fast_image_resize 6.0.0

use core::arch::x86_64::*;

use crate::convolution::optimisations::Normalizer16;
use crate::pixels::U8x2;
use crate::{simd_utils, ImageView, ImageViewMut};

#[inline]
pub(crate) fn horiz_convolution(
    src_view: &impl ImageView<Pixel = U8x2>,
    dst_view: &mut impl ImageViewMut<Pixel = U8x2>,
    offset: u32,
    normalizer: &Normalizer16,
) {
    let dst_height = dst_view.height();

    let src_iter = src_view.iter_4_rows(offset, dst_height + offset);
    let dst_iter = dst_view.iter_4_rows_mut();
    for (src_rows, dst_rows) in src_iter.zip(dst_iter) {
        unsafe {
            horiz_convolution_four_rows(src_rows, dst_rows, normalizer);
        }
    }

    let yy = dst_height - dst_height % 4;
    let src_rows = src_view.iter_rows(yy + offset);
    let dst_rows = dst_view.iter_rows_mut(yy);
    for (src_row, dst_row) in src_rows.zip(dst_rows) {
        unsafe {
            horiz_convolution_one_row(src_row, dst_row, normalizer);
        }
    }
}

/// For safety, it is necessary to ensure the following conditions:
/// - length of all rows in src_rows must be equal
/// - length of all rows in dst_rows must be equal
/// - coefficients_chunks.len() == dst_rows.0.len()
/// - max(chunk.start + chunk.values.len() for chunk in coefficients_chunks) <= src_row.0.len()
/// - precision <= MAX_COEFS_PRECISION
#[inline]
#[target_feature(enable = "sse4.1")]
unsafe fn horiz_convolution_four_rows(
    src_rows: [&[U8x2]; 4],
    dst_rows: [&mut [U8x2]; 4],
    normalizer: &Normalizer16,
) {
    let precision = normalizer.precision();
    let initial = _mm_set1_epi32(1 << (precision - 2));

    /*
        |L  A | |L  A | |L  A | |L  A | |L  A | |L  A | |L  A | |L  A |
        |00 01| |02 03| |04 05| |06 07| |08 09| |10 11| |12 13| |14 15|

        Shuffle components with converting from u8 into i16:

        A: |-1 07| |-1 05| |-1 03| |-1 01|
        L: |-1 06| |-1 04| |-1 02| |-1 00|
    */
    #[rustfmt::skip]
    let sh1 = _mm_set_epi8(
        -1, 7, -1, 5, -1, 3, -1, 1, -1, 6, -1, 4, -1, 2, -1, 0,
    );
    /*
        A: |-1 15| |-1 13| |-1 11| |-1 09|
        L: |-1 14| |-1 12| |-1 10| |-1 08|
    */
    #[rustfmt::skip]
    let sh2 = _mm_set_epi8(
        -1, 15, -1, 13, -1, 11, -1, 9, -1, 14, -1, 12, -1, 10, -1, 8,
    );

    for (dst_x, chunk) in normalizer.chunks().iter().enumerate() {
        let mut x = chunk.start as usize;
        let coeffs = chunk.values();

        let mut sss: [__m128i; 4] = [initial; 4];

        let coeffs_by_8 = coeffs.chunks_exact(8);
        let reminder = coeffs_by_8.remainder();

        for k in coeffs_by_8 {
            let mmk0 = simd_utils::ptr_i16_to_set1_epi64x(k, 0);
            let mmk1 = simd_utils::ptr_i16_to_set1_epi64x(k, 4);

            for i in 0..4 {
                let source = simd_utils::loadu_si128(src_rows[i], x);
                let pix = _mm_shuffle_epi8(source, sh1);
                let tmp_sum = _mm_add_epi32(sss[i], _mm_madd_epi16(pix, mmk0));
                let pix = _mm_shuffle_epi8(source, sh2);
                sss[i] = _mm_add_epi32(tmp_sum, _mm_madd_epi16(pix, mmk1));
            }
            x += 8;
        }

        let coeffs_by_4 = reminder.chunks_exact(4);
        let reminder = coeffs_by_4.remainder();

        for k in coeffs_by_4 {
            let mmk = simd_utils::ptr_i16_to_set1_epi64x(k, 0);

