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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291

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

use crate::convolution::optimisations::Normalizer32;
use crate::pixels::U16x3;
use crate::{simd_utils, ImageView, ImageViewMut};

#[inline]
pub(crate) fn horiz_convolution(
    src_view: &impl ImageView<Pixel = U16x3>,
    dst_view: &mut impl ImageViewMut<Pixel = U16x3>,
    offset: u32,
    normalizer: &Normalizer32,
) {
    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()
#[target_feature(enable = "avx2")]
unsafe fn horiz_convolution_four_rows(
    src_rows: [&[U16x3]; 4],
    dst_rows: [&mut [U16x3]; 4],
    normalizer: &Normalizer32,
) {
    let precision = normalizer.precision();
    let half_error = 1i64 << (precision - 1);
    let mut rg_buf = [0i64; 4];
    let mut rg_bb_buf = [0i64; 4];
    let mut bbb_buf = [0i64; 4];

    /*
        |R    G    B   | |R    G    B   | |R    G   | - |B   | |R    G    B   | |R    G    B   | |R   |
        |0001 0203 0405| |0607 0809 1011| |1213 1415| - |0001| |0203 0405 0607| |0809 1011 1213| |1415|

        Shuffle to extract RG components of pixels 0 and 3 as i64:
        lo: -1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0
        hi: -1, -1, -1, -1, -1, -1, 5, 4, -1, -1, -1, -1, -1, -1, 3, 2

        Shuffle to extract RG components of pixels 1 and 4 as i64:
        lo: -1, -1, -1, -1, -1, -1, 9, 8, -1, -1, -1, -1, -1, -1, 7, 6
        hi: -1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8

        Shuffle to extract RG components of pixel 2 and BB of pixels 2-3 as i64:
        lo: -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12
        hi: -1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 1, 0

        Shuffle to extract BB components of pixels 0, 1 and 4 as i64:
        lo: -1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 5, 4
        hi: -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 13, 12
    */

    let rg03_shuffle = _mm256_set_m128i(
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 5, 4, -1, -1, -1, -1, -1, -1, 3, 2),
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0),
    );
    let rg14_shuffle = _mm256_set_m128i(
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8),
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 9, 8, -1, -1, -1, -1, -1, -1, 7, 6),
    );
    let rg3_b3b4_shuffle = _mm256_set_m128i(
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 1, 0),
        _mm_set_epi8(
            -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12,
        ),
    );
    let b1b2_b5_shuffle = _mm256_set_m128i(
        _mm_set_epi8(
            -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 13, 12,
        ),
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 5, 4),
    );

    let width = src_rows[0].len();

    for (dst_x, chunk) in normalizer.chunks().iter().enumerate() {
        let mut x = chunk.start as usize;
        let mut rg_sum = [_mm256_set1_epi8(0); 4];
        let mut rg_bb_sum = [_mm256_set1_epi8(0); 4];
        let mut bbb_sum = [_mm256_set1_epi8(0); 4];

        let mut coeffs = chunk.values();
        let end_x = x + coeffs.len();

        if width - end_x >= 1 {
            let coeffs_by_5 = coeffs.chunks_exact(5);
            coeffs = coeffs_by_5.remainder();

            for k in coeffs_by_5 {
                let coeff0033_i64x4 =
                    _mm256_set_epi64x(k[3] as i64, k[3] as i64, k[0] as i64, k[0] as i64);
                let coeff1144_i64x2 =
                    _mm256_set_epi64x(k[4] as i64, k[4] as i64, k[1] as i64, k[1] as i64);
                let coeff2223_i64x2 =
                    _mm256_set_epi64x(k[3] as i64, k[2] as i64, k[2] as i64, k[2] as i64);
                let coeff014_i64x2 = _mm256_set_epi64x(0, k[4] as i64, k[1] as i64, k[0] as i64);

                for i in 0..4 {
                    let source = simd_utils::loadu_si256(src_rows[i], x);

                    let rg03_i64x4 = _mm256_shuffle_epi8(source, rg03_shuffle);
                    rg_sum[i] =
                        _mm256_add_epi64(rg_sum[i], _mm256_mul_epi32(rg03_i64x4, coeff0033_i64x4));

