Bindless variable length buffers

[azureblaze]

I am working on a voxel rendering system, where a mesh has a "texture" that is an array of sparse voxel octree nodes in a storage buffer of varying length (ideally up to 64k):

// WGSL
struct VoxelTexture {
    nodes: array<SparseVoxeOctreeNode >}

struct SparseVoxeOctreeNode {
    // 8 u16 indices packed into UVec4
    children: vec4<u32>,
}

I am trying to to make this texture bindless:

// Rust
#[derive(ShaderType)]
struct VoxelTexture {
    #[shader(size(runtime))]
    nodes: Vec<SparseVoxelOctreeNode>,
}

#[derive(Asset, TypePath, AsBindGroup, Debug, Clone, Default)]
#[bindless]
struct VoxelMaterial {
    #[storage(1, read_only, binding_array(10))]
    texture: Handle<ShaderStorageBuffer>,
}

struct MaterialBindings {
    material: u32,              // 0
}
@group(#{MATERIAL_BIND_GROUP}) @binding(0) var<storage> materials: array<MaterialBindings>;
@group(#{MATERIAL_BIND_GROUP}) @binding(10) var<storage> textures: binding_array<VoxelTexture>;

However, when trying to index into the texture array I think the shader compiler just gives up:

error: failed to build a valid final module: Entry point fragment at Fragment is invalid
   ┌─ shaders/shader.wgsl:86:19
   │
86 │     let texture = textures[materials[slot].material];
   │                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ naga::ir::Expression [15]
   │
   = Expression [15] is invalid
   = Loading of [14] can't be done

I guess because textures[] is now an array of items with unknown size. I thought binding_array is meant to handle this as it should contain pointers to the buffers, instead of being huge combined buffer.

Solutions I can think of to get around this:

• Give up on bindless
  • :(
• Make the node count in the array fixed
  • shader initialization seems to slow down at 2k nodes and nothing happens after 16k.
• use storage textures instead
  • bindless is not supported for storage textures?
• convert the octree and store them in unfiltered u32 textures
  • very ugly

Is there some way that can get this to work?