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

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

use crate::convolution::optimisations::{CoefficientsI32Chunk, Normalizer32};
use crate::convolution::vertical_u16::native::convolution_by_u16;
use crate::pixels::InnerPixel;
use crate::{simd_utils, ImageView, ImageViewMut};

pub(crate) fn vert_convolution<T>(
    src_view: &impl ImageView<Pixel = T>,
    dst_view: &mut impl ImageViewMut<Pixel = T>,
    offset: u32,
    normalizer: &Normalizer32,
) where
    T: InnerPixel<Component = u16>,
{
    let coefficients_chunks = normalizer.chunks();
    let src_x = offset as usize * T::count_of_components();

    let dst_rows = dst_view.iter_rows_mut(0);
    for (dst_row, coeffs_chunk) in dst_rows.zip(coefficients_chunks) {
        unsafe {
            vert_convolution_into_one_row_u16(src_view, dst_row, src_x, coeffs_chunk, normalizer);
        }
    }
}

#[target_feature(enable = "sse4.1")]
unsafe fn vert_convolution_into_one_row_u16<T: InnerPixel<Component = u16>>(
    src_view: &impl ImageView<Pixel = T>,
    dst_row: &mut [T],
    mut src_x: usize,
    coeffs_chunk: &CoefficientsI32Chunk,
    normalizer: &Normalizer32,
) {
    let y_start = coeffs_chunk.start;
    let coeffs = coeffs_chunk.values();
    let max_rows = coeffs.len() as u32;
    let mut dst_u16 = T::components_mut(dst_row);

    /*
        |0    1    2    3    4    5    6    7   |
        |0001 0203 0405 0607 0809 1011 1213 1415|

        Shuffle to extract 0-1 components as i64:
        -1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0

        Shuffle to extract 2-3 components as i64:
        -1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 5, 4

        Shuffle to extract 4-5 components as i64:
        -1, -1, -1, -1, -1, -1, 11, 10, -1, -1, -1, -1, -1, -1, 9, 8

        Shuffle to extract 6-7 components as i64:
        -1, -1, -1, -1, -1, -1, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12

    */

    let c_shuffles = [
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 3, 2, -1, -1, -1, -1, -1, -1, 1, 0),
        _mm_set_epi8(-1, -1, -1, -1, -1, -1, 7, 6, -1, -1, -1, -1, -1, -1, 5, 4),
        _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, 15, 14, -1, -1, -1, -1, -1, -1, 13, 12,
        ),
    ];

    let precision = normalizer.precision();
    let initial = _mm_set1_epi64x(1 << (precision - 1));
    let mut c_buf = [0i64; 2];

    let mut dst_chunks_16 = dst_u16.chunks_exact_mut(16);
    for dst_chunk in &mut dst_chunks_16 {
        let mut sums = [[initial; 2], [initial; 2], [initial; 2], [initial; 2]];

        let mut y: u32 = 0;
        let coeffs_2 = coeffs.chunks_exact(2);
        let coeffs_reminder = coeffs_2.remainder();

        for (src_rows, two_coeffs) in src_view.iter_2_rows(y_start, max_rows).zip(coeffs_2) {
            let src_rows = src_rows.map(|row| T::components(row));

            for r in 0..2 {
                let coeff_i64x2 = _mm_set1_epi64x(two_coeffs[r] as i64);
                for x in 0..2 {
                    let source = simd_utils::loadu_si128(src_rows[r], src_x + x * 8);
                    for i in 0..4 {
                        let c_i64x2 = _mm_shuffle_epi8(source, c_shuffles[i]);
                        sums[i][x] = _mm_add_epi64(sums[i][x], _mm_mul_epi32(c_i64x2, coeff_i64x2));
                    }
                }
            }
            y += 2;
        }

        if let Some(&k) = coeffs_reminder.first() {
            if let Some(s_row) = src_view.iter_rows(y_start + y).next() {
                let components = T::components(s_row);
                let coeff_i64x2 = _mm_set1_epi64x(k as i64);

                for x in 0..2 {
                    let source = simd_utils::loadu_si128(components, src_x + x * 8);
                    for i in 0..4 {
                        let c_i64x2 = _mm_shuffle_epi8(source, c_shuffles[i]);
                        sums[i][x] = _mm_add_epi64(sums[i][x], _mm_mul_epi32(c_i64x2, coeff_i64x2));
                    }
                }
            }
        }

        let mut dst_ptr = dst_chunk.as_mut_ptr();
        for x in 0..2 {
            for sum in sums {
                _mm_storeu_si128(c_buf.as_mut_ptr() as *mut __m128i, sum[x]);
                *dst_ptr = normalizer.clip(c_buf[0]);
                dst_ptr = dst_ptr.add(1);
                *dst_ptr = normalizer.clip(c_buf[1]);
                dst_ptr = dst_ptr.add(1);
            }
        }

        src_x += 16;
    }

    dst_u16 = dst_chunks_16.into_remainder();
    let mut dst_chunks_8 = dst_u16.chunks_exact_mut(8);
    if let Some(dst_chunk) = dst_chunks_8.next() {
        let mut sums = [initial, initial, initial, initial];

        let mut y: u32 = 0;
        let coeffs_2 = coeffs.chunks_exact(2);
        let coeffs_reminder = coeffs_2.remainder();