            for i in 0..4 {
                let source = simd_utils::loadl_epi64(src_rows[i], x);
                let pix = _mm_shuffle_epi8(source, sh1);
                sss[i] = _mm_add_epi32(sss[i], _mm_madd_epi16(pix, mmk));
            }
            x += 4;
        }

        let coeffs_by_2 = reminder.chunks_exact(2);
        let reminder = coeffs_by_2.remainder();

        for k in coeffs_by_2 {
            let mmk = simd_utils::mm_load_and_clone_i16x2(k);

            for i in 0..4 {
                let source = simd_utils::loadl_epi32(src_rows[i], x);
                let pix = _mm_shuffle_epi8(source, sh1);
                sss[i] = _mm_add_epi32(sss[i], _mm_madd_epi16(pix, mmk));
            }
            x += 2;
        }

        if let Some(&k) = reminder.first() {
            let mmk = _mm_set1_epi32(k as i32);

            for i in 0..4 {
                let source = simd_utils::loadl_epi16(src_rows[i], x);
                let pix = _mm_shuffle_epi8(source, sh1);
                sss[i] = _mm_add_epi32(sss[i], _mm_madd_epi16(pix, mmk));
            }
        }

        for i in 0..4 {
            set_dst_pixel(sss[i], dst_rows[i], dst_x, normalizer);
        }
    }
}

#[inline]
#[target_feature(enable = "sse4.1")]
unsafe fn set_dst_pixel(raw: __m128i, d_row: &mut [U8x2], dst_x: usize, normalizer: &Normalizer16) {
    let l32x2 = _mm_extract_epi64::<0>(raw);
    let a32x2 = _mm_extract_epi64::<1>(raw);
    let l32 = ((l32x2 >> 32) as i32).saturating_add((l32x2 & 0xffffffff) as i32);
    let a32 = ((a32x2 >> 32) as i32).saturating_add((a32x2 & 0xffffffff) as i32);
    let l8 = normalizer.clip(l32);
    let a8 = normalizer.clip(a32);
    d_row.get_unchecked_mut(dst_x).0 = [l8, a8];
}

/// For safety, it is necessary to ensure the following conditions:
/// - bounds.len() == dst_row.len()
/// - coeffs.len() == dst_rows.0.len() * window_size
/// - max(bound.start + bound.size for bound in bounds) <= src_row.len()
/// - precision <= MAX_COEFS_PRECISION
#[inline]
#[target_feature(enable = "sse4.1")]
unsafe fn horiz_convolution_one_row(
    src_row: &[U8x2],
    dst_row: &mut [U8x2],
    normalizer: &Normalizer16,
) {
    let precision = normalizer.precision();
    /*
       |L  A | |L  A | |L  A | |L  A | |L  A | |L  A | |L  A | |L  A |
       |00 01| |02 03| |04 05| |06 07| |08 09| |10 11| |12 13| |14 15|

       Scale first four pixels into i16:

       A: |-1 07| |-1 05|
       L: |-1 06| |-1 04|
       A: |-1 03| |-1 01|
       L: |-1 02| |-1 00|
    */
    #[rustfmt::skip]
    let pix_sh1 = _mm_set_epi8(
        -1, 7, -1, 5, -1, 6, -1, 4, -1, 3, -1, 1, -1, 2, -1, 0,
    );
    /*
       |C0   | |C1   | |C2   | |C3   | |C4   | |C5   | |C6   | |C7   |
       |00 01| |02 03| |04 05| |06 07| |08 09| |10 11| |12 13| |14 15|

       Duplicate first four coefficients for A and L components of pixels:

       CA: |07 06| |05 04|
       CL: |07 06| |05 04|
       CA: |03 02| |01 00|
       CL: |03 02| |01 00|
    */
    #[rustfmt::skip]
    let coeff_sh1 = _mm_set_epi8(
        7, 6, 5, 4, 7, 6, 5, 4, 3, 2, 1, 0, 3, 2, 1, 0,
    );

    /*
       |L  A | |L  A | |L  A | |L  A | |L  A | |L  A | |L  A | |L  A |
       |00 01| |02 03| |04 05| |06 07| |08 09| |10 11| |12 13| |14 15|

       Scale second four pixels into i16:

