use super::{Point, SPAN_STEPS};
use std::arch::aarch64::*;
#[target_feature(enable = "neon")]
pub(super) fn sample_span_neon(
controls: &[Point<f32>; 4],
distances: [f32; 3],
knots: [f32; 4],
out: &mut Vec<Point<f32>>,
) {
let [d0, d1, d2] = distances;
let sums = [vdupq_n_f32(d0 + d1), vdupq_n_f32(d1 + d2)];
let distances = distances.map(|d| vdupq_n_f32(d));
let knots = knots.map(|t| vdupq_n_f32(t));
let q = controls.map(|p| [vdupq_n_f32(p.x), vdupq_n_f32(p.y)]);
let mut points = [Point::default(); SPAN_STEPS.len()];
let lerp = |a: [float32x4_t; 2], b: [float32x4_t; 2], f| {
[
vaddq_f32(a[0], vmulq_f32(f, vsubq_f32(b[0], a[0]))),
vaddq_f32(a[1], vmulq_f32(f, vsubq_f32(b[1], a[1]))),
]
};
for (points, steps) in points
.as_chunks_mut::<4>()
.0
.iter_mut()
.zip(SPAN_STEPS.as_chunks::<4>().0)
{
let step = unsafe { vld1q_f32(steps.as_ptr()) };
let tt = vaddq_f32(distances[0], vmulq_f32(step, distances[1]));
let dt0 = vsubq_f32(tt, knots[0]);
let dt1 = vsubq_f32(tt, knots[1]);
let dt2 = vsubq_f32(tt, knots[2]);
let f1 = vdivq_f32(dt1, distances[1]);
let a0 = lerp(q[0], q[1], vdivq_f32(dt0, distances[0]));
let a1 = lerp(q[1], q[2], f1);
let a2 = lerp(q[2], q[3], vdivq_f32(dt2, distances[2]));
let b0 = lerp(a0, a1, vdivq_f32(dt0, sums[0]));
let b1 = lerp(a1, a2, vdivq_f32(dt1, sums[1]));
let [x, y] = lerp(b0, b1, f1);
let mut xy = [0.0; 8];
unsafe { vst2q_f32(xy.as_mut_ptr(), float32x4x2_t(x, y)) };
for (point, &[x, y]) in points.iter_mut().zip(xy.as_chunks::<2>().0) {
*point = Point::new(x, y);
}
}
out.extend_from_slice(&points[1..]);
}