Struct u32 field is always 0

[Steveplays28]

Hi, I'm trying to move this struct from the CPU to the GPU, but the material field is always 0. On the other hand, the position and radius fields get sent over correctly.
I tried moving the padding fields around, but the order doesn't matter. This same struct fully works as intended if I use position: Vec4, radius: Vec4 and material: i32 instead, which is what I had previously (I wanted to change it to Vec3 for ease of use).

Why would this occur?

#[derive(Debug, Clone, Copy, Default, Pod, Zeroable)]
#[repr(C)]
pub struct Voxel {
    pub position: Vec3,
    pub radius: Vec3,
    pub material: u32,
    _padding_0: u32,
    _padding_1: u32,
    _padding_2: u32,
}

[LegNeato]

Hello! Do you have a public repo I can look at to try to debug?

[Steveplays28]

Hi, currently my repository is private. I could invite you to it, if that's fine by you?

EDIT: I've sent you an invite, I'll also commit my changes to a new branch (simplify-gpu-structs)

[LegNeato]

Great, I'll try to take a look at it tomorrow

[LegNeato]

Poking at this now.

Edit: Ended up not having enough time tonight, will look into it later.

[Firestar99]

Currently, Vec3 has a different representation between the CPU (12 bytes) and GPU (16 bytes). So for now, please just use Vec4 for any structs that are communicated between host and device.

@LegNeato maybe we should rebase this PR, which iirc makes Vec3 be 12 bytes on the GPU as well. (I've got a different workaround in my project, basically I convert all glam Vec's to [f32; 3] and then layout always matches)

[LegNeato]

I think by the spec it has to be aligned? Isn't this why glam has the cuda trait we use, same reason on cuda? A better fix would be use cuda feature on CPU?

[Firestar99]

This extension enables C-like structure layout for SPIR-V blocks. It modifies the alignment rules for uniform buffers, storage buffers and push constants, allowing non-scalar types to be aligned solely based on the size of their components, without additional requirements.

https://registry.khronos.org/vulkan/specs/latest/man/html/VK_EXT_scalar_block_layout.html

This extension is ubiquitous nowadays, don't know about the wgpu side though

[Steveplays28]

thanks both!

[nazar-pc]

Is this still considered a bug that should be fixed?

IMO this is a very surprising behavior that is not obvious even after reading the type defintion.