GLTFExporter: Ensure normalized normal attribute

[takahirox]

This PR lets GLTFExporter ensure normalized normal attribute. See #11951 (comment)

I still need to evaluate to know if this change (quite) slows the performance because this's our first computation with each vertex but let me first share the PR with you. You can test by yourselves, too, if you want.

@fernandojsg Can you share the model file on which this change can affect the performance if you have? I tested some standard models but I didn't see performance degradation (from same to just a few percents worse).

[WestLangley]

Normals in three.js are assumed to have unit length. The library does not validate inputs, however.

We already have BufferGeometry.normalizeNormals(). In that method, zero-vectors are left unchanged. Zero-vectors are considered a user error.

[takahirox]

We already concluded in another thread that GLTFExporter should ensure exporting valid glTF even if model has unexpected values, like zero-vectors or non-unit length normal.

[fernandojsg]

Looking good so far, just two requests:

• I'd add a checkNormalizedNormals (or so) parameter :)
• What about checking just the normals that we're going to actually export? We already use drawRange to detect the number of attributes to export: https://github.com/mrdoob/three.js/blob/dev/examples/js/exporters/GLTFExporter.js#L467-L479 so we could save some time computing unused normals on very big buffers (like the vr paint examples).

[takahirox]

I'd add a checkNormalizedNormals (or so) parameter :)

I wanna think that option if we see the actual use case/model where the checking (quite) slows the performance. I still a bit doubt the checking does... So evaluating first.

What about checking just the normals that we're going to actually export?

If we do that, perhaps it might be better to check/ensure in processBuffer(View) as you first suggested. Let me think about it...

[fernandojsg]

I wanna think that option if we see the actual use case/model where checking (quite) slows the performance. I still a bit doubt checking does... So evaluating first.

You are right, I just did a quick test with a big enough buffer and it was quick to normalize:

var geometry = new THREE.BufferGeometry();
var numElements = 10000000;
var positions = new THREE.BufferAttribute( new Float32Array( numElements * 3 ), 3 );
positions.dynamic = true;
geometry.addAttribute( 'position', positions );

var normals = new THREE.BufferAttribute( new Float32Array( numElements * 3 ), 3 );
normals.dynamic = true;
geometry.addAttribute( 'normal', normals );
geometry.normalizeNormals();

[fernandojsg]

I've just tried with a-painter and works as expected 👍

[takahirox]

Cool.

Let's think those two options if/when we see this checking actually affects the performance.

[mrdoob]

Thanks!