Support for arrays of textures

[muni-corn]

I just wanted to get the ball rolling on support for arrays of textures in shaders. Coming from this issue on the wgpu-rs repo: gfx-rs/wgpu#106

The idea is to be able to use an array of 2d textures in shaders, as originally shown in the issue above:

layout(set = 0, binding = 1) uniform texture2D textures[1024];

[krogovin]

What are the expectations for the index into the array? Can it be anything, or must it be "dynamically uniform", i.e. dependent only on values that are derived from uniforms only?

[kvark]

wgpu on native exposes the restrictions on the index as different extra features:

• SAMPLED_TEXTURE_ARRAY_DYNAMIC_INDEXING
• SAMPLED_TEXTURE_ARRAY_NON_UNIFORM_INDEXING

cc @cwfitzgerald

[jozanza]

In WGSL, how are we able to access a single texture in an array of textures? A binding such as this seems valid:

[[group(1), binding(0)]] var textures: array<texture_2d<f32>, 2>;

However, any attempt to access an index via textures[n] throws various compiler errors. Is there some special array access notation or am I just missing something?

The texture arrays example only has glsl shaders, and I couldn't find info about this in the spec other than the basic array notation syntax.

Edit: This non-uniform shader is the one I'm interested in writing in WGSL. That "flat" specifier also seems to cause errors everywhere I've tried to place it.

[litherum]

I remember that we had a resolution in the group to not allow arrays of textures (as distinct from texture arrays) in the first version of the language, because of uniformity rules. Texture arrays are allowed, though.

[jozanza]

I did see that this was supported by wgpu on native with certain extensions. But you're saying there's currently no valid syntax for working with an array of textures in WGSL?

[Kangz]

I think you're referring to bindless, and that hasn't been discussed for WebGPU at this point. Because the API doesn't support bindless (even as an extension), it doesn't make sense to add support for it in WGSL. Experimental syntax can be added to wgpu-rs / Naga though.

[kvark]

Correct, we'll need to work on WGSL syntax for this internally (prototyping in Naga), and then potentially returning to the group with a proposal.

[jozanza]

Thanks for the comments!

I'm not sure what "bindless" means tbh 👀, but to clarify, is it correct to say there's no way to achieve dynamic non-uniform texture indexing/sampling in WGSL at this point? Like those example shaders cannot be converted to WGSL yet?

[kvark]

That's correct, yes. At least, not in the "standard" WGSL.

[kainino0x]

You should be able to use conditionals to choose dynamically between branches that sample different textures (or perhaps even assign different textures into a pointer variable?). But we don't have a way to bind arrays of textures at the API surface, so you still have to have a separate binding point for each texture.

[jozanza]

@kainino0x I had trouble taking that approach because, when compiling my shader, it apparently violated some non-uniformity restriction to selectively use a texture based on an in var from the vertex shader. So I ended up just rolling with a texture atlas.

[kainino0x]

Got it. I think it should* be valid if the conditional is uniform, but otherwise it will not be.

* I don't know if the uniformity rules are actually written out yet, but ultimately I think this will be the case.