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

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

use crate::convolution::optimisations::{CoefficientsI16Chunk, Normalizer16};
use crate::convolution::vertical_u8::native;
use crate::image_view::ImageViewMut;
use crate::pixels::InnerPixel;
use crate::{simd_utils, ImageView};

#[inline]
pub(crate) fn vert_convolution<T>(
    src_view: &impl ImageView<Pixel = T>,
    dst_view: &mut impl ImageViewMut<Pixel = T>,
    offset: u32,
    normalizer: &Normalizer16,
) where
    T: InnerPixel<Component = u8>,
{
    let precision = normalizer.precision();

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

fn vert_convolution_p<T, const PRECISION: i32>(
    src_view: &impl ImageView<Pixel = T>,
    dst_view: &mut impl ImageViewMut<Pixel = T>,
    offset: u32,
    normalizer: &Normalizer16,
) where
    T: InnerPixel<Component = u8>,
{
    let coefficients_chunks = normalizer.chunks();
    let src_x = offset as usize * T::count_of_components();
    let dst_rows = dst_view.iter_rows_mut(0);
    let dst_row_and_coefs = dst_rows.zip(coefficients_chunks);

    for (dst_row, coeffs_chunk) in dst_row_and_coefs {
        unsafe {
            vert_convolution_into_one_row::<T, PRECISION>(
                src_view,
                dst_row,
                src_x,
                coeffs_chunk,
                normalizer,
            );
        }
    }
}

#[inline]
#[target_feature(enable = "avx2")]
unsafe fn vert_convolution_into_one_row<T, const PRECISION: i32>(
    src_view: &impl ImageView<Pixel = T>,
    dst_row: &mut [T],
    mut src_x: usize,
    coeffs_chunk: &CoefficientsI16Chunk,
    normalizer: &Normalizer16,
) where
    T: InnerPixel<Component = u8>,
{
    let y_start = coeffs_chunk.start;
    let coeffs = coeffs_chunk.values();
    let max_rows = coeffs.len() as u32;
    let y_last = (y_start + max_rows).max(1) - 1;

    let initial = _mm_set1_epi32(1 << (PRECISION as u8 - 1));
    let initial_256 = _mm256_set1_epi32(1 << (PRECISION as u8 - 1));

    let mut dst_u8 = T::components_mut(dst_row);

    // 32 components in one register
    let mut dst_chunks_32 = dst_u8.chunks_exact_mut(32);
    for dst_chunk in &mut dst_chunks_32 {
        let mut sss0 = initial_256;
        let mut sss1 = initial_256;
        let mut sss2 = initial_256;
        let mut sss3 = initial_256;

        let coeffs_chunks = coeffs.chunks_exact(2);
        let coeffs_reminder = coeffs_chunks.remainder();

        for (src_rows, two_coeffs) in src_view.iter_2_rows(y_start, max_rows).zip(coeffs_chunks) {
            let components1 = T::components(src_rows[0]);
            let components2 = T::components(src_rows[1]);

            // Load two coefficients at once
            let mmk = simd_utils::mm256_load_and_clone_i16x2(two_coeffs);

            let source1 = simd_utils::loadu_si256(components1, src_x); // top line
            let source2 = simd_utils::loadu_si256(components2, src_x); // bottom line

            let source = _mm256_unpacklo_epi8(source1, source2);
            let pix = _mm256_unpacklo_epi8(source, _mm256_setzero_si256());
            sss0 = _mm256_add_epi32(sss0, _mm256_madd_epi16(pix, mmk));
            let pix = _mm256_unpackhi_epi8(source, _mm256_setzero_si256());
            sss1 = _mm256_add_epi32(sss1, _mm256_madd_epi16(pix, mmk));

            let source = _mm256_unpackhi_epi8(source1, source2);
            let pix = _mm256_unpacklo_epi8(source, _mm256_setzero_si256());
            sss2 = _mm256_add_epi32(sss2, _mm256_madd_epi16(pix, mmk));
            let pix = _mm256_unpackhi_epi8(source, _mm256_setzero_si256());
            sss3 = _mm256_add_epi32(sss3, _mm256_madd_epi16(pix, mmk));
        }

        if let Some(&k) = coeffs_reminder.first() {
            if let Some(s_row) = src_view.iter_rows(y_last).next() {
                let components = T::components(s_row);
                let mmk = _mm256_set1_epi32(k as i32);

                let source1 = simd_utils::loadu_si256(components, src_x); // top line
                let source2 = _mm256_setzero_si256(); // bottom line is empty

                let source = _mm256_unpacklo_epi8(source1, source2);
                let pix = _mm256_unpacklo_epi8(source, _mm256_setzero_si256());
                sss0 = _mm256_add_epi32(sss0, _mm256_madd_epi16(pix, mmk));
                let pix = _mm256_unpackhi_epi8(source, _mm256_setzero_si256());
                sss1 = _mm256_add_epi32(sss1, _mm256_madd_epi16(pix, mmk));

                let source = _mm256_unpackhi_epi8(source1, _mm256_setzero_si256());
                let pix = _mm256_unpacklo_epi8(source, _mm256_setzero_si256());
                sss2 = _mm256_add_epi32(sss2, _mm256_madd_epi16(pix, mmk));
                let pix = _mm256_unpackhi_epi8(source, _mm256_setzero_si256());
                sss3 = _mm256_add_epi32(sss3, _mm256_madd_epi16(pix, mmk));
            }
        }

