Consider supporting CPython's random compatibility

[ignassew]

In order to support CPython there are only a couple small changes required. I would like to suggest a MtPython struct that would generate values according to the CPython source code.

I am willing to write the code myself and submit a PR.

If you think that this doesn't suit your library, please consider at least adding this information to the README or the documentation.

Rust reference implementation:

use rand_mt::Mt;

const SEED: u32 = 2137;

fn main() {
    // Every python seed is converted, and initialized as a key: https://github.com/python/cpython/blob/ce558e69d4087dd3653207de78345fbb8a2c7835/Modules/_randommodule.c#LL355C8-L355C8
    let mut rng = Mt::new_with_key([SEED]);

    // CPython does this to get 64 bit float: https://github.com/python/cpython/blob/ce558e69d4087dd3653207de78345fbb8a2c7835/Modules/_randommodule.c#L181
    let a = rng.next_u32() >> 5;
    let b = rng.next_u32() >> 6;
    let n_rust = (a as f64 * 67108864.0 + b as f64) * (1.0 / 9007199254740992.0);

    // Python:
    // #!/bin/python3
    // import random
    //
    // SEED = 2137
    //
    // random.seed(SEED)
    // n_python = random.random()
    // assert n_python == 0.571786152497536
    let n_python = 0.571786152497536;

    assert_eq!(n_rust, n_python);
}

CPython random seed using bytearray: https://github.com/python/cpython/blob/ce558e69d4087dd3653207de78345fbb8a2c7835/Modules/_randommodule.c#LL355C8-L355C8
CPython random.random() implementation: https://github.com/python/cpython/blob/ce558e69d4087dd3653207de78345fbb8a2c7835/Modules/_randommodule.c#L181

Related to #192

[lopopolo]

Thanks for doing the legwork to find the differences between CPython's and CRuby's RNGs @ignassew! This doesn't look like that much of a burden to support. I'd love to accept a contribution to make this happen.

Let's add the CPython support and gate it behind an optional cpython feature. Can you structure the source code so it looks like this:

$ tree src
src
├── lib.rs
├── mt
│   └── rand.rs
├── mt.rs
├── mt64
│   └── rand.rs
├── mt64.rs
├── python
│   ├── mt
│   │   └── rand.rs
│   └── mt.rs
├── python.rs
├── vectors
│   ├── mt.rs
│   └── mt64.rs
└── vectors.rs

7 directories, 10 files

Let's also add a tests/python_reproducability.rs test to match the Ruby one and capture your test case in the issue description.

There's some breaking changes I've been meaning to make (for example, I think I've revisited my stance on #31 since in practice everyone, myself included, prefers the Mt type alias in code) which means it should be no problem packaging this up to crates.io.

I do like your idea on adding documentation to the README about what other languages' RNGs this crate has reproducibility guarantees with.

I'll assign this ticket to you. Looking forward to a PR 🙇