Storage buffer: multidimensional array

[Alphapage]

Hello,

Is it possible to create a multidimensional array like this:
let mut storage: Vec<Vec<u32>> = Vec::new();

let storage_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
            label: None,
            contents: bytemuck::cast_slice(&storage[..]), // don't know how to deal with this here.                   
            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
        });

I would like to get from wgsl:
var<storage,read> storage: array<array<u32>>;

Thank you in advance for your answers.

[kpreid]

A Vec<Vec<u32>> is a pointer to a memory allocation containing pointers to memory allocations containing u32s. You can't copy it directly into a GPU buffer for the same reason bytemuck doesn't straightforwardly accept it: it's not a single allocation, and it contains pointers which would be meaningless (or at least unsafe) on the GPU. For the same kind of reason, array<array<u32>> is not a valid WGSL array type.

What you need to do is design a data format which does use a single allocation for the data.

• If the data is rectangular (each inner Vec is the same length) then this is simple: use a single Vec<u32>, and compute “flattened” indices in it as x + width * y — the same way you'd access the texels in a texture. (Rust libraries like ndarray can help do this for you.) This is not just for the GPU's sake — the CPU side can also have increased performance by not needing to follow many different pointers to read or write the data, and so Vec<Vec<T>> is an inferior data structure for almost all rectangular storage applications.
• If the data is not rectangular (“ragged”) then you will still have to use a single vector, but also separately store offsets (in a separate Vec and a separate wgpu::Buffer) to be able to discover where each row starts and compute the flattened index from that information.

[Alphapage]

Yes, I will use this workaround...
But I hope it will be possible in the future.
Wait and see.
Many thanks.