samplecmp-vs-manual-compare
Status: stub. The full-length analysis is queued for a v1.0.x patch release per ADR 0018, section 5, criterion #6. The companion rule page at docs/rules/samplecmp-vs-manual-compare.md contains the canonical detection logic + GPU reasoning.
TL;DR
The TMU on all current GPU architectures (AMD RDNA / RDNA 2 / RDNA 3, NVIDIA Turing / Ada Lovelace, Intel Xe-HPG) includes a dedicated hardware comparison unit on the texture pipeline. SampleCmp with a SamplerComparisonState routes the reference depth into this unit, which performs a bilinear comparison across the 2x2 texel footprint in a single TMU issue. On RDNA 3, a SampleCmp completes in the same number of TMU cycles as a plain Sample for many formats, because the comparison is fused into the filter stage at no additional ALU cost. The output is a single float in [0, 1] representing the filtered shadow term.
What the rule fires on
A pattern where a shader samples a depth or shadow texture with a standard Sample or Load call and then performs a scalar comparison (<, >, <=, >=) on the fetched depth value against a reference depth variable. The rule fires when both conditions hold: the sampled texture is a declared depth resource (reflection type Texture2D<float> or Texture2D<float4> bound as a shadow map), and the comparison result drives a branch or is used as a shadow weight without passing through SampleCmp. It does not fire on uses of SampleCmp or SampleCmpLevelZero directly, nor when the comparison is against a constant rather than a per-pixel reference depth.
See the What it detects section of the rule page for the full pattern definition.
Why it matters
The full GPU-mechanism analysis lives in the Why it matters on a GPU section of the companion rule page.
Examples
The bad / good code snippets are kept canonical on the rule page; see samplecmp-vs-manual-compare.md -> Examples.
See also
- Rule page -- canonical reference + change log.
- texture overview -- broader context.
- ADR 0018 -- v1.0 readiness plan.
This is a v1.0-ship stub. Full analysis pending; track issue link TBD.