pic-scale 0.7.8

/*
 * Copyright (c) Radzivon Bartoshyk. 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::utils::{_mm_prefer_fma_ps, _mm256_prefer_fma_ps, shuffle};
use crate::filter_weights::FilterWeights;
use std::arch::x86_64::*;

#[inline(always)]
fn conv_horiz_rgba_1_u16<const FMA: bool>(
    start_x: usize,
    src: &[u16],
    w0: __m128,
    store: __m128,
) -> __m128 {
    unsafe {
        const CN: usize = 4;
        let src_ptr = src.get_unchecked((start_x * CN)..);
        let rgba_pixel = _mm_loadu_si64(src_ptr.as_ptr().cast());
        _mm_prefer_fma_ps::<FMA>(
            store,
            _mm_cvtepi32_ps(_mm_unpacklo_epi16(rgba_pixel, _mm_setzero_si128())),
            w0,
        )
    }
}

#[inline(always)]
fn conv_horiz_rgba_2_u16<const FMA: bool>(
    start_x: usize,
    src: &[u16],
    w0: __m128,
    w1: __m128,
    store: __m128,
) -> __m128 {
    unsafe {
        const CN: usize = 4;
        let src_ptr = src.get_unchecked((start_x * CN)..);

        let rgba_pixel = _mm_loadu_si128(src_ptr.as_ptr().cast());

        let acc = _mm_prefer_fma_ps::<FMA>(
            store,
            _mm_cvtepi32_ps(_mm_unpacklo_epi16(rgba_pixel, _mm_setzero_si128())),
            w0,
        );
        _mm_prefer_fma_ps::<FMA>(
            acc,
            _mm_cvtepi32_ps(_mm_unpackhi_epi16(rgba_pixel, _mm_setzero_si128())),
            w1,
        )
    }
}

#[inline(always)]
fn conv_horiz_rgba_4_u16<const FMA: bool>(
    start_x: usize,
    src: &[u16],
    w0: __m256,
    w1: __m256,
    store: __m256,
) -> __m256 {
    unsafe {
        const CN: usize = 4;
        let src_ptr = src.get_unchecked((start_x * CN)..);

        let rgba_pixel = _mm256_loadu_si256(src_ptr.as_ptr().cast());

        let acc = _mm256_prefer_fma_ps::<FMA>(
            store,
            _mm256_cvtepi32_ps(_mm256_unpackhi_epi16(rgba_pixel, _mm256_setzero_si256())),
            w0,
        );
        _mm256_prefer_fma_ps::<FMA>(
            acc,
            _mm256_cvtepi32_ps(_mm256_unpacklo_epi16(rgba_pixel, _mm256_setzero_si256())),
            w1,
        )
    }
}

#[inline(always)]
fn conv_horiz_rgba_8_u16<const FMA: bool>(
    start_x: usize,
    src: &[u16],
    w01: __m256,
    w23: __m256,
    w45: __m256,
    w67: __m256,
    store: __m256,
) -> __m256 {
    unsafe {
        const CN: usize = 4;
        let src_ptr = src.get_unchecked((start_x * CN)..);

        let z = _mm256_setzero_si256();

        let rgba_pixel0 = _mm256_loadu_si256(src_ptr.as_ptr().cast());
        let rgba_pixel1 = _mm256_loadu_si256(src_ptr.get_unchecked(16..).as_ptr().cast());

        let mut acc = _mm256_prefer_fma_ps::<FMA>(
            store,
            _mm256_cvtepi32_ps(_mm256_unpackhi_epi16(rgba_pixel1, z)),
            w67,
        );
        acc = _mm256_prefer_fma_ps::<FMA>(
            acc,
            _mm256_cvtepi32_ps(_mm256_unpacklo_epi16(rgba_pixel1, z)),
            w45,
        );
        acc = _mm256_prefer_fma_ps::<FMA>(
            acc,
            _mm256_cvtepi32_ps(_mm256_unpackhi_epi16(rgba_pixel0, z)),
            w23,
        );
        _mm256_prefer_fma_ps::<FMA>(
            acc,
            _mm256_cvtepi32_ps(_mm256_unpacklo_epi16(rgba_pixel0, z)),
            w01,
        )
    }
}

