libblur 0.24.0

// 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::EdgeMode;
use crate::edge_mode::clamp_edge;
use crate::neon::fast_gaussian::NeonI32x4;
use crate::neon::{load_u16_s32_fast, store_u16_s32_x4, store_u16x4, vmulq_by_3_s32};
use crate::unsafe_slice::UnsafeSlice;
use crate::util::ScratchBuffer;
use std::arch::aarch64::*;

pub(crate) fn fgn_vertical_pass_neon_u16<const CN: usize>(
    bytes: &UnsafeSlice<u16>,
    stride: u32,
    width: u32,
    height: u32,
    radius: u32,
    start: u32,
    end: u32,
    edge_mode: EdgeMode,
) {
    unsafe {
        let mut buffer = ScratchBuffer::<[NeonI32x4; 4], 1024>::new(1024);
        let buffer = buffer.as_mut_slice();

        let height_wide = height as i64;

        let radius_64 = radius as i64;
        let weight = 1.0f32 / ((radius as f32) * (radius as f32) * (radius as f32));
        let v_weight = vdupq_n_f32(weight);

        let mut xx = start;

        while xx + 4 <= width.min(end) {
            let mut diffs0 = vdupq_n_s32(0);
            let mut diffs1 = vdupq_n_s32(0);
            let mut diffs2 = vdupq_n_s32(0);
            let mut diffs3 = vdupq_n_s32(0);

            let mut ders0 = vdupq_n_s32(0);
            let mut ders1 = vdupq_n_s32(0);
            let mut ders2 = vdupq_n_s32(0);
            let mut ders3 = vdupq_n_s32(0);

            let mut summs0 = vdupq_n_s32(0);
            let mut summs1 = vdupq_n_s32(0);
            let mut summs2 = vdupq_n_s32(0);
            let mut summs3 = vdupq_n_s32(0);

            let start_y = -3 * radius as i64;

            let current_px0 = (xx * CN as u32) as usize;
            let current_px1 = ((xx + 1) * CN as u32) as usize;
            let current_px2 = ((xx + 2) * CN as u32) as usize;
            let current_px3 = ((xx + 3) * CN as u32) as usize;

            for y in start_y..height_wide {
                let current_y = (y * (stride as i64)) as usize;

                if y >= 0 {
                    let c0 = vcvtq_f32_s32(summs0);
                    let c1 = vcvtq_f32_s32(summs1);
                    let c2 = vcvtq_f32_s32(summs2);
                    let c3 = vcvtq_f32_s32(summs3);

                    let p0 = vmulq_f32(c0, v_weight);
                    let p1 = vmulq_f32(c1, v_weight);
                    let p2 = vmulq_f32(c2, v_weight);
                    let p3 = vmulq_f32(c3, v_weight);

                    let prepared_px0 = vcvtaq_s32_f32(p0);
                    let prepared_px1 = vcvtaq_s32_f32(p1);
                    let prepared_px2 = vcvtaq_s32_f32(p2);
                    let prepared_px3 = vcvtaq_s32_f32(p3);

                    let dst_ptr0 = bytes.get_ptr(current_y + current_px0);
                    let dst_ptr1 = bytes.get_ptr(current_y + current_px1);
                    let dst_ptr2 = bytes.get_ptr(current_y + current_px2);
                    let dst_ptr3 = bytes.get_ptr(current_y + current_px3);

                    store_u16_s32_x4::<CN>(
                        (dst_ptr0, dst_ptr1, dst_ptr2, dst_ptr3),
                        int32x4x4_t(prepared_px0, prepared_px1, prepared_px2, prepared_px3),
                    );

                    let d_arr_index_1 = ((y + radius_64) & 1023) as usize;
                    let d_arr_index_2 = ((y - radius_64) & 1023) as usize;
                    let d_arr_index = (y & 1023) as usize;

