Optimizing element-wise operations

[cuihantao]

Hi,

I'm new to Rust and ndarray and have been tinkering with element-wise array operations. Below is the code snippet to multiply five arrays and write to arr_out.

The code in the for loop takes 783 ns, and the commented-out line takes 1.24 µs. I checked the generated asm for the for loop; it does not use vmulpd but instead uses mulsd. These vectors are of the same length, but I'm not sure how to inform the compiler of it; adding assert! worsens the performance.

I also tried itertools.izip! but it's not as fast as the for loop. What is the best way to do element-wise operations? My expressions will have many variables so a brief syntax is preferred, but performance is the top priority.

pub fn mul_1d_inplace<'a>(arr_out: &'a mut Array1::<f64>,
                          arr1: &'a Array1::<f64>,
                          arr2: &'a Array1::<f64>,
                          arr3: &'a Array1::<f64>,
                          arr4: &'a Array1::<f64>,
                          arr5: &'a Array1::<f64>,
                          ) -> Option<f64> {
    for i in 0..arr1.len() {
        arr_out[i] = arr1[i] * arr2[i] * arr3[i] * arr4[i] * arr5[i];
    }

    // *arr_out = arr1 * arr2 * arr3 * arr4 * arr5;  // slower

    None
}

[adamreichold]

Please have a look at the Zip type and for convenience the azip! macro.

[cuihantao]

That really helps. With azip!, the assembly now uses mulpd.