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
292
293
294
295
296
297
298
299

use core::arch::x86_64::*;
use core::mem::transmute;

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

// This code is based on C-implementation from Pillow-SIMD package for Python
// https://github.com/uploadcare/pillow-simd

#[inline]
pub(crate) fn horiz_convolution(
    src_view: &impl ImageView<Pixel = U8x4>,
    dst_view: &mut impl ImageViewMut<Pixel = U8x4>,
    offset: u32,
    normalizer: &Normalizer16,
) {
    let precision = normalizer.precision();

    macro_rules! call {
        ($imm8:expr) => {{
            horiz_convolution_p::<$imm8>(src_view, dst_view, offset, normalizer);
        }};
    }
    constify_imm8!(precision, call);
}

fn horiz_convolution_p<const PRECISION: i32>(
    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::<PRECISION>(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::<PRECISION>(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
#[inline]
#[target_feature(enable = "avx2")]
unsafe fn horiz_convolution_four_rows<const PRECISION: i32>(
    src_rows: [&[U8x4]; 4],
    dst_rows: [&mut [U8x4]; 4],
    normalizer: &Normalizer16,
) {
    let zero = _mm256_setzero_si256();
    let initial = _mm256_set1_epi32(1 << (PRECISION - 1));

    #[rustfmt::skip]
    let sh1 = _mm256_set_epi8(
        -1, 7, -1, 3, -1, 6, -1, 2, -1, 5, -1, 1, -1, 4, -1, 0,
        -1, 7, -1, 3, -1, 6, -1, 2, -1, 5, -1, 1, -1, 4, -1, 0,
    );
    #[rustfmt::skip]
    let sh2 = _mm256_set_epi8(
        -1, 15, -1, 11, -1, 14, -1, 10, -1, 13, -1, 9, -1, 12, -1, 8,
        -1, 15, -1, 11, -1, 14, -1, 10, -1, 13, -1, 9, -1, 12, -1, 8,
    );

    let coefficients_chunks = normalizer.chunks();

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

        let mut sss0 = initial;
        let mut sss1 = 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 mmk0 = simd_utils::mm256_load_and_clone_i16x2(k);
            let mmk1 = simd_utils::mm256_load_and_clone_i16x2(&k[2..]);

            let source = _mm256_inserti128_si256::<1>(
                _mm256_castsi128_si256(simd_utils::loadu_si128(src_rows[0], x)),
                simd_utils::loadu_si128(src_rows[1], x),
            );
            let pix = _mm256_shuffle_epi8(source, sh1);
            sss0 = _mm256_add_epi32(sss0, _mm256_madd_epi16(pix, mmk0));
            let pix = _mm256_shuffle_epi8(source, sh2);
            sss0 = _mm256_add_epi32(sss0, _mm256_madd_epi16(pix, mmk1));

            let source = _mm256_inserti128_si256::<1>(
                _mm256_castsi128_si256(simd_utils::loadu_si128(src_rows[2], x)),
                simd_utils::loadu_si128(src_rows[3], x),
            );
            let pix = _mm256_shuffle_epi8(source, sh1);
            sss1 = _mm256_add_epi32(sss1, _mm256_madd_epi16(pix, mmk0));
            let pix = _mm256_shuffle_epi8(source, sh2);
            sss1 = _mm256_add_epi32(sss1, _mm256_madd_epi16(pix, mmk1));

            x += 4;
        }

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

        for k in coeffs_by_2 {
            let mmk = simd_utils::mm256_load_and_clone_i16x2(k);

            let source = _mm256_inserti128_si256::<1>(
                _mm256_castsi128_si256(simd_utils::loadl_epi64(src_rows[0], x)),
                simd_utils::loadl_epi64(src_rows[1], x),
            );
            let pix = _mm256_shuffle_epi8(source, sh1);
            sss0 = _mm256_add_epi32(sss0, _mm256_madd_epi16(pix, mmk));

            let source = _mm256_inserti128_si256::<1>(
                _mm256_castsi128_si256(simd_utils::loadl_epi64(src_rows[2], x)),
                simd_utils::loadl_epi64(src_rows[3], x),
            );
            let pix = _mm256_shuffle_epi8(source, sh1);
            sss1 = _mm256_add_epi32(sss1, _mm256_madd_epi16(pix, mmk));

            x += 2;
        }

        if let Some(&k) = reminder2.first() {
            // [16] xx k0 xx k0 xx k0 xx k0 xx k0 xx k0 xx k0 xx k0
            let mmk = _mm256_set1_epi32(k as i32);