                    let rg14_i64x4 = _mm256_shuffle_epi8(source, rg14_shuffle);
                    rg_sum[i] =
                        _mm256_add_epi64(rg_sum[i], _mm256_mul_epi32(rg14_i64x4, coeff1144_i64x2));

                    let rg_bb_i64x4 = _mm256_shuffle_epi8(source, rg3_b3b4_shuffle);
                    rg_bb_sum[i] = _mm256_add_epi64(
                        rg_bb_sum[i],
                        _mm256_mul_epi32(rg_bb_i64x4, coeff2223_i64x2),
                    );

                    let bbb_i64x4 = _mm256_shuffle_epi8(source, b1b2_b5_shuffle);
                    bbb_sum[i] =
                        _mm256_add_epi64(bbb_sum[i], _mm256_mul_epi32(bbb_i64x4, coeff014_i64x2));
                }
                x += 5;
            }
        }

        for &k in coeffs {
            let coeff_i64x4 = _mm256_set1_epi64x(k as i64);

            for i in 0..4 {
                let &pixel = src_rows[i].get_unchecked(x);
                let rgb_i64x4 =
                    _mm256_set_epi64x(0, pixel.0[2] as i64, pixel.0[1] as i64, pixel.0[0] as i64);
                rg_bb_sum[i] =
                    _mm256_add_epi64(rg_bb_sum[i], _mm256_mul_epi32(rgb_i64x4, coeff_i64x4));
            }
            x += 1;
        }

        for i in 0..4 {
            _mm256_storeu_si256(rg_buf.as_mut_ptr() as *mut __m256i, rg_sum[i]);
            _mm256_storeu_si256(rg_bb_buf.as_mut_ptr() as *mut __m256i, rg_bb_sum[i]);
            _mm256_storeu_si256(bbb_buf.as_mut_ptr() as *mut __m256i, bbb_sum[i]);
            let dst_pixel = dst_rows[i].get_unchecked_mut(dst_x);
            dst_pixel.0[0] = normalizer.clip(rg_buf[0] + rg_buf[2] + rg_bb_buf[0] + half_error);
            dst_pixel.0[1] = normalizer.clip(rg_buf[1] + rg_buf[3] + rg_bb_buf[1] + half_error);
            dst_pixel.0[2] = normalizer.clip(
                rg_bb_buf[2] + rg_bb_buf[3] + bbb_buf[0] + bbb_buf[1] + bbb_buf[2] + half_error,
            );
        }
    }
}

/// For safety, it is necessary to ensure the following conditions:
/// - bounds.len() == dst_row.len()
/// - coefficients_chunks.len() == dst_row.len()
/// - max(chunk.start + chunk.values.len() for chunk in coefficients_chunks) <= src_row.len()
#[target_feature(enable = "avx2")]
unsafe fn horiz_convolution_one_row(
    src_row: &[U16x3],
    dst_row: &mut [U16x3],
    normalizer: &Normalizer32,
) {
    let precision = normalizer.precision();
    let half_error = 1i64 << (precision - 1);
    let mut rg_buf = [0i64; 4];
    let mut rg_bb_buf = [0i64; 4];
    let mut bbb_buf = [0i64; 4];

    /*
        |R    G    B   | |R    G    B   | |R    G   | - |B   | |R    G    B   | |R    G    B   | |R   |
        |0001 0203 0405| |0607 0809 1011| |1213 1415| - |0001| |0203 0405 0607| |0809 1011 1213| |1415|

        Shuffle to extract RG components of pixels 0 and 3 as i64:
        lo: -1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0
        hi: -1, -1, -1, -1, -1, -1, 5, 4, -1, -1, -1, -1, -1, -1, 3, 2

        Shuffle to extract RG components of pixels 1 and 4 as i64:
        lo: -1, -1, -1, -1, -1, -1, 9, 8, -1, -1, -1, -1, -1, -1, 7, 6
        hi: -1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8

        Shuffle to extract RG components of pixel 2 and BB of pixels 2-3 as i64:
        lo: -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12
        hi: -1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 1, 0

