Dawoodoz
commented
4 years ago An alternative to dividing algorithm responsibility can be to create a custom rendering pipeline to test each individual function in isolation.
- Rasterize flat color triangles from integer coordinates. Then compare with both a few example images and a trivial reference implementation using millions of seemingly random triangles (using a determined seed).
- Render a surface of many small triangles and check that there are no holes. (As formally proven for non-clipped triangles in the rasterization algorithm)
- Sample textures from square shapes using different mip levels. If both source and target resolution is low, it should be easy to find any bad symmetry.
- Render a whole floating-point model placed with vertices far away from pixel edges.
The 3D camera is a bit strange and nested into 3D rendering instead of just having a standard 4x4 matrix. This is because being allowed to have an infinite far clip plane means not being able to normalize the camera space. Everything would become zero after dividing by infinity if applying the traditional GPU math. If there is another clean mathematical representation capable of expressing the camera's projection with the clip planes, it could be a simple pre-transforming step. This would then allow regression testing the pre-transformed triangle rasterization without the camera and make the core methods easier to understand.
Powerful transform One possible use of a more powerful camera transformation is stereoscopic views where the center of perspective is shifted in 2D, without having to crop on side of the image. Global functions for generating camera transformation packets can then return the perspective center point within some kind of flexible yet fast transformation. Maybe normalize X and Y to -1..+1 while the Z depth remains in the original length system.
Optional near and far clip planes For orthogonal systems, the near clip plane can also be disabled.