use std::arch::aarch64::*;
#[target_feature(enable = "neon")]
pub(crate) fn chroma_gradient_sums_neon(
x_plane: &[f32],
y_plane: &[f32],
b_plane: &[f32],
width: usize,
) -> [f32; 4] {
if width < 2 {
return [0.0; 4];
}
let mut sum_x = vdupq_n_f32(0.0);
let mut sum_y_x = vdupq_n_f32(0.0);
let mut sum_by = vdupq_n_f32(0.0);
let mut sum_y_b = vdupq_n_f32(0.0);
let mut tail_x = 0.0f32;
let mut tail_y_x = 0.0f32;
let mut tail_by = 0.0f32;
let mut tail_y_b = 0.0f32;
let diff_count = width - 1;
let rows = x_plane
.chunks_exact(width)
.zip(y_plane.chunks_exact(width))
.zip(b_plane.chunks_exact(width))
.step_by(4);
for ((x, y), b) in rows {
let (x0_chunks, x0_tail) = x[..diff_count].as_chunks::<4>();
let (x1_chunks, x1_tail) = x[1..].as_chunks::<4>();
let (y0_chunks, y0_tail) = y[..diff_count].as_chunks::<4>();
let (y1_chunks, y1_tail) = y[1..].as_chunks::<4>();
let (b0_chunks, b0_tail) = b[..diff_count].as_chunks::<4>();
let (b1_chunks, b1_tail) = b[1..].as_chunks::<4>();
for (((((x0, x1), y0), y1), b0), b1) in x0_chunks
.iter()
.zip(x1_chunks)
.zip(y0_chunks)
.zip(y1_chunks)
.zip(b0_chunks)
.zip(b1_chunks)
{
let x0 = unsafe { vld1q_f32(x0.as_ptr()) };
let x1 = unsafe { vld1q_f32(x1.as_ptr()) };
let y0 = unsafe { vld1q_f32(y0.as_ptr()) };
let y1 = unsafe { vld1q_f32(y1.as_ptr()) };
let b0 = unsafe { vld1q_f32(b0.as_ptr()) };
let b1 = unsafe { vld1q_f32(b1.as_ptr()) };
let dy = vabsq_f32(vsubq_f32(y1, y0));
sum_x = vaddq_f32(sum_x, vabsq_f32(vsubq_f32(x1, x0)));
sum_y_x = vaddq_f32(sum_y_x, dy);
sum_by = vaddq_f32(
sum_by,
vabsq_f32(vsubq_f32(vsubq_f32(b1, y1), vsubq_f32(b0, y0))),
);
sum_y_b = vaddq_f32(sum_y_b, dy);
}
for (((((&x0, &x1), &y0), &y1), &b0), &b1) in x0_tail
.iter()
.zip(x1_tail)
.zip(y0_tail)
.zip(y1_tail)
.zip(b0_tail)
.zip(b1_tail)
{
let dy = (y1 - y0).abs();
tail_x += (x1 - x0).abs();
tail_y_x += dy;
tail_by += ((b1 - y1) - (b0 - y0)).abs();
tail_y_b += dy;
}
}
[
vaddvq_f32(sum_x) + tail_x,
vaddvq_f32(sum_y_x) + tail_y_x,
vaddvq_f32(sum_by) + tail_by,
vaddvq_f32(sum_y_b) + tail_y_b,
]
}