Normals flipped on some parts of the mesh that have double sided materials

[pushmatrix]

Description of the problem

Upgrading from 107 to 108, I've noticed that parts of my model now look like they have flipped normals. This only occurs with double sided materials.
See here: https://vaulted-jonquil.glitch.me/

This is what it should look like, with the pattern being indented on both wheels. That's how it was in 107

In 108 though one of the wheels has the pattern looking like it's popping out:

Three.js version

• r108

Browser

• All of them

OS

• All of them
• Windows
• macOS
• Linux
• Android
• iOS

[elalish]

@pushmatrix Can you confirm if this model does or does not supply its own tangents? I'm curious if this may be related to #11438.

[pushmatrix]

I don't believe it specifies its own tangents

[WestLangley]

Using a different test case, git bisect points to #17586 as the PR that initiated the problem if the mesh is double-sided and has negative scale X. That was in r.109, not r.108, however.

This was not tested on the model from this PR because that model is not available.

[pushmatrix]

Here's the test glb
wheels.glb.zip

[WestLangley]

From 3 years ago: #10331 (comment)

Basically it seems that ... gl_FrontFace is not working properly with double sided materials on Adreno GPUs.

Is this still an issue that we need to accommodate?

//

This model appears to render correctly on my machine if we remove the Adreno workaround in normalmap_pars_fragment:

#ifdef DOUBLE_SIDED

	// Workaround for Adreno GPUs gl_FrontFacing bug. See #15850 and #10331

	bool frontFacing = dot( cross( S, T ), N ) > 0.0;

	mapN.xy *= ( float( frontFacing ) * 2.0 - 1.0 );

#else

	mapN.xy *= ( float( gl_FrontFacing ) * 2.0 - 1.0 );

#endif

[elalish]

Hmm, I remember reading that comment too; I think I came across it while looking at this bug, which was apparently fixed in r108 (at least upgrading fixed it for us): google/model-viewer#740

Maybe it's related? FWIW, I have gotten some feedback about old Adreno GPUs, so apparently they're still fairly common.

[mrdoob]

Just to confirm, this still happens in dev:

[WestLangley]

Just to confirm, this still happens in dev:

Right. But not if you remove the Adreno bug workaround.

[Mugen87]

But not if you remove the Adreno bug workaround.

How about removing the workaround for now and then search for a different solution? The current implementation leads to wrong visuals no matter what platform you are using.

[WestLangley]

@Mugen87 In this particular case, I believe the geometries are mirrors of each other, but they share the same normal map.

I agree with you. I have invested many hours trying to find an Adreno-compatible workaround that is correct in every use case.