fast_image_resize 6.0.0

Library for fast image resizing with using of SIMD instructions
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149

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

use crate::compat::*;
use crate::convolution::optimisations::Normalizer16;
use crate::pixels::U8;
use crate::{simd_utils, ImageView, ImageViewMut};

#[inline]
pub(crate) fn horiz_convolution(
    src_view: &impl ImageView<Pixel = U8>,
    dst_view: &mut impl ImageViewMut<Pixel = U8>,
    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: [&[U8]; 4],
    dst_rows: [&mut [U8]; 4],
    normalizer: &Normalizer16,
) {
    let zero = _mm_setzero_si128();
    let initial = 1 << (normalizer.precision() - 1);
    let mut buf = [0, 0, 0, 0, initial];

    for (dst_x, chunk) in normalizer.chunks().iter().enumerate() {
        let mut x = chunk.start as usize;
        let mut result_i32x4 = [zero, zero, zero, zero];

        let coeffs_by_8 = chunk.values().chunks_exact(8);
        let reminder8 = coeffs_by_8.remainder();
        for k in coeffs_by_8 {
            let coeffs_i16x8 = _mm_loadu_si128(k.as_ptr() as *const __m128i);
            for i in 0..4 {
                let pixels_u8x8 = simd_utils::loadl_epi64(src_rows[i], x);
                let pixels_i16x8 = _mm_cvtepu8_epi16(pixels_u8x8);
                result_i32x4[i] =
                    _mm_add_epi32(result_i32x4[i], _mm_madd_epi16(pixels_i16x8, coeffs_i16x8));
            }
            x += 8;
        }

        let mut coeffs_by_4 = reminder8.chunks_exact(4);
        let reminder4 = coeffs_by_4.remainder();
        if let Some(k) = coeffs_by_4.next() {
            let coeffs_i16x4 = simd_utils::loadl_epi64(k, 0);
            for i in 0..4 {
                let pixels_u8x4 = simd_utils::loadl_epi32(src_rows[i], x);
                let pixels_i16x4 = _mm_cvtepu8_epi16(pixels_u8x4);
                result_i32x4[i] =
                    _mm_add_epi32(result_i32x4[i], _mm_madd_epi16(pixels_i16x4, coeffs_i16x4));
            }
            x += 4;
        }

        let mut result_i32x4 = result_i32x4.map(|v| {
            _mm_storeu_si128(buf.as_mut_ptr() as *mut __m128i, v);
            buf.iter().sum()
        });

        for &coeff in reminder4 {
            let coeff_i32 = coeff as i32;
            for i in 0..4 {
                result_i32x4[i] += src_rows[i].get_unchecked(x).0.to_owned() as i32 * coeff_i32;
            }
            x += 1;
        }

        let result_u8x4 = result_i32x4.map(|v| normalizer.clip(v));
        for i in 0..4 {
            dst_rows[i].get_unchecked_mut(dst_x).0 = result_u8x4[i];
        }
    }
}

/// 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: &[U8], dst_row: &mut [U8], normalizer: &Normalizer16) {
    let zero = _mm_setzero_si128();
    let initial = 1 << (normalizer.precision() - 1);
    let mut buf = [0, 0, 0, 0, initial];

    for (dst_x, chunk) in normalizer.chunks().iter().enumerate() {
        let mut x = chunk.start as usize;
        let mut result_i32x4 = zero;

        let coeffs_by_8 = chunk.values().chunks_exact(8);
        let reminder8 = coeffs_by_8.remainder();
        for k in coeffs_by_8 {
            let coeffs_i16x8 = _mm_loadu_si128(k.as_ptr() as *const __m128i);
            let pixels_u8x8 = simd_utils::loadl_epi64(src_row, x);
            let pixels_i16x8 = _mm_cvtepu8_epi16(pixels_u8x8);
            result_i32x4 = _mm_add_epi32(result_i32x4, _mm_madd_epi16(pixels_i16x8, coeffs_i16x8));
            x += 8;
        }

        let mut coeffs_by_4 = reminder8.chunks_exact(4);
        let reminder4 = coeffs_by_4.remainder();
        if let Some(k) = coeffs_by_4.next() {
            let coeffs_i16x4 = simd_utils::loadl_epi64(k, 0);
            let pixels_u8x4 = simd_utils::loadl_epi32(src_row, x);
            let pixels_i16x4 = _mm_cvtepu8_epi16(pixels_u8x4);
            result_i32x4 = _mm_add_epi32(result_i32x4, _mm_madd_epi16(pixels_i16x4, coeffs_i16x4));
            x += 4;
        }

        _mm_storeu_si128(buf.as_mut_ptr() as *mut __m128i, result_i32x4);
        let mut result_i32 = buf.iter().sum();

        for &coeff in reminder4 {
            let coeff_i32 = coeff as i32;
            result_i32 += src_row.get_unchecked(x).0 as i32 * coeff_i32;
            x += 1;
        }

        dst_row.get_unchecked_mut(dst_x).0 = normalizer.clip(result_i32);
    }
}