Use PEXT/PDEP intrinsics

[jgraef]

Rust recently added a nightly feature that adds methods to integer types to enable use of PEXT/PDEP intrinsics: uint_gather_scatter_bits

I did a quick benchmark encoding a [u16; 2] and it's ~20x faster (code)

[jgraef]

I missed that LLVM doesn't yet support the intrinsics, so this doesn't even use PEXT/PDEP instructions (generated code).

So, I guess you can close this issue. I'll keep it open in case you want to investigate this feature. Though I'm curious why this is so much faster then.

[jgraef]

You can use core::arch::x86_64::_pdep_u64 on x86_64.

Benchmark results encoding a [u16; 2]:

• morton_encoding::morton_encode: 60 ns
• uint_gather_scatter_bits: 3.02 ns
• _pdep_u64: 2.1 ns

[DoubleHyphen]

Alright, I'll investigate and get back to you.

That said, I'm fairly surprised: when I'd benchmarked _pdep_u64 a few years ago, the speed was consistently slightly lower than my own; the crate's documentation mentions as much. Could I ask what code you used to benchmark it?

[jgraef]

This is the benchmark code.

I don't remember exactly, but I think I got the uint_gather_scatter_bits by just modifying the impls here to use the new methods temporarily.