yuv 0.8.13

/*
 * Copyright (c) Radzivon Bartoshyk, 10/2024. All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without modification,
 * are permitted provided that the following conditions are met:
 *
 * 1.  Redistributions of source code must retain the above copyright notice, this
 * list of conditions and the following disclaimer.
 *
 * 2.  Redistributions in binary form must reproduce the above copyright notice,
 * this list of conditions and the following disclaimer in the documentation
 * and/or other materials provided with the distribution.
 *
 * 3.  Neither the name of the copyright holder nor the names of its
 * contributors may be used to endorse or promote products derived from
 * this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

use crate::avx2::avx2_utils::{
    _mm256_expand_rgb_to_rgba, _mm256_interleave_epi8, _mm256_set4r_epi8, avx2_pack_u16, shuffle,
};
use crate::internals::ProcessedOffset;
use crate::yuv_support::{
    CbCrForwardTransform, YuvChromaRange, YuvChromaSubsampling, YuvNVOrder, YuvSourceChannels,
};
#[cfg(target_arch = "x86")]
use std::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;

pub(crate) fn avx2_rgba_to_nv_fast_rgba<
    const ORIGIN_CHANNELS: u8,
    const UV_ORDER: u8,
    const SAMPLING: u8,
>(
    y_plane: &mut [u8],
    uv_plane: &mut [u8],
    rgba: &[u8],
    width: u32,
    range: &YuvChromaRange,
    transform: &CbCrForwardTransform<i32>,
    start_cx: usize,
    start_ux: usize,
) -> ProcessedOffset {
    unsafe {
        #[cfg(feature = "nightly_avx512")]
        {
            #[allow(clippy::incompatible_msrv)]
            if std::arch::is_x86_feature_detected!("avxvnni") {
                return avx2_rgba_to_nv_fast_rgba_impl_dot::<ORIGIN_CHANNELS, UV_ORDER, SAMPLING>(
                    y_plane, uv_plane, rgba, width, range, transform, start_cx, start_ux,
                );
            }
        }
        avx2_rgba_to_nv_fast_rgba_impl_ubs::<ORIGIN_CHANNELS, UV_ORDER, SAMPLING>(
            y_plane, uv_plane, rgba, width, range, transform, start_cx, start_ux,
        )
    }
}

#[target_feature(enable = "avx2")]
unsafe fn avx2_rgba_to_nv_fast_rgba_impl_ubs<
    const ORIGIN_CHANNELS: u8,
    const UV_ORDER: u8,
    const SAMPLING: u8,
>(
    y_plane: &mut [u8],
    uv_plane: &mut [u8],
    rgba: &[u8],
    width: u32,
    range: &YuvChromaRange,
    transform: &CbCrForwardTransform<i32>,
    start_cx: usize,
    start_ux: usize,
) -> ProcessedOffset {
    let chroma_subsampling: YuvChromaSubsampling = SAMPLING.into();
    let source_channels: YuvSourceChannels = ORIGIN_CHANNELS.into();
    let uv_order: YuvNVOrder = UV_ORDER.into();
    let channels = source_channels.get_channels_count();

    let y_ptr = y_plane;

    const A_E: i32 = 7;
    let ones = _mm256_set1_epi16(1);
    let y_bias32 = _mm256_set1_epi32(range.bias_y as i32 * (1 << A_E) + (1 << (A_E - 1)) - 1);
    let uv_bias32 = _mm256_set1_epi32(range.bias_uv as i32 * (1 << A_E) + (1 << (A_E - 1)) - 1);

    let y_weights = if source_channels == YuvSourceChannels::Rgba
        || source_channels == YuvSourceChannels::Rgb
    {
        _mm256_set4r_epi8(
            transform.yr as i8,
            transform.yg as i8,
            transform.yb as i8,
            0,
        )
    } else {
        _mm256_set4r_epi8(
            transform.yb as i8,
            transform.yg as i8,
            transform.yr as i8,
            0,
        )
    };
    let cb_weights = if source_channels == YuvSourceChannels::Rgba
        || source_channels == YuvSourceChannels::Rgb
    {
        _mm256_set4r_epi8(
            transform.cb_r as i8,
            transform.cb_g as i8,
            transform.cb_b as i8,
            0,
        )
    } else {
        _mm256_set4r_epi8(
            transform.cb_b as i8,
            transform.cb_g as i8,
            transform.cb_r as i8,
            0,
        )
    };
    let cr_weights = if source_channels == YuvSourceChannels::Rgba
        || source_channels == YuvSourceChannels::Rgb
    {
        _mm256_set4r_epi8(
            transform.cr_r as i8,
            transform.cr_g as i8,
            transform.cr_b as i8,
            0,
        )
    } else {
        _mm256_set4r_epi8(
            transform.cr_b as i8,
            transform.cr_g as i8,
            transform.cr_r as i8,
            0,
        )
    };

    let combined_fixup = _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7);
    let v422_shuffle = _mm256_setr_epi32(0, 2, 4, 6, 1, 3, 5, 7);

    let mut cx = start_cx;
    let mut ux = start_ux;

    while cx + 32 <= width as usize {
        let src = rgba.get_unchecked(cx * channels..).as_ptr();

