Due to having different approximation algorithms for floating-point division between Intel and ARM architecture, it is difficult to write test cases that are 100% guaranteed not off by one pixel compared to reference images when a polygon edge ends up crossing the center of a pixel. If any regression test begins to randomly fail from relying on contemporary floating-point precision, future maintainers will not trust the tests.
One might have to render the same scene with different camera offsets, to generate minimum and maximum intensity images for an acceptable loss of spatial precision, and then add and subtract an acceptable luma tolerance to give some room for future optimization. If the center of perspective can be moved using a 2D offset in camera constructors, the offset can be used to generate minimum and maximum images. The view frustums for culling and clipping will have to shift together with the 2D offset to make sure that triangles are not missing.
Due to having different approximation algorithms for floating-point division between Intel and ARM architecture, it is difficult to write test cases that are 100% guaranteed not off by one pixel compared to reference images when a polygon edge ends up crossing the center of a pixel. If any regression test begins to randomly fail from relying on contemporary floating-point precision, future maintainers will not trust the tests.
One might have to render the same scene with different camera offsets, to generate minimum and maximum intensity images for an acceptable loss of spatial precision, and then add and subtract an acceptable luma tolerance to give some room for future optimization. If the center of perspective can be moved using a 2D offset in camera constructors, the offset can be used to generate minimum and maximum images. The view frustums for culling and clipping will have to shift together with the 2D offset to make sure that triangles are not missing.