Neon implementation of linear to sRGB transfer function

neon_to_srgb.rs

// Copyright 2024 the Color Authors
// SPDX-License-Identifier: Apache-2.0 OR MIT

#[cfg(target_arch = "aarch64")]
#[target_feature(enable = "neon")]
#[inline(never)]
pub unsafe fn to_srgb(rgba: [f32; 4]) -> [f32; 4] {
    let v = aarch64::vld1q_f32(rgba.as_ptr());
    let vabs = aarch64::vabsq_f32(v);
    let bias = aarch64::vdupq_n_f32(-5.35862651e-04);
    let x = aarch64::vaddq_f32(vabs, bias);
    let y = aarch64::vsqrtq_f32(x);
    let x2 = aarch64::vmulq_f32(x, x);
    // y = sqrt(v + bias)
    let even1 = aarch64::vfmaq_f32(aarch64::vdupq_n_f32(-2.88143143e-02), x, aarch64::vdupq_n_f32(-9.12795913e-01));
    let even2 = aarch64::vfmaq_f32(even1, x2, aarch64::vdupq_n_f32(-7.29192910e-01));
    let odd1 = aarch64::vfmaq_f32(aarch64::vdupq_n_f32(1.40194533e+00), x, aarch64::vdupq_n_f32(1.06133172e+00));
    let odd2 = aarch64::vfmaq_f32(odd1, x2, aarch64::vdupq_n_f32(2.07758287e-01));
    let poly = aarch64::vfmaq_f32(even2, odd2, y);

    let mask = aarch64::vcgtq_f32(vabs, aarch64::vdupq_n_f32(0.0031308));
    let lin = aarch64::vmulq_f32(vabs, aarch64::vdupq_n_f32(12.92));
    let z = aarch64::vbslq_f32(mask, poly, lin);
    let sign_mask = aarch64::vdupq_n_u32(0x80000000);
    let z_signed = aarch64::vbslq_f32(sign_mask, v, z);
    let out = aarch64::vcopyq_laneq_f32(z_signed, 3, v, 3);
    let mut result = [0.0f32; 4];
    aarch64::vst1q_f32(result.as_mut_ptr(), out);
    result
}