wgsl-compute-comparison-sampling

name: wgsl-compute-comparison-sampling description: textureSampleCompare is fragment-stage-only — compute shaders must use textureSampleCompareLevel with 4 params (no LOD, no offset). Also: bind group layouts must match pipeline stage visibility. When adding passes to shared RONs, all benches inherit them. source: auto-skill extracted_at: '2026-06-16T01:35:34.704Z'

WGSL Compute Comparison Sampling + Pass Inheritance

Two distinct traps found during scene culling integration (2026-06-15).

Trap 1: textureSampleCompare in compute shaders

textureSampleCompare(t, s, coords, depth_ref) is fragment-stage-only — it uses implicit LOD which requires fragment derivative hardware. Compute shaders must use the explicit-LOD variant:

// WRONG — fragment-only, will fail naga validation:
return textureSampleCompare(vsm_physical_pool, vsm_sampler, phys_uv, depth_ref);

// CORRECT — compute-safe, implicit LOD 0:
return textureSampleCompareLevel(vsm_physical_pool, vsm_sampler, phys_uv, depth_ref);

Why: WGSL § texture builtins: comparison sampling with implicit LOD (textureSampleCompare) requires fragment stage. The Level variant takes an explicit LOD that the return type supports, but with 4 params it defaults to level 0. Do NOT add a 5th 0i param — WGSL interprets the 5th param as a vec2<i32> offset for texture_sample_2d, not as the LOD, and naga rejects it with "Sample offset constant doesn't match the image dimension."

Verification: textureSampleCompare → textureSampleCompareLevel in all .wgsl files under crates/. The commit hook at .qwen/local-scripts/rendering_naming_guard.py does not catch this — it blocks proprietary names, not WGSL API misuse. A manual grep at review time is the only gate.

Trap 2: Bind group layout stage visibility mismatch

When creating a simplified pipeline (e.g. depth-only prepass) that reuses a bind group from the main rendering pipeline, the bind group's BGL must match the pipeline's BGL stage visibility:

// WRONG — frame_bind_group() was created with VERTEX|FRAGMENT visibility,
// but the depth prepass pipeline expects VERTEX-only at group 0:
rp.set_bind_group(0, self.viewport.frame_bind_group(), &[]);

// CORRECT — create a dedicated camera BG matching the pipeline's BGL:
let cam_bgl = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
    entries: &[wgpu::BindGroupLayoutEntry {
        binding: 0,
        visibility: wgpu::ShaderStages::VERTEX,  // match the pipeline
        ty: wgpu::BindingType::Buffer {
            ty: wgpu::BufferBindingType::Uniform,
            has_dynamic_offset: false,
            min_binding_size: None,
        },
        count: None,
    }],
    label: Some("depth_prepass camera bgl (per-frame)"),
});

Why: wgpu validates BGL compatibility at set_bind_group time. A BGL with VERTEX|FRAGMENT is not compatible with one that has only VERTEX, even if both have the same binding count and types.

Trap 3: Shared RON pass inheritance

When new passes are added to shared RON configs (pipeline_deferred_stratum_mid.ron, etc.), ALL benches that use those configs inherit the passes. If a bench doesn't populate the pass-specific data structures (e.g. SceneCullData), the dispatch arm must gracefully fall through:

// Required pattern for every new pass dispatch arm:
if let Some(ref data) = self.pass_specific_data {
    // GPU-driven path
} else {
    // Fallback: warn once and skip, or use legacy path
    self.warn_once_unimplemented_pass("PassName (data not populated)", "boot");
    continue;
}

Verification checklist when adding new passes:

1. Run cargo run -p titan-render-bench-probe -- --capture-fps 60 --capture-start 0 --capture-end 1 to smoke-test the probe bench
2. If it crashes, the new pass is interfering with an existing bench
3. Add fallback path in the dispatch arm