pub(crate) fn convolve_horizontal_rgba_avx_rows_4_u16_default(
    src: &[u16],
    src_stride: usize,
    dst: &mut [u16],
    dst_stride: usize,
    filter_weights: &FilterWeights<f32>,
    bit_depth: u32,
) {
    unsafe {
        convolve_horizontal_rgba_avx_rows_4_u16_def(
            src,
            src_stride,
            dst,
            dst_stride,
            filter_weights,
            bit_depth,
        );
    }
}

pub(crate) fn convolve_horizontal_rgba_avx_rows_4_u16_fma(
    src: &[u16],
    src_stride: usize,
    dst: &mut [u16],
    dst_stride: usize,
    filter_weights: &FilterWeights<f32>,
    bit_depth: u32,
) {
    unsafe {
        convolve_horizontal_rgba_avx_rows_4_u16_fma_impl(
            src,
            src_stride,
            dst,
            dst_stride,
            filter_weights,
            bit_depth,
        );
    }
}

#[target_feature(enable = "avx2")]
/// This inlining is required to activate all features for runtime dispatch.
fn convolve_horizontal_rgba_avx_rows_4_u16_def(
    src: &[u16],
    src_stride: usize,
    dst: &mut [u16],
    dst_stride: usize,
    filter_weights: &FilterWeights<f32>,
    bit_depth: u32,
) {
    let unit = Row4ExecutionHandler::<false>::default();
    unit.pass(src, src_stride, dst, dst_stride, filter_weights, bit_depth);
}

#[target_feature(enable = "avx2", enable = "fma")]
/// This inlining is required to activate all features for runtime dispatch.
fn convolve_horizontal_rgba_avx_rows_4_u16_fma_impl(
    src: &[u16],
    src_stride: usize,
    dst: &mut [u16],
    dst_stride: usize,
    filter_weights: &FilterWeights<f32>,
    bit_depth: u32,
) {
    let unit = Row4ExecutionHandler::<true>::default();
    unit.pass(src, src_stride, dst, dst_stride, filter_weights, bit_depth);
}

#[derive(Copy, Clone, Default)]
struct Row4ExecutionHandler<const FMA: bool> {}

impl<const FMA: bool> Row4ExecutionHandler<FMA> {
    #[inline(always)]
    fn rgba_2_u16_avx(
        &self,
        start_x: usize,
        src0: &[u16],
        src1: &[u16],
        w0: __m256,
        w1: __m256,
        store: __m256,
    ) -> __m256 {
        unsafe {
            const CN: usize = 4;
            let src_ptr0 = src0.get_unchecked((start_x * CN)..);
            let src_ptr1 = src1.get_unchecked((start_x * CN)..);

            let rgba_pixel0 = _mm256_cvtepi32_ps(_mm256_cvtepu16_epi32(_mm_loadu_si128(
                src_ptr0.as_ptr().cast(),
            )));
            let rgba_pixel1 = _mm256_cvtepi32_ps(_mm256_cvtepu16_epi32(_mm_loadu_si128(
                src_ptr1.as_ptr().cast(),
            )));

            let lo = _mm256_insertf128_ps::<1>(
                _mm256_castps128_ps256(_mm256_castps256_ps128(rgba_pixel0)),
                _mm256_castps256_ps128(rgba_pixel1),
            );
            let hi = _mm256_insertf128_ps::<1>(
                _mm256_castps128_ps256(_mm256_extractf128_ps::<1>(rgba_pixel0)),
                _mm256_extractf128_ps::<1>(rgba_pixel1),
            );

            let acc = _mm256_prefer_fma_ps::<FMA>(store, lo, w0);
            _mm256_prefer_fma_ps::<FMA>(acc, hi, w1)
        }
    }