       A: |-1 15| |-1 13|
       L: |-1 14| |-1 12|
       A: |-1 11| |-1 09|
       L: |-1 10| |-1 08|
    */
    #[rustfmt::skip]
    let pix_sh2 = _mm_set_epi8(
        -1, 15, -1, 13, -1, 14, -1, 12, -1, 11, -1, 9, -1, 10, -1, 8,
    );
    /*
       |C0   | |C1   | |C2   | |C3   | |C4   | |C5   | |C6   | |C7   |
       |00 01| |02 03| |04 05| |06 07| |08 09| |10 11| |12 13| |14 15|

       Duplicate second four coefficients for A and L components of pixels:

       CA: |15 14| |13 12|
       CL: |15 14| |13 12|
       CA: |11 10| |09 08|
       CL: |11 10| |09 08|
    */
    #[rustfmt::skip]
    let coeff_sh2 = _mm_set_epi8(
        15, 14, 13, 12, 15, 14, 13, 12, 11, 10, 9, 8, 11, 10, 9, 8,
    );

    /*
       |L  A | |L  A | |L  A | |L  A |
       |00 01| |02 03| |04 05| |06 07| |08 09| |10 11| |12 13| |14 15|

       Scale four pixels into i16:

       A: |-1 07| |-1 05|
       L: |-1 06| |-1 04|
       A: |-1 03| |-1 01|
       L: |-1 02| |-1 00|
    */
    let pix_sh3 = _mm_set_epi8(-1, 7, -1, 5, -1, 6, -1, 4, -1, 3, -1, 1, -1, 2, -1, 0);

    for (dst_x, chunk) in normalizer.chunks().iter().enumerate() {
        let mut x = chunk.start as usize;
        let mut coeffs = chunk.values();

        // Lower part will be added to higher, use only half of the error
        let mut sss = _mm_set1_epi32(1 << (precision - 2));

        let coeffs_by_8 = coeffs.chunks_exact(8);
        coeffs = coeffs_by_8.remainder();

        for k in coeffs_by_8 {
            let ksource = simd_utils::loadu_si128(k, 0);
            let source = simd_utils::loadu_si128(src_row, x);

            let pix = _mm_shuffle_epi8(source, pix_sh1);
            let mmk = _mm_shuffle_epi8(ksource, coeff_sh1);
            sss = _mm_add_epi32(sss, _mm_madd_epi16(pix, mmk));

            let pix = _mm_shuffle_epi8(source, pix_sh2);
            let mmk = _mm_shuffle_epi8(ksource, coeff_sh2);
            sss = _mm_add_epi32(sss, _mm_madd_epi16(pix, mmk));

            x += 8;
        }

        let coeffs_by_4 = coeffs.chunks_exact(4);
        let reminder1 = coeffs_by_4.remainder();

        for k in coeffs_by_4 {
            let mmk = _mm_set_epi16(k[3], k[2], k[3], k[2], k[1], k[0], k[1], k[0]);
            let source = simd_utils::loadl_epi64(src_row, x);
            let pix = _mm_shuffle_epi8(source, pix_sh3);
            sss = _mm_add_epi32(sss, _mm_madd_epi16(pix, mmk));

            x += 4
        }

        if !reminder1.is_empty() {
            let mut pixels: [i16; 6] = [0; 6];
            let mut coeffs: [i16; 3] = [0; 3];
            for (i, &coeff) in reminder1.iter().enumerate() {
                coeffs[i] = coeff;
                let pixel: [u8; 2] = src_row.get_unchecked(x).0;
                pixels[i * 2] = pixel[0] as i16;
                pixels[i * 2 + 1] = pixel[1] as i16;
                x += 1;
            }

            let pix = _mm_set_epi16(
                0, pixels[5], 0, pixels[4], pixels[3], pixels[1], pixels[2], pixels[0],
            );
            let mmk = _mm_set_epi16(
                0, coeffs[2], 0, coeffs[2], coeffs[1], coeffs[0], coeffs[1], coeffs[0],
            );

            sss = _mm_add_epi32(sss, _mm_madd_epi16(pix, mmk));
        }

        let lo = _mm_extract_epi64::<0>(sss);
        let hi = _mm_extract_epi64::<1>(sss);

        let a32 = ((lo >> 32) as i32).saturating_add((hi >> 32) as i32);
        let l32 = ((lo & 0xffffffff) as i32).saturating_add((hi & 0xffffffff) as i32);
        let a8 = normalizer.clip(a32);
        let l8 = normalizer.clip(l32);
        dst_row.get_unchecked_mut(dst_x).0 = [l8, a8];
    }
}