Repository navigation
[PROPOSAL] The C-vs-Rust dual-build comparison from #2207 never runs; the C-only build drops 708 and HEVC captions #2398
Description
#2207 added two builds so that "both must produce identical output on the sample platform; divergence means a Rust port introduced a behavioral difference". That comparison never got wired up: sample-platform#1122 pointed the platform at the "(with migrations)" builds only (mod_ci/controllers.py:264-268), and no workflow compares the two artifacts. In #2109 you mentioned Rust ports won't be merged until this kind of C-vs-Rust check exists.
I built both variants from master on macOS (same source, only -DDISABLE_RUST differs) and compared them on public samples from https://streams.videolan.org/streams/ts/ with 5 option sets each:
| Sample | Result |
|---|---|
Closed Caption EIA.ts |
identical |
atsc-cea-708-dtvcc-and-cea-608.ts (first 20 MB) |
608 identical; the min-rust build writes no 708 file (Rust: 5 cues), both exit 0 |
CC/wabc-4pm-ota.clip1.ts (US broadcast) |
608 identical (40 cues); the min-rust build writes no 708 file (Rust: 42 cues), both exit 0 |
CC/hevc_cc.ts |
min-rust build writes nothing (exit 10) although it logs "Total caption blocks received: 762"; Rust extracts 608 + 708 |
I'd like to build the comparison:
- A CI job that runs both Linux artifacts on a small public sample set and reports differences (non-blocking at first).
- Triage each difference (fix the C path, or record it as expected).
- Then use it as the check for Rust ports.
Two questions before I start:
- Should this live in this repo's GitHub Actions, or in the sample platform?
- Where should the samples come from (download from VideoLAN with caching, or the sample platform)?
Activity
Deepak-negi11 commentedon Oct 18, 2026
Update: I found why the C-only (min-rust) build loses all 708 captions.
In #1738 the C code that sends 708 data to the C decoder was removed. Since then, with DISABLE_RUST, the data goes to the Rust decoder instead, which isn't running in that build, so the captions are silently dropped.
Fixing that showed two more problems in the C-only build:
I've opened a #2399 that fixes all three (the normal Rust build's output doesn't change). With it, the US broadcast sample gives identical output from both builds. Two differences remain: HEVC captions (only the Rust code handles them, #1852) and a 1 ms timing difference on one sample.
I also have the comparison job working locally. I'll wait for your answer on where it should live before opening it.