    #[inline(always)]
    fn rgba_1_u16(
        &self,
        start_x: usize,
        src0: &[u16],
        src1: &[u16],
        w0: __m256,
        store: __m256,
    ) -> __m256 {
        unsafe {
            const CN: usize = 4;
            let src_ptr0 = src0.get_unchecked((start_x * CN)..);
            let src_ptr1 = src1.get_unchecked((start_x * CN)..);

            let rgba_pixel0 = _mm_loadu_si64(src_ptr0.as_ptr().cast());
            let rgba_pixel1 = _mm_loadu_si64(src_ptr1.as_ptr().cast());

            let full_pixel = _mm256_cvtepu16_epi32(_mm_unpacklo_epi64(rgba_pixel0, rgba_pixel1));

            let f_pixel = _mm256_cvtepi32_ps(full_pixel);

            _mm256_prefer_fma_ps::<FMA>(store, f_pixel, w0)
        }
    }

    #[inline(always)]
    fn pass(
        &self,
        src: &[u16],
        src_stride: usize,
        dst: &mut [u16],
        dst_stride: usize,
        filter_weights: &FilterWeights<f32>,
        bit_depth: u32,
    ) {
        unsafe {
            const CN: usize = 4;

            let v_cap_colors = _mm256_set1_epi16((((1i32 << bit_depth) - 1) as u16) as i16);

            let (row0_ref, rest) = dst.split_at_mut(dst_stride);
            let (row1_ref, rest) = rest.split_at_mut(dst_stride);
            let (row2_ref, row3_ref) = rest.split_at_mut(dst_stride);

            let iter_row0 = row0_ref.as_chunks_mut::<CN>().0.iter_mut();
            let iter_row1 = row1_ref.as_chunks_mut::<CN>().0.iter_mut();
            let iter_row2 = row2_ref.as_chunks_mut::<CN>().0.iter_mut();
            let iter_row3 = row3_ref.as_chunks_mut::<CN>().0.iter_mut();

            for (((((chunk0, chunk1), chunk2), chunk3), &bounds), weights) in iter_row0
                .zip(iter_row1)
                .zip(iter_row2)
                .zip(iter_row3)
                .zip(filter_weights.bounds.iter())
                .zip(
                    filter_weights
                        .weights
                        .chunks_exact(filter_weights.aligned_size),
                )
            {
                let mut jx = 0usize;

                let mut store_0 = _mm256_setzero_ps();
                let mut store_1 = _mm256_setzero_ps();

                let bounds_size = bounds.size;

                let src0 = src;
                let src1 = src0.get_unchecked(src_stride..);
                let src2 = src1.get_unchecked(src_stride..);
                let src3 = src2.get_unchecked(src_stride..);

                if jx >= 4 {
                    let mut astore_0 = _mm256_setzero_ps();
                    let mut astore_1 = _mm256_setzero_ps();
                    let mut astore_2 = _mm256_setzero_ps();
                    let mut astore_3 = _mm256_setzero_ps();

                    while jx + 8 <= bounds_size {
                        let bounds_start = bounds.start + jx;
                        let w_ptr = weights.get_unchecked(jx..);

                        let weights = _mm256_loadu_ps(w_ptr.as_ptr());
                        let w_lo = _mm256_castps256_ps128(weights);
                        let w_hi = _mm256_extractf128_ps::<1>(weights);

                        let w0 = _mm_shuffle_ps::<{ shuffle(0, 0, 0, 0) }>(w_lo, w_lo);
                        let w1 = _mm_shuffle_ps::<{ shuffle(1, 1, 1, 1) }>(w_lo, w_lo);
                        let w2 = _mm_shuffle_ps::<{ shuffle(2, 2, 2, 2) }>(w_lo, w_lo);
                        let w3 = _mm_shuffle_ps::<{ shuffle(3, 3, 3, 3) }>(w_lo, w_lo);
                        let w4 = _mm_shuffle_ps::<{ shuffle(0, 0, 0, 0) }>(w_hi, w_hi);
                        let w5 = _mm_shuffle_ps::<{ shuffle(1, 1, 1, 1) }>(w_hi, w_hi);
                        let w6 = _mm_shuffle_ps::<{ shuffle(2, 2, 2, 2) }>(w_hi, w_hi);
                        let w7 = _mm_shuffle_ps::<{ shuffle(3, 3, 3, 3) }>(w_hi, w_hi);

