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

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

use crate::convolution::optimisations::Normalizer32;
use crate::pixels::U16;
use crate::{simd_utils, ImageView, ImageViewMut};

#[inline]
pub(crate) fn horiz_convolution(
    src_view: &impl ImageView<Pixel = U16>,
    dst_view: &mut impl ImageViewMut<Pixel = U16>,
    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 = "sse4.1")]
unsafe fn horiz_convolution_four_rows(
    src_rows: [&[U16]; 4],
    dst_rows: [&mut [U16]; 4],
    normalizer: &Normalizer32,
) {
    let precision = normalizer.precision();
    let half_error = 1i64 << (precision - 1);
    let mut ll_buf = [0i64; 2];

    /*
        |L0  | |L1  | |L2  | |L3  | |L4  | |L5  | |L6  | |L7  |
        |0001| |0203| |0405| |0607| |0809| |1011| |1213| |1415|

        Shuffle to extract L0 and L1 as i64:
        -1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0

        Shuffle to extract L2 and L3 as i64:
        -1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 5, 4

        Shuffle to extract L4 and L5 as i64:
        -1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8

        Shuffle to extract L6 and L7 as i64:
        -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12
    */

    let l0l1_shuffle = _mm_set_epi8(-1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0);
    let l2l3_shuffle = _mm_set_epi8(-1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 5, 4);
    let l4l5_shuffle = _mm_set_epi8(-1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8);
    let l6l7_shuffle = _mm_set_epi8(
        -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12,
    );

    for (dst_x, chunk) in normalizer.chunks().iter().enumerate() {
        let mut x = chunk.start as usize;
        let mut ll_sum = [_mm_set1_epi64x(0); 4];

        let mut coeffs = chunk.values();

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

        for k in coeffs_by_8 {
            let coeff01_i64x2 = _mm_set_epi64x(k[1] as i64, k[0] as i64);
            let coeff23_i64x2 = _mm_set_epi64x(k[3] as i64, k[2] as i64);
            let coeff45_i64x2 = _mm_set_epi64x(k[5] as i64, k[4] as i64);
            let coeff67_i64x2 = _mm_set_epi64x(k[7] as i64, k[6] as i64);

            for i in 0..4 {
                let mut sum = ll_sum[i];
                let source = simd_utils::loadu_si128(src_rows[i], x);

                let l0l1_i64x2 = _mm_shuffle_epi8(source, l0l1_shuffle);
                sum = _mm_add_epi64(sum, _mm_mul_epi32(l0l1_i64x2, coeff01_i64x2));

                let l2l3_i64x2 = _mm_shuffle_epi8(source, l2l3_shuffle);
                sum = _mm_add_epi64(sum, _mm_mul_epi32(l2l3_i64x2, coeff23_i64x2));

                let l4l5_i64x2 = _mm_shuffle_epi8(source, l4l5_shuffle);
                sum = _mm_add_epi64(sum, _mm_mul_epi32(l4l5_i64x2, coeff45_i64x2));

                let l6l7_i64x2 = _mm_shuffle_epi8(source, l6l7_shuffle);
                sum = _mm_add_epi64(sum, _mm_mul_epi32(l6l7_i64x2, coeff67_i64x2));

                ll_sum[i] = sum;
            }
            x += 8;
        }

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

        for k in coeffs_by_4 {
            let coeff01_i64x2 = _mm_set_epi64x(k[1] as i64, k[0] as i64);
            let coeff23_i64x2 = _mm_set_epi64x(k[3] as i64, k[2] as i64);

            for i in 0..4 {
                let mut sum = ll_sum[i];
                let source = simd_utils::loadl_epi64(src_rows[i], x);

                let l0l1_i64x2 = _mm_shuffle_epi8(source, l0l1_shuffle);
                sum = _mm_add_epi64(sum, _mm_mul_epi32(l0l1_i64x2, coeff01_i64x2));

                let l2l3_i64x2 = _mm_shuffle_epi8(source, l2l3_shuffle);
                sum = _mm_add_epi64(sum, _mm_mul_epi32(l2l3_i64x2, coeff23_i64x2));

                ll_sum[i] = sum;
            }
            x += 4;
        }

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

        for k in coeffs_by_2 {
            let coeff01_i64x2 = _mm_set_epi64x(k[1] as i64, k[0] as i64);
            for i in 0..4 {
                let source = simd_utils::loadl_epi32(src_rows[i], x);
                let l_i64x2 = _mm_shuffle_epi8(source, l0l1_shuffle);
                ll_sum[i] = _mm_add_epi64(ll_sum[i], _mm_mul_epi32(l_i64x2, coeff01_i64x2));
            }
            x += 2;
        }

        if let Some(&k) = coeffs.first() {
            let coeff01_i64x2 = _mm_set_epi64x(0, k as i64);
            for i in 0..4 {
                let pixel = src_rows[i].get_unchecked(x).0 as i64;
                let source = _mm_set_epi64x(0, pixel);
                ll_sum[i] = _mm_add_epi64(ll_sum[i], _mm_mul_epi32(source, coeff01_i64x2));
            }
        }

        for i in 0..4 {
            _mm_storeu_si128(ll_buf.as_mut_ptr() as *mut __m128i, ll_sum[i]);
            let dst_pixel = dst_rows[i].get_unchecked_mut(dst_x);
            dst_pixel.0 = normalizer.clip(ll_buf.iter().sum::<i64>() + 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 = "sse4.1")]
unsafe fn horiz_convolution_one_row(
    src_row: &[U16],
    dst_row: &mut [U16],
    normalizer: &Normalizer32,
) {
    let precision = normalizer.precision();
    let half_error = 1i64 << (precision - 1);
    let mut ll_buf = [0i64; 2];