            // [16] xx a0 xx b0 xx g0 xx r0 xx a0 xx b0 xx g0 xx r0
            let pix = _mm256_inserti128_si256::<1>(
                _mm256_castsi128_si256(simd_utils::mm_cvtepu8_epi32(src_rows[0], x)),
                simd_utils::mm_cvtepu8_epi32(src_rows[1], x),
            );
            sss0 = _mm256_add_epi32(sss0, _mm256_madd_epi16(pix, mmk));

            let pix = _mm256_inserti128_si256::<1>(
                _mm256_castsi128_si256(simd_utils::mm_cvtepu8_epi32(src_rows[2], x)),
                simd_utils::mm_cvtepu8_epi32(src_rows[3], x),
            );
            sss1 = _mm256_add_epi32(sss1, _mm256_madd_epi16(pix, mmk));
        }

        sss0 = _mm256_srai_epi32::<PRECISION>(sss0);
        sss1 = _mm256_srai_epi32::<PRECISION>(sss1);

        sss0 = _mm256_packs_epi32(sss0, zero);
        sss1 = _mm256_packs_epi32(sss1, zero);
        sss0 = _mm256_packus_epi16(sss0, zero);
        sss1 = _mm256_packus_epi16(sss1, zero);
        *dst_rows[0].get_unchecked_mut(dst_x) =
            transmute::<i32, U8x4>(_mm_cvtsi128_si32(_mm256_extracti128_si256::<0>(sss0)));
        *dst_rows[1].get_unchecked_mut(dst_x) =
            transmute::<i32, U8x4>(_mm_cvtsi128_si32(_mm256_extracti128_si256::<1>(sss0)));
        *dst_rows[2].get_unchecked_mut(dst_x) =
            transmute::<i32, U8x4>(_mm_cvtsi128_si32(_mm256_extracti128_si256::<0>(sss1)));
        *dst_rows[3].get_unchecked_mut(dst_x) =
            transmute::<i32, U8x4>(_mm_cvtsi128_si32(_mm256_extracti128_si256::<1>(sss1)));
    }
}

/// For safety, it is necessary to ensure the following conditions:
/// - bounds.len() == dst_row.len()
/// - coeffs.len() == dst_rows.0.len() * window_size
/// - max(bound.start + bound.size for bound in bounds) <= src_row.len()
/// - precision <= MAX_COEFS_PRECISION
#[inline]
#[target_feature(enable = "avx2")]
unsafe fn horiz_convolution_one_row<const PRECISION: i32>(
    src_row: &[U8x4],
    dst_row: &mut [U8x4],
    normalizer: &Normalizer16,
) {
    #[rustfmt::skip]
    let sh1 = _mm256_set_epi8(
        -1, 7, -1, 3, -1, 6, -1, 2, -1, 5, -1, 1, -1, 4, -1, 0,
        -1, 7, -1, 3, -1, 6, -1, 2, -1, 5, -1, 1, -1, 4, -1, 0,
    );
    #[rustfmt::skip]
    let sh2 = _mm256_set_epi8(
        11, 10, 9, 8, 11, 10, 9, 8, 11, 10, 9, 8, 11, 10, 9, 8,
        3, 2, 1, 0, 3, 2, 1, 0, 3, 2, 1, 0, 3, 2, 1, 0,
    );
    #[rustfmt::skip]
    let sh3 = _mm256_set_epi8(
        -1, 15, -1, 11, -1, 14, -1, 10, -1, 13, -1, 9, -1, 12, -1, 8,
        -1, 15, -1, 11, -1, 14, -1, 10, -1, 13, -1, 9, -1, 12, -1, 8,
    );
    #[rustfmt::skip]
    let sh4 = _mm256_set_epi8(
        15, 14, 13, 12, 15, 14, 13, 12, 15, 14, 13, 12, 15, 14, 13, 12,
        7, 6, 5, 4, 7, 6, 5, 4, 7, 6, 5, 4, 7, 6, 5, 4,
    );
    #[rustfmt::skip]
    let sh5 = _mm256_set_epi8(
        -1, 15, -1, 11, -1, 14, -1, 10, -1, 13, -1, 9, -1, 12, -1, 8,
        -1, 7, -1, 3, -1, 6, -1, 2, -1, 5, -1, 1, -1, 4, -1, 0,
    );
    #[rustfmt::skip]
    let sh6 = _mm256_set_epi8(
        7, 6, 5, 4, 7, 6, 5, 4, 7, 6, 5, 4, 7, 6, 5, 4,
        3, 2, 1, 0, 3, 2, 1, 0, 3, 2, 1, 0, 3, 2, 1, 0,
    );
    let sh7 = _mm_set_epi8(-1, 7, -1, 3, -1, 6, -1, 2, -1, 5, -1, 1, -1, 4, -1, 0);