                    let da_b = buffer.get_unchecked(d_arr_index);
                    let da_b1 = buffer.get_unchecked(d_arr_index_1);
                    let da_b2 = buffer.get_unchecked(d_arr_index_2);

                    let stored0 = vld1q_s32(da_b.as_ptr().cast());
                    let stored1 = vld1q_s32(da_b[1..].as_ptr().cast());
                    let stored2 = vld1q_s32(da_b[2..].as_ptr().cast());
                    let stored3 = vld1q_s32(da_b[3..].as_ptr().cast());

                    let stored_10 = vld1q_s32(da_b1.as_ptr().cast());
                    let stored_11 = vld1q_s32(da_b1[1..].as_ptr().cast());
                    let stored_12 = vld1q_s32(da_b1[2..].as_ptr().cast());
                    let stored_13 = vld1q_s32(da_b1[3..].as_ptr().cast());

                    let j0 = vsubq_s32(stored0, stored_10);
                    let j1 = vsubq_s32(stored1, stored_11);
                    let j2 = vsubq_s32(stored2, stored_12);
                    let j3 = vsubq_s32(stored3, stored_13);

                    let stored_20 = vld1q_s32(da_b2.as_ptr().cast());
                    let stored_21 = vld1q_s32(da_b2[1..].as_ptr().cast());
                    let stored_22 = vld1q_s32(da_b2[2..].as_ptr().cast());
                    let stored_23 = vld1q_s32(da_b2[3..].as_ptr().cast());

                    let k0 = vmulq_by_3_s32(j0);
                    let k1 = vmulq_by_3_s32(j1);
                    let k2 = vmulq_by_3_s32(j2);
                    let k3 = vmulq_by_3_s32(j3);

                    let new_diff0 = vsubq_s32(k0, stored_20);
                    let new_diff1 = vsubq_s32(k1, stored_21);
                    let new_diff2 = vsubq_s32(k2, stored_22);
                    let new_diff3 = vsubq_s32(k3, stored_23);

                    diffs0 = vaddq_s32(diffs0, new_diff0);
                    diffs1 = vaddq_s32(diffs1, new_diff1);
                    diffs2 = vaddq_s32(diffs2, new_diff2);
                    diffs3 = vaddq_s32(diffs3, new_diff3);
                } else if y + radius_64 >= 0 {
                    let arr_index = (y & 1023) as usize;
                    let arr_index_1 = ((y + radius_64) & 1023) as usize;
                    let da_b = buffer.get_unchecked(arr_index);
                    let da_b1 = buffer.get_unchecked(arr_index_1);

                    let stored0 = vld1q_s32(da_b.as_ptr().cast());
                    let stored1 = vld1q_s32(da_b[1..].as_ptr().cast());
                    let stored2 = vld1q_s32(da_b[2..].as_ptr().cast());
                    let stored3 = vld1q_s32(da_b[3..].as_ptr().cast());

                    let stored_10 = vld1q_s32(da_b1.as_ptr().cast());
                    let stored_11 = vld1q_s32(da_b1[1..].as_ptr().cast());
                    let stored_12 = vld1q_s32(da_b1[2..].as_ptr().cast());
                    let stored_13 = vld1q_s32(da_b1[3..].as_ptr().cast());

                    diffs0 = vmlaq_n_s32(diffs0, vsubq_s32(stored0, stored_10), 3);
                    diffs1 = vmlaq_n_s32(diffs1, vsubq_s32(stored1, stored_11), 3);
                    diffs2 = vmlaq_n_s32(diffs2, vsubq_s32(stored2, stored_12), 3);
                    diffs3 = vmlaq_n_s32(diffs3, vsubq_s32(stored3, stored_13), 3);
                } else if y + 2 * radius_64 >= 0 {
                    let arr_index = ((y + radius_64) & 1023) as usize;
                    let da_b = buffer.get_unchecked(arr_index);