        for (src_rows, two_coeffs) in src_view.iter_2_rows(y_start, max_rows).zip(coeffs_2) {
            let src_rows = src_rows.map(|row| T::components(row));
            let coeffs_i64 = [
                _mm_set1_epi64x(two_coeffs[0] as i64),
                _mm_set1_epi64x(two_coeffs[1] as i64),
            ];

            for r in 0..2 {
                let source = simd_utils::loadu_si128(src_rows[r], src_x);
                for i in 0..4 {
                    let c_i64x2 = _mm_shuffle_epi8(source, c_shuffles[i]);
                    sums[i] = _mm_add_epi64(sums[i], _mm_mul_epi32(c_i64x2, coeffs_i64[r]));
                }
            }
            y += 2;
        }

        if let Some(&k) = coeffs_reminder.first() {
            if let Some(s_row) = src_view.iter_rows(y_start + y).next() {
                let components = T::components(s_row);
                let coeff_i64x2 = _mm_set1_epi64x(k as i64);
                let source = simd_utils::loadu_si128(components, src_x);
                for i in 0..4 {
                    let c_i64x2 = _mm_shuffle_epi8(source, c_shuffles[i]);
                    sums[i] = _mm_add_epi64(sums[i], _mm_mul_epi32(c_i64x2, coeff_i64x2));
                }
            }
        }

        let mut dst_ptr = dst_chunk.as_mut_ptr();
        for sum in sums {
            // let mask = _mm_cmpgt_epi64(sums[i], zero);
            // sums[i] = _mm_and_si128(sums[i] , mask);
            // sums[i] = _mm_srl_epi64(sums[i] , precision_i64);
            // _mm_packus_epi32(sums[i] , sums[i] );
            _mm_storeu_si128(c_buf.as_mut_ptr() as *mut __m128i, sum);
            *dst_ptr = normalizer.clip(c_buf[0]);
            dst_ptr = dst_ptr.add(1);
            *dst_ptr = normalizer.clip(c_buf[1]);
            dst_ptr = dst_ptr.add(1);
        }

        src_x += 8;
    }

    dst_u16 = dst_chunks_8.into_remainder();
    let mut dst_chunks_4 = dst_u16.chunks_exact_mut(4);
    if let Some(dst_chunk) = dst_chunks_4.next() {
        let mut c01 = initial;
        let mut c23 = initial;
        let mut y: u32 = 0;
        let coeffs_2 = coeffs.chunks_exact(2);
        let coeffs_reminder = coeffs_2.remainder();

        for (src_rows, two_coeffs) in src_view.iter_2_rows(y_start, max_rows).zip(coeffs_2) {
            let src_rows = src_rows.map(|row| T::components(row));
            let coeffs_i64 = [
                _mm_set1_epi64x(two_coeffs[0] as i64),
                _mm_set1_epi64x(two_coeffs[1] as i64),
            ];
            for r in 0..2 {
                let comp_x4 = src_rows[r].get_unchecked(src_x..src_x + 4);
                let c_i64x2 = _mm_set_epi64x(comp_x4[1] as i64, comp_x4[0] as i64);
                c01 = _mm_add_epi64(c01, _mm_mul_epi32(c_i64x2, coeffs_i64[r]));
                let c_i64x2 = _mm_set_epi64x(comp_x4[3] as i64, comp_x4[2] as i64);
                c23 = _mm_add_epi64(c23, _mm_mul_epi32(c_i64x2, coeffs_i64[r]));
            }
            y += 2;
        }

        if let Some(&k) = coeffs_reminder.first() {
            if let Some(s_row) = src_view.iter_rows(y_start + y).next() {
                let components = T::components(s_row);
                let coeff_i64x2 = _mm_set1_epi64x(k as i64);

                let comp_x4 = components.get_unchecked(src_x..src_x + 4);
                let c_i64x2 = _mm_set_epi64x(comp_x4[1] as i64, comp_x4[0] as i64);
                c01 = _mm_add_epi64(c01, _mm_mul_epi32(c_i64x2, coeff_i64x2));
                let c_i64x2 = _mm_set_epi64x(comp_x4[3] as i64, comp_x4[2] as i64);
                c23 = _mm_add_epi64(c23, _mm_mul_epi32(c_i64x2, coeff_i64x2));
            }
        }

        let mut dst_ptr = dst_chunk.as_mut_ptr();
        _mm_storeu_si128(c_buf.as_mut_ptr() as *mut __m128i, c01);
        *dst_ptr = normalizer.clip(c_buf[0]);
        dst_ptr = dst_ptr.add(1);
        *dst_ptr = normalizer.clip(c_buf[1]);
        dst_ptr = dst_ptr.add(1);
        _mm_storeu_si128(c_buf.as_mut_ptr() as *mut __m128i, c23);
        *dst_ptr = normalizer.clip(c_buf[0]);
        dst_ptr = dst_ptr.add(1);
        *dst_ptr = normalizer.clip(c_buf[1]);

        src_x += 4;
    }

    dst_u16 = dst_chunks_4.into_remainder();
    if !dst_u16.is_empty() {
        let initial = 1 << (precision - 1);
        convolution_by_u16(
            src_view, normalizer, initial, dst_u16, src_x, y_start, coeffs,
        );
    }
}