                        let w01 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w0), w1);
                        let w23 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w2), w3);
                        let w45 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w4), w5);
                        let w67 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w6), w7);

                        astore_0 = conv_horiz_rgba_8_u16::<FMA>(
                            bounds_start,
                            src0,
                            w01,
                            w23,
                            w45,
                            w67,
                            astore_0,
                        );
                        astore_1 = conv_horiz_rgba_8_u16::<FMA>(
                            bounds_start,
                            src1,
                            w01,
                            w23,
                            w45,
                            w67,
                            astore_1,
                        );
                        astore_2 = conv_horiz_rgba_8_u16::<FMA>(
                            bounds_start,
                            src2,
                            w01,
                            w23,
                            w45,
                            w67,
                            astore_2,
                        );
                        astore_3 = conv_horiz_rgba_8_u16::<FMA>(
                            bounds_start,
                            src3,
                            w01,
                            w23,
                            w45,
                            w67,
                            astore_3,
                        );
                        jx += 8;
                    }

                    while jx + 4 <= bounds_size {
                        let bounds_start = bounds.start + jx;
                        let w_ptr = weights.get_unchecked(jx..);

                        let weights = _mm_loadu_ps(w_ptr.as_ptr());

                        let w0 = _mm_shuffle_ps::<{ shuffle(0, 0, 0, 0) }>(weights, weights);
                        let w1 = _mm_shuffle_ps::<{ shuffle(1, 1, 1, 1) }>(weights, weights);
                        let w2 = _mm_shuffle_ps::<{ shuffle(2, 2, 2, 2) }>(weights, weights);
                        let w3 = _mm_shuffle_ps::<{ shuffle(3, 3, 3, 3) }>(weights, weights);

                        let w01 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w0), w1);
                        let w23 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w2), w3);

                        astore_0 =
                            conv_horiz_rgba_4_u16::<FMA>(bounds_start, src0, w01, w23, astore_0);
                        astore_1 =
                            conv_horiz_rgba_4_u16::<FMA>(bounds_start, src1, w01, w23, astore_1);
                        astore_2 =
                            conv_horiz_rgba_4_u16::<FMA>(bounds_start, src2, w01, w23, astore_2);
                        astore_3 =
                            conv_horiz_rgba_4_u16::<FMA>(bounds_start, src3, w01, w23, astore_3);
                        jx += 4;
                    }

                    store_0 = _mm256_add_ps(
                        _mm256_permute2f128_ps::<0x20>(astore_0, astore_1),
                        _mm256_permute2f128_ps::<0x31>(astore_0, astore_1),
                    );

                    store_1 = _mm256_add_ps(
                        _mm256_permute2f128_ps::<0x20>(astore_2, astore_3),
                        _mm256_permute2f128_ps::<0x31>(astore_2, astore_3),
                    );
                }

                while jx + 2 <= bounds_size {
                    let w_ptr = weights.get_unchecked(jx..);
                    let bounds_start = bounds.start + jx;
                    let w0 = _mm256_broadcast_ss(w_ptr.get_unchecked(0));
                    let w1 = _mm256_broadcast_ss(w_ptr.get_unchecked(1));
                    store_0 = self.rgba_2_u16_avx(bounds_start, src0, src1, w0, w1, store_0);
                    store_1 = self.rgba_2_u16_avx(bounds_start, src2, src3, w0, w1, store_1);
                    jx += 2;
                }

                while jx < bounds_size {
                    let w_ptr = weights.get_unchecked(jx..);
                    let bounds_start = bounds.start + jx;
                    let w0 = _mm256_broadcast_ss(w_ptr.get_unchecked(0));
                    store_0 = self.rgba_1_u16(bounds_start, src0, src1, w0, store_0);
                    store_1 = self.rgba_1_u16(bounds_start, src2, src3, w0, store_1);
                    jx += 1;
                }

                let v_st0 = _mm256_cvtps_epi32(store_0);
                let v_st1 = _mm256_cvtps_epi32(store_1);

                let store_16_0 = _mm256_min_epu16(_mm256_packus_epi32(v_st0, v_st0), v_cap_colors);
                let store_16_1 = _mm256_min_epu16(_mm256_packus_epi32(v_st1, v_st1), v_cap_colors);