                    let stored0 = vld1q_s32(da_b.as_ptr().cast());
                    let stored1 = vld1q_s32(da_b[1..].as_ptr().cast());
                    let stored2 = vld1q_s32(da_b[2..].as_ptr().cast());
                    let stored3 = vld1q_s32(da_b[3..].as_ptr().cast());

                    diffs0 = vmlaq_n_s32(diffs0, stored0, -3);
                    diffs1 = vmlaq_n_s32(diffs1, stored1, -3);
                    diffs2 = vmlaq_n_s32(diffs2, stored2, -3);
                    diffs3 = vmlaq_n_s32(diffs3, stored3, -3);
                }

                let next_row_y = clamp_edge!(edge_mode, y + ((3 * radius_64) >> 1), 0, height_wide)
                    * (stride as usize);

                let s_ptr0 = bytes.get_ptr(next_row_y + current_px0);
                let s_ptr1 = bytes.get_ptr(next_row_y + current_px1);
                let s_ptr2 = bytes.get_ptr(next_row_y + current_px2);
                let s_ptr3 = bytes.get_ptr(next_row_y + current_px3);

                let pixel_color0 = load_u16_s32_fast::<CN>(s_ptr0);
                let pixel_color1 = load_u16_s32_fast::<CN>(s_ptr1);
                let pixel_color2 = load_u16_s32_fast::<CN>(s_ptr2);
                let pixel_color3 = load_u16_s32_fast::<CN>(s_ptr3);

                let arr_index = ((y + 2 * radius_64) & 1023) as usize;
                let da_b = buffer.get_unchecked_mut(arr_index);

                diffs0 = vaddq_s32(diffs0, pixel_color0);
                diffs1 = vaddq_s32(diffs1, pixel_color1);
                diffs2 = vaddq_s32(diffs2, pixel_color2);
                diffs3 = vaddq_s32(diffs3, pixel_color3);

                vst1q_s32(da_b.as_mut_ptr().cast(), pixel_color0);
                vst1q_s32(da_b[1..].as_mut_ptr().cast(), pixel_color1);
                vst1q_s32(da_b[2..].as_mut_ptr().cast(), pixel_color2);
                vst1q_s32(da_b[3..].as_mut_ptr().cast(), pixel_color3);

                ders0 = vaddq_s32(ders0, diffs0);
                ders1 = vaddq_s32(ders1, diffs1);
                ders2 = vaddq_s32(ders2, diffs2);
                ders3 = vaddq_s32(ders3, diffs3);

                summs0 = vaddq_s32(summs0, ders0);
                summs1 = vaddq_s32(summs1, ders1);
                summs2 = vaddq_s32(summs2, ders2);
                summs3 = vaddq_s32(summs3, ders3);
            }
            xx += 4;
        }

        for x in xx..width.min(end) {
            let mut diffs = vdupq_n_s32(0);
            let mut ders = vdupq_n_s32(0);
            let mut summs = vdupq_n_s32(0);

            let current_px = (x * CN as u32) as usize;

            let start_y = 0 - 3 * radius as i64;
            for y in start_y..height_wide {
                let current_y = (y * (stride as i64)) as usize;

                if y >= 0 {
                    let prepared_px_s32 = vcvtaq_s32_f32(vmulq_f32(vcvtq_f32_s32(summs), v_weight));
                    let prepared_u16 = vqmovun_s32(prepared_px_s32);

                    let bytes_offset = current_y + current_px;

                    let dst_ptr = bytes.get_ptr(bytes_offset);
                    store_u16x4::<CN>(dst_ptr, prepared_u16);

                    let d_arr_index_1 = ((y + radius_64) & 1023) as usize;
                    let d_arr_index_2 = ((y - radius_64) & 1023) as usize;
                    let d_arr_index = (y & 1023) as usize;

                    let buf_ptr = buffer.get_unchecked(d_arr_index);
                    let stored = vld1q_s32(buf_ptr.as_ptr().cast());

                    let buf_ptr_1 = buffer.get_unchecked(d_arr_index_1);
                    let stored_1 = vld1q_s32(buf_ptr_1.as_ptr().cast());

