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

use core::arch::wasm32::*;
use core::mem::transmute;

use crate::convolution::optimisations::Normalizer16;
use crate::pixels::U8x4;
use crate::{wasm32_utils, ImageView, ImageViewMut};

#[inline]
pub(crate) fn horiz_convolution(
    src_view: &impl ImageView<Pixel = U8x4>,
    dst_view: &mut impl ImageViewMut<Pixel = U8x4>,
    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
#[target_feature(enable = "simd128")]
unsafe fn horiz_convolution_four_rows(
    src_rows: [&[U8x4]; 4],
    dst_rows: [&mut [U8x4]; 4],
    normalizer: &Normalizer16,
) {
    let precision = normalizer.precision() as u32;
    let initial = i32x4_splat(1 << (precision - 1));
    const MASK_LO: v128 = i8x16(0, -1, 4, -1, 1, -1, 5, -1, 2, -1, 6, -1, 3, -1, 7, -1);
    const MASK_HI: v128 = i8x16(8, -1, 12, -1, 9, -1, 13, -1, 10, -1, 14, -1, 11, -1, 15, -1);
    const MASK: v128 = i8x16(0, -1, 4, -1, 1, -1, 5, -1, 2, -1, 6, -1, 3, -1, 7, -1);

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

        let mut sss0 = initial;
        let mut sss1 = initial;
        let mut sss2 = initial;
        let mut sss3 = initial;

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

        for k in coeffs_by_4 {
            let mmk_lo = wasm32_utils::ptr_i16_to_set1_i32(k, 0);
            let mmk_hi = wasm32_utils::ptr_i16_to_set1_i32(k, 2);

            // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
            let mut source = wasm32_utils::load_v128(src_rows[0], x);
            // [16] a1 a0 b1 b0 g1 g0 r1 r0
            let mut pix = i8x16_swizzle(source, MASK_LO);
            sss0 = i32x4_add(sss0, i32x4_dot_i16x8(pix, mmk_lo));
            // [16] a3 a2 b3 b2 g3 g2 r3 r2
            pix = i8x16_swizzle(source, MASK_HI);
            sss0 = i32x4_add(sss0, i32x4_dot_i16x8(pix, mmk_hi));

            source = wasm32_utils::load_v128(src_rows[1], x);
            pix = i8x16_swizzle(source, MASK_LO);
            sss1 = i32x4_add(sss1, i32x4_dot_i16x8(pix, mmk_lo));
            pix = i8x16_swizzle(source, MASK_HI);
            sss1 = i32x4_add(sss1, i32x4_dot_i16x8(pix, mmk_hi));

            source = wasm32_utils::load_v128(src_rows[2], x);
            pix = i8x16_swizzle(source, MASK_LO);
            sss2 = i32x4_add(sss2, i32x4_dot_i16x8(pix, mmk_lo));
            pix = i8x16_swizzle(source, MASK_HI);
            sss2 = i32x4_add(sss2, i32x4_dot_i16x8(pix, mmk_hi));

            source = wasm32_utils::load_v128(src_rows[3], x);
            pix = i8x16_swizzle(source, MASK_LO);
            sss3 = i32x4_add(sss3, i32x4_dot_i16x8(pix, mmk_lo));
            pix = i8x16_swizzle(source, MASK_HI);
            sss3 = i32x4_add(sss3, i32x4_dot_i16x8(pix, mmk_hi));
            x += 4;
        }

        let coeffs_by_2 = reminder1.chunks_exact(2);
        let reminder2 = coeffs_by_2.remainder();

        for k in coeffs_by_2 {
            // [16] k1 k0 k1 k0 k1 k0 k1 k0
            let mmk = wasm32_utils::ptr_i16_to_set1_i32(k, 0);

            // [8] x x x x x x x x a1 b1 g1 r1 a0 b0 g0 r0
            let mut pix = wasm32_utils::loadl_i64(src_rows[0], x);
            // [16] a1 a0 b1 b0 g1 g0 r1 r0
            pix = i8x16_swizzle(pix, MASK);
            sss0 = i32x4_add(sss0, i32x4_dot_i16x8(pix, mmk));

            pix = wasm32_utils::loadl_i64(src_rows[1], x);
            pix = i8x16_swizzle(pix, MASK);
            sss1 = i32x4_add(sss1, i32x4_dot_i16x8(pix, mmk));

            pix = wasm32_utils::loadl_i64(src_rows[2], x);
            pix = i8x16_swizzle(pix, MASK);
            sss2 = i32x4_add(sss2, i32x4_dot_i16x8(pix, mmk));

            pix = wasm32_utils::loadl_i64(src_rows[3], x);
            pix = i8x16_swizzle(pix, MASK);
            sss3 = i32x4_add(sss3, i32x4_dot_i16x8(pix, mmk));

            x += 2;
        }

        if let Some(&k) = reminder2.first() {
            // [16] xx k0 xx k0 xx k0 xx k0
            let mmk = i32x4_splat(k as i32);
            // [16] xx a0 xx b0 xx g0 xx r0
            let mut pix = wasm32_utils::i32x4_extend_low_ptr_u8x4(src_rows[0], x);
            sss0 = i32x4_add(sss0, i32x4_dot_i16x8(pix, mmk));

            pix = wasm32_utils::i32x4_extend_low_ptr_u8x4(src_rows[1], x);
            sss1 = i32x4_add(sss1, i32x4_dot_i16x8(pix, mmk));