                _mm_storeu_si64(
                    chunk0.as_mut_ptr().cast(),
                    _mm256_castsi256_si128(store_16_0),
                );
                _mm_storeu_si64(
                    chunk1.as_mut_ptr().cast(),
                    _mm256_extracti128_si256::<1>(store_16_0),
                );
                _mm_storeu_si64(
                    chunk2.as_mut_ptr().cast(),
                    _mm256_castsi256_si128(store_16_1),
                );
                _mm_storeu_si64(
                    chunk3.as_mut_ptr().cast(),
                    _mm256_extracti128_si256::<1>(store_16_1),
                );
            }
        }
    }
}

pub(crate) fn convolve_horizontal_rgba_avx_u16_row_default(
    src: &[u16],
    dst: &mut [u16],
    filter_weights: &FilterWeights<f32>,
    bit_depth: u32,
) {
    unsafe {
        convolve_horizontal_rgba_avx_u16_row_def(src, dst, filter_weights, bit_depth);
    }
}

///SAFETY: checking FMA is required before the call
pub(crate) fn convolve_horizontal_rgba_avx_u16_row_fma(
    src: &[u16],
    dst: &mut [u16],
    filter_weights: &FilterWeights<f32>,
    bit_depth: u32,
) {
    unsafe {
        convolve_horizontal_rgba_avx_u16_row_fma_impl(src, dst, filter_weights, bit_depth);
    }
}

#[target_feature(enable = "avx2")]
/// This inlining is required to activate all features for runtime dispatch.
fn convolve_horizontal_rgba_avx_u16_row_def(
    src: &[u16],
    dst: &mut [u16],
    filter_weights: &FilterWeights<f32>,
    bit_depth: u32,
) {
    let unit = OneRowExecutionHandler::<false>::default();
    unit.pass(src, dst, filter_weights, bit_depth);
}

#[target_feature(enable = "avx2", enable = "fma")]
/// This inlining is required to activate all features for runtime dispatch.
fn convolve_horizontal_rgba_avx_u16_row_fma_impl(
    src: &[u16],
    dst: &mut [u16],
    filter_weights: &FilterWeights<f32>,
    bit_depth: u32,
) {
    let unit = OneRowExecutionHandler::<true>::default();
    unit.pass(src, dst, filter_weights, bit_depth);
}

#[derive(Copy, Clone, Default)]
struct OneRowExecutionHandler<const FMA: bool> {}

impl<const FMA: bool> OneRowExecutionHandler<FMA> {
    #[inline(always)]
    fn pass(
        &self,
        src: &[u16],
        dst: &mut [u16],
        filter_weights: &FilterWeights<f32>,
        bit_depth: u32,
    ) {
        unsafe {
            const CN: usize = 4;

            let v_cap_colors = _mm_set1_epi16((((1i32 << bit_depth) - 1) as u16) as i16);

            for ((dst, bounds), weights) in dst
                .as_chunks_mut::<CN>()
                .0
                .iter_mut()
                .zip(filter_weights.bounds.iter())
                .zip(
                    filter_weights
                        .weights
                        .chunks_exact(filter_weights.aligned_size),
                )
            {
                let bounds_size = bounds.size;
                let mut jx = 0usize;
                let mut store = _mm_setzero_ps();

                if jx >= 4 {
                    let mut astore = _mm256_setzero_ps();

                    while jx + 8 <= bounds_size {
                        let bounds_start = bounds.start + jx;
                        let w_ptr = weights.get_unchecked(jx..);
                        let weights = _mm256_loadu_ps(w_ptr.as_ptr());
                        let w_lo = _mm256_castps256_ps128(weights);
                        let w_hi = _mm256_extractf128_ps::<1>(weights);