        let (v0, v1, v2, v3);
        if source_channels == YuvSourceChannels::Rgba || source_channels == YuvSourceChannels::Bgra
        {
            v0 = _mm256_loadu_si256(src as *const __m256i);
            v1 = _mm256_loadu_si256(src.add(32) as *const __m256i);
            v2 = _mm256_loadu_si256(src.add(64) as *const __m256i);
            v3 = _mm256_loadu_si256(src.add(96) as *const __m256i);
        } else if source_channels == YuvSourceChannels::Bgr
            || source_channels == YuvSourceChannels::Rgb
        {
            let m0 = _mm256_loadu_si256(src as *const __m256i);
            let m1 = _mm256_loadu_si256(src.add(32) as *const __m256i);
            let m2 = _mm256_loadu_si256(src.add(64) as *const __m256i);
            (v0, v1, v2, v3) = _mm256_expand_rgb_to_rgba(m0, m1, m2);
        } else {
            unimplemented!()
        }

        let y0s = _mm256_maddubs_epi16(v0, y_weights);
        let y1s = _mm256_maddubs_epi16(v1, y_weights);
        let y2s = _mm256_maddubs_epi16(v2, y_weights);
        let y3s = _mm256_maddubs_epi16(v3, y_weights);

        let v0_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v0, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v1_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v1, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v2_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v2, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v3_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v3, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };

        let y0_32 = _mm256_madd_epi16(y0s, ones);
        let y1_32 = _mm256_madd_epi16(y1s, ones);
        let y2_32 = _mm256_madd_epi16(y2s, ones);
        let y3_32 = _mm256_madd_epi16(y3s, ones);

        let y0_32 = _mm256_add_epi32(y0_32, y_bias32);
        let y1_32 = _mm256_add_epi32(y1_32, y_bias32);
        let y2_32 = _mm256_add_epi32(y2_32, y_bias32);
        let y3_32 = _mm256_add_epi32(y3_32, y_bias32);

        let y0_32 = _mm256_srai_epi32::<A_E>(y0_32);
        let y1_32 = _mm256_srai_epi32::<A_E>(y1_32);
        let y2_32 = _mm256_srai_epi32::<A_E>(y2_32);
        let y3_32 = _mm256_srai_epi32::<A_E>(y3_32);

        let y0m = _mm256_packs_epi32(y0_32, y1_32);
        let y1m = _mm256_packs_epi32(y2_32, y3_32);

        let y_vl0 = _mm256_permutevar8x32_epi32(_mm256_packus_epi16(y0m, y1m), combined_fixup);

        _mm256_storeu_si256(y_ptr.get_unchecked_mut(cx..).as_mut_ptr() as *mut _, y_vl0);

        if chroma_subsampling == YuvChromaSubsampling::Yuv444 {
            let cb0 = _mm256_maddubs_epi16(v0, cb_weights);
            let cb1 = _mm256_maddubs_epi16(v1, cb_weights);
            let cb2 = _mm256_maddubs_epi16(v2, cb_weights);
            let cb3 = _mm256_maddubs_epi16(v3, cb_weights);

            let cr0 = _mm256_maddubs_epi16(v0, cr_weights);
            let cr1 = _mm256_maddubs_epi16(v1, cr_weights);
            let cr2 = _mm256_maddubs_epi16(v2, cr_weights);
            let cr3 = _mm256_maddubs_epi16(v3, cr_weights);

            let cb0_32 = _mm256_madd_epi16(cb0, ones);
            let cb1_32 = _mm256_madd_epi16(cb1, ones);
            let cb2_32 = _mm256_madd_epi16(cb2, ones);
            let cb3_32 = _mm256_madd_epi16(cb3, ones);

            let cr0_32 = _mm256_madd_epi16(cr0, ones);
            let cr1_32 = _mm256_madd_epi16(cr1, ones);
            let cr2_32 = _mm256_madd_epi16(cr2, ones);
            let cr3_32 = _mm256_madd_epi16(cr3, ones);

            let cb0_32 = _mm256_add_epi32(cb0_32, uv_bias32);
            let cb1_32 = _mm256_add_epi32(cb1_32, uv_bias32);
            let cb2_32 = _mm256_add_epi32(cb2_32, uv_bias32);
            let cb3_32 = _mm256_add_epi32(cb3_32, uv_bias32);

            let cr0_32 = _mm256_add_epi32(cr0_32, uv_bias32);
            let cr1_32 = _mm256_add_epi32(cr1_32, uv_bias32);
            let cr2_32 = _mm256_add_epi32(cr2_32, uv_bias32);
            let cr3_32 = _mm256_add_epi32(cr3_32, uv_bias32);

            let cb0_32 = _mm256_srai_epi32::<A_E>(cb0_32);
            let cb1_32 = _mm256_srai_epi32::<A_E>(cb1_32);
            let cb2_32 = _mm256_srai_epi32::<A_E>(cb2_32);
            let cb3_32 = _mm256_srai_epi32::<A_E>(cb3_32);

            let cr0_32 = _mm256_srai_epi32::<A_E>(cr0_32);
            let cr1_32 = _mm256_srai_epi32::<A_E>(cr1_32);
            let cr2_32 = _mm256_srai_epi32::<A_E>(cr2_32);
            let cr3_32 = _mm256_srai_epi32::<A_E>(cr3_32);

            let mut cb00 = _mm256_packs_epi32(cb0_32, cb1_32);
            let mut cb01 = _mm256_packs_epi32(cb2_32, cb3_32);

            let mut cr00 = _mm256_packs_epi32(cr0_32, cr1_32);
            let mut cr01 = _mm256_packs_epi32(cr2_32, cr3_32);

            const MASK: i32 = shuffle(3, 1, 2, 0);
            cb00 = _mm256_permute4x64_epi64::<MASK>(cb00);
            cb01 = _mm256_permute4x64_epi64::<MASK>(cb01);
            cr00 = _mm256_permute4x64_epi64::<MASK>(cr00);
            cr01 = _mm256_permute4x64_epi64::<MASK>(cr01);

            let mut cb_vl = avx2_pack_u16(cb00, cb01);
            let mut cr_vl = avx2_pack_u16(cr00, cr01);

            if uv_order == YuvNVOrder::VU {
                std::mem::swap(&mut cb_vl, &mut cr_vl);
            }

            let uv_interleaved = _mm256_interleave_epi8(cb_vl, cr_vl);

