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

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

use super::sse4;
use crate::pixels::U8x4;
use crate::utils::foreach_with_pre_reading;
use crate::{simd_utils, ImageView, ImageViewMut};

#[target_feature(enable = "avx2")]
pub(crate) unsafe fn multiply_alpha(
    src_view: &impl ImageView<Pixel = U8x4>,
    dst_view: &mut impl ImageViewMut<Pixel = U8x4>,
) {
    let src_rows = src_view.iter_rows(0);
    let dst_rows = dst_view.iter_rows_mut(0);
    for (src_row, dst_row) in src_rows.zip(dst_rows) {
        multiply_alpha_row(src_row, dst_row);
    }
}

#[target_feature(enable = "avx2")]
pub(crate) unsafe fn multiply_alpha_inplace(image_view: &mut impl ImageViewMut<Pixel = U8x4>) {
    let rows = image_view.iter_rows_mut(0);
    for row in rows {
        multiply_alpha_row_inplace(row);
    }
}

#[inline]
#[target_feature(enable = "avx2")]
unsafe fn multiply_alpha_row(src_row: &[U8x4], dst_row: &mut [U8x4]) {
    let src_chunks = src_row.chunks_exact(8);
    let src_tail = src_chunks.remainder();
    let mut dst_chunks = dst_row.chunks_exact_mut(8);
    let src_dst = src_chunks.zip(&mut dst_chunks);
    foreach_with_pre_reading(
        src_dst,
        |(src, dst)| {
            let pixels = simd_utils::loadu_si256(src, 0);
            let dst_ptr = dst.as_mut_ptr() as *mut __m256i;
            (pixels, dst_ptr)
        },
        |(mut pixels, dst_ptr)| {
            pixels = multiply_alpha_8_pixels(pixels);
            _mm256_storeu_si256(dst_ptr, pixels);
        },
    );
    if !src_tail.is_empty() {
        let dst_tail = dst_chunks.into_remainder();
        sse4::multiply_alpha_row(src_tail, dst_tail);
    }
}

#[inline]
#[target_feature(enable = "avx2")]
unsafe fn multiply_alpha_row_inplace(row: &mut [U8x4]) {
    let mut chunks = row.chunks_exact_mut(8);
    foreach_with_pre_reading(
        &mut chunks,
        |chunk| {
            let pixels = simd_utils::loadu_si256(chunk, 0);
            let dst_ptr = chunk.as_mut_ptr() as *mut __m256i;
            (pixels, dst_ptr)
        },
        |(mut pixels, dst_ptr)| {
            pixels = multiply_alpha_8_pixels(pixels);
            _mm256_storeu_si256(dst_ptr, pixels);
        },
    );
    let tail = chunks.into_remainder();
    if !tail.is_empty() {
        sse4::multiply_alpha_row_inplace(tail);
    }
}

#[inline]
#[target_feature(enable = "avx2")]
unsafe fn multiply_alpha_8_pixels(pixels: __m256i) -> __m256i {
    let zero = _mm256_setzero_si256();
    let half = _mm256_set1_epi16(128);

    const MAX_A: i32 = 0xff000000u32 as i32;
    let max_alpha = _mm256_set1_epi32(MAX_A);
    #[rustfmt::skip]
    let factor_mask = _mm256_set_epi8(
        15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3,
        15, 15, 15, 15, 11, 11, 11, 11, 7, 7, 7, 7, 3, 3, 3, 3,
    );

    let factor_pixels = _mm256_shuffle_epi8(pixels, factor_mask);
    let factor_pixels = _mm256_or_si256(factor_pixels, max_alpha);

    let pix1 = _mm256_unpacklo_epi8(pixels, zero);
    let factors = _mm256_unpacklo_epi8(factor_pixels, zero);
    let pix1 = _mm256_add_epi16(_mm256_mullo_epi16(pix1, factors), half);
    let pix1 = _mm256_add_epi16(pix1, _mm256_srli_epi16::<8>(pix1));
    let pix1 = _mm256_srli_epi16::<8>(pix1);

    let pix2 = _mm256_unpackhi_epi8(pixels, zero);
    let factors = _mm256_unpackhi_epi8(factor_pixels, zero);
    let pix2 = _mm256_add_epi16(_mm256_mullo_epi16(pix2, factors), half);
    let pix2 = _mm256_add_epi16(pix2, _mm256_srli_epi16::<8>(pix2));
    let pix2 = _mm256_srli_epi16::<8>(pix2);

    _mm256_packus_epi16(pix1, pix2)
}

// Divide

#[target_feature(enable = "avx2")]
pub(crate) unsafe fn divide_alpha(
    src_view: &impl ImageView<Pixel = U8x4>,
    dst_view: &mut impl ImageViewMut<Pixel = U8x4>,
) {
    let src_rows = src_view.iter_rows(0);
    let dst_rows = dst_view.iter_rows_mut(0);
    let rows = src_rows.zip(dst_rows);
    for (src_row, dst_row) in rows {
        divide_alpha_row(src_row, dst_row);
    }
}