                    let buf_ptr_2 = buffer.get_unchecked(d_arr_index_2);
                    let stored_2 = vld1q_s32(buf_ptr_2.as_ptr().cast());

                    let new_diff = vsubq_s32(vmulq_by_3_s32(vsubq_s32(stored, stored_1)), stored_2);
                    diffs = vaddq_s32(diffs, new_diff);
                } else if y + radius_64 >= 0 {
                    let arr_index = (y & 1023) as usize;
                    let arr_index_1 = ((y + radius_64) & 1023) as usize;
                    let buf_ptr = buffer.get_unchecked(arr_index);
                    let stored = vld1q_s32(buf_ptr.as_ptr().cast());

                    let buf_ptr_1 = buffer.get_unchecked(arr_index_1);
                    let stored_1 = vld1q_s32(buf_ptr_1.as_ptr().cast());

                    diffs = vmlaq_n_s32(diffs, vsubq_s32(stored, stored_1), 3);
                } else if y + 2 * radius_64 >= 0 {
                    let arr_index = ((y + radius_64) & 1023) as usize;
                    let buf_ptr = buffer.get_unchecked(arr_index);
                    let stored = vld1q_s32(buf_ptr.as_ptr().cast());
                    diffs = vmlaq_n_s32(diffs, stored, -3);
                }

                let next_row_y = clamp_edge!(edge_mode, y + ((3 * radius_64) >> 1), 0, height_wide)
                    * (stride as usize);
                let next_row_x = (x * CN as u32) as usize;

                let s_ptr = bytes.get_ptr(next_row_y + next_row_x);

                let pixel_color = load_u16_s32_fast::<CN>(s_ptr);

                let arr_index = ((y + 2 * radius_64) & 1023) as usize;
                let buf_ptr = buffer.get_unchecked_mut(arr_index);

                diffs = vaddq_s32(diffs, pixel_color);
                ders = vaddq_s32(ders, diffs);
                summs = vaddq_s32(summs, ders);
                vst1q_s32(buf_ptr.as_mut_ptr().cast(), pixel_color);
            }
        }
    }
}

pub(crate) fn fgn_horizontal_pass_neon_u16<const CN: usize>(
    bytes: &UnsafeSlice<u16>,
    stride: u32,
    width: u32,
    height: u32,
    radius: u32,
    start: u32,
    end: u32,
    edge_mode: EdgeMode,
) {
    unsafe {
        let mut buffer = ScratchBuffer::<[NeonI32x4; 4], 1024>::new(1024);
        let buffer = buffer.as_mut_slice();

        let width_wide = width as i64;

        let radius_64 = radius as i64;
        let weight = 1.0f32 / ((radius as f32) * (radius as f32) * (radius as f32));
        let v_weight = vdupq_n_f32(weight);

        let mut yy = start;

        while yy + 4 <= height.min(end) {
            let mut diffs0 = vdupq_n_s32(0);
            let mut diffs1 = vdupq_n_s32(0);
            let mut diffs2 = vdupq_n_s32(0);
            let mut diffs3 = vdupq_n_s32(0);

            let mut ders0 = vdupq_n_s32(0);
            let mut ders1 = vdupq_n_s32(0);
            let mut ders2 = vdupq_n_s32(0);
            let mut ders3 = vdupq_n_s32(0);

            let mut summs0 = vdupq_n_s32(0);
            let mut summs1 = vdupq_n_s32(0);
            let mut summs2 = vdupq_n_s32(0);
            let mut summs3 = vdupq_n_s32(0);

            let current_y0 = ((yy as i64) * (stride as i64)) as usize;
            let current_y1 = ((yy as i64 + 1) * (stride as i64)) as usize;
            let current_y2 = ((yy as i64 + 2) * (stride as i64)) as usize;
            let current_y3 = ((yy as i64 + 3) * (stride as i64)) as usize;

