[glsl-out] fix usage of gsamplerCubeArrayShadow

[kvark]

from bevyengine/bevy#841 (comment)

Internal error in FRAGMENT shader: gsamplerCubeArrayShadow isn't supported in textureGrad, textureLod or texture with bias

[magcius]

This is literally impossible -- the GLSL spec doesn't define any of those functions for gsamplerCubeArrayShadow. https://www.khronos.org/registry/OpenGL/specs/gl/GLSLangSpec.4.60.pdf#page=183

Neither does WGSL: https://gpuweb.github.io/gpuweb/wgsl/#texturesamplecompare

What does the input shader look like?

[kvark]

We have a workaround that uses textureGrad instead of textureLod on cubemaps. It's probably involved:
https://github.com/gfx-rs/naga/blob/54178dede2a8b3f6304e3c71081db31189f786ee/src/back/glsl/mod.rs#L1963

[magcius]

That's after the code that's outputting that error.

[JCapucho]

@kvark as magicus pointed out this is impossible, this workaround doesn't work on gsamplerCubeArrayShadow and the only functions that do support it are textureSize, texture and textureGather, but those would could potentially violate uniform control flow rules so I don't think there's anything we can do here.

[kvark]

If we can't do this, we should return an error from the backend (instead of generating bad code).

[magcius]

We're not generating bad code -- we're returning an error (admittedly a confusing one). What is the WGSL that's generating this error?

[kvark]

Damn, you are right.
@jakobhellermann could you provide the WGSL code for the shader triggering this? Or was it SPIR-V/GLSL source?

[jakobhellermann]

The code in question is https://github.com/bevyengine/bevy/blob/963e2f08a2f762d896eb28ef4a80ba2ec124655b/crates/bevy_pbr/src/render/pbr.wgsl#L387.

Minimal example:

[[group(0), binding(2)]]
var point_shadow_textures: texture_depth_cube_array;
[[group(0), binding(3)]]
var point_shadow_textures_sampler: sampler_comparison;

[[stage(fragment)]]
fn main() {
    let frag_ls = vec3<f32>(0.0);
    let depth: f32 = 0.0;
    let value = textureSampleCompareLevel(point_shadow_textures, point_shadow_textures_sampler, frag_ls, i32(0), depth);
}

> naga repro.wgsl out.frag
gsamplerCubeArrayShadow isn't supported in textureGrad, textureLod or texture with bias

[magcius]

This could be supported if with this OpenGL extension: https://www.khronos.org/registry/OpenGL/extensions/EXT/EXT_texture_shadow_lod.txt

Obviously not going to work with WebGL, though. If you only have a single mipmap in your cube (not sure why you'd have more?), consider using textureSampleCompare instead. Another workaround would be to pack them into a depth texture array and do the 3D -> 2D array coordinate conversion in the shader.

[jakobhellermann]

I believe the reason for using the textureSampleCompareLevel function instead of textureSampleCompare is that the latter must be called in uniform control flow, which isn't the case here.

// NOTE: Due to the non-uniform control flow above, we must use the Level variant of
// textureSampleCompare to avoid undefined behaviour due to some of the fragments in
// a quad (2x2 fragments) being processed not being sampled, and this messing with
// mip-mapping functionality. The shadow maps have no mipmaps so Level just samples
// from LOD 0.

[magcius]

boo... I understand the WGSL compiler can't possibly know that the texture has no mipmaps, but this sure is unfortunate. As a hack, Naga might be able to generate texture() in this case and pray it has no mipmaps? All options are unfortunate.

[magcius]

I realize immediately below the code that errors out is code that checks for TEXTURE_SHADOW_LOD. Assuming the user has support for TEXTURE_SHADOW_LOD, this shouldn't be impossible to fix.

[teoxoy]

#5171 should have resolved this, closing.