            _mm256_storeu_si256(
                uv_plane.get_unchecked_mut(ux..).as_mut_ptr() as *mut _,
                uv_interleaved.0,
            );
            _mm256_storeu_si256(
                uv_plane.get_unchecked_mut((ux + 32)..).as_mut_ptr() as *mut _,
                uv_interleaved.1,
            );

            ux += 64;
        } else if (chroma_subsampling == YuvChromaSubsampling::Yuv420)
            || (chroma_subsampling == YuvChromaSubsampling::Yuv422)
        {
            let h0 = _mm256_extracti128_si256::<1>(v0_s);
            let h1 = _mm256_extracti128_si256::<1>(v1_s);
            let h2 = _mm256_extracti128_si256::<1>(v2_s);
            let h3 = _mm256_extracti128_si256::<1>(v3_s);

            let vh0 = _mm_avg_epu8(_mm256_castsi256_si128(v0_s), h0);
            let vh1 = _mm_avg_epu8(_mm256_castsi256_si128(v1_s), h1);
            let vh2 = _mm_avg_epu8(_mm256_castsi256_si128(v2_s), h2);
            let vh3 = _mm_avg_epu8(_mm256_castsi256_si128(v3_s), h3);

            let v0_f = _mm256_set_m128i(vh1, vh0);
            let v1_f = _mm256_set_m128i(vh3, vh2);

            let cb0 = _mm256_maddubs_epi16(v0_f, cb_weights);
            let cb1 = _mm256_maddubs_epi16(v1_f, cb_weights);

            let cr0 = _mm256_maddubs_epi16(v0_f, cr_weights);
            let cr1 = _mm256_maddubs_epi16(v1_f, cr_weights);

            let cb0_32 = _mm256_madd_epi16(cb0, ones);
            let cb1_32 = _mm256_madd_epi16(cb1, ones);
            let cr0_32 = _mm256_madd_epi16(cr0, ones);
            let cr1_32 = _mm256_madd_epi16(cr1, ones);

            let cb0_32 = _mm256_add_epi32(cb0_32, uv_bias32);
            let cb1_32 = _mm256_add_epi32(cb1_32, uv_bias32);
            let cr0_32 = _mm256_add_epi32(cr0_32, uv_bias32);
            let cr1_32 = _mm256_add_epi32(cr1_32, uv_bias32);

            let cb0_32 = _mm256_srai_epi32::<A_E>(cb0_32);
            let cb1_32 = _mm256_srai_epi32::<A_E>(cb1_32);
            let cr0_32 = _mm256_srai_epi32::<A_E>(cr0_32);
            let cr1_32 = _mm256_srai_epi32::<A_E>(cr1_32);

            let mut cb00 = _mm256_packs_epi32(cb0_32, cb1_32);
            let mut cr00 = _mm256_packs_epi32(cr0_32, cr1_32);

            const MASK: i32 = shuffle(3, 1, 2, 0);
            cb00 = _mm256_permute4x64_epi64::<MASK>(cb00);
            cr00 = _mm256_permute4x64_epi64::<MASK>(cr00);

            let mut cb_vl = avx2_pack_u16(cb00, cb00);
            let mut cr_vl = avx2_pack_u16(cr00, cr00);

            if uv_order == YuvNVOrder::VU {
                std::mem::swap(&mut cb_vl, &mut cr_vl);
            }

            let uv_interleaved = _mm256_interleave_epi8(cb_vl, cr_vl);

            _mm256_storeu_si256(
                uv_plane.get_unchecked_mut(ux..).as_mut_ptr() as *mut _,
                uv_interleaved.0,
            );

            ux += 32;
        }

        cx += 32;
    }

    if cx < width as usize {
        let diff = width as usize - cx;
        assert!(diff <= 32);

        let mut src_buffer: [u8; 32 * 4] = [0; 32 * 4];
        let mut y_buffer: [u8; 32] = [0; 32];
        let mut uv_buffer: [u8; 64] = [0; 64];

        std::ptr::copy_nonoverlapping(
            rgba.get_unchecked(cx * channels..).as_ptr(),
            src_buffer.as_mut_ptr().cast(),
            diff * channels,
        );

        // Replicate last item to one more position for subsampling
        if chroma_subsampling != YuvChromaSubsampling::Yuv444 && diff % 2 != 0 {
            let lst = (width as usize - 1) * channels;
            let last_items = rgba.get_unchecked(lst..(lst + channels));
            let dvb = diff * channels;
            let dst = src_buffer.get_unchecked_mut(dvb..(dvb + channels));
            for (dst, src) in dst.iter_mut().zip(last_items) {
                *dst = *src;
            }
        }

