use std::arch::aarch64::*;
#[target_feature(enable = "neon")]
pub(crate) fn quantize_xyb_channels_neon(
input: [&[f32]; 3],
output: [&mut [i32]; 3],
scales: [f32; 3],
) {
let [src_x, src_y, src_b] = input;
let [dst_y, dst_x, dst_b] = output;
let [scale_x, scale_y, scale_b] = scales;
let (x4s, x_tail) = src_x.as_chunks::<4>();
let (y4s, y_tail) = src_y.as_chunks::<4>();
let (b4s, b_tail) = src_b.as_chunks::<4>();
let (yq4s, yq_tail) = dst_y.as_chunks_mut::<4>();
let (xq4s, xq_tail) = dst_x.as_chunks_mut::<4>();
let (bq4s, bq_tail) = dst_b.as_chunks_mut::<4>();
let sx = vdupq_n_f32(scale_x);
let sy = vdupq_n_f32(scale_y);
let sb = vdupq_n_f32(scale_b);
for (((((x4, y4), b4), yq4), xq4), bq4) in
x4s.iter().zip(y4s).zip(b4s).zip(yq4s).zip(xq4s).zip(bq4s)
{
let x = unsafe { vld1q_f32(x4.as_ptr()) };
let y = unsafe { vld1q_f32(y4.as_ptr()) };
let b = unsafe { vld1q_f32(b4.as_ptr()) };
let yq = vcvtaq_s32_f32(vmulq_f32(y, sy));
let xq = vcvtaq_s32_f32(vmulq_f32(x, sx));
let bq = vcvtaq_s32_f32(vmulq_f32(b, sb));
unsafe {
vst1q_s32(yq4.as_mut_ptr(), yq);
vst1q_s32(xq4.as_mut_ptr(), xq);
vst1q_s32(bq4.as_mut_ptr(), vsubq_s32(bq, yq));
}
}
let src = x_tail.iter().zip(y_tail).zip(b_tail);
let dst = yq_tail.iter_mut().zip(xq_tail).zip(bq_tail);
for (((x, y), b), ((yq, xq), bq)) in src.zip(dst) {
let quantized_y = (*y * scale_y).round() as i32;
*yq = quantized_y;
*xq = (*x * scale_x).round() as i32;
*bq = (*b * scale_b).round() as i32 - quantized_y;
}
}
#[target_feature(enable = "neon")]
pub(crate) fn quantize_xyb_tile_colors_neon(
input: [&[f32]; 3],
output: &mut [[i32; 3]],
scales: [f32; 3],
) {
let [src_x, src_y, src_b] = input;
let [scale_x, scale_y, scale_b] = scales;
let (x4s, x_tail) = src_x.as_chunks::<4>();
let (y4s, y_tail) = src_y.as_chunks::<4>();
let (b4s, b_tail) = src_b.as_chunks::<4>();
let (out4s, out_tail) = output.as_chunks_mut::<4>();
let sx = vdupq_n_f32(scale_x);
let sy = vdupq_n_f32(scale_y);
let sb = vdupq_n_f32(scale_b);
for (((x4, y4), b4), out4) in x4s.iter().zip(y4s).zip(b4s).zip(out4s) {
let x = unsafe { vld1q_f32(x4.as_ptr()) };
let y = unsafe { vld1q_f32(y4.as_ptr()) };
let b = unsafe { vld1q_f32(b4.as_ptr()) };
let yq = vcvtaq_s32_f32(vmulq_f32(y, sy));
let xq = vcvtaq_s32_f32(vmulq_f32(x, sx));
let bq = vcvtaq_s32_f32(vmulq_f32(b, sb));
unsafe {
vst3q_s32(out4.as_mut_ptr().cast(), int32x4x3_t(yq, xq, bq));
}
}
let src = x_tail.iter().zip(y_tail).zip(b_tail);
for (((x, y), b), out) in src.zip(out_tail) {
*out = [
(*y * scale_y).round() as i32,
(*x * scale_x).round() as i32,
(*b * scale_b).round() as i32,
];
}
}
#[target_feature(enable = "neon")]
pub(crate) fn round_lattice_pow2_neon(plane: &mut [i32], shift: u32) {
let half = vdupq_n_s32((1 << shift) >> 1);
let mask = vdupq_n_s32(-(1 << shift));
let (v4s, tail) = plane.as_chunks_mut::<4>();
for v4 in v4s {
let v = unsafe { vld1q_s32(v4.as_ptr()) };
let r = vandq_s32(vaddq_s32(vabsq_s32(v), half), mask);
let out = vbslq_s32(vcltzq_s32(v), vnegq_s32(r), r);
unsafe { vst1q_s32(v4.as_mut_ptr(), out) };
}
crate::xyb::round_lattice_pow2_scalar(tail, shift);
}