Fix scaling of czm_normal

[jjhembd]

Normals should be transformed by the inverse transpose of the model-view matrix. See lighthouse3d.com and Jason McKesson's Learning Modern 3D Graphics Programming.

The automatic uniform czm_normal, as computed in UniformState, is currently derived from the transpose of the rotation component of the inverse model-view matrix. This approach is only valid when the model-view matrix has uniform scale.

I think UniformState was actually correct before #8182. In that PR, we renamed Matrix4.getRotation asMatrix4.getMatrix3 to better describe what it was actually doing. But then in UniformState, we mistakenly replaced Matrix4.getRotation with a Matrix4.getMatrix3 followed by Matrix3.getRotation.

This PR removes the spurious Matrix3.getRotation call.

One caution: ever since #8182, czm_normal has been unitary. This PR will leave a scale component in the matrix, so normals will need to be re-normalized after the transform.

[cesium-concierge]

Thanks for the pull request @jjhembd!

• ✔️ Signed CLA found.
• ❔ CHANGES.md was not updated.
  • If this change updates the public API in any way, please add a bullet point to CHANGES.md.
• ❔ Unit tests were not updated.
  • Make sure you've updated tests to reflect your changes, added tests for any new code, and ran the code coverage tool.

Reviewers, don't forget to make sure that:

• Cesium Viewer works.
• Works in 2D/CV.

[jjhembd]

I pushed a similar fix for czm_normal3D.

I'm not sure yet what is the correct approach for czm_inverseNormal and czm_inverseNormal3D.

[lilleyse]

Excellent catch @jjhembd.

I'm not sure yet what is the correct approach for czm_inverseNormal and czm_inverseNormal3D.

Shouldn't the getRotation calls be removed there too? Or is it more involved than that?

[jjhembd]

@lilleyse I think czm_inverseNormal, to fit the definition, should be the transpose of the model-view matrix. This would make it the true inverse of czm_normal, which is the transpose of the inverse model-view matrix.

We were computing it as the rotation component of the inverse model-view matrix, without any transpose. This was actually valid after #8182 for model-view matrices with uniform scale, since the rotation component is orthogonal, and for orthogonal matrices the transpose and the inverse are equal.

To honor non-uniform scale throughout, I changed czm_inverseNormal to be the transpose of czm_modelView, and czm_inverseNormal3D to be the transpose of czm_modelView3D.

[lilleyse]

I ran the equations on godbolt just to confirm: https://godbolt.org/z/ar7z8dYa5. It all checks out. Thanks!