Support BindGroups with fewer elements in Binding Array than in BindGroupLayoutEntry

[igaryhe]

Is your feature request related to a problem? Please describe.
Current wgpu could support unsized binding array via Features::UNSIZED_BINDING_ARRAY. However, users still have to manually set the array size of the binding in BindGroupLayoutEntry. Is it possible for wgpu to fully support unsized binding array?

Describe the solution you'd like
Maybe we could provide an API to let users specify a maximum array size

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

[cwfitzgerald]

Unfortunately that requirement is there because that's what vulkan (I'd need to check dx12) requires, so we actually can't implement it.

[igaryhe]

As far as I know, if descriptor indexing feature is enabled in Vulkan, the descriptorCount field in a VkDescriptorSetLayoutBinding means the maximum possible count

[kvark]

@igaryhe

Maybe we could provide an API to let users specify a maximum array size

There is already count in BindGroupLayoutEntry. Is anything else needed?

[cwfitzgerald]

@kvark it would be loosening validation to allow the bind group to be smaller than the BGL. It would be quite the improvement in usability actually, as it would mean shader recompiles wouldn't be necessary when your bind group count changes. On the other hand, it does mean that when we bounds check the access into the descriptor array, the bound would have to be a push constant vs a compile time constant.

[kvark]

We can keep this issue then to track the feature, in case anybody wants to see how this may turn out.

[cwfitzgerald]

I just realized this would vastly reduce the complexity of rend3. Currently I need to be able to deal with JIT pipeline recompiles to adjust the BGL size, but if I could just automatically max it out then use the BG to set the real size (which has to be recreated anyway), it would mean I would never need to do JIT recompiles.

[cwfitzgerald]

We should check to see if this is portable to both DX12 and metal.

[cwfitzgerald]

Metal doesn't support unsized binding arrays anyway so it's out of the picture. DX12 should support this feature fine, so we can go ahead and implement this on Vulkan.

[kvark]

Why are we talking about "unsized" here? I thought it's still sized, just with a maximum instead of the exact value. In Metal, the argument buffer would just have some uninitialized textures in there. Seems like it should work.