        let (v0, v1, v2, v3);
        if source_channels == YuvSourceChannels::Rgba || source_channels == YuvSourceChannels::Bgra
        {
            v0 = _mm256_loadu_si256(src_buffer.as_ptr() as *const __m256i);
            v1 = _mm256_loadu_si256(src_buffer.as_ptr().add(32) as *const __m256i);
            v2 = _mm256_loadu_si256(src_buffer.as_ptr().add(64) as *const __m256i);
            v3 = _mm256_loadu_si256(src_buffer.as_ptr().add(96) as *const __m256i);
        } else if source_channels == YuvSourceChannels::Bgr
            || source_channels == YuvSourceChannels::Rgb
        {
            let m0 = _mm256_loadu_si256(src_buffer.as_ptr() as *const __m256i);
            let m1 = _mm256_loadu_si256(src_buffer.as_ptr().add(32) as *const __m256i);
            let m2 = _mm256_loadu_si256(src_buffer.as_ptr().add(64) as *const __m256i);
            (v0, v1, v2, v3) = _mm256_expand_rgb_to_rgba(m0, m1, m2);
        } else {
            unimplemented!()
        }

        let y0s = _mm256_maddubs_epi16(v0, y_weights);
        let y1s = _mm256_maddubs_epi16(v1, y_weights);
        let y2s = _mm256_maddubs_epi16(v2, y_weights);
        let y3s = _mm256_maddubs_epi16(v3, y_weights);

        let v0_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v0, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v1_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v1, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v2_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v2, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v3_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v3, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };

        let y0_32 = _mm256_madd_epi16(y0s, ones);
        let y1_32 = _mm256_madd_epi16(y1s, ones);
        let y2_32 = _mm256_madd_epi16(y2s, ones);
        let y3_32 = _mm256_madd_epi16(y3s, ones);

        let y0_32 = _mm256_add_epi32(y0_32, y_bias32);
        let y1_32 = _mm256_add_epi32(y1_32, y_bias32);
        let y2_32 = _mm256_add_epi32(y2_32, y_bias32);
        let y3_32 = _mm256_add_epi32(y3_32, y_bias32);

        let y0_32 = _mm256_srai_epi32::<A_E>(y0_32);
        let y1_32 = _mm256_srai_epi32::<A_E>(y1_32);
        let y2_32 = _mm256_srai_epi32::<A_E>(y2_32);
        let y3_32 = _mm256_srai_epi32::<A_E>(y3_32);

        let y0m = _mm256_packs_epi32(y0_32, y1_32);
        let y1m = _mm256_packs_epi32(y2_32, y3_32);

        let y_vl0 = _mm256_permutevar8x32_epi32(_mm256_packus_epi16(y0m, y1m), combined_fixup);

        _mm256_storeu_si256(y_buffer.as_mut_ptr() as *mut _, y_vl0);

        if chroma_subsampling == YuvChromaSubsampling::Yuv444 {
            let cb0 = _mm256_maddubs_epi16(v0, cb_weights);
            let cb1 = _mm256_maddubs_epi16(v1, cb_weights);
            let cb2 = _mm256_maddubs_epi16(v2, cb_weights);
            let cb3 = _mm256_maddubs_epi16(v3, cb_weights);

            let cr0 = _mm256_maddubs_epi16(v0, cr_weights);
            let cr1 = _mm256_maddubs_epi16(v1, cr_weights);
            let cr2 = _mm256_maddubs_epi16(v2, cr_weights);
            let cr3 = _mm256_maddubs_epi16(v3, cr_weights);

            let cb0_32 = _mm256_madd_epi16(cb0, ones);
            let cb1_32 = _mm256_madd_epi16(cb1, ones);
            let cb2_32 = _mm256_madd_epi16(cb2, ones);
            let cb3_32 = _mm256_madd_epi16(cb3, ones);

            let cr0_32 = _mm256_madd_epi16(cr0, ones);
            let cr1_32 = _mm256_madd_epi16(cr1, ones);
            let cr2_32 = _mm256_madd_epi16(cr2, ones);
            let cr3_32 = _mm256_madd_epi16(cr3, ones);

            let cb0_32 = _mm256_add_epi32(cb0_32, uv_bias32);
            let cb1_32 = _mm256_add_epi32(cb1_32, uv_bias32);
            let cb2_32 = _mm256_add_epi32(cb2_32, uv_bias32);
            let cb3_32 = _mm256_add_epi32(cb3_32, uv_bias32);

            let cr0_32 = _mm256_add_epi32(cr0_32, uv_bias32);
            let cr1_32 = _mm256_add_epi32(cr1_32, uv_bias32);
            let cr2_32 = _mm256_add_epi32(cr2_32, uv_bias32);
            let cr3_32 = _mm256_add_epi32(cr3_32, uv_bias32);

            let cb0_32 = _mm256_srai_epi32::<A_E>(cb0_32);
            let cb1_32 = _mm256_srai_epi32::<A_E>(cb1_32);
            let cb2_32 = _mm256_srai_epi32::<A_E>(cb2_32);
            let cb3_32 = _mm256_srai_epi32::<A_E>(cb3_32);

            let cr0_32 = _mm256_srai_epi32::<A_E>(cr0_32);
            let cr1_32 = _mm256_srai_epi32::<A_E>(cr1_32);
            let cr2_32 = _mm256_srai_epi32::<A_E>(cr2_32);
            let cr3_32 = _mm256_srai_epi32::<A_E>(cr3_32);

            let mut cb00 = _mm256_packs_epi32(cb0_32, cb1_32);
            let mut cb01 = _mm256_packs_epi32(cb2_32, cb3_32);

            let mut cr00 = _mm256_packs_epi32(cr0_32, cr1_32);
            let mut cr01 = _mm256_packs_epi32(cr2_32, cr3_32);

            const MASK: i32 = shuffle(3, 1, 2, 0);
            cb00 = _mm256_permute4x64_epi64::<MASK>(cb00);
            cb01 = _mm256_permute4x64_epi64::<MASK>(cb01);
            cr00 = _mm256_permute4x64_epi64::<MASK>(cr00);
            cr01 = _mm256_permute4x64_epi64::<MASK>(cr01);

            let mut cb_vl = avx2_pack_u16(cb00, cb01);
            let mut cr_vl = avx2_pack_u16(cr00, cr01);

            if uv_order == YuvNVOrder::VU {
                std::mem::swap(&mut cb_vl, &mut cr_vl);
            }

            let uv_interleaved = _mm256_interleave_epi8(cb_vl, cr_vl);