        sss0 = _mm256_srai_epi32::<PRECISION>(sss0);
        sss1 = _mm256_srai_epi32::<PRECISION>(sss1);
        sss2 = _mm256_srai_epi32::<PRECISION>(sss2);
        sss3 = _mm256_srai_epi32::<PRECISION>(sss3);

        sss0 = _mm256_packs_epi32(sss0, sss1);
        sss2 = _mm256_packs_epi32(sss2, sss3);
        sss0 = _mm256_packus_epi16(sss0, sss2);

        let dst_ptr = dst_chunk.as_mut_ptr() as *mut __m256i;
        _mm256_storeu_si256(dst_ptr, sss0);

        src_x += 32;
    }

    // 8 components in half of SSE register
    dst_u8 = dst_chunks_32.into_remainder();
    let mut dst_chunks_8 = dst_u8.chunks_exact_mut(8);
    for dst_chunk in &mut dst_chunks_8 {
        let mut sss0 = initial; // left row
        let mut sss1 = initial; // right row

        let coeffs_chunks = coeffs.chunks_exact(2);
        let coeffs_reminder = coeffs_chunks.remainder();

        for (src_rows, two_coeffs) in src_view.iter_2_rows(y_start, max_rows).zip(coeffs_chunks) {
            let components1 = T::components(src_rows[0]);
            let components2 = T::components(src_rows[1]);
            // Load two coefficients at once
            let mmk = simd_utils::mm_load_and_clone_i16x2(two_coeffs);

            let source1 = simd_utils::loadl_epi64(components1, src_x); // top line
            let source2 = simd_utils::loadl_epi64(components2, src_x); // bottom line

            let source = _mm_unpacklo_epi8(source1, source2);
            let pix = _mm_unpacklo_epi8(source, _mm_setzero_si128());
            sss0 = _mm_add_epi32(sss0, _mm_madd_epi16(pix, mmk));
            let pix = _mm_unpackhi_epi8(source, _mm_setzero_si128());
            sss1 = _mm_add_epi32(sss1, _mm_madd_epi16(pix, mmk));
        }

        if let Some(&k) = coeffs_reminder.first() {
            if let Some(s_row) = src_view.iter_rows(y_last).next() {
                let components = T::components(s_row);
                let mmk = _mm_set1_epi32(k as i32);

                let source1 = simd_utils::loadl_epi64(components, src_x); // top line
                let source2 = _mm_setzero_si128(); // bottom line is empty

                let source = _mm_unpacklo_epi8(source1, source2);
                let pix = _mm_unpacklo_epi8(source, _mm_setzero_si128());
                sss0 = _mm_add_epi32(sss0, _mm_madd_epi16(pix, mmk));
                let pix = _mm_unpackhi_epi8(source, _mm_setzero_si128());
                sss1 = _mm_add_epi32(sss1, _mm_madd_epi16(pix, mmk));
            }
        }

        sss0 = _mm_srai_epi32::<PRECISION>(sss0);
        sss1 = _mm_srai_epi32::<PRECISION>(sss1);

        sss0 = _mm_packs_epi32(sss0, sss1);
        sss0 = _mm_packus_epi16(sss0, sss0);

        let dst_ptr = dst_chunk.as_mut_ptr() as *mut __m128i;
        _mm_storel_epi64(dst_ptr, sss0);

        src_x += 8;
    }

    dst_u8 = dst_chunks_8.into_remainder();
    let mut dst_chunks_4 = dst_u8.chunks_exact_mut(4);
    if let Some(dst_chunk) = dst_chunks_4.next() {
        let mut sss = initial;

        let coeffs_chunks = coeffs.chunks_exact(2);
        let coeffs_reminder = coeffs_chunks.remainder();

        for (src_rows, two_coeffs) in src_view.iter_2_rows(y_start, max_rows).zip(coeffs_chunks) {
            let components1 = T::components(src_rows[0]);
            let components2 = T::components(src_rows[1]);
            // Load two coefficients at once
            let two_coeffs = simd_utils::mm_load_and_clone_i16x2(two_coeffs);

            let row1 = simd_utils::mm_cvtsi32_si128_from_u8(components1, src_x); // top line
            let row2 = simd_utils::mm_cvtsi32_si128_from_u8(components2, src_x); // bottom line

            let pixels_u8 = _mm_unpacklo_epi8(row1, row2);
            let pixels_i16 = _mm_unpacklo_epi8(pixels_u8, _mm_setzero_si128());
            sss = _mm_add_epi32(sss, _mm_madd_epi16(pixels_i16, two_coeffs));
        }

        if let Some(&k) = coeffs_reminder.first() {
            if let Some(s_row) = src_view.iter_rows(y_last).next() {
                let components = T::components(s_row);
                let pix = simd_utils::mm_cvtepu8_epi32_from_u8(components, src_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);
        let dst_ptr = dst_chunk.as_mut_ptr() as *mut i32;
        dst_ptr.write_unaligned(_mm_cvtsi128_si32(_mm_packus_epi16(sss, sss)));

        src_x += 4;
    }

    dst_u8 = dst_chunks_4.into_remainder();
    if !dst_u8.is_empty() {
        native::convolution_by_u8(
            src_view,
            normalizer,
            1 << (PRECISION as u8 - 1),
            dst_u8,
            src_x,
            y_start,
            coeffs,
        );
    }
}