long-vector-typed-buffer-load

Status: shipped (Phase 3) — see CHANGELOG.

(via ADR 0010)

What it detects

A Buffer<vector<T, N>> declaration with N >= 5, or a typed-buffer Load/operator[] whose returned type is a long vector. The SM 6.9 long-vector spec restricts typed-buffer (Buffer<T> / RWBuffer<T>) views to the legacy 1/2/3/4-wide vector types because typed-buffer fetches go through the texture cache, which expects DXGI-format-compatible types. Long vectors must be loaded through ByteAddressBuffer or StructuredBuffer. Slang reflection identifies the resource type at binding sites; the rule fires when the typed-view's element type is a long vector.

Why it matters on a GPU

Typed buffers (Buffer<float4>, RWBuffer<uint2>, etc.) are texture-cache-backed views: the texture units on every IHV (NVIDIA's L1, AMD RDNA's TC, Intel Xe-HPG's L1 sampler cache) accept the load with a DXGI format descriptor, which the hardware uses to pick the right fetch shape and the right format conversion. Long vectors have no DXGI format equivalent — there is no DXGI_FORMAT_R32G32B32A32B32C32D32E32F32_FLOAT because the texture cache hardware was never designed for 32-byte fetches. The DXC validator rejects the type combination at PSO compile.

Slang and DXC both surface this as a clean error; the lint catches it earlier and points the author at the right resource type. ByteAddressBuffer (raw, untyped) and StructuredBuffer<MyStruct> (struct-typed, raw) both go through the K$ / scalar-data cache rather than the texture cache and have no DXGI-format restriction. They are the right targets for long-vector data.

The performance trade-off (texture cache vs. scalar cache) is real and is captured by the sibling rule byteaddressbuffer-narrow-when-typed-fits (perf-tier suggestion to move 1/2/3/4-vec data to typed views when it fits). For long vectors, the choice is forced: typed views are not an option.

Examples

Bad

hlsl

// Typed buffer cannot hold a long vector.
Buffer<vector<float, 8>> g_LongVecData : register(t0);  // ERROR: typed long vec

[numthreads(32, 1, 1)]
void main(uint tid : SV_DispatchThreadID) {
    vector<float, 8> v = g_LongVecData[tid];   // load via typed view
}

Good

hlsl

// StructuredBuffer / ByteAddressBuffer for long-vector data.
struct LongVec { vector<float, 8> v; };
StructuredBuffer<LongVec> g_LongVecData : register(t0);

[numthreads(32, 1, 1)]
void main(uint tid : SV_DispatchThreadID) {
    vector<float, 8> v = g_LongVecData[tid].v;
}

Options

none

Fix availability

suggestion — Switching the resource type requires a corresponding application-side change to the descriptor; the diagnostic names the offending declaration and the recommended replacement.

long-vector-typed-buffer-load ​

What it detects ​

Why it matters on a GPU ​

Examples ​

Bad ​

Good ​

Options ​

Fix availability ​

See also ​