            _mm256_storeu_si256(uv_buffer.as_mut_ptr() as *mut _, uv_interleaved.0);
            _mm256_storeu_si256(
                uv_buffer.get_unchecked_mut(32..).as_mut_ptr() as *mut _,
                uv_interleaved.1,
            );
        } else if (chroma_subsampling == YuvChromaSubsampling::Yuv420)
            || (chroma_subsampling == YuvChromaSubsampling::Yuv422)
        {
            let h0 = _mm256_extracti128_si256::<1>(v0_s);
            let h1 = _mm256_extracti128_si256::<1>(v1_s);
            let h2 = _mm256_extracti128_si256::<1>(v2_s);
            let h3 = _mm256_extracti128_si256::<1>(v3_s);

            let v0_f = _mm256_set_m128i(h1, h0);
            let v1_f = _mm256_set_m128i(h3, h2);

            let cb0 = _mm256_maddubs_epi16(v0_f, cb_weights);
            let cb1 = _mm256_maddubs_epi16(v1_f, cb_weights);

            let cr0 = _mm256_maddubs_epi16(v0_f, cr_weights);
            let cr1 = _mm256_maddubs_epi16(v1_f, cr_weights);

            let cb0_32 = _mm256_madd_epi16(cb0, ones);
            let cb1_32 = _mm256_madd_epi16(cb1, ones);
            let cr0_32 = _mm256_madd_epi16(cr0, ones);
            let cr1_32 = _mm256_madd_epi16(cr1, ones);

            let cb0_32 = _mm256_add_epi32(cb0_32, uv_bias32);
            let cb1_32 = _mm256_add_epi32(cb1_32, uv_bias32);
            let cr0_32 = _mm256_add_epi32(cr0_32, uv_bias32);
            let cr1_32 = _mm256_add_epi32(cr1_32, uv_bias32);

            let cb0_32 = _mm256_srai_epi32::<A_E>(cb0_32);
            let cb1_32 = _mm256_srai_epi32::<A_E>(cb1_32);
            let cr0_32 = _mm256_srai_epi32::<A_E>(cr0_32);
            let cr1_32 = _mm256_srai_epi32::<A_E>(cr1_32);

            let mut cb00 = _mm256_packs_epi32(cb0_32, cb1_32);
            let mut cr00 = _mm256_packs_epi32(cr0_32, cr1_32);

            const MASK: i32 = shuffle(3, 1, 2, 0);
            cb00 = _mm256_permute4x64_epi64::<MASK>(cb00);
            cr00 = _mm256_permute4x64_epi64::<MASK>(cr00);

            let mut cb_vl = avx2_pack_u16(cb00, cb00);
            let mut cr_vl = avx2_pack_u16(cr00, cr00);

            if uv_order == YuvNVOrder::VU {
                std::mem::swap(&mut cb_vl, &mut cr_vl);
            }

            let uv_interleaved = _mm256_interleave_epi8(cb_vl, cr_vl);

            _mm256_storeu_si256(uv_buffer.as_mut_ptr() as *mut _, uv_interleaved.0);
        }

        std::ptr::copy_nonoverlapping(
            y_buffer.as_ptr().cast(),
            y_ptr.get_unchecked_mut(cx..).as_mut_ptr(),
            diff,
        );

        cx += diff;

        if chroma_subsampling == YuvChromaSubsampling::Yuv444 {
            std::ptr::copy_nonoverlapping(
                uv_buffer.as_ptr().cast(),
                uv_plane.get_unchecked_mut(ux..).as_mut_ptr(),
                diff * 2,
            );

            ux += diff * 2;
        } else if (chroma_subsampling == YuvChromaSubsampling::Yuv420)
            || (chroma_subsampling == YuvChromaSubsampling::Yuv422)
        {
            let hv = diff.div_ceil(2) * 2;
            std::ptr::copy_nonoverlapping(
                uv_buffer.as_ptr().cast(),
                uv_plane.get_unchecked_mut(ux..).as_mut_ptr(),
                hv,
            );
            ux += hv;
        }
    }

    ProcessedOffset { cx, ux }
}

#[cfg(feature = "nightly_avx512")]
#[target_feature(enable = "avx2", enable = "avxvnni")]
unsafe fn avx2_rgba_to_nv_fast_rgba_impl_dot<
    const ORIGIN_CHANNELS: u8,
    const UV_ORDER: u8,
    const SAMPLING: u8,
>(
    y_plane: &mut [u8],
    uv_plane: &mut [u8],
    rgba: &[u8],
    width: u32,
    range: &YuvChromaRange,
    transform: &CbCrForwardTransform<i32>,
    start_cx: usize,
    start_ux: usize,
) -> ProcessedOffset {
    let chroma_subsampling: YuvChromaSubsampling = SAMPLING.into();
    let source_channels: YuvSourceChannels = ORIGIN_CHANNELS.into();
    let channels = source_channels.get_channels_count();
    let uv_order: YuvNVOrder = UV_ORDER.into();

    let y_ptr = y_plane;

    const A_E: i32 = 7;
    let y_bias = _mm256_set1_epi32(range.bias_y as i32 * (1 << A_E) + (1 << (A_E - 1)) - 1);
    let uv_bias = _mm256_set1_epi32(range.bias_uv as i32 * (1 << A_E) + (1 << (A_E - 1)) - 1);

