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

use core::arch::aarch64::*;

use crate::convolution::optimisations::{CoefficientsI32Chunk, Normalizer32};
use crate::pixels::U16x4;
use crate::{neon_utils, ImageView, ImageViewMut};

#[inline]
pub(crate) fn horiz_convolution(
    src_view: &impl ImageView<Pixel = U16x4>,
    dst_view: &mut impl ImageViewMut<Pixel = U16x4>,
    offset: u32,
    normalizer: &Normalizer32,
) {
    let precision = normalizer.precision();

    macro_rules! call {
        ($imm8:expr) => {{
            horiz_convolution_p::<$imm8>(src_view, dst_view, offset, normalizer);
        }};
    }
    constify_64_imm8!(precision, call);
}

fn horiz_convolution_p<const PRECISION: i32>(
    src_view: &impl ImageView<Pixel = U16x4>,
    dst_view: &mut impl ImageViewMut<Pixel = U16x4>,
    offset: u32,
    normalizer: &Normalizer32,
) {
    let coefficients_chunks = normalizer.chunks();
    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::<PRECISION>(src_rows, dst_rows, coefficients_chunks);
        }
    }

    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::<PRECISION>(src_row, dst_row, coefficients_chunks);
        }
    }
}

/// 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
#[target_feature(enable = "neon")]
unsafe fn horiz_convolution_four_rows<const PRECISION: i32>(
    src_rows: [&[U16x4]; 4],
    dst_rows: [&mut [U16x4]; 4],
    coefficients_chunks: &[CoefficientsI32Chunk],
) {
    let initial = vdupq_n_s64(1i64 << (PRECISION - 1));
    let zero_u16x8 = vdupq_n_u16(0);
    let zero_u16x4 = vdup_n_u16(0);

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

        let mut sss_a = [int64x2x2_t(initial, initial); 4];

        let mut coeffs = coeffs_chunk.values();

        let coeffs_by_8 = coeffs.chunks_exact(4);
        coeffs = coeffs_by_8.remainder();
        for k in coeffs_by_8 {
            let coeffs_i32x4 = neon_utils::load_i32x4(k, 0);
            let coeff0 = vdup_laneq_s32::<0>(coeffs_i32x4);
            let coeff1 = vdup_laneq_s32::<1>(coeffs_i32x4);
            let coeff2 = vdup_laneq_s32::<2>(coeffs_i32x4);
            let coeff3 = vdup_laneq_s32::<3>(coeffs_i32x4);

            for i in 0..4 {
                let mut sss = sss_a[i];
                let source = neon_utils::load_u16x8(src_rows[i], x);

                let pix_i32 = vreinterpretq_s32_u16(vzip1q_u16(source, zero_u16x8));
                sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff0);
                sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff0);

                let pix_i32 = vreinterpretq_s32_u16(vzip2q_u16(source, zero_u16x8));
                sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff1);
                sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff1);

                let source = neon_utils::load_u16x8(src_rows[i], x + 2);

                let pix_i32 = vreinterpretq_s32_u16(vzip1q_u16(source, zero_u16x8));
                sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff2);
                sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff2);

                let pix_i32 = vreinterpretq_s32_u16(vzip2q_u16(source, zero_u16x8));
                sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff3);
                sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff3);

                sss_a[i] = sss;
            }

            x += 4;
        }

        let coeffs_by_4 = coeffs.chunks_exact(2);
        coeffs = coeffs_by_4.remainder();

        for k in coeffs_by_4 {
            let coeffs_i32x2 = neon_utils::load_i32x2(k, 0);
            let coeff0 = vdup_lane_s32::<0>(coeffs_i32x2);
            let coeff1 = vdup_lane_s32::<1>(coeffs_i32x2);

            for i in 0..4 {
                let mut sss = sss_a[i];
                let source = neon_utils::load_u16x8(src_rows[i], x);

                let pix_i32 = vreinterpretq_s32_u16(vzip1q_u16(source, zero_u16x8));
                sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff0);
                sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff0);

                let pix_i32 = vreinterpretq_s32_u16(vzip2q_u16(source, zero_u16x8));
                sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff1);
                sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff1);

                sss_a[i] = sss;
            }
            x += 2;
        }

        if let Some(&k) = coeffs.first() {
            let coeff = vdup_n_s32(k);

            for i in 0..4 {
                let mut sss = sss_a[i];
                let source = vcombine_u16(neon_utils::load_u16x4(src_rows[i], x), zero_u16x4);

                let pix_i32 = vreinterpretq_s32_u16(vzip1q_u16(source, zero_u16x8));
                sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff);
                sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff);

