use crate::avx2::avx2_utils::{
_mm256_deinterleave_rgba_epi8, _mm256_interleave_epi8, _mm256_interleave_x2_epi8,
};
use crate::yuv_support::{YuvChromaSubsampling, Yuy2Description};
use crate::yuv_to_yuy2::YuvToYuy2Navigation;
#[cfg(target_arch = "x86")]
use std::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;
pub(crate) fn yuy2_to_yuv_avx<const SAMPLING: u8, const YUY2_TARGET: usize>(
y_plane: &mut [u8],
u_plane: &mut [u8],
v_plane: &mut [u8],
yuy2_store: &[u8],
width: u32,
nav: YuvToYuy2Navigation,
) -> YuvToYuy2Navigation {
unsafe {
yuy2_to_yuv_avx_impl::<SAMPLING, YUY2_TARGET>(
y_plane, u_plane, v_plane, yuy2_store, width, nav,
)
}
}
#[target_feature(enable = "avx2")]
unsafe fn yuy2_to_yuv_avx_impl<const SAMPLING: u8, const YUY2_TARGET: usize>(
y_plane: &mut [u8],
u_plane: &mut [u8],
v_plane: &mut [u8],
yuy2_store: &[u8],
width: u32,
nav: YuvToYuy2Navigation,
) -> YuvToYuy2Navigation {
let yuy2_source: Yuy2Description = YUY2_TARGET.into();
let chroma_subsampling: YuvChromaSubsampling = SAMPLING.into();
let mut cx = nav.cx;
let mut uv_x = nav.uv_x;
let mut _yuy2_x = nav.x;
while cx + 64 < width as usize {
let dst_offset = cx * 2;
let u_pos = uv_x;
let v_pos = uv_x;
let y_pos = cx;
let yuy2_ptr = yuy2_store.get_unchecked(dst_offset..);
let j0 = _mm256_loadu_si256(yuy2_ptr.as_ptr() as *const __m256i);
let j1 = _mm256_loadu_si256(yuy2_ptr.get_unchecked(32..).as_ptr() as *const __m256i);
let j2 = _mm256_loadu_si256(yuy2_ptr.get_unchecked(64..).as_ptr() as *const __m256i);
let j3 = _mm256_loadu_si256(yuy2_ptr.get_unchecked(96..).as_ptr() as *const __m256i);
let pixel_set = _mm256_deinterleave_rgba_epi8(j0, j1, j2, j3);
let mut y_first = match yuy2_source {
Yuy2Description::YUYV | Yuy2Description::YVYU => pixel_set.0,
Yuy2Description::UYVY | Yuy2Description::VYUY => pixel_set.1,
};
let mut y_second = match yuy2_source {
Yuy2Description::YUYV | Yuy2Description::YVYU => pixel_set.2,
Yuy2Description::UYVY | Yuy2Description::VYUY => pixel_set.3,
};
(y_first, y_second) = _mm256_interleave_epi8(y_first, y_second);
let u_value = match yuy2_source {
Yuy2Description::YUYV => pixel_set.1,
Yuy2Description::UYVY => pixel_set.0,
Yuy2Description::YVYU => pixel_set.3,
Yuy2Description::VYUY => pixel_set.2,
};
let v_value = match yuy2_source {
Yuy2Description::YUYV => pixel_set.3,
Yuy2Description::UYVY => pixel_set.2,
Yuy2Description::YVYU => pixel_set.1,
Yuy2Description::VYUY => pixel_set.0,
};
if chroma_subsampling == YuvChromaSubsampling::Yuv444 {
let (low_u_value, high_u_value) = _mm256_interleave_x2_epi8(u_value, u_value);
let (low_v_value, high_v_value) = _mm256_interleave_x2_epi8(v_value, v_value);
let u_plane_ptr = u_plane.get_unchecked_mut(u_pos..);
let v_plane_ptr = v_plane.get_unchecked_mut(v_pos..);
_mm256_storeu_si256(u_plane_ptr.as_mut_ptr() as *mut __m256i, low_u_value);
_mm256_storeu_si256(
u_plane_ptr.get_unchecked_mut(32..).as_mut_ptr() as *mut __m256i,
high_u_value,
);
_mm256_storeu_si256(v_plane_ptr.as_mut_ptr() as *mut __m256i, low_v_value);
_mm256_storeu_si256(
v_plane_ptr.get_unchecked_mut(32..).as_mut_ptr() as *mut __m256i,
high_v_value,
);
} else {
_mm256_storeu_si256(
u_plane.get_unchecked_mut(u_pos..).as_mut_ptr() as *mut __m256i,
u_value,
);
_mm256_storeu_si256(
v_plane.get_unchecked_mut(v_pos..).as_mut_ptr() as *mut __m256i,
v_value,
);
}
let y_plane_ptr = y_plane.get_unchecked_mut(y_pos..);
_mm256_storeu_si256(y_plane_ptr.as_mut_ptr() as *mut __m256i, y_first);
_mm256_storeu_si256(
y_plane_ptr.get_unchecked_mut(32..).as_mut_ptr() as *mut __m256i,
y_second,
);
uv_x += match chroma_subsampling {
YuvChromaSubsampling::Yuv420 | YuvChromaSubsampling::Yuv422 => 32,
YuvChromaSubsampling::Yuv444 => 64,
};
cx += 64;
}
_yuy2_x = cx;
YuvToYuy2Navigation {
cx,
uv_x,
x: _yuy2_x,
}
}