jcmoyer
commented
8 years ago - Implementing individual, small math functions in C is a no-go; calling into C has an incredibly large overhead. The time saved in C must be greater than the
- Jumping into C once and performing a bunch of math ops before returning control to Lua is very fast.
- Implementing the same thing in pure Lua with ffi-backed data structures is just as fast.
- C API is somewhat clumsy because userdata also has a huge overhead.
- C is only optimal in the bulk-op case, how would it fit in a world where there are e.g. many scattered vectors with no data locality?
- Around C boundaries, LuaJIT 2.0 aborts traces and 2.1 does stitched traces leading to performance degradation.
- vector2 backed by cdef is the same speed as pure Lua in the trivial case and much faster for many operations over multiple vectors
- cdef allows users to instantiate dense arrays of vector2 which is useful for data-oriented designs