            for x in (0 - 3 * radius_64)..(width as i64) {
                if x >= 0 {
                    let current_px = x as usize * CN;

                    let c0 = vcvtq_f32_s32(summs0);
                    let c1 = vcvtq_f32_s32(summs1);
                    let c2 = vcvtq_f32_s32(summs2);
                    let c3 = vcvtq_f32_s32(summs3);

                    let p0 = vmulq_f32(c0, v_weight);
                    let p1 = vmulq_f32(c1, v_weight);
                    let p2 = vmulq_f32(c2, v_weight);
                    let p3 = vmulq_f32(c3, v_weight);

                    let prepared_px0 = vcvtaq_s32_f32(p0);
                    let prepared_px1 = vcvtaq_s32_f32(p1);
                    let prepared_px2 = vcvtaq_s32_f32(p2);
                    let prepared_px3 = vcvtaq_s32_f32(p3);

                    let dst_ptr0 = bytes.get_ptr(current_y0 + current_px);
                    let dst_ptr1 = bytes.get_ptr(current_y1 + current_px);
                    let dst_ptr2 = bytes.get_ptr(current_y2 + current_px);
                    let dst_ptr3 = bytes.get_ptr(current_y3 + current_px);

                    store_u16_s32_x4::<CN>(
                        (dst_ptr0, dst_ptr1, dst_ptr2, dst_ptr3),
                        int32x4x4_t(prepared_px0, prepared_px1, prepared_px2, prepared_px3),
                    );

                    let d_arr_index_1 = ((x + radius_64) & 1023) as usize;
                    let d_arr_index_2 = ((x - radius_64) & 1023) as usize;
                    let d_arr_index = (x & 1023) as usize;

                    let da_b = buffer.get_unchecked(d_arr_index);
                    let da_b1 = buffer.get_unchecked(d_arr_index_1);
                    let da_b2 = buffer.get_unchecked(d_arr_index_2);

                    let stored0 = vld1q_s32(da_b.as_ptr().cast());
                    let stored1 = vld1q_s32(da_b[1..].as_ptr().cast());
                    let stored2 = vld1q_s32(da_b[2..].as_ptr().cast());
                    let stored3 = vld1q_s32(da_b[3..].as_ptr().cast());

                    let stored_10 = vld1q_s32(da_b1.as_ptr().cast());
                    let stored_11 = vld1q_s32(da_b1[1..].as_ptr().cast());
                    let stored_12 = vld1q_s32(da_b1[2..].as_ptr().cast());
                    let stored_13 = vld1q_s32(da_b1[3..].as_ptr().cast());

                    let j0 = vsubq_s32(stored0, stored_10);
                    let j1 = vsubq_s32(stored1, stored_11);
                    let j2 = vsubq_s32(stored2, stored_12);
                    let j3 = vsubq_s32(stored3, stored_13);

                    let stored_20 = vld1q_s32(da_b2.as_ptr().cast());
                    let stored_21 = vld1q_s32(da_b2[1..].as_ptr().cast());
                    let stored_22 = vld1q_s32(da_b2[2..].as_ptr().cast());
                    let stored_23 = vld1q_s32(da_b2[3..].as_ptr().cast());

                    let k0 = vmulq_by_3_s32(j0);
                    let k1 = vmulq_by_3_s32(j1);
                    let k2 = vmulq_by_3_s32(j2);
                    let k3 = vmulq_by_3_s32(j3);

                    let new_diff0 = vsubq_s32(k0, stored_20);
                    let new_diff1 = vsubq_s32(k1, stored_21);
                    let new_diff2 = vsubq_s32(k2, stored_22);
                    let new_diff3 = vsubq_s32(k3, stored_23);

