coopvec-transpose-without-feature-check

Status: shipped (Phase 3) — see CHANGELOG.

(via ADR 0010)

What it detects

A cooperative-vector matrix-multiply call whose transposed flag is true (MATRIX_FLAG_TRANSPOSED set) without a corresponding application-side feature check on D3D12_FEATURE_DATA_D3D12_OPTIONS18.CooperativeVectorTier. The SM 6.9 cooperative-vector spec marks the transpose path as conditionally supported: tier 1 implementations may not honour it, and a runtime call with transpose set on an unsupporting device fails. The rule surfaces every transpose use so the developer can confirm the tier check is in place.

Why it matters on a GPU

The transpose path lets the application swap the matrix's effective row/column orientation at the matmul site without re-uploading the data. On NVIDIA Ada Lovelace, the tensor cores expose hardware transpose as a free-rate option in the inference path; on AMD RDNA 3/4 WMMA, transpose support is gated behind a driver feature flag because the WMMA hardware lacks the on-the-fly swizzle and the driver emulates it via a software path; on Intel Xe-HPG XMX engines, transpose is supported but the cost varies by component type. The runtime exposes all of this through a single tier value: if the device is tier 0 or below the per-IHV transpose tier threshold, the transpose call fails.

The failure mode is a runtime device-removed event — the DXR-style "your dispatch crashed the device" error — not a clean rejection. Surfacing the transpose use at lint time gives the developer a chance to add the feature check, supply a fallback path that uploads a transposed copy of the matrix, or document the minimum-tier requirement in the project's README.

The rule is suggestion-tier because legitimate transpose use is part of the cooperative-vector design surface; the linter cannot tell whether the application has the feature check elsewhere. The diagnostic includes a one-line snippet of the recommended CheckFeatureSupport call.

Examples

Bad

hlsl

// Transpose without an application-side tier check — fails on tier-0 devices.
ByteAddressBuffer g_Weights : register(t0);

[numthreads(32, 1, 1)]
void main(uint tid : SV_DispatchThreadID) {
    using namespace dx::linalg;
    vector<float, 16> input  = LoadInput(tid);
    vector<float, 16> output;
    MatrixVectorMul(output, input,
                    g_Weights, 0, 64,
                    MATRIX_LAYOUT_INFERENCING_OPTIMAL,
                    MATRIX_FLAG_TRANSPOSED);     // requires tier check
}

Good

hlsl

// Application-side, before dispatch:
//   D3D12_FEATURE_DATA_D3D12_OPTIONS18 opts = {};
//   device->CheckFeatureSupport(D3D12_FEATURE_D3D12_OPTIONS18, &opts, sizeof(opts));
//   if (opts.CooperativeVectorTier >= D3D12_COOPERATIVE_VECTOR_TIER_1_1) { ... }
//
// Shader path with transpose is unchanged; an upload-time fallback copies
// the matrix transposed when the tier is below the threshold.

Options

none

Fix availability

suggestion — The fix is an application-side feature check + a fallback upload path. The diagnostic names the call site and emits the recommended CheckFeatureSupport snippet as a comment.

coopvec-transpose-without-feature-check ​

What it detects ​

Why it matters on a GPU ​

Examples ​

Bad ​

Good ​

Options ​

Fix availability ​

See also ​