groupshared-over-32k-without-attribute

Status: shipped (Phase 8) — see CHANGELOG.

What it detects

Total groupshared allocation in a translation unit exceeding the SM 6.10 default of 32 KB without a [GroupSharedLimit(<bytes>)] attribute on the entry point.

Why it matters on a GPU

Per HLSL Specs proposal 0049 (Accepted), the SM 6.10 default groupshared cap is 32 KB; exceeding it requires the [GroupSharedLimit(N)] attribute. On SM 6.10 retail this compile-errors; on SM <= 6.9 the LDS allocation is silently truncated, producing out-of-bounds writes in groupshared. RDNA 3 and Turing/Ada have 64 KB / 100 KB / 128 KB caps respectively that the attribute opts the kernel into; without it the driver enforces the conservative ceiling.

Examples

Bad

groupshared float lds[9000]; // 36 KB
[numthreads(64, 1, 1)]
void cs_main() { lds[0] = 0; }

Good

groupshared float lds[9000]; // 36 KB
[GroupSharedLimit(65536)]
[numthreads(64, 1, 1)]
void cs_main() { lds[0] = 0; }

Options

none

Fix availability

machine-applicable — shader-clippy fix inserts [GroupSharedLimit(<round-up-to-64KB>)] before the entry-point declaration.

See also

• Related rule: groupshared-too-large
• HLSL Specs proposal 0049