    let y_weights = if source_channels == YuvSourceChannels::Rgba
        || source_channels == YuvSourceChannels::Rgb
    {
        _mm256_set4r_epi8(
            transform.yr as i8,
            transform.yg as i8,
            transform.yb as i8,
            0,
        )
    } else {
        _mm256_set4r_epi8(
            transform.yb as i8,
            transform.yg as i8,
            transform.yr as i8,
            0,
        )
    };
    let cb_weights = if source_channels == YuvSourceChannels::Rgba
        || source_channels == YuvSourceChannels::Rgb
    {
        _mm256_set4r_epi8(
            transform.cb_r as i8,
            transform.cb_g as i8,
            transform.cb_b as i8,
            0,
        )
    } else {
        _mm256_set4r_epi8(
            transform.cb_b as i8,
            transform.cb_g as i8,
            transform.cb_r as i8,
            0,
        )
    };
    let cr_weights = if source_channels == YuvSourceChannels::Rgba
        || source_channels == YuvSourceChannels::Rgb
    {
        _mm256_set4r_epi8(
            transform.cr_r as i8,
            transform.cr_g as i8,
            transform.cr_b as i8,
            0,
        )
    } else {
        _mm256_set4r_epi8(
            transform.cr_b as i8,
            transform.cr_g as i8,
            transform.cr_r as i8,
            0,
        )
    };

    use crate::avx2::avx2_utils::avx2_pack_u32;

    let v422_shuffle = _mm256_setr_epi32(0, 2, 4, 6, 1, 3, 5, 7);

    let mut cx = start_cx;
    let mut ux = start_ux;

    while cx + 32 < width as usize {
        let src = rgba.get_unchecked(cx * channels..).as_ptr();

        let (v0, v1, v2, v3);
        if source_channels == YuvSourceChannels::Rgba || source_channels == YuvSourceChannels::Bgra
        {
            v0 = _mm256_loadu_si256(src as *const __m256i);
            v1 = _mm256_loadu_si256(src.add(32) as *const __m256i);
            v2 = _mm256_loadu_si256(src.add(64) as *const __m256i);
            v3 = _mm256_loadu_si256(src.add(96) as *const __m256i);
        } else if source_channels == YuvSourceChannels::Bgr
            || source_channels == YuvSourceChannels::Rgb
        {
            let m0 = _mm256_loadu_si256(src as *const __m256i);
            let m1 = _mm256_loadu_si256(src.add(32) as *const __m256i);
            let m2 = _mm256_loadu_si256(src.add(64) as *const __m256i);
            (v0, v1, v2, v3) = _mm256_expand_rgb_to_rgba(m0, m1, m2);
        } else {
            unimplemented!()
        }

        let y0s = _mm256_dpbusd_avx_epi32(y_bias, v0, y_weights);
        let y1s = _mm256_dpbusd_avx_epi32(y_bias, v1, y_weights);
        let y2s = _mm256_dpbusd_avx_epi32(y_bias, v2, y_weights);
        let y3s = _mm256_dpbusd_avx_epi32(y_bias, v3, y_weights);

        let v0_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v0, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v1_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v1, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v2_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v2, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v3_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v3, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };

        let mut y0m = avx2_pack_u32(y0s, y1s);
        let mut y1m = avx2_pack_u32(y2s, y3s);

        y0m = _mm256_srai_epi16::<A_E>(y0m);
        y1m = _mm256_srai_epi16::<A_E>(y1m);

        let y_vl = avx2_pack_u16(y0m, y1m);

        _mm256_storeu_si256(y_ptr.get_unchecked_mut(cx..).as_mut_ptr() as *mut _, y_vl);

        if chroma_subsampling == YuvChromaSubsampling::Yuv444 {
            let cb0 = _mm256_dpbusd_avx_epi32(uv_bias, v0, cb_weights);
            let cb1 = _mm256_dpbusd_avx_epi32(uv_bias, v1, cb_weights);
            let cb2 = _mm256_dpbusd_avx_epi32(uv_bias, v2, cb_weights);
            let cb3 = _mm256_dpbusd_avx_epi32(uv_bias, v3, cb_weights);

            let cr0 = _mm256_dpbusd_avx_epi32(uv_bias, v0, cr_weights);
            let cr1 = _mm256_dpbusd_avx_epi32(uv_bias, v1, cr_weights);
            let cr2 = _mm256_dpbusd_avx_epi32(uv_bias, v2, cr_weights);
            let cr3 = _mm256_dpbusd_avx_epi32(uv_bias, v3, cr_weights);

            let mut cb00 = avx2_pack_u32(cb0, cb1);
            let mut cb01 = avx2_pack_u32(cb2, cb3);

            let mut cr00 = avx2_pack_u32(cr0, cr1);
            let mut cr01 = avx2_pack_u32(cr2, cr3);

            cb00 = _mm256_srai_epi16::<A_E>(cb00);
            cb01 = _mm256_srai_epi16::<A_E>(cb01);
            cr00 = _mm256_srai_epi16::<A_E>(cr00);
            cr01 = _mm256_srai_epi16::<A_E>(cr01);

            let mut cb_vl = avx2_pack_u16(cb00, cb01);
            let mut cr_vl = avx2_pack_u16(cr00, cr01);

            if uv_order == YuvNVOrder::VU {
                std::mem::swap(&mut cb_vl, &mut cr_vl);
            }

            let uv_interleaved = _mm256_interleave_epi8(cb_vl, cr_vl);

