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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220

use core::arch::wasm32::*;

use crate::convolution::optimisations::Normalizer16;
use crate::pixels::U8x2;
use crate::{wasm32_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 = "simd128")]
unsafe fn horiz_convolution_four_rows(
    src_rows: [&[U8x2]; 4],
    dst_rows: [&mut [U8x2]; 4],
    normalizer: &Normalizer16,
) {
    let precision = normalizer.precision();
    let initial = i32x4_splat(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|
    */
    const SH1: v128 = i8x16(0, -1, 2, -1, 4, -1, 6, -1, 1, -1, 3, -1, 5, -1, 7, -1);
    /*
        A: |-1 15| |-1 13| |-1 11| |-1 09|
        L: |-1 14| |-1 12| |-1 10| |-1 08|
    */
    const SH2: v128 = i8x16(8, -1, 10, -1, 12, -1, 14, -1, 9, -1, 11, -1, 13, -1, 15, -1);

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

        let mut sss: [v128; 4] = [initial; 4];
        let coeffs = coeffs_chunk.values();

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

        for k in coeffs_by_8 {
            let mmk0 = wasm32_utils::ptr_i16_to_set1_i64(k, 0);
            let mmk1 = wasm32_utils::ptr_i16_to_set1_i64(k, 4);

            for i in 0..4 {
                let source = wasm32_utils::load_v128(src_rows[i], x);
                let pix = i8x16_swizzle(source, SH1);
                let tmp_sum = i32x4_add(sss[i], i32x4_dot_i16x8(pix, mmk0));
                let pix = i8x16_swizzle(source, SH2);
                sss[i] = i32x4_add(tmp_sum, i32x4_dot_i16x8(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 = wasm32_utils::ptr_i16_to_set1_i64(k, 0);

            for i in 0..4 {
                let source = wasm32_utils::loadl_i64(src_rows[i], x);
                let pix = i8x16_swizzle(source, SH1);
                sss[i] = i32x4_add(sss[i], i32x4_dot_i16x8(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 = wasm32_utils::ptr_i16_to_set1_i32(k, 0);

            for i in 0..4 {
                let source = wasm32_utils::loadl_i32(src_rows[i], x);
                let pix = i8x16_swizzle(source, SH1);
                sss[i] = i32x4_add(sss[i], i32x4_dot_i16x8(pix, mmk));
            }
            x += 2;
        }

        if let Some(&k) = reminder.first() {
            let mmk = i32x4_splat(k as i32);

            for i in 0..4 {
                let source = wasm32_utils::loadl_i16(src_rows[i], x);
                let pix = i8x16_swizzle(source, SH1);
                sss[i] = i32x4_add(sss[i], i32x4_dot_i16x8(pix, mmk));
            }
        }

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

/// 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 = "simd128")]
unsafe fn horiz_convolution_one_row(
    src_row: &[U8x2],
    dst_row: &mut [U8x2],
    normalizer: &Normalizer16,
) {
    const SH1: v128 = i8x16(0, -1, 2, -1, 4, -1, 6, -1, 1, -1, 3, -1, 5, -1, 7, -1);
    /*
        A: |-1 15| |-1 13| |-1 11| |-1 09|
        L: |-1 14| |-1 12| |-1 10| |-1 08|
    */
    const SH2: v128 = i8x16(8, -1, 10, -1, 12, -1, 14, -1, 9, -1, 11, -1, 13, -1, 15, -1);

    // Lower part will be added to higher, use only half of the error
    let precision = normalizer.precision();
    let initial = i32x4_splat(1 << (precision - 2));

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

        let mut sss = initial;

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

        for k in coeffs_by_8 {
            let mmk0 = wasm32_utils::ptr_i16_to_set1_i64(k, 0);
            let mmk1 = wasm32_utils::ptr_i16_to_set1_i64(k, 4);

            let source = wasm32_utils::load_v128(src_row, x);
            let pix = i8x16_swizzle(source, SH1);
            let tmp_sum = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk0));
            let pix = i8x16_swizzle(source, SH2);
            sss = i32x4_add(tmp_sum, i32x4_dot_i16x8(pix, mmk1));

            x += 8;
        }

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

        for k in coeffs_by_4 {
            let mmk = wasm32_utils::ptr_i16_to_set1_i64(k, 0);
            let source = wasm32_utils::loadl_i64(src_row, x);
            let pix = i8x16_swizzle(source, SH1);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));
            x += 4
        }

        let coeffs_by_2 = reminder1.chunks_exact(2);
        let reminder = coeffs_by_2.remainder();
        for k in coeffs_by_2 {
            let mmk = wasm32_utils::ptr_i16_to_set1_i32(k, 0);

            let source = wasm32_utils::loadl_i32(src_row, x);
            let pix = i8x16_swizzle(source, SH1);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));

            x += 2;
        }

        if let Some(&k) = reminder.first() {
            let mmk = i32x4_splat(k as i32);
            let source = wasm32_utils::loadl_i16(src_row, x);
            let pix = i8x16_swizzle(source, SH1);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));
        }

        set_dst_pixel(sss, dst_row, dst_x, normalizer);
    }
}

#[inline]
#[target_feature(enable = "simd128")]
unsafe fn set_dst_pixel(raw: v128, d_row: &mut [U8x2], dst_x: usize, normalizer: &Normalizer16) {
    let mut buf = [0i32; 4];
    v128_store(buf.as_mut_ptr() as _, raw);
    let l32 = buf[0].saturating_add(buf[1]);
    let a32 = buf[2].saturating_add(buf[3]);
    let l8 = normalizer.clip(l32);
    let a8 = normalizer.clip(a32);
    d_row.get_unchecked_mut(dst_x).0 = [l8, a8];
}