cubek-random 0.3.0-pre.2

use cubecl::{CubeType, Runtime, prelude::*, std::tensor::ViewMut};

use crate::RandomFamily;

use super::{
    PrngArgs, PrngRuntime, lcg_step, random, taus_step_0, taus_step_1, taus_step_2,
    to_unit_interval_closed_open,
};

#[derive(CubeLaunch, CubeType)]
pub(crate) struct Bernoulli {
    probability: f32,
}

#[derive(Debug)]
struct BernoulliFamily;

impl RandomFamily for BernoulliFamily {
    type Runtime = Bernoulli;
}

#[cube]
impl PrngRuntime for Bernoulli {
    fn inner_loop<E: Numeric, N: Size>(
        args: Bernoulli,
        write_index_base: usize,
        n_invocations: u32,
        #[comptime] n_values_per_thread: usize,
        state_0: &mut u32,
        state_1: &mut u32,
        state_2: &mut u32,
        state_3: &mut u32,
        output: &mut ViewMut<'_, Vector<E, N>, usize>,
    ) {
        let prob = args.probability;

        let mut output_vector = Vector::empty();

        let num_iterations = n_values_per_thread / N::value();
        #[unroll(num_iterations <=8)]
        for vector_index in 0..num_iterations {
            // vectorization
            #[unroll]
            for i in 0..N::value() {
                *state_0 = taus_step_0(*state_0);
                *state_1 = taus_step_1(*state_1);
                *state_2 = taus_step_2(*state_2);
                *state_3 = lcg_step(*state_3);

                let int_random = *state_0 ^ *state_1 ^ *state_2 ^ *state_3;
                let float_random = to_unit_interval_closed_open(int_random);
                output_vector.insert(i, E::cast_from(float_random < prob));
            }
            let write_index = vector_index * n_invocations as usize + write_index_base;

            output.write_checked(write_index, output_vector);
        }
    }
}

impl PrngArgs for Bernoulli {
    type Args = Self;

    fn args<R: Runtime>(self) -> BernoulliLaunch<R> {
        BernoulliLaunch::new(self.probability)
    }
}

/// Pseudo-random generator with bernoulli distribution
pub fn random_bernoulli<R: Runtime>(
    client: &ComputeClient<R>,
    probability: f32,
    out: TensorBinding<R>,
    dtype: StorageType,
) -> Result<(), LaunchError> {
    random::<BernoulliFamily, R>(client, Bernoulli { probability }, out, dtype)
}