Implements bitonic parallel sort in CUDA, using rust for the host side.
For now, it only sorts f32s and it only works on power-of-two arrays.
Benchmarks on a Ubuntu 23.10 desktop PC
- CPU: 24 core 12th Gen Intel(R) Core(TM) i9-12900KS
- GPU: NVIDIA GeForce RTX 3090 Ti
~ cargo bench
test bitonic_32768 ... bench: 100,796 ns/iter (+/- 2,366)
test bitonic_65536 ... bench: 152,260 ns/iter (+/- 2,395)
test bitonic_1048576 ... bench: 2,428,581 ns/iter (+/- 201,018)
test bitonic_33554432 ... bench: 85,287,101 ns/iter (+/- 6,548,327)
test rayon_stable_32768 ... bench: 231,308 ns/iter (+/- 5,226)
test rayon_stable_65536 ... bench: 345,339 ns/iter (+/- 8,769)
test rayon_stable_1048576 ... bench: 4,682,997 ns/iter (+/- 295,796)
test rayon_stable_33554432 ... bench: 155,202,823 ns/iter (+/- 7,103,368)
test rayon_unstable_32768 ... bench: 154,876 ns/iter (+/- 7,472)
test rayon_unstable_65536 ... bench: 283,243 ns/iter (+/- 21,623)
test rayon_unstable_1048576 ... bench: 4,112,566 ns/iter (+/- 839,600)
test rayon_unstable_33554432 ... bench: 111,669,328 ns/iter (+/- 7,156,464)
test std_stable_32768 ... bench: 1,278,452 ns/iter (+/- 75,445)
test std_stable_65536 ... bench: 2,732,487 ns/iter (+/- 90,811)
test std_stable_1048576 ... bench: 55,422,265 ns/iter (+/- 2,664,924)
test std_stable_33554432 ... bench: 2,247,221,910 ns/iter (+/- 11,668,572)
test std_unstable_32768 ... bench: 554,653 ns/iter (+/- 11,537)
test std_unstable_65536 ... bench: 1,152,406 ns/iter (+/- 43,448)
test std_unstable_1048576 ... bench: 20,855,117 ns/iter (+/- 190,036)
test std_unstable_33554432 ... bench: 754,761,743 ns/iter (+/- 21,142,560)