Allow for setting the array length in a shader using a specialization constant

[rohdealx]

To allow for something like this using vulkano-shaders:

layout(constant_id = 0) const uint n = 64;
layout(std430, set = 0, binding = 0) uniform layout_sample { float sample[n]; };

Fixes #1956

Changelog:

### Addditions
- Vulkano-shaders: support for specialization-constant-sized arrays in structs (they are generated with the size specified as fallback in the specialization constant initializer).

[Rua]

That should be already implemented by #2329. Does it not work for you?

[marc0246]

I'm assuming what they meant to say is to allow this usage in vulkano-shaders. In other words, it's attempting to fix #1956. Using the default value of the specialization constant as the length is certainly a solution and one that I considered, but personally I would prefer if the struct wasn't generated at all. It's going to be useless anyway and I bet cause more confusion when it's generated like that. Nevertheless this would be the simplest solution.

[Rua]

What about generating a struct with a const parameter for the length?

[marc0246]

That's also a good option. shaderc vomits up array length 1 in places where runtime lengths are illegal without warning. We could do something similar, length 0 ideally.

[marc0246]

Wait I misread your question. I thought you meant generating it with a const length. Generating a const parameter won't make it runtime-sized.

[marc0246]

Also you would have to propagate the const parameter(s) throughout all uses of the structs.

[rohdealx]

I'm assuming what they meant to say is to allow this usage in vulkano-shaders. In other words, it's attempting to fix #1956. Using the default value of the specialization constant as the length is certainly a solution and one that I considered, but personally I would prefer if the struct wasn't generated at all. It's going to be useless anyway and I bet cause more confusion when it's generated like that. Nevertheless this would be the simplest solution.

Yes, sorry i did not check if this issue existed already. If you would rather not generate a struct at all you can just close this pull request.

[marc0246]

I think it would be best to use 0 as the array length. Hopefully it won't cause confusion, as SPIR-V doesn't allow it, so it fits nicely as a kind of sentinel value. The issue with the other options is that it would require a lot of logic because things need to be propagated through structs that use the problematic structs. Like not generating a struct using spec-constant-sized arrays, we would also have to not generate any structs using those structs, and structs using those... This on the other hand requires little to no effort, also for any kind of spec constant.

This would of course also require changing the Option<NonZeroUsize> to an Option<usize>. Would you be willing to make the change? All that's needed is to add a separate arm for the spec constant and return 0 as the length.

[rohdealx]

@marc0246 Sure, in that case, does https://github.com/vulkano-rs/vulkano/blob/master/vulkano-macros/src/derive_buffer_contents.rs#L250C50-L250C50 need to be changed too?
Or should BufferContents just not be derived for 0 sized structs?
https://github.com/vulkano-rs/vulkano/blob/master/vulkano-shaders/src/structs.rs#L130

[marc0246]

You make a very good point, I had forgotten about that. That would complicate things quite a bit. Knowing this, I think your PR is perfect just the way it is. :D

Thank you for looking into this!