        Shuffle to extract BB components of pixels 0, 1 and 4 as i64:
        lo: -1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 5, 4
        hi: -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 13, 12
    */

    let rg03_shuffle = _mm256_set_m128i(
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 5, 4, -1, -1, -1, -1, -1, -1, 3, 2),
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0),
    );
    let rg14_shuffle = _mm256_set_m128i(
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8),
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 9, 8, -1, -1, -1, -1, -1, -1, 7, 6),
    );
    let rg3_b3b4_shuffle = _mm256_set_m128i(
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 1, 0),
        _mm_set_epi8(
            -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12,
        ),
    );
    let b1b2_b5_shuffle = _mm256_set_m128i(
        _mm_set_epi8(
            -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 13, 12,
        ),
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 5, 4),
    );

    let zero_i64x4 = _mm256_set1_epi8(0);

    let width = src_row.len();

    for (dst_x, chunk) in normalizer.chunks().iter().enumerate() {
        let mut x = chunk.start as usize;
        let mut rg_sum = zero_i64x4;
        let mut rg_bb_sum = zero_i64x4;
        let mut bbb_sum = zero_i64x4;

        let mut coeffs = chunk.values();
        let end_x = x + coeffs.len();

        if width - end_x >= 1 {
            let coeffs_by_5 = coeffs.chunks_exact(5);
            coeffs = coeffs_by_5.remainder();

            for k in coeffs_by_5 {
                let coeff0033_i64x4 =
                    _mm256_set_epi64x(k[3] as i64, k[3] as i64, k[0] as i64, k[0] as i64);
                let coeff1144_i64x2 =
                    _mm256_set_epi64x(k[4] as i64, k[4] as i64, k[1] as i64, k[1] as i64);
                let coeff2223_i64x2 =
                    _mm256_set_epi64x(k[3] as i64, k[2] as i64, k[2] as i64, k[2] as i64);
                let coeff014_i64x2 = _mm256_set_epi64x(0, k[4] as i64, k[1] as i64, k[0] as i64);

                let source = simd_utils::loadu_si256(src_row, x);

                let rg03_i64x4 = _mm256_shuffle_epi8(source, rg03_shuffle);
                rg_sum = _mm256_add_epi64(rg_sum, _mm256_mul_epi32(rg03_i64x4, coeff0033_i64x4));

                let rg14_i64x4 = _mm256_shuffle_epi8(source, rg14_shuffle);
                rg_sum = _mm256_add_epi64(rg_sum, _mm256_mul_epi32(rg14_i64x4, coeff1144_i64x2));

                let rg_bb_i64x4 = _mm256_shuffle_epi8(source, rg3_b3b4_shuffle);
                rg_bb_sum =
                    _mm256_add_epi64(rg_bb_sum, _mm256_mul_epi32(rg_bb_i64x4, coeff2223_i64x2));

                let bbb_i64x4 = _mm256_shuffle_epi8(source, b1b2_b5_shuffle);
                bbb_sum = _mm256_add_epi64(bbb_sum, _mm256_mul_epi32(bbb_i64x4, coeff014_i64x2));

                x += 5;
            }
        }

        for &k in coeffs {
            let coeff_i64x4 = _mm256_set1_epi64x(k as i64);
            let &pixel = src_row.get_unchecked(x);
            let rgb_i64x4 =
                _mm256_set_epi64x(0, pixel.0[2] as i64, pixel.0[1] as i64, pixel.0[0] as i64);
            rg_bb_sum = _mm256_add_epi64(rg_bb_sum, _mm256_mul_epi32(rgb_i64x4, coeff_i64x4));

            x += 1;
        }

        _mm256_storeu_si256(rg_buf.as_mut_ptr() as *mut __m256i, rg_sum);
        _mm256_storeu_si256(rg_bb_buf.as_mut_ptr() as *mut __m256i, rg_bb_sum);
        _mm256_storeu_si256(bbb_buf.as_mut_ptr() as *mut __m256i, bbb_sum);
        let dst_pixel = dst_row.get_unchecked_mut(dst_x);
        dst_pixel.0[0] = normalizer.clip(rg_buf[0] + rg_buf[2] + rg_bb_buf[0] + half_error);
        dst_pixel.0[1] = normalizer.clip(rg_buf[1] + rg_buf[3] + rg_bb_buf[1] + half_error);
        dst_pixel.0[2] = normalizer
            .clip(rg_bb_buf[2] + rg_bb_buf[3] + bbb_buf[0] + bbb_buf[1] + bbb_buf[2] + half_error);
    }
}