Dump of things that come up during development, that we might want to look at in the future.
- Can we avoid some normalize() calls here?
- Ray differential computation
- Some projection and view matrix elements are zero.
- Wald triangle would be more memory efficient
- No need to store vertex coordinates for direct light sampling
- No need to store face normal separately
- Compute differentials from precomputed triangle
- Store smooth normals only when used
- Store quads instead of triangles
- Store CDF only for emitting faces
- Sample according to light power or custom importance
- Surface area computation shared with CDF
- Surface area quick computation for uniform scale
- Direct sampling of background
- For instancing storage is much too big.
- Don't store vectors/colors in float4
- Support uchar colors.
- Convert CORNER to VERTEX attributes by splitting.
- Prune unused nodes for OSL
- Constant folding
- Test if direct use of default texture coordinate is faster
- Optimize fresnel/ior conversion for fixed values
- Triangle wastes quarter of storage
- Leaf bounding box is unused
- Object transforms don't need last row of matrix
- Reduce register pressure by loading object transform row by row
- SIMD optimization for CPU
- QBVH for CPU
- Members need to dereference each time
- Path tracing termination could maybe keep going anyway as long as any thread in the warp is going.
- glDrawPixels does format conversion at least on Mac, with GL_BGRA we could avoid it.
- We do a lot of push_back on vectors, this seems to be quick on mac/linux, but could be optimized
- Diced patches could share vertices