Remove isInf, isNan

[dneto0]

Fixes: #2270

[github-actions]

Previews, as seen when this build job started (6b5d2a1):
WebGPU | IDL
WGSL
Explainer

[alan-baker]

Shouldn't we remove isFinite too?

[dneto0]

Discussed in #2309.

We have agreement on isNan and isFinite. I'd rather land this and then do a second MR for isFinite.

[almarklein]

In one of our shaders (in pygfx) there can be nans in the input data (storage buffer). We need to detect these nans and make the shader behave differently if a nan is found. Right now this works (code here):

    let is_nan = min(w, 1.0) == 1.0 && max(w, -1.0) == -1.0

We have three questions related to this:

1. what would be the recommended way to check for nans in input data (no math applied yet).

2. what happens with nans in the input data in a fastmath implementation? Would it result in bogus data?

3. I read that on Metal fastmath is on by default. Is this also the case with wgpu/naga?

PS: our use-case is rendering lines, and allowing users to insert gaps by using nans, which is a common trick in dataviz.

[ben-clayton]

@almarklein

1. what would be the recommended way to check for nans in input data (no math applied yet).

WGSL currently makes no guarantees that non-finite f32s can be used in any form. So, even the process of loading a non-finite f32 could result in an undefined value, even if you were to immediately bitcast the result to a u32 and checked the bits.
While intrusive in your implementation, you could replace the use of f32 in the storage buffer with u32, load that value, check for inf / nan bitpatterns, and then if finite, bitcast that to f32.

2. what happens with nans in the input data in a fastmath implementation? Would it result in bogus data?

It's not just fastmath. For backends that do not explicitly state they support non-finites, you can get undefined values, and we've also seen cases where the backend shader compilers have done some very inexplicable things with flow control. We will likely offer a future extension to support non-finites and expose it for devices that have been tested and shown to work correctly.

3. I read that on Metal fastmath is on by default. Is this also the case with wgpu/naga?

I can't answer that, I'm afraid.

[almarklein]

Thanks for the answers @ben-clayton!