            _mm256_storeu_si256(
                uv_plane.get_unchecked_mut(ux..).as_mut_ptr() as *mut _,
                uv_interleaved.0,
            );
            _mm256_storeu_si256(
                uv_plane.get_unchecked_mut((ux + 32)..).as_mut_ptr() as *mut _,
                uv_interleaved.1,
            );

            ux += 64;
        } else if (chroma_subsampling == YuvChromaSubsampling::Yuv420)
            || (chroma_subsampling == YuvChromaSubsampling::Yuv422)
        {
            let h0 = _mm256_extracti128_si256::<1>(v0_s);
            let h1 = _mm256_extracti128_si256::<1>(v1_s);
            let h2 = _mm256_extracti128_si256::<1>(v2_s);
            let h3 = _mm256_extracti128_si256::<1>(v3_s);

            let vh0 = _mm_avg_epu8(_mm256_castsi256_si128(v0_s), h0);
            let vh1 = _mm_avg_epu8(_mm256_castsi256_si128(v1_s), h1);
            let vh2 = _mm_avg_epu8(_mm256_castsi256_si128(v2_s), h2);
            let vh3 = _mm_avg_epu8(_mm256_castsi256_si128(v3_s), h3);

            let v0_f = _mm256_set_m128i(vh1, vh0);
            let v1_f = _mm256_set_m128i(vh3, vh2);

            let cb0 = _mm256_dpbusd_avx_epi32(uv_bias, v0_f, cb_weights);
            let cb1 = _mm256_dpbusd_avx_epi32(uv_bias, v1_f, cb_weights);

            let cr0 = _mm256_dpbusd_avx_epi32(uv_bias, v0_f, cr_weights);
            let cr1 = _mm256_dpbusd_avx_epi32(uv_bias, v1_f, cr_weights);

            let mut cb00 = avx2_pack_u32(cb0, cb1);
            let mut cr00 = avx2_pack_u32(cr0, cr1);

            cb00 = _mm256_srai_epi16::<A_E>(cb00);
            cr00 = _mm256_srai_epi16::<A_E>(cr00);

            let mut cb_vl = avx2_pack_u16(cb00, cb00);
            let mut cr_vl = avx2_pack_u16(cr00, cr00);

            if uv_order == YuvNVOrder::VU {
                std::mem::swap(&mut cb_vl, &mut cr_vl);
            }

            let uv_interleaved = _mm256_interleave_epi8(cb_vl, cr_vl);

            _mm256_storeu_si256(
                uv_plane.get_unchecked_mut(ux..).as_mut_ptr() as *mut _,
                uv_interleaved.0,
            );

            ux += 32;
        }

        cx += 32;
    }

    if cx < width as usize {
        let diff = width as usize - cx;
        assert!(diff <= 32);

        let mut src_buffer: [u8; 32 * 4] = [0; 32 * 4];
        let mut y_buffer: [u8; 32] = [0; 32];
        let mut uv_buffer: [u8; 64] = [0; 64];

        std::ptr::copy_nonoverlapping(
            rgba.get_unchecked(cx * channels..).as_ptr(),
            src_buffer.as_mut_ptr(),
            diff * channels,
        );

        // Replicate last item to one more position for subsampling
        if chroma_subsampling != YuvChromaSubsampling::Yuv444 && diff % 2 != 0 {
            let lst = (width as usize - 1) * channels;
            let last_items = rgba.get_unchecked(lst..(lst + channels));
            let dvb = diff * channels;
            let dst = src_buffer.get_unchecked_mut(dvb..(dvb + channels));
            for (dst, src) in dst.iter_mut().zip(last_items) {
                *dst = *src;
            }
        }

        let (v0, v1, v2, v3);
        if source_channels == YuvSourceChannels::Rgba || source_channels == YuvSourceChannels::Bgra
        {
            v0 = _mm256_loadu_si256(src_buffer.as_ptr() as *const __m256i);
            v1 = _mm256_loadu_si256(src_buffer.as_ptr().add(32) as *const __m256i);
            v2 = _mm256_loadu_si256(src_buffer.as_ptr().add(64) as *const __m256i);
            v3 = _mm256_loadu_si256(src_buffer.as_ptr().add(96) as *const __m256i);
        } else if source_channels == YuvSourceChannels::Bgr
            || source_channels == YuvSourceChannels::Rgb
        {
            let m0 = _mm256_loadu_si256(src_buffer.as_ptr() as *const __m256i);
            let m1 = _mm256_loadu_si256(src_buffer.as_ptr().add(32) as *const __m256i);
            let m2 = _mm256_loadu_si256(src_buffer.as_ptr().add(64) as *const __m256i);
            (v0, v1, v2, v3) = _mm256_expand_rgb_to_rgba(m0, m1, m2);
        } else {
            unimplemented!()
        }

        let y0s = _mm256_dpbusd_avx_epi32(y_bias, v0, y_weights);
        let y1s = _mm256_dpbusd_avx_epi32(y_bias, v1, y_weights);
        let y2s = _mm256_dpbusd_avx_epi32(y_bias, v2, y_weights);
        let y3s = _mm256_dpbusd_avx_epi32(y_bias, v3, y_weights);

        let v0_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v0, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v1_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v1, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v2_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v2, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };
        let v3_s = if chroma_subsampling != YuvChromaSubsampling::Yuv444 {
            _mm256_permutevar8x32_epi32(v3, v422_shuffle)
        } else {
            _mm256_setzero_si256()
        };

        let mut y0m = avx2_pack_u32(y0s, y1s);
        let mut y1m = avx2_pack_u32(y2s, y3s);

        y0m = _mm256_srai_epi16::<A_E>(y0m);
        y1m = _mm256_srai_epi16::<A_E>(y1m);

        let y_vl = avx2_pack_u16(y0m, y1m);

