WebGPU / standalone diagnostic
WebGPU subgroup challenges
Compare advertised subgroup bounds with shader acceptance, executed results and exact diagnostics. Run the same grid in this page and a dedicated worker.
Starts on request and saves no session. The latest capture stays in this browser for recovery after a reload. Download each run before replacing it.
Waiting for JavaScript. No checks have run.
Challenge results
No capture yet. Default bounds are supplied expectations, not observations.
A finding can be a contract violation or an execution pattern. WGSL permits partially occupied fragment groups; repeated occupancy of at most four lanes is a pattern, not proof of a conformance failure.
| Context | Challenge | Pipeline | Requested | Observed sizes | Result | Evidence |
|---|
Environment and capture events
No capture yet.
Confidence and limits
This grid is a regression diagnostic, not WebGPU conformance certification. Strict divergent-execution checks describe the candidate backend; other valid implementations may differ. Per-case wall time includes compilation, pipeline creation, execution and readback. It is not a GPU benchmark.
For broader confidence, run a pinned WebGPU CTS revision against baseline and candidate builds, then compare repeated warm and cold workloads on the same machine. Keep failures, skips, device loss and resource use alongside timings. A fast failed workload is not a performance result.
PR reproduction links can select cases and context: fragment and depth checks · override and diagnostic checks. Add &autorun to start on navigation. ?json=true exposes the same export after an explicit run.