LukasKalbertodt
Use optimized, non-trigonometric version as mandelbulb DE
There is a version of the mandelbulb DE that doesn't use trigonometric functions, such as sin or atan, which should boost performance a bit. The formulas can be found here:
Mirrored images (as the original was sorta down):
TODO:
- [ ] Power 1
- [ ] Power 2
- [ ] Power 3
- [ ] Power 4
- [ ] Power 5
- [ ] Power 6
- [ ] Power 7
- [X] Power 8
- [ ] Maybe power -1
- [ ] Maybe power -2
Before we implement all those powers, we should wait for the DSL to be finished. We don't want to implement this in both, Rust and GLSL...
I quickly tested this in https://github.com/LukasKalbertodt/cantucci/commit/f56fc9794a22454dc75c4463577a053e1f792ef3 and https://github.com/LukasKalbertodt/cantucci/commit/97fcef60648ecec558faa1390c5ed2bd59a81520 : the optimized version was a lot slower :open_mouth:
~~I hand optimized the non-trig version by replacing all powf() calls with repeated multiplication in #30 (luckily the only powers we raise to are 2, 4, 6 and 8). Turns out it's a lot faster to do x * x * x * x than x.powf(4.0) (if I remember correctly, around factor 100). However, LLVM does not perform this optimization (only for a power of 2). I am not sure why, but the x * x * ... * x version probably looses more precision in the process than with powf().~~
~~Anyway, as you can read in #30, the new version is around twice as fast as the general trigonometric version. I don't think there are many possible improvements left, this code should be pretty optimal (I'd love to be proven wrong, though!).~~
This was just stupid. Of course there is powi() which is a lot faster for variable powers and is optimized perfectly for constant powers.