dot() with blas backend

Reproduce the two matrix examples with the openblas-static backend, then compare the results with openblas-system and native ndarray::dot(). Trace the BLAS-backed multiplication entry point and verify that the 8x3 case returns the expected nonzero matrix; add a regression check if the relevant test location is identified.

When I want to multiply two matrices that exceed a certain size with the ndarray-linalg blas backend enabled, it returns a matrix filled with zeros. With the native dot() from ndarray (without blas) it works as expected.

Example:

This works, if the size of a is 7x3:

fn main() {
    let a = array![
        [1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1.],
    ];
    let b = array![
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
    ];
    println!("{}", a.dot(&b));
}

now but if the size of a is 8x3 (or bigger) it returns a matrix filled with zeros: (with ndarray-linalg and blas backend)

fn main() {
    let a = array![
        [1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.],
    ];
    let b = array![
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.],
    ];
    println!("{}", a.dot(&b));
}

however, the last example works as expected without the ndarray-linalg crate. (directly implemented in ndarray itself)

Edit:

The issue only occures with openblas-static, not openblas-system.