        _mm256_storeu_si256(y_buffer.as_mut_ptr() as *mut _, y_vl);

        if chroma_subsampling == YuvChromaSubsampling::Yuv444 {
            let cb0 = _mm256_dpbusd_avx_epi32(uv_bias, v0, cb_weights);
            let cb1 = _mm256_dpbusd_avx_epi32(uv_bias, v1, cb_weights);
            let cb2 = _mm256_dpbusd_avx_epi32(uv_bias, v2, cb_weights);
            let cb3 = _mm256_dpbusd_avx_epi32(uv_bias, v3, cb_weights);

            let cr0 = _mm256_dpbusd_avx_epi32(uv_bias, v0, cr_weights);
            let cr1 = _mm256_dpbusd_avx_epi32(uv_bias, v1, cr_weights);
            let cr2 = _mm256_dpbusd_avx_epi32(uv_bias, v2, cr_weights);
            let cr3 = _mm256_dpbusd_avx_epi32(uv_bias, v3, cr_weights);

            let mut cb00 = avx2_pack_u32(cb0, cb1);
            let mut cb01 = avx2_pack_u32(cb2, cb3);

            let mut cr00 = avx2_pack_u32(cr0, cr1);
            let mut cr01 = avx2_pack_u32(cr2, cr3);

            cb00 = _mm256_srai_epi16::<A_E>(cb00);
            cb01 = _mm256_srai_epi16::<A_E>(cb01);
            cr00 = _mm256_srai_epi16::<A_E>(cr00);
            cr01 = _mm256_srai_epi16::<A_E>(cr01);

            let mut cb_vl = avx2_pack_u16(cb00, cb01);
            let mut cr_vl = avx2_pack_u16(cr00, cr01);

            if uv_order == YuvNVOrder::VU {
                std::mem::swap(&mut cb_vl, &mut cr_vl);
            }

            let uv_interleaved = _mm256_interleave_epi8(cb_vl, cr_vl);

            _mm256_storeu_si256(uv_buffer.as_mut_ptr() as *mut _, uv_interleaved.0);
            _mm256_storeu_si256(
                uv_buffer.get_unchecked_mut(32..).as_mut_ptr() as *mut _,
                uv_interleaved.1,
            );
        } else if (chroma_subsampling == YuvChromaSubsampling::Yuv420)
            || (chroma_subsampling == YuvChromaSubsampling::Yuv422)
        {
            let h0 = _mm256_extracti128_si256::<1>(v0_s);
            let h1 = _mm256_extracti128_si256::<1>(v1_s);
            let h2 = _mm256_extracti128_si256::<1>(v2_s);
            let h3 = _mm256_extracti128_si256::<1>(v3_s);

            let vh0 = _mm_avg_epu8(_mm256_castsi256_si128(v0_s), h0);
            let vh1 = _mm_avg_epu8(_mm256_castsi256_si128(v1_s), h1);
            let vh2 = _mm_avg_epu8(_mm256_castsi256_si128(v2_s), h2);
            let vh3 = _mm_avg_epu8(_mm256_castsi256_si128(v3_s), h3);

            let v0_f = _mm256_set_m128i(vh1, vh0);
            let v1_f = _mm256_set_m128i(vh3, vh2);

            let cb0 = _mm256_dpbusd_avx_epi32(uv_bias, v0_f, cb_weights);
            let cb1 = _mm256_dpbusd_avx_epi32(uv_bias, v1_f, cb_weights);

            let cr0 = _mm256_dpbusd_avx_epi32(uv_bias, v0_f, cr_weights);
            let cr1 = _mm256_dpbusd_avx_epi32(uv_bias, v1_f, cr_weights);

            let mut cb00 = avx2_pack_u32(cb0, cb1);
            let mut cr00 = avx2_pack_u32(cr0, cr1);

            cb00 = _mm256_srai_epi16::<A_E>(cb00);
            cr00 = _mm256_srai_epi16::<A_E>(cr00);

            let mut cb_vl = avx2_pack_u16(cb00, cb00);
            let mut cr_vl = avx2_pack_u16(cr00, cr00);

            if uv_order == YuvNVOrder::VU {
                std::mem::swap(&mut cb_vl, &mut cr_vl);
            }

            let uv_interleaved = _mm256_interleave_epi8(cb_vl, cr_vl);

            _mm256_storeu_si256(uv_buffer.as_mut_ptr() as *mut _, uv_interleaved.0);
        }

        std::ptr::copy_nonoverlapping(
            y_buffer.as_ptr(),
            y_ptr.get_unchecked_mut(cx..).as_mut_ptr(),
            diff,
        );

        cx += diff;

        if chroma_subsampling == YuvChromaSubsampling::Yuv444 {
            std::ptr::copy_nonoverlapping(
                uv_buffer.as_ptr(),
                uv_plane.get_unchecked_mut(ux..).as_mut_ptr(),
                diff * 2,
            );

            ux += diff * 2;
        } else if (chroma_subsampling == YuvChromaSubsampling::Yuv420)
            || (chroma_subsampling == YuvChromaSubsampling::Yuv422)
        {
            let hv = diff.div_ceil(2) * 2;
            std::ptr::copy_nonoverlapping(
                uv_buffer.as_ptr(),
                uv_plane.get_unchecked_mut(ux..).as_mut_ptr(),
                hv,
            );
            ux += hv;
        }
    }

    ProcessedOffset { cx, ux }
}