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

use core::arch::aarch64::*;

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

#[target_feature(enable = "neon")]
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 = "neon")]
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(always)]
unsafe fn multiply_alpha_row(src_row: &[U16x2], dst_row: &mut [U16x2]) {
    let src_chunks = src_row.chunks_exact(8);
    let src_remainder = src_chunks.remainder();
    let mut dst_chunks = dst_row.chunks_exact_mut(8);
    // Using a simple for-loop in this case is faster than implementation with pre-reading
    for (src, dst) in src_chunks.zip(&mut dst_chunks) {
        let mut pixels = neon_utils::load_deintrel_u16x8x2(src, 0);
        pixels.0 = neon_utils::multiply_color_to_alpha_u16x8(pixels.0, pixels.1);
        let dst_ptr = dst.as_mut_ptr() as *mut u16;
        vst2q_u16(dst_ptr, pixels);
    }

    if !src_remainder.is_empty() {
        let src_chunks = src_remainder.chunks_exact(4);
        let src_remainder = src_chunks.remainder();
        let dst_reminder = dst_chunks.into_remainder();
        let mut dst_chunks = dst_reminder.chunks_exact_mut(4);
        let mut src_dst = src_chunks.zip(&mut dst_chunks);
        if let Some((src, dst)) = src_dst.next() {
            let mut pixels = neon_utils::load_deintrel_u16x4x2(src, 0);
            pixels.0 = neon_utils::multiply_color_to_alpha_u16x4(pixels.0, pixels.1);
            let dst_ptr = dst.as_mut_ptr() as *mut u16;
            vst2_u16(dst_ptr, pixels);
        }

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

#[inline(always)]
unsafe fn multiply_alpha_row_inplace(row: &mut [U16x2]) {
    let mut chunks = row.chunks_exact_mut(8);
    // Using a simple for-loop in this case is faster than implementation with pre-reading
    for chunk in &mut chunks {
        let mut pixels = neon_utils::load_deintrel_u16x8x2(chunk, 0);
        pixels.0 = neon_utils::multiply_color_to_alpha_u16x8(pixels.0, pixels.1);
        let dst_ptr = chunk.as_mut_ptr() as *mut u16;
        vst2q_u16(dst_ptr, pixels);
    }

    let reminder = chunks.into_remainder();
    if !reminder.is_empty() {
        let mut chunks = reminder.chunks_exact_mut(4);
        if let Some(chunk) = chunks.next() {
            let mut pixels = neon_utils::load_deintrel_u16x4x2(chunk, 0);
            pixels.0 = neon_utils::multiply_color_to_alpha_u16x4(pixels.0, pixels.1);
            let dst_ptr = chunk.as_mut_ptr() as *mut u16;
            vst2_u16(dst_ptr, pixels);
        }

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

// Divide

#[target_feature(enable = "neon")]
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 = "neon")]
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);
    }
}

#[inline(always)]
pub(crate) unsafe fn divide_alpha_row(src_row: &[U16x2], dst_row: &mut [U16x2]) {
    let src_chunks = src_row.chunks_exact(8);
    let src_remainder = src_chunks.remainder();
    let mut dst_chunks = dst_row.chunks_exact_mut(8);
    // Using a simple for-loop in this case is faster than implementation with pre-reading
    for (src, dst) in src_chunks.zip(&mut dst_chunks) {
        let mut pixels = neon_utils::load_deintrel_u16x8x2(src, 0);
        pixels = divide_alpha_8_pixels(pixels);
        let dst_ptr = dst.as_mut_ptr() as *mut u16;
        vst2q_u16(dst_ptr, pixels);
    }

    if !src_remainder.is_empty() {
        let dst_reminder = dst_chunks.into_remainder();
        let mut src_pixels = [U16x2::new([0, 0]); 8];
        src_pixels
            .iter_mut()
            .zip(src_remainder)
            .for_each(|(d, s)| *d = *s);

        let mut dst_pixels = [U16x2::new([0, 0]); 8];
        let mut pixels = neon_utils::load_deintrel_u16x8x2(&src_pixels, 0);
        pixels = divide_alpha_8_pixels(pixels);
        let dst_ptr = dst_pixels.as_mut_ptr() as *mut u16;
        vst2q_u16(dst_ptr, pixels);

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

#[inline(always)]
pub(crate) unsafe fn divide_alpha_row_inplace(row: &mut [U16x2]) {
    let mut chunks = row.chunks_exact_mut(8);
    // Using a simple for-loop in this case is faster than implementation with pre-reading
    for chunk in &mut chunks {
        let mut pixels = neon_utils::load_deintrel_u16x8x2(chunk, 0);
        pixels = divide_alpha_8_pixels(pixels);
        let dst_ptr = chunk.as_mut_ptr() as *mut u16;
        vst2q_u16(dst_ptr, pixels);
    }

    let reminder = chunks.into_remainder();
    if !reminder.is_empty() {
        let mut src_pixels = [U16x2::new([0, 0]); 8];
        src_pixels
            .iter_mut()
            .zip(reminder.iter())
            .for_each(|(d, s)| *d = *s);

        let mut dst_pixels = [U16x2::new([0, 0]); 8];
        let mut pixels = neon_utils::load_deintrel_u16x8x2(&src_pixels, 0);
        pixels = divide_alpha_8_pixels(pixels);
        let dst_ptr = dst_pixels.as_mut_ptr() as *mut u16;
        vst2q_u16(dst_ptr, pixels);

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

#[inline(always)]
unsafe fn divide_alpha_8_pixels(mut pixels: uint16x8x2_t) -> uint16x8x2_t {
    let zero = vdupq_n_u16(0);
    let alpha_scale = vdupq_n_f32(65535.0);
    let nonzero_alpha_mask = vmvnq_u16(vceqzq_u16(pixels.1));

    // Low
    let alpha_scaled_u32 = vreinterpretq_u32_u16(vzip1q_u16(pixels.1, zero));
    let alpha_scaled_f32 = vcvtq_f32_u32(alpha_scaled_u32);
    let recip_alpha_lo_f32 = vdivq_f32(alpha_scale, alpha_scaled_f32);

    // High
    let alpha_scaled_u32 = vreinterpretq_u32_u16(vzip2q_u16(pixels.1, zero));
    let alpha_scaled_f32 = vcvtq_f32_u32(alpha_scaled_u32);
    let recip_alpha_hi_f32 = vdivq_f32(alpha_scale, alpha_scaled_f32);

    pixels.0 = neon_utils::mul_color_recip_alpha_u16x8(
        pixels.0,
        recip_alpha_lo_f32,
        recip_alpha_hi_f32,
        zero,
    );
    pixels.0 = vandq_u16(pixels.0, nonzero_alpha_mask);
    pixels
}