                    diffs0 = vaddq_s32(diffs0, new_diff0);
                    diffs1 = vaddq_s32(diffs1, new_diff1);
                    diffs2 = vaddq_s32(diffs2, new_diff2);
                    diffs3 = vaddq_s32(diffs3, new_diff3);
                } else if x + radius_64 >= 0 {
                    let arr_index = (x & 1023) as usize;
                    let arr_index_1 = ((x + radius_64) & 1023) as usize;
                    let da_b = buffer.get_unchecked(arr_index);
                    let da_b1 = buffer.get_unchecked(arr_index_1);

                    let stored0 = vld1q_s32(da_b.as_ptr().cast());
                    let stored1 = vld1q_s32(da_b[1..].as_ptr().cast());
                    let stored2 = vld1q_s32(da_b[2..].as_ptr().cast());
                    let stored3 = vld1q_s32(da_b[3..].as_ptr().cast());

                    let stored_10 = vld1q_s32(da_b1.as_ptr().cast());
                    let stored_11 = vld1q_s32(da_b1[1..].as_ptr().cast());
                    let stored_12 = vld1q_s32(da_b1[2..].as_ptr().cast());
                    let stored_13 = vld1q_s32(da_b1[3..].as_ptr().cast());

                    diffs0 = vmlaq_n_s32(diffs0, vsubq_s32(stored0, stored_10), 3);
                    diffs1 = vmlaq_n_s32(diffs1, vsubq_s32(stored1, stored_11), 3);
                    diffs2 = vmlaq_n_s32(diffs2, vsubq_s32(stored2, stored_12), 3);
                    diffs3 = vmlaq_n_s32(diffs3, vsubq_s32(stored3, stored_13), 3);
                } else if x + 2 * radius_64 >= 0 {
                    let arr_index = ((x + radius_64) & 1023) as usize;
                    let da_b = buffer.get_unchecked(arr_index);

                    let stored0 = vld1q_s32(da_b.as_ptr().cast());
                    let stored1 = vld1q_s32(da_b[1..].as_ptr().cast());
                    let stored2 = vld1q_s32(da_b[2..].as_ptr().cast());
                    let stored3 = vld1q_s32(da_b[3..].as_ptr().cast());

                    diffs0 = vmlaq_n_s32(diffs0, stored0, -3);
                    diffs1 = vmlaq_n_s32(diffs1, stored1, -3);
                    diffs2 = vmlaq_n_s32(diffs2, stored2, -3);
                    diffs3 = vmlaq_n_s32(diffs3, stored3, -3);
                }

                let next_row_x = clamp_edge!(edge_mode, x + 3 * radius_64 / 2, 0, width_wide);
                let next_row_px = next_row_x * CN;

                let s_ptr0 = bytes.get_ptr(current_y0 + next_row_px);
                let s_ptr1 = bytes.get_ptr(current_y1 + next_row_px);
                let s_ptr2 = bytes.get_ptr(current_y2 + next_row_px);
                let s_ptr3 = bytes.get_ptr(current_y3 + next_row_px);

                let pixel_color0 = load_u16_s32_fast::<CN>(s_ptr0);
                let pixel_color1 = load_u16_s32_fast::<CN>(s_ptr1);
                let pixel_color2 = load_u16_s32_fast::<CN>(s_ptr2);
                let pixel_color3 = load_u16_s32_fast::<CN>(s_ptr3);

                let arr_index = ((x + 2 * radius_64) & 1023) as usize;
                let da_b = buffer.get_unchecked_mut(arr_index);

                diffs0 = vaddq_s32(diffs0, pixel_color0);
                diffs1 = vaddq_s32(diffs1, pixel_color1);
                diffs2 = vaddq_s32(diffs2, pixel_color2);
                diffs3 = vaddq_s32(diffs3, pixel_color3);

                vst1q_s32(da_b.as_mut_ptr().cast(), pixel_color0);
                vst1q_s32(da_b[1..].as_mut_ptr().cast(), pixel_color1);
                vst1q_s32(da_b[2..].as_mut_ptr().cast(), pixel_color2);
                vst1q_s32(da_b[3..].as_mut_ptr().cast(), pixel_color3);

