numwaves-anchored-cap

Status: shipped (Phase 8) — see CHANGELOG.

What it detects

A [numthreads(X, Y, Z)] declaration where X * Y * Z > 1024 (the current per-thread-group lane cap). Defensive rule for HLSL Specs proposal 0054 (numWaves, under-consideration).

Why it matters on a GPU

The current per-group lane cap is 1024 across every modern IHV (D3D12 spec limit). Exceeding it is a hard validator error in DXC -- but the proposal 0054 horizon adds a numWaves attribute that may relax this. Until 0054 ships, exceedance is an error worth flagging early.

Options

none

Fix availability

suggestion — Reduce the dispatch shape or wait for proposal 0054 to ship.