Function launch_copy_perpendicular_ref 

pub fn launch_copy_perpendicular_ref(
    client: &Client,
    input: TensorBinding,
    output: TensorBinding,
    dtype: ElemType,
)

Launches the perpendicular contiguous kernel.

This is used when the input tensor’s memory layout is such that the last dimension is not the one with a stride of 1 (the vectorized dimension). It optimizes the copy by using hardware vectorization (Vectors) and an in-register transpose.