                sss_a[i] = sss;
            }
        }

        sss_a[0].0 = vshrq_n_s64::<PRECISION>(sss_a[0].0);
        sss_a[0].1 = vshrq_n_s64::<PRECISION>(sss_a[0].1);
        sss_a[1].0 = vshrq_n_s64::<PRECISION>(sss_a[1].0);
        sss_a[1].1 = vshrq_n_s64::<PRECISION>(sss_a[1].1);
        sss_a[2].0 = vshrq_n_s64::<PRECISION>(sss_a[2].0);
        sss_a[2].1 = vshrq_n_s64::<PRECISION>(sss_a[2].1);
        sss_a[3].0 = vshrq_n_s64::<PRECISION>(sss_a[3].0);
        sss_a[3].1 = vshrq_n_s64::<PRECISION>(sss_a[3].1);

        for i in 0..4 {
            let sss = sss_a[i];
            let sss_i32x4 = vcombine_s32(vqmovn_s64(sss.0), vqmovn_s64(sss.1));
            let sss_u16x4 = vqmovun_s32(sss_i32x4);
            let dst_pix = dst_rows[i].get_unchecked_mut(dst_x);
            let ptr = dst_pix as *mut U16x4 as *mut u16;
            vst1_u16(ptr, sss_u16x4);
        }
    }
}

/// 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()
/// - precision <= MAX_COEFS_PRECISION
#[target_feature(enable = "neon")]
unsafe fn horiz_convolution_one_row<const PRECISION: i32>(
    src_row: &[U16x4],
    dst_row: &mut [U16x4],
    coefficients_chunks: &[CoefficientsI32Chunk],
) {
    let initial = vdupq_n_s64(1i64 << (PRECISION - 1));
    let zero_u16x8 = vdupq_n_u16(0);
    let zero_u16x4 = vdup_n_u16(0);

    for (coeffs_chunk, dst_pix) in coefficients_chunks.iter().zip(dst_row) {
        let mut x: usize = coeffs_chunk.start as usize;
        let mut sss = int64x2x2_t(initial, initial);
        let mut coeffs = coeffs_chunk.values();

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

        for k in coeffs_by_4 {
            let coeffs_i32x4 = neon_utils::load_i32x4(k, 0);
            let source = neon_utils::load_u16x8(src_row, x);

            let coeff = vdup_laneq_s32::<0>(coeffs_i32x4);
            let pix_i32 = vreinterpretq_s32_u16(vzip1q_u16(source, zero_u16x8));
            sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff);
            sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff);

            let coeff = vdup_laneq_s32::<1>(coeffs_i32x4);
            let pix_i32 = vreinterpretq_s32_u16(vzip2q_u16(source, zero_u16x8));
            sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff);
            sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff);

            let source = neon_utils::load_u16x8(src_row, x + 2);

            let coeff = vdup_laneq_s32::<2>(coeffs_i32x4);
            let pix_i32 = vreinterpretq_s32_u16(vzip1q_u16(source, zero_u16x8));
            sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff);
            sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff);

            let coeff = vdup_laneq_s32::<3>(coeffs_i32x4);
            let pix_i32 = vreinterpretq_s32_u16(vzip2q_u16(source, zero_u16x8));
            sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff);
            sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff);

            x += 4;
        }

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

        for k in coeffs_by_2 {
            let coeffs_i32x2 = neon_utils::load_i32x2(k, 0);
            let source = neon_utils::load_u16x8(src_row, x);

            let coeff = vdup_lane_s32::<0>(coeffs_i32x2);
            let pix_i32 = vreinterpretq_s32_u16(vzip1q_u16(source, zero_u16x8));
            sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff);
            sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff);

            let coeff = vdup_lane_s32::<1>(coeffs_i32x2);
            let pix_i32 = vreinterpretq_s32_u16(vzip2q_u16(source, zero_u16x8));
            sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff);
            sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff);

            x += 2;
        }

        if let Some(&k) = coeffs.first() {
            let source = vcombine_u16(neon_utils::load_u16x4(src_row, x), zero_u16x4);

            let coeff = vdup_n_s32(k);
            let pix_i32 = vreinterpretq_s32_u16(vzip1q_u16(source, zero_u16x8));
            sss.0 = vmlal_s32(sss.0, vget_low_s32(pix_i32), coeff);
            sss.1 = vmlal_s32(sss.1, vget_high_s32(pix_i32), coeff);
        }

        sss.0 = vshrq_n_s64::<PRECISION>(sss.0);
        sss.1 = vshrq_n_s64::<PRECISION>(sss.1);

        let sss_i32x4 = vcombine_s32(vqmovn_s64(sss.0), vqmovn_s64(sss.1));
        let sss_u16x4 = vqmovun_s32(sss_i32x4);
        let ptr = dst_pix as *mut U16x4 as *mut u16;
        vst1_u16(ptr, sss_u16x4);
    }
}