Metal passthrough shaders have limitations

[inner-daemons]

Metal passthrough shaders don't expose options like wg_memory_sizes, which is required for threadgroup memory to my knowledge. Therefore, passthrough shaders can't use threadgroup memory.

It comes from this code in wgpu_hal::metal::Device::create_compute_pipeline. This will also have an impact on mesh shaders when they arrive.

let cs = if let ShaderModuleSource::Passthrough(desc) = &module.source {
    CompiledShader {
        library: desc.library.clone(),
        function: desc.function.clone(),
        wg_size: MTLSize::new(
            desc.num_workgroups.0 as u64,
            desc.num_workgroups.1 as u64,
            desc.num_workgroups.2 as u64,
        ),
        wg_memory_sizes: vec![],
        sized_bindings: vec![],
        immutable_buffer_mask: 0,
    }
} else {
    self.load_shader(
        &desc.stage,
        &[],
        desc.layout,
        MTLPrimitiveTopologyClass::Unspecified,
        naga::ShaderStage::Compute,
    )?
};

I'm not exactly sure what sized_bindings or immutable_buffer_mask refer to, but they might cause their own limitations as well.