Fixing Nan errors in Camera Frustum Mesh

[emily-howell]

Issue was that when normalized, the look vector had a slight floating point imprecision that caused this to try and calculate a new up vector. Because the up vector was being rotated based on that floating point imprecision, it was trying to rotate around a (0,0,0) axis which caused nan values

[cyrush]

@emily-howell can we add a test that exercises one of these cases?

Worried about the divide + mult impact on other cases.

Is it a case where the look at and position are very clost?

Is it possible to fix this before the normalize?

[emily-howell]

For sure. Let me put together some test cases for this

[emily-howell]

@cyrush To answer your above questions:

1. Q: can we add a test that exercises one of these cases?
   A: I just pushed up changes with a test added

2. Q: Worried about the divide + mult impact on other cases.
   A: I don't think it would have negatively impacted other cases. Since it was after the normalization step, all of the values being divided and multiplied are [0,1] which helps avoid potential overflow. That being said, I modified the handling of this case to avoid any mult/div/rounding.

3. Q: Is it a case where the look at and position are very clost?
   A: Not completely, it is a case where the vtkm Normalization of the look vector has some additional floating point values. E.g. instead of normalizing to [0.0, 0.0, 1.0] it is normalizing to [0.0, 0.0, 1.000000000003]. Which is a problem because when you try to find the angle between the forward vector [0.0, 0.0, -1.0] and that look vector you get a non-zero angle. But because it is such a small non zero angle the 3D vector transform of the up vector is trying to transform by essentially 0 degrees which causes nans.

4. Q: Is it possible to fix this before the normalize?
   A: No because it is a byproduct of that normalization step being imperfect. But there are plenty of ways of handling that imperfection and I pushed up an alternative approach just now that you can look at.

[cyrush]

Looks great, thanks!

[cyrush]

Awesome, this looks like a robust approach.

[cyrush]

Thanks @emily-howell !