A attempt at the Ray Tracing in One Weekend book, in C.
$ make
$ ./main > cover.ppm
The resulting image is the cover image of the book: 1200x675px, 500 samples/pixel.
The original, single-threaded implementation ran in 18m 41s with optimization flags enabled. Further optimizations reduced the time to 3m 31s, a 5.3x speedup. Here's a summary of optimizations with a smaller test image (./main -small) - 400x225px, 100 samples/pixel.
| Optimization | real | user | sys | cpu% |
|---|---|---|---|---|
Optimization flags (-O2, -flto) |
24.568 | 24.52 | 0.01 | 99 |
| Multi-threaded pixel rendering (4 threads) | 13.817 | 33.93 | 8.31 | 305 |
| Multi-threaded scanline rendering | 9.459 | 36.43 | 0.02 | 385 |
| Multi-threaded scanlines rendering | 9.261 | 36.68 | 0.01 | 396 |
| Changing function pointer to enums + union | 7.855 | 31.10 | 0.01 | 396 |
Use custom inline random instead of stdlib rand() |
7.690 | 30.40 | 0.01 | 395 |
Change raycolor() to be iterative instead of recursive |
7.519 | 29.72 | 0.01 | 395 |
| Use ARM Neon SIMD instructions in vector functions | 7.066 | 27.91 | 0.01 | 395 |
| Use SIMD instructions to calculate 4 spheres at once | 6.085 | 24.01 | 0.01 | 394 |
| Improved data loading for "4 spheres at once" | 5.847 | 23.09 | 0.01 | 395 |
Vectorized discrimiant check in spherelisthit() |
4.902 | 19.32 | 0.01 | 394 |
Small improvements in spherelisthit() |
4.743 | 18.70 | 0.01 | 394 |
- jacobvosmaer/raytracingweekend + accompanying blog post
- jfeintzeig/ray_tracer + accompanying blog post
- ARM Neon Instruction Set