entry point fn with too many params fails

[galop1n]

I am trying to convert a project of mine from wgsl to rust-gpu. I got the basics done, but I have a shader I can't get to compile

This is the entry point

#[spirv(compute(threads(8, 8, 1)))]
pub fn cs_main(
    #[spirv(global_invocation_id)] gid: UVec3,
    #[spirv(uniform, descriptor_set = 1, binding = 0)] globals: &SBGlobalData,
    #[spirv(uniform, descriptor_set = 2, binding = 0)] camera: &SBCameraData,
    #[spirv(storage_buffer, descriptor_set = 3, binding = 0)] tlas_nodes: &[BvhNode],
    #[spirv(storage_buffer, descriptor_set = 3, binding = 1)] tlas_instances: &[TlasInstance],
    #[spirv(storage_buffer, descriptor_set = 3, binding = 2)] blas_nodes: &[BvhNode],
    #[spirv(storage_buffer, descriptor_set = 3, binding = 3)] blas_tris: &[Triangle],
    #[spirv(descriptor_set = 3, binding = 4)] rt_out: &Image!(2D, format = rgba16f, sampled = false, arrayed = false, multisampled = false),
    #[spirv(uniform, descriptor_set = 3, binding = 5)] rt_consts: &RaytracerConsts,
    #[spirv(uniform, descriptor_set = 3, binding = 6)] mat_consts: &RtLightConsts,
    #[spirv(storage_buffer, descriptor_set = 3, binding = 7)] rt_materials: &[RtMaterialEntry],
    #[spirv(descriptor_set = 3, binding = 8)] albedo_sampler: &Sampler,
    #[spirv(descriptor_set = 4, binding = 0)] albedo_tex_table: &RuntimeArray<Image!(2D, type = f32, sampled = true, arrayed = false, multisampled = false)>,
) {

Expected Behaviour

Shader compiles fine

Example & Steps To Reproduce

This is the errors I am getting

info: latest update on 2026-04-11 for version 1.96.0-nightly (02c7f9bec 2026-04-10)
info: component rust-src is up to date
warn: skipping unavailable component rust-std for target spirv-unknown-vulkan1.2
   Compiling shader-bench-rsgpu-user v0.1.0 (F:\rust\ideas\shader-bench\projects\raytracer-rs\.shader-bench\shader-cache\rust-gpu-build\ce910f26f4188d71)
warning: an unknown error occurred
  |
  = note: spirv-opt failed, leaving as unoptimized
error: [VUID-StandaloneSpirv-OpTypeRuntimeArray-04680] OpVariable, <id> '14[%albedo_tex_table]', is attempting to create memory for an illegal type, OpTypeRuntimeArray.

For Vulkan OpTypeRuntimeArray can only appear as the final member of an OpTypeStruct, thus cannot be instantiated via OpVariable
         %albedo_tex_table = OpVariable %_ptr_UniformConstant__runtimearr_94 UniformConstant

Bonus: u32.saturated_sub is not implemented btw

System Info

Windows + custom compiler wrap to compile rust-gpu shader at runtime without having my whole workspace on the specific nightly toolchain

[Firestar99]

For Vulkan OpTypeRuntimeArray can only appear as the final member of an OpTypeStruct, thus cannot be instantiated via OpVariable

So usually this is indicative of you declaring a struct containing a slice or RuntimeArray, which we currently don't properly support for silly reasons. Now I don't believe based on your screenshot in discord that you're doing that.

I've seen rustc rewriting function ABI from passing each argument individually to putting all arguments into a struct and changing the function ABI to just passing a ptr to that struct, when a function has a lot of arguments. I suspect this is happening to your entry point function, which I haven't seen happen yet. Unfortunately this won't be an easy fix...

Some months ago I've been investigating using inline assembly for declaring inputs, outputs, built-ins and storage buffer declarations, which are an alternative way to declare entry points and shouldn't suffer from function ABI limitations. I want to eventually make a new entry point systen to use inline asm like that, but haven't gotten around to finishing it off.

[galop1n]

I can close the bug for now, I found a fix! I needed to set Capability::RuntimeDescriptorArray on the spirv build object. Coming from HLSL+PSSL shader compilers. I find very unintuitive and cumbersome that you have to set dozens of flags for every single little GPU feature your shader may use ! Be it with NAGA/WGSL or now rust-gpu !

Is there a reason for it to be so granular, rather than let's say, a more general feature tier. Example, in HLSL you build for a given shader model, like 6.7, and you get all it encompass. Most features been mandatory for that tier, and optional feature uses just set flags in the reflection data you can compare to your device capabilities ( ex: u64 atomic )