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

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

use super::native;
use crate::pixels::U16x2;
use crate::utils::foreach_with_pre_reading;
use crate::{ImageView, ImageViewMut};

#[target_feature(enable = "sse4.1")]
pub(crate) unsafe fn multiply_alpha(
    src_view: &impl ImageView<Pixel = U16x2>,
    dst_view: &mut impl ImageViewMut<Pixel = U16x2>,
) {
    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 = "sse4.1")]
pub(crate) unsafe fn multiply_alpha_inplace(image_view: &mut impl ImageViewMut<Pixel = U16x2>) {
    for row in image_view.iter_rows_mut(0) {
        multiply_alpha_row_inplace(row);
    }
}

#[inline]
#[target_feature(enable = "sse4.1")]
pub(crate) unsafe fn multiply_alpha_row(src_row: &[U16x2], dst_row: &mut [U16x2]) {
    let src_chunks = src_row.chunks_exact(4);
    let src_remainder = src_chunks.remainder();
    let mut dst_chunks = dst_row.chunks_exact_mut(4);
    let src_dst = src_chunks.zip(&mut dst_chunks);
    foreach_with_pre_reading(
        src_dst,
        |(src, dst)| {
            let pixels = _mm_loadu_si128(src.as_ptr() as *const __m128i);
            let dst_ptr = dst.as_mut_ptr() as *mut __m128i;
            (pixels, dst_ptr)
        },
        |(mut pixels, dst_ptr)| {
            pixels = multiply_alpha_4_pixels(pixels);
            _mm_storeu_si128(dst_ptr, pixels);
        },
    );

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

#[inline]
#[target_feature(enable = "sse4.1")]
pub(crate) unsafe fn multiply_alpha_row_inplace(row: &mut [U16x2]) {
    let mut chunks = row.chunks_exact_mut(4);
    foreach_with_pre_reading(
        &mut chunks,
        |chunk| {
            let pixels = _mm_loadu_si128(chunk.as_ptr() as *const __m128i);
            let dst_ptr = chunk.as_mut_ptr() as *mut __m128i;
            (pixels, dst_ptr)
        },
        |(mut pixels, dst_ptr)| {
            pixels = multiply_alpha_4_pixels(pixels);
            _mm_storeu_si128(dst_ptr, pixels);
        },
    );

    let reminder = chunks.into_remainder();
    if !reminder.is_empty() {
        native::multiply_alpha_row_inplace(reminder);
    }
}

#[inline]
#[target_feature(enable = "sse4.1")]
unsafe fn multiply_alpha_4_pixels(pixels: __m128i) -> __m128i {
    let zero = _mm_setzero_si128();
    let half = _mm_set1_epi32(0x8000);

    const MAX_A: i32 = 0xffff0000u32 as i32;
    let max_alpha = _mm_set1_epi32(MAX_A);
    /*
       |L0   A0  | |L1   A1  | |L2   A2  | |L3   A3  |
       |0001 0203| |0405 0607| |0809 1011| |1213 1415|
    */
    let factor_mask = _mm_set_epi8(15, 14, 15, 14, 11, 10, 11, 10, 7, 6, 7, 6, 3, 2, 3, 2);

    let factor_pixels = _mm_shuffle_epi8(pixels, factor_mask);
    let factor_pixels = _mm_or_si128(factor_pixels, max_alpha);

    let src_i32_lo = _mm_unpacklo_epi16(pixels, zero);
    let factors = _mm_unpacklo_epi16(factor_pixels, zero);
    let src_i32_lo = _mm_add_epi32(_mm_mullo_epi32(src_i32_lo, factors), half);
    let dst_i32_lo = _mm_add_epi32(src_i32_lo, _mm_srli_epi32::<16>(src_i32_lo));
    let dst_i32_lo = _mm_srli_epi32::<16>(dst_i32_lo);

    let src_i32_hi = _mm_unpackhi_epi16(pixels, zero);
    let factors = _mm_unpackhi_epi16(factor_pixels, zero);
    let src_i32_hi = _mm_add_epi32(_mm_mullo_epi32(src_i32_hi, factors), half);
    let dst_i32_hi = _mm_add_epi32(src_i32_hi, _mm_srli_epi32::<16>(src_i32_hi));
    let dst_i32_hi = _mm_srli_epi32::<16>(dst_i32_hi);

    _mm_packus_epi32(dst_i32_lo, dst_i32_hi)
}

// Divide

#[target_feature(enable = "sse4.1")]
pub(crate) unsafe fn divide_alpha(
    src_view: &impl ImageView<Pixel = U16x2>,
    dst_view: &mut impl ImageViewMut<Pixel = U16x2>,
) {
    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) {
        divide_alpha_row(src_row, dst_row);
    }
}