            pix = wasm32_utils::i32x4_extend_low_ptr_u8x4(src_rows[2], x);
            sss2 = i32x4_add(sss2, i32x4_dot_i16x8(pix, mmk));

            pix = wasm32_utils::i32x4_extend_low_ptr_u8x4(src_rows[3], x);
            sss3 = i32x4_add(sss3, i32x4_dot_i16x8(pix, mmk));
        }

        sss0 = i32x4_shr(sss0, precision);
        sss1 = i32x4_shr(sss1, precision);
        sss2 = i32x4_shr(sss2, precision);
        sss3 = i32x4_shr(sss3, precision);

        sss0 = i16x8_narrow_i32x4(sss0, sss0);
        sss1 = i16x8_narrow_i32x4(sss1, sss1);
        sss2 = i16x8_narrow_i32x4(sss2, sss2);
        sss3 = i16x8_narrow_i32x4(sss3, sss3);
        *dst_rows[0].get_unchecked_mut(dst_x) =
            transmute(i32x4_extract_lane::<0>(u8x16_narrow_i16x8(sss0, sss0)));
        *dst_rows[1].get_unchecked_mut(dst_x) =
            transmute(i32x4_extract_lane::<0>(u8x16_narrow_i16x8(sss1, sss1)));
        *dst_rows[2].get_unchecked_mut(dst_x) =
            transmute(i32x4_extract_lane::<0>(u8x16_narrow_i16x8(sss2, sss2)));
        *dst_rows[3].get_unchecked_mut(dst_x) =
            transmute(i32x4_extract_lane::<0>(u8x16_narrow_i16x8(sss3, sss3)));
    }
}

/// 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 = "simd128")]
unsafe fn horiz_convolution_one_row(
    src_row: &[U8x4],
    dst_row: &mut [U8x4],
    normalizer: &Normalizer16,
) {
    let precision = normalizer.precision() as u32;
    let initial = i32x4_splat(1 << (precision - 1));
    const SH1: v128 = i8x16(0, -1, 8, -1, 1, -1, 9, -1, 2, -1, 10, -1, 3, -1, 11, -1);
    const SH2: v128 = i8x16(0, 1, 4, 5, 0, 1, 4, 5, 0, 1, 4, 5, 0, 1, 4, 5);
    const SH3: v128 = i8x16(4, -1, 12, -1, 5, -1, 13, -1, 6, -1, 14, -1, 7, -1, 15, -1);
    const SH4: v128 = i8x16(2, 3, 6, 7, 2, 3, 6, 7, 2, 3, 6, 7, 2, 3, 6, 7);
    const SH5: v128 = i8x16(8, 9, 12, 13, 8, 9, 12, 13, 8, 9, 12, 13, 8, 9, 12, 13);
    const SH6: v128 = i8x16(
        10, 11, 14, 15, 10, 11, 14, 15, 10, 11, 14, 15, 10, 11, 14, 15,
    );
    const SH7: v128 = i8x16(0, -1, 4, -1, 1, -1, 5, -1, 2, -1, 6, -1, 3, -1, 7, -1);

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

        let coeffs_by_8 = coeffs_chunk.values().chunks_exact(8);
        let reminder8 = coeffs_by_8.remainder();

        for k in coeffs_by_8 {
            let ksource = wasm32_utils::load_v128(k, 0);

            let mut source = wasm32_utils::load_v128(src_row, x);

            let mut pix = i8x16_swizzle(source, SH1);
            let mut mmk = i8x16_swizzle(ksource, SH2);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));

            pix = i8x16_swizzle(source, SH3);
            mmk = i8x16_swizzle(ksource, SH4);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));

            source = wasm32_utils::load_v128(src_row, x + 4);

            pix = i8x16_swizzle(source, SH1);
            mmk = i8x16_swizzle(ksource, SH5);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));

            pix = i8x16_swizzle(source, SH3);
            mmk = i8x16_swizzle(ksource, SH6);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));

            x += 8;
        }

        let coeffs_by_4 = reminder8.chunks_exact(4);
        let reminder4 = coeffs_by_4.remainder();

        for k in coeffs_by_4 {
            let source = wasm32_utils::load_v128(src_row, x);
            let ksource = wasm32_utils::loadl_i64(k, 0);

            let mut pix = i8x16_swizzle(source, SH1);
            let mut mmk = i8x16_swizzle(ksource, SH2);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));

            pix = i8x16_swizzle(source, SH3);
            mmk = i8x16_swizzle(ksource, SH4);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));

            x += 4;
        }

        let coeffs_by_2 = reminder4.chunks_exact(2);
        let reminder2 = 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_i64(src_row, x);
            let pix = i8x16_swizzle(source, SH7);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));

            x += 2
        }

        if let Some(&k) = reminder2.first() {
            let pix = wasm32_utils::i32x4_extend_low_ptr_u8x4(src_row, x);
            let mmk = i32x4_splat(k as i32);
            sss = i32x4_add(sss, i32x4_dot_i16x8(pix, mmk));
        }

        sss = i32x4_shr(sss, precision);

        sss = i16x8_narrow_i32x4(sss, sss);
        *dst_row.get_unchecked_mut(dst_x) =
            transmute(i32x4_extract_lane::<0>(u8x16_narrow_i16x8(sss, sss)));
    }
}