mip-clamp-zero-on-mipped-texture

Status: shipped (Phase 3) — see CHANGELOG.

(via ADR 0011)

What it detects

A SamplerState whose descriptor pins MaxLOD = 0 (or a Sample/SampleLevel call with clamp = 0 arg, or a MinMipLevel = 0 clamp on a sampler-feedback path) bound against a Texture2D / Texture2DArray / TextureCube resource that reflection reports as carrying more than one mip level. The detector cross-references the sampler's MaxLOD field via reflection with the resource's mip count via reflection. It does not fire on textures that are genuinely single-mip (loading a single-level surface with MaxLOD = 0 is correct and intended); it fires only when mips exist and the sampler clamps them off.

Why it matters on a GPU

Mipmaps on every desktop GPU exist for two reasons: anti-aliasing minified samples (the higher-frequency texels of mip 0 alias under minification, so the hardware blends in coarser mips when the screen-space derivatives indicate minification) and bandwidth amortisation (a coarser mip is 1/4 the texel count and 1/4 the bandwidth). The texture sampler unit on AMD RDNA 2/3 (TMU), NVIDIA Turing/Ada (TEX/L1), and Intel Xe-HPG samples mips by computing a fractional LOD from the screen-space derivatives, then trilinearly blending two adjacent mips. When MaxLOD = 0, the hardware clamps the LOD selection to mip 0 regardless of derivatives — minified samples re-enter the aliasing regime and bandwidth scales with screen footprint instead of texel footprint.

The bandwidth cost is the easier to quantify. A terrain or streaming-virtual-texture pass that loads a 4096x4096 surface with MaxLOD = 0 reads from the full mip 0 working set on every minified sample. With trilinear filtering across mip 0 and mip 1, the same minified pixel reads at 1/4 + 1/16 = ~31% of the mip-0 bandwidth and the L1 cache hit rate climbs because adjacent pixels touch overlapping mip-1 footprints. On RDNA 3 with its 16 KB L1 per CU, the difference shows up as a measurable cache hit rate change in any minified-texture-bound pass.

The aliasing cost is the silent one. MaxLOD = 0 re-introduces shimmering and crawling on minified geometry that the asset's mip chain was authored to suppress. It is also a common copy-paste accident: a developer pins MaxLOD = 0 to debug LOD selection, then forgets to reset it. The rule fires whenever reflection confirms the texture has mips and the sampler discards them.

Examples

Bad

hlsl

// Sampler clamps MaxLOD to 0 — discards every mip the texture carries.
SamplerState ClampedSampler : register(s0);  // MaxLOD = 0 in descriptor
Texture2D<float4> Diffuse  : register(t0);   // 11 mips: 1024 -> 1

float4 sample_terrain(float2 uv) {
    return Diffuse.Sample(ClampedSampler, uv);  // always reads mip 0
}

Good

hlsl

// Default sampler — MaxLOD = D3D12_FLOAT32_MAX; trilinear blend across
// mips driven by ddx/ddy of uv.
SamplerState TrilinearSampler : register(s0);  // MaxLOD = FLT_MAX
Texture2D<float4> Diffuse     : register(t0);

float4 sample_terrain(float2 uv) {
    return Diffuse.Sample(TrilinearSampler, uv);
}

Options

none

Fix availability

suggestion — Removing the MaxLOD = 0 clamp changes the visual output (mip filtering takes over). The diagnostic flags the mismatch; the author confirms the visual change is intended.

mip-clamp-zero-on-mipped-texture ​

What it detects ​

Why it matters on a GPU ​

Examples ​

Bad ​

Good ​

Options ​

Fix availability ​

See also ​