellings. QualityBasis::Absent is NotMeasured.

wai_encode_measure run no longer fails after encoding when a decode gives nothing to measure. It
writes the record with null and says why in quality_not_measured, and --no-decode asks for that.
An Apple-gated test now drives run itself for each codec: a verifier's own decode recomputes
what it signed, with the verdict decoded, and the not-measured path keeps its record.

record::sign writes quality at its section 6.1 place, as its comment says, and the change note no
longer states a count of cases.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-11-07 03:01:50 -05:00

44 lines
4.3 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# `encode-conformance/`: the corpus of `wai.video.encode`
The corpus of [platform-encode](../extensions/platform-encode.md) §10.1.
| path | what | how it is made |
|---|---|---|
| `streams/` | AVC and HEVC elementary streams made through each encoder interface (§3): `platform-*` by the media framework's hardware engine, `va-1-*` and `va-2-*` by two GPUs' encode engines through the video acceleration API, `gpu-driver-*` by a GPU driver's encode interface, `software-*` by the runtime-loaded software library | fixtures: made once by `wai encode avc|hevc source.y4m --idr-plan plan.txt --kbps 400 --es …` (`software-avc-128x72` with `--size 128x72`); no encoder is a pure function, so they are never regenerated |
| `streams.json` | each fixture's SHA-256 | `wai_encode_vectors gen` |
| `decoded/` | the pictures a conforming decoder made from a fixture stream (`software-avc-128x72.yuv`, from `streams/software-avc-128x72.h264`): each picture's luma, Cb and Cr planes, one picture after another (§8.3) | a fixture: decoded once by the platform's decoder (`wai_encode_measure quality --decoded-out`); every conforming decoder outputs these pictures from that stream, so they are kept, not decoded again |
| `decoded.json` | each decoded fixture's SHA-256 | `wai_encode_vectors gen` |
| `handed/` | for each decoded fixture, the pictures its stream's encoder was handed (the source resampled to the rung), in the same layout: a perfect score no decode of the stream gives, for the case that hands them over (§7) | generated |
| `source.y4m`, `plan.txt` | the source every fixture was made from (256×144, 12 frames, 30 fps) and its IDR plan | generated |
| `mutated/` | streams derived from a fixture by one edit, each breaking one rule of §4.3 | generated |
| `records/`, `cases.json` | a record per case, signed with the corpus key (32 bytes of `0x45`), and every case's files and outcome; the `quality` cases state their figures over the source resampled to the 128×72 rung, hold pictures too (`decoded`, the file; `pictures`, how the verifier came by them, `decoded` or `handed`), and, where accepted, give §7's verdict (`quality`: `decoded`, `handed` or `stated`) | generated |
| `rendition/` | encoded renditions (§9) of the packaging golden `pack-conformance/packs/video-base-01`: `streams/` fixtures made from its source with its plan, `cases/` and `cases.json`, accepted and mutated | `wai-pack rendition-vectors gen` (key: 32 bytes of `0x46`) |
| `encode_verify.py` | the second implementation, written from the extension's text | — |
**Provenance.** Generated here, as the table states: `source.y4m` and `plan.txt` are
generated; each fixture under `streams/` and `rendition/streams/` was made once from that
source by `wai encode` through the encoder interface its name gives, and is kept, not
regenerated; the pictures under `decoded/` were decoded once from their fixture stream by the
platform's decoder, and are kept; `streams.json`, `decoded.json`, `mutated/`, `records/` and
`cases.json` are written by `wai_encode_vectors gen`, as is `handed/`, and `rendition/`'s cases
by `wai-pack rendition-vectors gen`. Every input is generated in this repository.
A case's `decoded` pictures stand in for a verifier's own decode, because every conforming
decoder outputs them from that stream (§10.2): neither `wai-rs` on a host with no platform
decoder nor `encode_verify.py` decodes anything, and a verifier with a decoder may decode the
stream instead. On `accept-quality-own-pictures-*` the figures hold over pictures handed over
and say nothing of the stream, which `quality-own-pictures-decoded-*` shows. On
`accept-quality-null-*` nothing was measured, so the pictures held have no figures to hold to
and the case gives no verdict; a record with no `quality` member reads the same.
The identities in the records are the corpus's, not the encoders' that made the streams.
The software library made the same bytes on three hosts of two architectures; nothing here
relies on that.
Check:
```
cd wai/wai-rs && cargo run --locked -q --no-default-features --features platform_encode --bin wai_encode_vectors -- verify ../encode-conformance
cd wai/wai-pack && cargo run --locked -q --bin wai-pack -- rendition-vectors check ../encode-conformance/rendition
python3 -I wai/encode-conformance/encode_verify.py wai/encode-conformance
```