in crates/ratchet-core/src/ops/cast.rs [174:183]
fn kernel_element(&self, dst: &Tensor) -> KernelElement {
let numel = dst.shape().numel();
if numel % 4 == 0 {
KernelElement::Vec4
} else if numel % 2 == 0 {
KernelElement::Vec2
} else {
KernelElement::Scalar
}
}