#[target_feature(enable = "sse4.1")]
pub(crate) unsafe fn divide_alpha_inplace(image_view: &mut impl ImageViewMut<Pixel = U16x2>) {
    for row in image_view.iter_rows_mut(0) {
        divide_alpha_row_inplace(row);
    }
}

#[target_feature(enable = "sse4.1")]
pub(crate) unsafe fn divide_alpha_row(src_row: &[U16x2], dst_row: &mut [U16x2]) {
    let src_chunks = src_row.chunks_exact(4);
    let src_remainder = src_chunks.remainder();
    let mut dst_chunks = dst_row.chunks_exact_mut(4);
    let src_dst = src_chunks.zip(&mut dst_chunks);
    foreach_with_pre_reading(
        src_dst,
        |(src, dst)| {
            let pixels = _mm_loadu_si128(src.as_ptr() as *const __m128i);
            let dst_ptr = dst.as_mut_ptr() as *mut __m128i;
            (pixels, dst_ptr)
        },
        |(mut pixels, dst_ptr)| {
            pixels = divide_alpha_4_pixels(pixels);
            _mm_storeu_si128(dst_ptr, pixels);
        },
    );

    if !src_remainder.is_empty() {
        let dst_reminder = dst_chunks.into_remainder();
        let mut src_pixels = [U16x2::new([0, 0]); 4];
        src_pixels
            .iter_mut()
            .zip(src_remainder)
            .for_each(|(d, s)| *d = *s);

        let mut dst_pixels = [U16x2::new([0, 0]); 4];
        let mut pixels = _mm_loadu_si128(src_pixels.as_ptr() as *const __m128i);
        pixels = divide_alpha_4_pixels(pixels);
        _mm_storeu_si128(dst_pixels.as_mut_ptr() as *mut __m128i, pixels);

        dst_pixels
            .iter()
            .zip(dst_reminder)
            .for_each(|(s, d)| *d = *s);
    }
}

#[target_feature(enable = "sse4.1")]
pub(crate) unsafe fn divide_alpha_row_inplace(row: &mut [U16x2]) {
    let mut chunks = row.chunks_exact_mut(4);
    // Using a simple for-loop in this case is faster than implementation with pre-reading
    for chunk in &mut chunks {
        let mut pixels = _mm_loadu_si128(chunk.as_ptr() as *const __m128i);
        pixels = divide_alpha_4_pixels(pixels);
        _mm_storeu_si128(chunk.as_mut_ptr() as *mut __m128i, pixels);
    }

    let reminder = chunks.into_remainder();
    if !reminder.is_empty() {
        let mut src_pixels = [U16x2::new([0, 0]); 4];
        src_pixels
            .iter_mut()
            .zip(reminder.iter())
            .for_each(|(d, s)| *d = *s);

        let mut dst_pixels = [U16x2::new([0, 0]); 4];
        let mut pixels = _mm_loadu_si128(src_pixels.as_ptr() as *const __m128i);
        pixels = divide_alpha_4_pixels(pixels);
        _mm_storeu_si128(dst_pixels.as_mut_ptr() as *mut __m128i, pixels);

        dst_pixels.iter().zip(reminder).for_each(|(s, d)| *d = *s);
    }
}

#[inline]
#[target_feature(enable = "sse4.1")]
unsafe fn divide_alpha_4_pixels(pixels: __m128i) -> __m128i {
    let alpha_mask = _mm_set1_epi32(0xffff0000u32 as i32);
    let luma_mask = _mm_set1_epi32(0xffff);
    let alpha_max = _mm_set1_ps(65535.0);
    /*
       |L0   A0  | |L1   A1  | |L2   A2  | |L3   A3  |
       |0001 0203| |0405 0607| |0809 1011| |1213 1415|
    */
    let alpha32_sh = _mm_set_epi8(-1, -1, 15, 14, -1, -1, 11, 10, -1, -1, 7, 6, -1, -1, 3, 2);

    let alpha_f32x4 = _mm_cvtepi32_ps(_mm_shuffle_epi8(pixels, alpha32_sh));
    let luma_f32x4 = _mm_cvtepi32_ps(_mm_and_si128(pixels, luma_mask));
    let scaled_luma_f32x4 = _mm_mul_ps(luma_f32x4, alpha_max);
    let mut divided_luma_i32x4 = _mm_cvtps_epi32(_mm_div_ps(scaled_luma_f32x4, alpha_f32x4));
    // Clamp result to [0..0xffff]
    divided_luma_i32x4 = _mm_min_epi32(divided_luma_i32x4, luma_mask);

    let alpha = _mm_and_si128(pixels, alpha_mask);
    _mm_blendv_epi8(divided_luma_i32x4, alpha, alpha_mask)
}