    /*
        |L0  | |L1  | |L2  | |L3  | |L4  | |L5  | |L6  | |L7  |
        |0001| |0203| |0405| |0607| |0809| |1011| |1213| |1415|

        Shuffle to extract L0 and L1 as i64:
        -1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0

        Shuffle to extract L2 and L3 as i64:
        -1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 5, 4

        Shuffle to extract L4 and L5 as i64:
        -1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8

        Shuffle to extract L6 and L7 as i64:
        -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12
    */

    let l01_shuffle = _mm_set_epi8(-1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0);
    let l23_shuffle = _mm_set_epi8(-1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 5, 4);
    let l45_shuffle = _mm_set_epi8(-1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8);
    let l67_shuffle = _mm_set_epi8(
        -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12,
    );

    for (dst_x, chunk) in normalizer.chunks().iter().enumerate() {
        let mut x = chunk.start as usize;
        let mut ll_sum = _mm_set1_epi64x(0);
        let mut coeffs = chunk.values();

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

        for k in coeffs_by_8 {
            let coeff01_i64x2 = _mm_set_epi64x(k[1] as i64, k[0] as i64);
            let coeff23_i64x2 = _mm_set_epi64x(k[3] as i64, k[2] as i64);
            let coeff45_i64x2 = _mm_set_epi64x(k[5] as i64, k[4] as i64);
            let coeff67_i64x2 = _mm_set_epi64x(k[7] as i64, k[6] as i64);

            let source = simd_utils::loadu_si128(src_row, x);

            let l_i64x2 = _mm_shuffle_epi8(source, l01_shuffle);
            ll_sum = _mm_add_epi64(ll_sum, _mm_mul_epi32(l_i64x2, coeff01_i64x2));

            let l_i64x2 = _mm_shuffle_epi8(source, l23_shuffle);
            ll_sum = _mm_add_epi64(ll_sum, _mm_mul_epi32(l_i64x2, coeff23_i64x2));

            let l_i64x2 = _mm_shuffle_epi8(source, l45_shuffle);
            ll_sum = _mm_add_epi64(ll_sum, _mm_mul_epi32(l_i64x2, coeff45_i64x2));

            let l_i64x2 = _mm_shuffle_epi8(source, l67_shuffle);
            ll_sum = _mm_add_epi64(ll_sum, _mm_mul_epi32(l_i64x2, coeff67_i64x2));

            x += 8;
        }

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

        for k in coeffs_by_4 {
            let coeff01_i64x2 = _mm_set_epi64x(k[1] as i64, k[0] as i64);
            let coeff23_i64x2 = _mm_set_epi64x(k[3] as i64, k[2] as i64);

            let source = simd_utils::loadl_epi64(src_row, x);

            let l_i64x2 = _mm_shuffle_epi8(source, l01_shuffle);
            ll_sum = _mm_add_epi64(ll_sum, _mm_mul_epi32(l_i64x2, coeff01_i64x2));

            let l_i64x2 = _mm_shuffle_epi8(source, l23_shuffle);
            ll_sum = _mm_add_epi64(ll_sum, _mm_mul_epi32(l_i64x2, coeff23_i64x2));

            x += 4;
        }

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

        for k in coeffs_by_2 {
            let coeff01_i64x2 = _mm_set_epi64x(k[1] as i64, k[0] as i64);
            let source = simd_utils::loadl_epi32(src_row, x);

            let l_i64x2 = _mm_shuffle_epi8(source, l01_shuffle);
            ll_sum = _mm_add_epi64(ll_sum, _mm_mul_epi32(l_i64x2, coeff01_i64x2));

            x += 2;
        }

        if let Some(&k) = coeffs.first() {
            let coeff01_i64x2 = _mm_set_epi64x(0, k as i64);
            let pixel = src_row.get_unchecked(x).0 as i64;
            let source = _mm_set_epi64x(0, pixel);
            ll_sum = _mm_add_epi64(ll_sum, _mm_mul_epi32(source, coeff01_i64x2));
        }

        _mm_storeu_si128(ll_buf.as_mut_ptr() as *mut __m128i, ll_sum);
        let dst_pixel = dst_row.get_unchecked_mut(dst_x);
        dst_pixel.0 = normalizer.clip(ll_buf[0] + ll_buf[1] + half_error);
    }
}