Read/write Rgba16Float texture

[andrewbaxter]

I just realized I was accidentally writing f32 (glam::Vec4) data into a Rgba16Float texture. When trying to fix it I realize I'm not exactly sure how to do this. Rust doesn't have a f16 on stable, and glam doesn't have a glam::F16Vec4. I'm not even sure what encoding of f16 the GPU needs - is it IEEE 754-2008 referenced by https://docs.rs/float16/latest/float16/ ?

Is Rgba16Float intended to be mostly an internal format and there's no simple way to do this out of the box, or is there some way to do it I missed?

I searched around and couldn't find any evidence of anyone else with the same issue.

[kpreid]

I'm not even sure what encoding of f16 the GPU needs - is it IEEE 754-2008 referenced by https://docs.rs/float16/latest/float16/ ?

Yes, it is. A closer-to-relevant reference is the WGSL specification; wgpu is founded on the WebGPU specification and so a WebGPU citation would be ideal but they don't seem to have a mention of exactly what type is used in texture formats. In general, if something says floating-point of a particular power-of-2 size and doesn’t specify further details, they mean IEEE 754 formats.

You can use float16 or half and expect to get correct results.

Is Rgba16Float intended to be mostly an internal format and there's no simple way to do this out of the box, or is there some way to do it I missed?

I would say that neither of those is true (though Rgba16Float is particularly useful for HDR intermediate framebuffers that never leave the GPU). f16 support is valuable for GPU work and so WebGPU includes support for it, but Rust’s original design is not particularly focused on GPU programming and so it didn’t start with f16 support. There’s no big picture in which this state of affairs is the most appropriate one; it is just what currently happens to be available.

[andrewbaxter]

Okay, glad I didn't miss anything. Thanks!