Working with uninitialized data

[grothesque]

The following complete program demonstrates two (hopefully correct) ways of working with uninitialized tensors:

use std::mem::MaybeUninit;

use mdarray::{Tensor, DTensor, DenseMapping, Shape};

fn use_with_capacity_mut() {
    let mut a = DTensor::<f64, 2>::with_capacity(9);
    let data = a.spare_capacity_mut();
    for (i, val) in data.iter_mut().enumerate() {
        val.write(i as f64);
    }
    // SAFETY: We initialized all elements.
    unsafe {
        a.set_mapping(DenseMapping::new((3, 3)));
    }
    dbg!(&a);
}

/// # Safety
/// All elements must have been initialized.
unsafe fn tensor_assume_init<T, S: Shape>(tensor: Tensor<MaybeUninit<T>, S>) -> Tensor<T, S>
{
    let (ptr, mapping, capacity) = tensor.into_raw_parts();
    unsafe {
        Tensor::from_raw_parts(ptr as *mut T, mapping, capacity)
    }
}

fn use_tensor_of_maybe_uninit() {
    // let mut b = tensor![[MaybeUninit::<f64>::uninit(); 3]; 3];
    let mut a = Tensor::from_elem((3, 3), MaybeUninit::<f64>::uninit());
    for (i, val) in a.iter_mut().enumerate() {
        val.write(i as f64);
    }
    let a = unsafe { tensor_assume_init(a) };
    dbg!(&a);
}

fn main() {
    use_with_capacity_mut();
    use_tensor_of_maybe_uninit();
}

Could you confirm that use_with_capacity_mut is OK? If it is, then I think that there is a typo in the documentation of Tensor::spare_capacity_mut. It refers to a method set_shape that no longer exists. It probably should refer to set_mapping.

The second function, use_tensor_of_maybe_uninit demonstrates another way of working with uninitialized data. If correct, this way seems safer/easier to me (because the user does not have to construct a correct mapping). It is also more compact.

Perhaps a method similar to tensor_assume_init would be an useful addition to Tensor? (But perhaps not – it is quickly written if needed.)

[fre-hu]

Yes use_with_capacity_mut is fine, and the documentation is just outdated and should be fixed.

I had also thought about having an assume_init function, and yes it seems useful to add.

[grothesque]

Perfect!