This is a test repository to compare Rust SIMD options on an array min/max function which is problematic for automatic vectorization.

I see using the portable SIMD methods around 4x faster than the basic implementation.

The intrinsics method is around 8x the basic implementation

This requires nightly rust to run the portable SIMD methods.

I also run with the rustflags for AVX2 and optimisation level 3 to max everything out on my system!

RUSTFLAGS="-C target-feature=+avx2 -C opt-level=3" cargo bench 

The basic naive method should be pretty self explainatory. This doesn't vectorise well since we have a 1 element width dependency between iterations.

I attempted a SIMD friendly version which works on 8 elements in parallel at a time which shows a speed up in some cases but was limited.

range_simd uses the portable SIMD methods on nightly. This was a significant improvement over other versions.

Finally range_simd_intrinsics calls directly to the x86 AVX instrinsics. Surprisingly this was double the speed of the range_simd`` method. Looking in Godbolt range_simd` uses some additional calls in min/max which seem unnecessary in our case which is the difference in performance.