clip-from-non-uniform-cf

Status: shipped (Phase 4) — see CHANGELOG.

(via ADR 0011)

What it detects

A call to clip(x) (the HLSL component-wise discard intrinsic that retires the lane when any component of x is negative) inside a pixel-shader entry point whose containing function lacks the [earlydepthstencil] attribute and is reachable from non-uniform control flow — that is, the path from entry to the clip call passes through at least one branch whose condition is per-lane varying. Distinct from the locked early-z-disabled-by-conditional-discard, which fires on discard; clip(x) has its own semantics (retire on negative component, threshold per channel) and an independent suppression scope.

Why it matters on a GPU

Pixel-shader early-Z (and stencil) testing is a critical bandwidth optimisation: the rasteriser tests depth before invoking the pixel shader, so occluded fragments never run. A pixel shader that may issue clip(x) (or discard) modifies the depth attachment as a side effect of the shader body, which forces the hardware to defer depth update until after shader execution. On AMD RDNA 2/3, the GE / SX hardware downgrades from "early-Z + early-stencil" to "late-Z" for the entire pipeline state when the shader contains a clip / discard reachable on any path — losing the bandwidth savings on every fragment, not just the clipped ones. NVIDIA Ada applies the same downgrade per-pipeline-state; Intel Xe-HPG behaves analogously. The shader author can opt back into early-Z by adding the [earlydepthstencil] attribute, which promises the hardware that the depth value is unaffected by the shader (the clip / discard then retires the lane after the early-Z test has already updated depth, accepting the resulting incorrect depth in exchange for the bandwidth recovery — a trade-off only the author can authorise).

The non-uniform-CF flavour is the worst case because the optimiser cannot hoist the clip to a uniform location: lanes diverge on the predicate and clip independently. Even on hardware that supports per-lane retirement without warp-level disruption, the resulting pixel-quad fragmentation costs derivative-uniformity — the surviving lanes in a quad may need helper-lane participation to maintain ddx/ddy accuracy, and the retired lanes are still invoked as helpers (consuming ALU). The bandwidth cost from losing early-Z is the dominant term; the helper-lane fragmentation is a second-order tax.

The fix is one of: (a) add [earlydepthstencil] if the algorithm tolerates the depth-test-before-clip ordering, (b) hoist the clip to a uniform predicate (typical: clip on a wave-uniform function-of-vertex-input), or (c) restructure to use alpha-test in the rasteriser pipeline state instead of an in-shader clip. Each has cost; the rule surfaces the hazard rather than mandating a resolution. The rule is intentionally distinct from early-z-disabled-by-conditional-discard because authors expect to suppress one without the other — clip and discard have different semantics in source even when they collapse to the same HW retirement.

Examples

Bad

hlsl

Texture2D<float4> g_Mask : register(t0);
SamplerState g_Samp : register(s0);

float4 ps_clipped(float2 uv : TEXCOORD0) : SV_Target {
    float4 mask = g_Mask.Sample(g_Samp, uv);
    // Per-pixel divergent clip — disables early-Z for the whole pipeline state
    // because no [earlydepthstencil] attribute on the entry point.
    clip(mask.a - 0.5);
    return float4(shade(uv), 1.0);
}

Good

hlsl

// Option A — opt into early-Z, accepting that depth updates before the clip
// retires the lane. Valid only if the depth value the shader writes is
// independent of the clip outcome.
Texture2D<float4> g_Mask : register(t0);
SamplerState g_Samp : register(s0);

[earlydepthstencil]
float4 ps_clipped_earlyz(float2 uv : TEXCOORD0) : SV_Target {
    float4 mask = g_Mask.Sample(g_Samp, uv);
    clip(mask.a - 0.5);
    return float4(shade(uv), 1.0);
}

// Option B — move the clip to alpha-test in the PSO and drop the in-shader clip.
// (Pseudocode — done in the application's pipeline-state setup, not the shader.)
float4 ps_no_clip(float2 uv : TEXCOORD0) : SV_Target {
    return float4(shade(uv), g_Mask.Sample(g_Samp, uv).a);
}

Options

none

Fix availability

suggestion — Adding [earlydepthstencil] is a semantic change (depth update reordering); moving to alpha-test is an application-side change. The diagnostic identifies the clip call, the divergent predicate path, and the missing attribute so the author can decide.

clip-from-non-uniform-cf ​

What it detects ​

Why it matters on a GPU ​

Examples ​

Bad ​

Good ​

Options ​

Fix availability ​

See also ​