                        let w0 = _mm_shuffle_ps::<{ shuffle(0, 0, 0, 0) }>(w_lo, w_lo);
                        let w1 = _mm_shuffle_ps::<{ shuffle(1, 1, 1, 1) }>(w_lo, w_lo);
                        let w2 = _mm_shuffle_ps::<{ shuffle(2, 2, 2, 2) }>(w_lo, w_lo);
                        let w3 = _mm_shuffle_ps::<{ shuffle(3, 3, 3, 3) }>(w_lo, w_lo);
                        let w4 = _mm_shuffle_ps::<{ shuffle(0, 0, 0, 0) }>(w_hi, w_hi);
                        let w5 = _mm_shuffle_ps::<{ shuffle(1, 1, 1, 1) }>(w_hi, w_hi);
                        let w6 = _mm_shuffle_ps::<{ shuffle(2, 2, 2, 2) }>(w_hi, w_hi);
                        let w7 = _mm_shuffle_ps::<{ shuffle(3, 3, 3, 3) }>(w_hi, w_hi);

                        let w01 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w0), w1);
                        let w23 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w2), w3);
                        let w45 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w4), w5);
                        let w67 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w6), w7);
                        astore = conv_horiz_rgba_8_u16::<FMA>(
                            bounds_start,
                            src,
                            w01,
                            w23,
                            w45,
                            w67,
                            astore,
                        );
                        jx += 8;
                    }

                    while jx + 4 <= bounds_size {
                        let w_ptr = weights.get_unchecked(jx..);
                        let weights = _mm_loadu_ps(w_ptr.as_ptr());

                        let w0 = _mm_shuffle_ps::<{ shuffle(0, 0, 0, 0) }>(weights, weights);
                        let w1 = _mm_shuffle_ps::<{ shuffle(1, 1, 1, 1) }>(weights, weights);
                        let w2 = _mm_shuffle_ps::<{ shuffle(2, 2, 2, 2) }>(weights, weights);
                        let w3 = _mm_shuffle_ps::<{ shuffle(3, 3, 3, 3) }>(weights, weights);

                        let bounds_start = bounds.start + jx;

                        let w01 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w0), w1);
                        let w23 = _mm256_insertf128_ps::<1>(_mm256_castps128_ps256(w2), w3);

                        astore = conv_horiz_rgba_4_u16::<FMA>(bounds_start, src, w01, w23, astore);
                        jx += 4;
                    }

                    store = _mm_add_ps(
                        _mm256_castps256_ps128(astore),
                        _mm256_extractf128_ps::<1>(astore),
                    );
                }

                while jx + 2 <= bounds_size {
                    let w_ptr = weights.get_unchecked(jx..);
                    let bounds_start = bounds.start + jx;
                    let w0 = _mm_broadcast_ss(w_ptr.get_unchecked(0));
                    let w1 = _mm_broadcast_ss(w_ptr.get_unchecked(1));
                    store = conv_horiz_rgba_2_u16::<FMA>(bounds_start, src, w0, w1, store);
                    jx += 2;
                }

                while jx < bounds_size {
                    let w_ptr = weights.get_unchecked(jx..);
                    let w0 = _mm_broadcast_ss(w_ptr.get_unchecked(0));
                    let bounds_start = bounds.start + jx;
                    store = conv_horiz_rgba_1_u16::<FMA>(bounds_start, src, w0, store);
                    jx += 1;
                }

                let v_st = _mm_cvtps_epi32(store);

                let store_16_0 = _mm_min_epu16(_mm_packus_epi32(v_st, v_st), v_cap_colors);
                _mm_storeu_si64(dst.as_mut_ptr().cast(), store_16_0);
            }
        }
    }
}