Improve SmallRng initialization performance

[arthurprs]

• Added a CHANGELOG.md entry

Summary

Improve initialization speed for SmallRng

Motivation

SmallRng is plenty fast, but the initialization speed was suboptimal.

Details

An example binary was added to check for code generation with cargo asm.

Benchmarks were also added. The results are shown below:

before
test init_from_seed_smallrng  ... bench:           0.84 ns/iter (+/- 0.06)
test init_from_u64_smallrng  ... bench:           6.19 ns/iter (+/- 0.58)
test init_smallrng           ... bench:           9.45 ns/iter (+/- 0.18)

after
test init_from_seed_smallrng  ... bench:           0.85 ns/iter (+/- 0.08)
test init_from_u64_smallrng  ... bench:           3.42 ns/iter (+/- 0.09)
test init_smallrng           ... bench:           6.65 ns/iter (+/- 0.15)

[arthurprs]

Seems like the min MSRV is problematic here, maybe this will have to be removed.

[dhardy]

Either that, or...

This example serves as a non-automated test to check code generation? Can we turn it into an automated test? Otherwise I suggest it just be removed.

[dhardy]

So this is just an optimisation which does not change any results?

[dhardy]

Either that, or...

This example serves as a non-automated test to check code generation? Can we turn it into an automated test? Otherwise I suggest it just be removed.

[dhardy]

Reminder: changes are needed to ensure the API remains consistent on 32-bit and 64-bit platforms.

Also, these new benchmarks can't be merged since #1490 moved all benchmarks to use Criterion.

[arthurprs]

Rebased on main

[arthurprs]

CI seems unhappy about a file I didn't change, not sure what's up.

[dhardy]

One minor 'bug'; otherwise looks good.

[dhardy]

We aren't round-tripping to LE bytes and back since your code now directly initializes state, hence we don't want .to_le() here.

This should trip a test, but it seems we never actually test seed_from_u64... because we don't guarantee value-stability for SmallRng anyway. (In other words, this isn't strictly a bug, but you should still remove .to_le(), and preferably also add some form of reference test for this method.)

[arthurprs]

If I remove it, won't the result be different than before on BE? I was trying to preserve that here. But yeah, kinda pointless w/o a test to ensure that it's correct.

[arthurprs]

Do you mean that this PR should add test for the expected state of the Rng after seed_from_u64?

[dhardy]

There's a reference test in the same file, but in this case testing a single sample should be enough. And yes, results should match from prior to your PR.

It doesn't really matter since we explicitly do not promise value stability of these RNGs, but feel like we should anyway.

[arthurprs]

Done in bee90d1 and 4d6b651

[dhardy]

Thanks!