Which AVX512 instructions sets are supported?

[TheIronBorn]

See https://en.wikipedia.org/wiki/AVX-512#CPUs_with_AVX-512.
Some instructions like _mm512_rol_epi64 are only available in avx512vl. Most CPUs have avx512vl but it would be good to report which instruction sets are assumed/supported

[martinothamar]

Hi, very good point, I've updated the README

simd-rand/README.md

Lines 5 to 10 in 20ddb7b

	- [`portable`] - portable implementations using `std::simd` (nightly required)
	- [`specific`] - implementations using architecture-specific hardware intrinsics
	- [`specific::avx2`] - AVX2 for x86_64 architecture (4 lanes for 64bit)
	- Requires `avx2` CPU flag, but has additional optimization if you have `avx512dq` and `avx512vl`
	- [`specific::avx512`] - AVX512 for x86_64 architecture (8 lanes for 64bit)
	- Requires `avx512f`, `avx512dq` CPU flags

The specific module looks like this:

simd-rand/src/specific/mod.rs

Lines 1 to 5 in 20ddb7b

	#[cfg(all(target_arch = "x86_64", target_feature = "avx2"))]
	pub mod avx2;
	#[cfg(all(target_arch = "x86_64", target_feature = "avx512f", target_feature = "avx512dq"))]
	pub mod avx512;

One thing I'm a bit unsure about though now that you mention it is _mm_cvtsi32_si128, it's used in basically all bitshifting code for Xoshiro variants. This is a sse2 intruction, but if I've understood correctly avx2 always implies sse2 as well, atleast in practice. I see that numpy makes the same assumptions bot for AVX2 and AVX512 codepaths:

https://github.com/numpy/numpy/blob/e41180d3fe0cd054d57ce446b20b22b95f206d85/numpy/core/src/common/simd/avx512/operators.h#L34

[TheIronBorn]

Yeah instruction subsets are confusing and annoying. To my best knowledge all AVX2/512 will also have sse2. (Also Rust assumes sse2 for any x86 64-bit hardware, so I'm not even sure if such hardware existed it could even use Rust)

[TheIronBorn]

Actually you can just use _mm_slli_epi64 and variants and thus skip _mm_cvtsi32_si128.
Note also you could provide variants which use _mm512_rol_epi64 if avx512vl is available

[martinothamar]

true! the only place where I can't use the i-suffix variant currently is the rotate_left generic function, since it has a const generic parameter. It was telling me to enable some nightly feature and it didn't seem functional at all

what do you mean regarding _mm512_rol_epi64? there are both _mm256_slli_epi64 and _mm512_slli_epi64. Are they functionally equivalent to the rol-variants? It looks like it to me. Latency and throughput numbers seem equal as well

[martinothamar]

ooo sorry for being slow, you're saying that rotate_left == _mm512_rol_epi64. Awesome, thanks 😄

[martinothamar]

Compiler was already doing it apparantly