                ders0 = vaddq_s32(ders0, diffs0);
                ders1 = vaddq_s32(ders1, diffs1);
                ders2 = vaddq_s32(ders2, diffs2);
                ders3 = vaddq_s32(ders3, diffs3);

                summs0 = vaddq_s32(summs0, ders0);
                summs1 = vaddq_s32(summs1, ders1);
                summs2 = vaddq_s32(summs2, ders2);
                summs3 = vaddq_s32(summs3, ders3);
            }

            yy += 4;
        }

        for y in yy..height.min(end) {
            let mut diffs: int32x4_t = vdupq_n_s32(0);
            let mut ders: int32x4_t = vdupq_n_s32(0);
            let mut summs: int32x4_t = vdupq_n_s32(0);

            let current_y = ((y as i64) * (stride as i64)) as usize;

            for x in (0 - 3 * radius_64)..(width as i64) {
                if x >= 0 {
                    let current_px = x as usize * CN;

                    let prepared_px_s32 = vcvtaq_s32_f32(vmulq_f32(vcvtq_f32_s32(summs), v_weight));
                    let prepared_u16 = vqmovun_s32(prepared_px_s32);

                    let bytes_offset = current_y + current_px;

                    store_u16x4::<CN>(bytes.get_ptr(bytes_offset), prepared_u16);

                    let d_arr_index_1 = ((x + radius_64) & 1023) as usize;
                    let d_arr_index_2 = ((x - radius_64) & 1023) as usize;
                    let d_arr_index = (x & 1023) as usize;

                    let buf_ptr = buffer.get_unchecked(d_arr_index);
                    let stored = vld1q_s32(buf_ptr.as_ptr().cast());

                    let buf_ptr_1 = buffer.get_unchecked(d_arr_index_1);
                    let stored_1 = vld1q_s32(buf_ptr_1.as_ptr().cast());

                    let buf_ptr_2 = buffer.get_unchecked(d_arr_index_2);
                    let stored_2 = vld1q_s32(buf_ptr_2.as_ptr().cast());

                    let new_diff = vsubq_s32(vmulq_by_3_s32(vsubq_s32(stored, stored_1)), stored_2);
                    diffs = vaddq_s32(diffs, new_diff);
                } else if x + radius_64 >= 0 {
                    let arr_index = (x & 1023) as usize;
                    let arr_index_1 = ((x + radius_64) & 1023) as usize;
                    let buf_ptr = buffer.get_unchecked(arr_index);
                    let stored = vld1q_s32(buf_ptr.as_ptr().cast());

                    let buf_ptr_1 = buffer.get_unchecked(arr_index_1);
                    let stored_1 = vld1q_s32(buf_ptr_1.as_ptr().cast());

                    diffs = vmlaq_n_s32(diffs, vsubq_s32(stored, stored_1), 3);
                } else if x + 2 * radius_64 >= 0 {
                    let arr_index = ((x + radius_64) & 1023) as usize;
                    let buf_ptr = buffer.get_unchecked(arr_index);
                    let stored = vld1q_s32(buf_ptr.as_ptr().cast());
                    diffs = vmlaq_n_s32(diffs, stored, -3);
                }

                let next_row_y = (y as usize) * (stride as usize);
                let next_row_x = clamp_edge!(edge_mode, x + 3 * radius_64 / 2, 0, width_wide);
                let next_row_px = next_row_x * CN;

                let s_ptr = bytes.get_ptr(next_row_y + next_row_px);

                let pixel_color = load_u16_s32_fast::<CN>(s_ptr);

                let arr_index = ((x + 2 * radius_64) & 1023) as usize;
                let buf_ptr = buffer.get_unchecked_mut(arr_index);

                diffs = vaddq_s32(diffs, pixel_color);
                ders = vaddq_s32(ders, diffs);
                summs = vaddq_s32(summs, ders);
                vst1q_s32(buf_ptr.as_mut_ptr().cast(), pixel_color);
            }
        }
    }
}