Uniform block sizes seem to be rounded up to the nearest 16 on WebGL

[Liamolucko]

Description
WebGL seems to be rounding the size of uniforms up to the nearest 16, causing errors to be thrown when the slightly-smaller buffer is passed to a draw call. Adding some padding to the buffer solves the problem.

Repro steps
None of wgpu's examples use uniforms with alignments less than 16, so I created a toy example which draws a vertical line: https://github.com/Liamolucko/wgpu-uniform-repro. Build with cargo build && wasm-bindgen target/wasm32-unknown-unknown/debug/wgpu-uniform-repro.wasm --target web --out-dir pkg, and then start a http server for the folder.

Expected vs observed behavior
Expected:

Observed:

WebGL warning: drawArraysInstanced: Buffer for uniform block is smaller than UNIFORM_BLOCK_DATA_SIZE. wgpu-uniform-repro.js:604:25

Platform
wgpu 0.11.0, Firefox 93 (also fails in Chromium), Manjaro Linux, nVidia GeForce GTX 780.

[kvark]

Interesting. So it sounds like a validation issue? Other backends do not require this.
Best we can do is to have some validation code active on WebGL/GLES specifically and checking for this.

[Liamolucko]

Apparently this used to be a layout requirement for uniform buffers, but was removed in gpuweb/gpuweb#1560.

[kvark]

This isn't a clear as we'd hope. See gpuweb/gpuweb#1558 and https://github.com/austinEng/webgpu-samples/issues/149, which notably has zero responses.
Finally, clarification from @Kangz confirms that we end up requiring 16-byte aligned uniform buffers in the end - gpuweb/gpuweb#425 (comment)

[grovesNL]

I think gpuweb/gpuweb#425 (comment) is deriving the alignment from the types of the fields though, so we wouldn't necessarily end up with 16-byte alignment for something like SizeOf(struct Block {float x;};), right?

Should we go ahead with having WebGL validation to enforce alignment to UNIFORM_BLOCK_DATA_SIZE (queried from the program)?

We probably need some way to surface this size too, because we don't know the size in advance (i.e. besides guessing 16-byte alignment/rounding up).

[cwfitzgerald]

Uniform buffers require that all structs be rounded up to 16, so that would still be align 16. The only potential issue is if someone had a uniform buffer of a bare u32. This can be a downlevel feature forcing all assignments to 16.