use crate::haar::{CHROMA_DC_SQUEEZE_DEADZONE, CHROMA_DC_SQUEEZE_STEP};
use std::arch::aarch64::*;
#[inline]
#[target_feature(enable = "neon")]
fn dzq(x: float32x4_t, q: float32x4_t, offset: float32x4_t) -> float32x4_t {
let magnitude = vmulq_f32(vrndmq_f32(vaddq_f32(vdivq_f32(vabsq_f32(x), q), offset)), q);
vbslq_f32(vdupq_n_u32(0x8000_0000), x, magnitude)
}
#[target_feature(enable = "neon")]
pub(crate) fn haar_rows_neon(
top: &[f32],
bottom: &[f32],
q_hf: f32,
top_out: &mut [f32],
bottom_out: &mut [f32],
) -> usize {
let quarter = vdupq_n_f32(0.25);
let low_scale = vdupq_n_f32(0.25 / CHROMA_DC_SQUEEZE_STEP);
let step = vdupq_n_f32(CHROMA_DC_SQUEEZE_STEP);
let q = vdupq_n_f32(q_hf);
let offset = vdupq_n_f32(1.0 - CHROMA_DC_SQUEEZE_DEADZONE);
let mut done = 0;
for ((top, bottom), (top_out, bottom_out)) in top
.as_chunks::<8>()
.0
.iter()
.zip(bottom.as_chunks::<8>().0)
.zip(
top_out
.as_chunks_mut::<8>()
.0
.iter_mut()
.zip(bottom_out.as_chunks_mut::<8>().0),
)
{
let float32x4x2_t(a, b) = unsafe { vld2q_f32(top.as_ptr()) };
let float32x4x2_t(c, d) = unsafe { vld2q_f32(bottom.as_ptr()) };
let ll = vmulq_f32(
vrndaq_f32(vmulq_f32(
vaddq_f32(vaddq_f32(vaddq_f32(a, b), c), d),
low_scale,
)),
step,
);
let hl = dzq(
vmulq_f32(vsubq_f32(vaddq_f32(vsubq_f32(a, b), c), d), quarter),
q,
offset,
);
let lh = dzq(
vmulq_f32(vsubq_f32(vsubq_f32(vaddq_f32(a, b), c), d), quarter),
q,
offset,
);
let hh = dzq(
vmulq_f32(vaddq_f32(vsubq_f32(vsubq_f32(a, b), c), d), quarter),
q,
offset,
);
let a_out = vaddq_f32(vaddq_f32(vaddq_f32(ll, hl), lh), hh);
let b_out = vsubq_f32(vaddq_f32(vsubq_f32(ll, hl), lh), hh);
let c_out = vsubq_f32(vsubq_f32(vaddq_f32(ll, hl), lh), hh);
let d_out = vaddq_f32(vsubq_f32(vsubq_f32(ll, hl), lh), hh);
unsafe {
vst2q_f32(top_out.as_mut_ptr(), float32x4x2_t(a_out, b_out));
vst2q_f32(bottom_out.as_mut_ptr(), float32x4x2_t(c_out, d_out));
}
done += 8;
}
done
}