    let coefficients_chunks = normalizer.chunks();

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

        let mut sss = if coeffs.len() < 8 {
            _mm_set1_epi32(1 << (PRECISION - 1))
        } else {
            // Lower part will be added to higher, use only half of the error
            let mut sss256 = _mm256_set1_epi32(1 << (PRECISION - 2));

            let coeffs_by_8 = coeffs.chunks_exact(8);
            let reminder1 = coeffs_by_8.remainder();

            for k in coeffs_by_8 {
                let tmp = simd_utils::loadu_si128(k, 0);
                let ksource = _mm256_insertf128_si256::<1>(_mm256_castsi128_si256(tmp), tmp);

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

                let pix = _mm256_shuffle_epi8(source, sh1);
                let mmk = _mm256_shuffle_epi8(ksource, sh2);
                sss256 = _mm256_add_epi32(sss256, _mm256_madd_epi16(pix, mmk));

                let pix = _mm256_shuffle_epi8(source, sh3);
                let mmk = _mm256_shuffle_epi8(ksource, sh4);
                sss256 = _mm256_add_epi32(sss256, _mm256_madd_epi16(pix, mmk));

                x += 8;
            }

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

            for k in coeffs_by_4 {
                let tmp = simd_utils::loadl_epi64(k, 0);
                let ksource = _mm256_insertf128_si256::<1>(_mm256_castsi128_si256(tmp), tmp);

                let tmp = simd_utils::loadu_si128(src_row, x);
                let source = _mm256_insertf128_si256::<1>(_mm256_castsi128_si256(tmp), tmp);

                let pix = _mm256_shuffle_epi8(source, sh5);
                let mmk = _mm256_shuffle_epi8(ksource, sh6);
                sss256 = _mm256_add_epi32(sss256, _mm256_madd_epi16(pix, mmk));

                x += 4;
            }

            _mm_add_epi32(
                _mm256_extracti128_si256::<0>(sss256),
                _mm256_extracti128_si256::<1>(sss256),
            )
        };

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

        for k in coeffs_by_2 {
            let mmk = simd_utils::mm_load_and_clone_i16x2(k);
            let source = simd_utils::loadl_epi64(src_row, x);
            let pix = _mm_shuffle_epi8(source, sh7);
            sss = _mm_add_epi32(sss, _mm_madd_epi16(pix, mmk));

            x += 2
        }

        if let Some(&k) = reminder1.first() {
            let pix = simd_utils::mm_cvtepu8_epi32(src_row, x);
            let mmk = _mm_set1_epi32(k as i32);
            sss = _mm_add_epi32(sss, _mm_madd_epi16(pix, mmk));
        }

        sss = _mm_srai_epi32::<PRECISION>(sss);

        sss = _mm_packs_epi32(sss, sss);
        *dst_row.get_unchecked_mut(dst_x) =
            transmute::<i32, U8x4>(_mm_cvtsi128_si32(_mm_packus_epi16(sss, sss)));
    }
}