Naga invariant error is confusing

[robtfm]

We can at least call out the most common case of ending an imported item with a number.

Ideally find a way to remap so that users don’t need to worry about it

[viridia]

I get the following error whenever I define a function that ends with a digit:

╭ fn foo2() -> f32 {
14 │ │   return 1.;
   │ ╰────────────^ Composable module identifiers must not require substitution according to naga writeback rules: `"shaderlib/permute.wgsl"::foo2`

The error happens whether I import the function or not.

This is annoying because there are a lot of functions that end with digits (permute2, permute3, etc.) due to the non-support of function overloading in WGSL.

[Neopallium]

I was getting that same error before bevy 0.11 release. But now I get this when trying to import a function that ends with a digit:

ERROR bevy_render::render_resource::pipeline_cache: failed to process shader:
error: no definition in scope for identifier: 'bevy_water::noise::vnoise::vnoise2'
  ┌─ /home/robert/projects/rust/bevy_water/assets/shaders/noise/fbm.wgsl:9:10
  │
9 │   return bevy_water::noise::vnoise::vnoise2(v);
  │          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ unknown identifier
  │
  = no definition in scope for identifier: 'bevy_water::noise::vnoise::vnoise2'

I can't remember how I cause that Composable module identifiers.... error before.

vnoise.wgsl:

#define_import_path bevy_water::noise::vnoise

#import bevy_water::noise::random random2di, cubic_hermite_curve_2d

fn vnoise2(v: vec2<f32>) -> f32 {
        let i = floor(v);
        let f = fract(v);

        // corners.
        let a = random2di(i);
        let b = random2di(i + vec2<f32>(1.0, 0.0));
        let c = random2di(i + vec2<f32>(0.0, 1.0));
        let d = random2di(i + vec2<f32>(1.0, 1.0));

        // Smooth
  let u = cubic_hermite_curve_2d(f);

        // Mix
        return mix(a, b, u.x) +
                (c - a) * u.y * (1.0 - u.x) +
                (d - b) * u.x * u.y;
}

fbm.wgsl:

#define_import_path bevy_water::noise::fbm

#import bevy_water::noise::vnoise vnoise2

fn noise2(v: vec2<f32>) -> f32 {
  return vnoise2(v);
}

fn fbm(v2: vec2<f32>) -> f32 {
  let m2 = mat2x2<f32>(vec2<f32>(0.8, 0.6), vec2<f32>(-0.6, 0.8));
  var p = v2;
  var f = 0.;
  f = f + 0.5000 * noise2(p); p = m2 * p * 2.02;
  f = f + 0.2500 * noise2(p); p = m2 * p * 2.03;
  f = f + 0.1250 * noise2(p); p = m2 * p * 2.01;
  f = f + 0.0625 * noise2(p);
  return f / 0.9375;
}

Those shaders were modified from: https://github.com/Neopallium/bevy_water/tree/main/assets/shaders/noise Adding d to the end of the vnoise2 function and the error goes away.

[Cifram]

I just encountered the exact same issue. Spent a couple hours trying in vain to debug it, before I looked up the naga_oil repo to see if there was a bug about this, and here it is! In my case, it was a random_f32 function, but I've currently renamed to random_float to sidestep this. Hopefully I won't need random f16 or f64 values any time soon.

I initially did a similar test, renaming it to random_f32x, and that also fixed it. From what I can tell, any attempt to import a function which ends in a number will hit this issue.

[Bcompartment]

We can at least call out the most common case of ending an imported item with a number.

Ideally find a way to remap so that users don’t need to worry about it

I found similar error to this thread #71