Wrong result of min() and max() in nalgebra::base::min_max::Matrix when f64::NAN is included

[yzobus]

There seems to be a bug for the min() and max() methods of nalgebra::base::min_max::Matrix when an f64::NAN is included in the matrix. Specifically, the entries before the NAN-entry are discarded when using these functions, which results in a wrong output.

For example in this simplified function:

fn main() {
    let mat_wrong_min = MatrixXx1::from_vec(vec![-3., f64::NAN,-2.,-1.,0.]);
    let mat_wrong_max = MatrixXx1::from_vec(vec![3., f64::NAN,2.,1.,0.]);

    println!("min:\t{}", mat_wrong_min.min());
    println!("min:\t{}", mat_wrong_max.max());
};

The first println! statement prints -2., instead of -3. and the second println! statement prints out 2. instead of 3.

[Andlon]

While this is definitely a bug, I do not necessarily agree that the first result should give -3, but instead NaN to indicate that the result is not well-defined.

[Andlon]

@yzobus: can you please indicate the version of nalgebra you're using? Is it the latest one?

[yzobus]

@Andlon You are right with your comment. However, I thought maybe this was intended, similar to the methods min_skipnan or max_skipnan of the ndarray crate.

I tested this with version "0.32.3" and "0.30"

[arscisca]

Both the min() and max() methods use simba crate's SimdPartialOrd methods, which I believe to be the source of the issue. I created the following: https://github.com/dimforge/simba/issues/51.