#[target_feature(enable = "avx2")]
pub(crate) unsafe fn divide_alpha_inplace(image_view: &mut impl ImageViewMut<Pixel = U8x4>) {
    let rows = image_view.iter_rows_mut(0);
    for row in rows {
        divide_alpha_row_inplace(row);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn divide_alpha_row(src_row: &[U8x4], dst_row: &mut [U8x4]) {
    let src_chunks = src_row.chunks_exact(8);
    let src_remainder = src_chunks.remainder();
    let mut dst_chunks = dst_row.chunks_exact_mut(8);
    let src_dst = src_chunks.zip(&mut dst_chunks);
    foreach_with_pre_reading(
        src_dst,
        |(src, dst)| {
            let pixels = simd_utils::loadu_si256(src, 0);
            let dst_ptr = dst.as_mut_ptr() as *mut __m256i;
            (pixels, dst_ptr)
        },
        |(mut pixels, dst_ptr)| {
            pixels = divide_alpha_8_pixels(pixels);
            _mm256_storeu_si256(dst_ptr, pixels);
        },
    );

    if !src_remainder.is_empty() {
        let dst_reminder = dst_chunks.into_remainder();
        sse4::divide_alpha_row(src_remainder, dst_reminder);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn divide_alpha_row_inplace(row: &mut [U8x4]) {
    let mut chunks = row.chunks_exact_mut(8);
    foreach_with_pre_reading(
        &mut chunks,
        |chunk| {
            let pixels = simd_utils::loadu_si256(chunk, 0);
            let dst_ptr = chunk.as_mut_ptr() as *mut __m256i;
            (pixels, dst_ptr)
        },
        |(mut pixels, dst_ptr)| {
            pixels = divide_alpha_8_pixels(pixels);
            _mm256_storeu_si256(dst_ptr, pixels);
        },
    );

    let tail = chunks.into_remainder();
    if !tail.is_empty() {
        sse4::divide_alpha_row_inplace(tail);
    }
}

#[inline]
#[target_feature(enable = "avx2")]
unsafe fn divide_alpha_8_pixels(pixels: __m256i) -> __m256i {
    let zero = _mm256_setzero_si256();
    let alpha_mask = _mm256_set1_epi32(0xff000000u32 as i32);
    #[rustfmt::skip]
    let shuffle1 = _mm256_set_epi8(
        5, 4, 5, 4, 5, 4, 5, 4, 1, 0, 1, 0, 1, 0, 1, 0,
        5, 4, 5, 4, 5, 4, 5, 4, 1, 0, 1, 0, 1, 0, 1, 0,
    );
    #[rustfmt::skip]
    let shuffle2 = _mm256_set_epi8(
        13, 12, 13, 12, 13, 12, 13, 12, 9, 8, 9, 8, 9, 8, 9, 8,
        13, 12, 13, 12, 13, 12, 13, 12, 9, 8, 9, 8, 9, 8, 9, 8,
    );
    let alpha_scale = _mm256_set1_ps(255.0 * 256.0);
    let max_value = _mm256_set1_epi16(0xff);

    let alpha_f32 = _mm256_cvtepi32_ps(_mm256_srli_epi32::<24>(pixels));
    let recip_alpha_i32 = _mm256_cvtps_epi32(_mm256_div_ps(alpha_scale, alpha_f32));

    // Recip alpha in Q8.8 format
    let recip_alpha_lo_q8_8 = _mm256_shuffle_epi8(recip_alpha_i32, shuffle1);
    let recip_alpha_hi_q8_8 = _mm256_shuffle_epi8(recip_alpha_i32, shuffle2);

    // Pixels components in format Q9.7
    let components_lo_q9_7 = _mm256_slli_epi16::<7>(_mm256_unpacklo_epi8(pixels, zero));
    let components_hi_q9_7 = _mm256_slli_epi16::<7>(_mm256_unpackhi_epi8(pixels, zero));

    // Multiplied pixels components as i16.
    //
    // fn _mm256_mulhrs_epi16(a: i16, b: i16) -> i16 {
    //   let tmp: i32 = ((a as i32 * b as i32) >> 14) + 1;
    //   (tmp >> 1) as i16
    // }
    let res_components_lo_i16 = _mm256_min_epu16(
        _mm256_mulhrs_epi16(components_lo_q9_7, recip_alpha_lo_q8_8),
        max_value,
    );
    let res_components_hi_i16 = _mm256_min_epu16(
        _mm256_mulhrs_epi16(components_hi_q9_7, recip_alpha_hi_q8_8),
        max_value,
    );

    let alpha = _mm256_and_si256(pixels, alpha_mask);
    let rgb = _mm256_packus_epi16(res_components_lo_i16, res_components_hi_i16);
    _mm256_blendv_epi8(rgb, alpha, alpha_mask)
}