xsor/src/util.rs

use std::collections::HashSet;

use super::tensor::Tensor;

pub fn get_flattened_indice(shape: &[usize], index: &[usize]) -> usize {
    assert_eq!(shape.len(), index.len());
    index.iter().enumerate().fold(0, |acc, (i, &v)| {
        acc + v * shape[i + 1..].iter().product::<usize>()
    })
}

pub fn get_nested_indice(shape: &[usize], index: usize) -> Vec<usize> {
    let mut index = index;
    let mut nested_index = vec![0; shape.len()];
    for i in (0..shape.len()).rev() {
        let size = shape[i];
        nested_index[i] = index % size;
        index /= size;
    }
    nested_index
}

pub fn broadcast_shapes(shapes: &[&[usize]]) -> Vec<usize> {
    let max_ndim = shapes.iter().map(|s| s.len()).max().unwrap();
    let shapes: Vec<Vec<usize>> = shapes
        .iter()
        .map(|s| {
            let mut s = s.to_vec();
            while s.len() < max_ndim {
                s.insert(0, 1);
            }
            s
        })
        .collect();
    let mut shape = vec![1; max_ndim];
    for dim in 0..max_ndim {
        let sizes: HashSet<usize> = shapes.iter().map(|s| s[dim]).collect();
        assert!(
            sizes.len() <= 2 || (sizes.len() == 2 && sizes.contains(&1)),
            "Shapes are not broadcastable"
        );
        let max_size = sizes.iter().max().unwrap();
        shape[dim] = *max_size;
    }
    shape
}

pub fn broadcast_to<T>(tensor: &Tensor<T>, shape: &[usize]) -> Tensor<T>
where
    T: Copy,
{
    assert!(tensor.shape.len() <= shape.len(), "Shape is too small");
    if tensor.shape == shape {
        return tensor.clone();
    }
    if tensor.shape.is_empty() {
        return Tensor {
            data: vec![tensor.data[0]; shape.iter().product()],
            shape: shape.to_vec(),
        };
    }
    let shape = broadcast_shapes(&[&tensor.shape, shape]);
    let diff_ndim = shape.len() - tensor.shape.len();
    let expanded_shape: Vec<usize> = shape
        .iter()
        .enumerate()
        .map(|(i, &v)| if i < diff_ndim { 1 } else { v })
        .collect();
    let dims_to_expand: HashSet<usize> = shape
        .iter()
        .enumerate()
        .filter(|(i, &v)| *i < diff_ndim || v != tensor.shape[*i - diff_ndim])
        .map(|(i, _)| i)
        .collect();
    let size = shape.iter().product();
    let mut data = vec![tensor.data[0]; size];
    for index in 0..size {
        let nested_index = get_nested_indice(&shape, index)
            .iter()
            .enumerate()
            .map(|(i, &v)| if dims_to_expand.contains(&i) { 0 } else { v })
            .collect::<Vec<usize>>();
        let flattened_index = get_flattened_indice(&expanded_shape, &nested_index);
        data[index] = tensor.data[flattened_index];
    }
    Tensor { data, shape }
}

pub fn broadcast_tensors<T>(tensors: &[&Tensor<T>]) -> Vec<Tensor<T>>
where
    T: Copy,
{
    let shape = broadcast_shapes(
        &tensors
            .iter()
            .map(|t| t.shape.as_slice())
            .collect::<Vec<_>>(),
    );
    tensors
        .iter()
        .map(|t| broadcast_to(t, &shape))
        .collect::<Vec<_>>()
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_get_flattened_indice() {
        assert_eq!(get_flattened_indice(&[2, 3], &[1, 2]), 5);
        assert_eq!(get_flattened_indice(&[2, 3], &[0, 0]), 0);
        assert_eq!(get_flattened_indice(&[2, 3], &[1, 0]), 3);
    }

    #[test]
    fn test_get_nested_indice() {
        assert_eq!(get_nested_indice(&[2, 3], 5), vec![1, 2]);
        assert_eq!(get_nested_indice(&[2, 3], 0), vec![0, 0]);
        assert_eq!(get_nested_indice(&[2, 3], 3), vec![1, 0]);
    }

    #[test]
    fn test_broadcast_shapes() {
        assert_eq!(broadcast_shapes(&[&[1, 2, 3], &[1, 2, 3]]), vec![1, 2, 3]);
        assert_eq!(broadcast_shapes(&[&[2, 3], &[3, 2, 1]]), vec![3, 2, 3]);
    }

    #[test]
    fn test_broadcast_to() {
        let tensor = Tensor::new(&[1, 2, 3], &[3]);
        assert_eq!(
            broadcast_to(&tensor, &[2, 3]),
            Tensor::new(&[1, 2, 3, 1, 2, 3], &[2, 3]),
        );
    }
}