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

use core::arch::wasm32::*;

use super::native;
use crate::pixels::U8x2;
use crate::wasm32_utils::u16x8_mul_add_shr16;
use crate::{ImageView, ImageViewMut};

pub(crate) unsafe fn multiply_alpha(
    src_view: &impl ImageView<Pixel = U8x2>,
    dst_view: &mut impl ImageViewMut<Pixel = U8x2>,
) {
    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);
    }
}

pub(crate) unsafe fn multiply_alpha_inplace(image_view: &mut impl ImageViewMut<Pixel = U8x2>) {
    for row in image_view.iter_rows_mut(0) {
        multiply_alpha_row_inplace(row);
    }
}

#[inline]
#[target_feature(enable = "simd128")]
unsafe fn multiply_alpha_row(src_row: &[U8x2], dst_row: &mut [U8x2]) {
    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);
    // A simple for-loop in this case is as fast as implementation with pre-reading
    for (src, dst) in src_dst {
        let src_pixels = v128_load(src.as_ptr() as *const v128);
        let dst_pixels = multiplies_alpha_8_pixels(src_pixels);
        v128_store(dst.as_mut_ptr() as *mut v128, dst_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 = "simd128")]
unsafe fn multiply_alpha_row_inplace(row: &mut [U8x2]) {
    let mut chunks = row.chunks_exact_mut(8);
    // Using a simple for-loop in this case is as fast as implementation with pre-reading
    for chunk in &mut chunks {
        let src_pixels = v128_load(chunk.as_ptr() as *const v128);
        let dst_pixels = multiplies_alpha_8_pixels(src_pixels);
        v128_store(chunk.as_mut_ptr() as *mut v128, dst_pixels);
    }

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

#[inline]
#[target_feature(enable = "simd128")]
unsafe fn multiplies_alpha_8_pixels(pixels: v128) -> v128 {
    let half = u16x8_splat(128);
    let max_alpha = u16x8_splat(0xff00);
    /*
       |L  A | |L  A | |L  A | |L  A | |L  A | |L  A | |L  A | |L  A |
       |00 01| |02 03| |04 05| |06 07| |08 09| |10 11| |12 13| |14 15|
    */
    const FACTOR_MASK: v128 = i8x16(1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11, 13, 13, 15, 15);

    let factor_pixels = u8x16_swizzle(pixels, FACTOR_MASK);
    let factor_pixels = v128_or(factor_pixels, max_alpha);

    let src_u16_lo = u16x8_extend_low_u8x16(pixels);
    let factors = u16x8_extend_low_u8x16(factor_pixels);
    let mut dst_u16_lo = u16x8_add(u16x8_mul(src_u16_lo, factors), half);
    dst_u16_lo = u16x8_add(dst_u16_lo, u16x8_shr(dst_u16_lo, 8));
    dst_u16_lo = u16x8_shr(dst_u16_lo, 8);

    let src_u16_hi = u16x8_extend_high_u8x16(pixels);
    let factors = u16x8_extend_high_u8x16(factor_pixels);
    let mut dst_u16_hi = u16x8_add(u16x8_mul(src_u16_hi, factors), half);
    dst_u16_hi = u16x8_add(dst_u16_hi, u16x8_shr(dst_u16_hi, 8));
    dst_u16_hi = u16x8_shr(dst_u16_hi, 8);

    u8x16_narrow_i16x8(dst_u16_lo, dst_u16_hi)
}

// Divide

pub(crate) unsafe fn divide_alpha(
    src_view: &impl ImageView<Pixel = U8x2>,
    dst_view: &mut impl ImageViewMut<Pixel = U8x2>,
) {
    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);
    }
}

pub(crate) unsafe fn divide_alpha_inplace(image_view: &mut impl ImageViewMut<Pixel = U8x2>) {
    for row in image_view.iter_rows_mut(0) {
        divide_alpha_row_inplace(row);
    }
}

#[inline]
#[target_feature(enable = "simd128")]
unsafe fn divide_alpha_row(src_row: &[U8x2], dst_row: &mut [U8x2]) {
    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);
    // Using a simple for-loop in this case is as fast as implementation with pre-reading
    for (src, dst) in src_dst {
        let src_pixels = v128_load(src.as_ptr() as *const v128);
        let dst_pixels = divide_alpha_8_pixels(src_pixels);
        v128_store(dst.as_mut_ptr() as *mut v128, dst_pixels);
    }

    if !src_remainder.is_empty() {
        let dst_reminder = dst_chunks.into_remainder();
        let mut src_pixels = [U8x2::new([0, 0]); 8];
        src_pixels
            .iter_mut()
            .zip(src_remainder)
            .for_each(|(d, s)| *d = *s);

        let mut dst_pixels = [U8x2::new([0; 2]); 8];
        let mut pixels = v128_load(src_pixels.as_ptr() as *const v128);
        pixels = divide_alpha_8_pixels(pixels);
        v128_store(dst_pixels.as_mut_ptr() as *mut v128, pixels);

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

#[inline]
#[target_feature(enable = "simd128")]
unsafe fn divide_alpha_row_inplace(row: &mut [U8x2]) {
    let mut chunks = row.chunks_exact_mut(8);
    // Using a simple for-loop in this case is as fast as implementation with pre-reading
    for chunk in &mut chunks {
        let src_pixels = v128_load(chunk.as_ptr() as *const v128);
        let dst_pixels = divide_alpha_8_pixels(src_pixels);
        v128_store(chunk.as_mut_ptr() as *mut v128, dst_pixels);
    }

    let reminder = chunks.into_remainder();
    if !reminder.is_empty() {
        let mut src_pixels = [U8x2::new([0; 2]); 8];
        src_pixels
            .iter_mut()
            .zip(reminder.iter())
            .for_each(|(d, s)| *d = *s);

        let mut dst_pixels = [U8x2::new([0; 2]); 8];
        let mut pixels = v128_load(src_pixels.as_ptr() as *const v128);
        pixels = divide_alpha_8_pixels(pixels);
        v128_store(dst_pixels.as_mut_ptr() as *mut v128, pixels);

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

#[inline]
#[target_feature(enable = "simd128")]
unsafe fn divide_alpha_8_pixels(pixels: v128) -> v128 {
    const ALPHA32_SH_LO: v128 = i8x16(1, -1, -1, -1, 3, -1, -1, -1, 5, -1, -1, -1, 7, -1, -1, -1);
    const ALPHA32_SH_HI: v128 = i8x16(
        9, -1, -1, -1, 11, -1, -1, -1, 13, -1, -1, -1, 15, -1, -1, -1,
    );
    let alpha_scale = f32x4_splat(255.0 * 256.0);

    let alpha_lo_f32 = f32x4_convert_u32x4(u8x16_swizzle(pixels, ALPHA32_SH_LO));
    // In case of zero division the result will be u32::MAX or 0.
    let scaled_alpha_lo_u32 = u32x4_trunc_sat_f32x4(f32x4_add(
        f32x4_div(alpha_scale, alpha_lo_f32),
        f32x4_splat(0.5),
    ));

    let alpha_hi_f32 = f32x4_convert_u32x4(u8x16_swizzle(pixels, ALPHA32_SH_HI));
    let scaled_alpha_hi_u32 = u32x4_trunc_sat_f32x4(f32x4_add(
        f32x4_div(alpha_scale, alpha_hi_f32),
        f32x4_splat(0.5),
    ));

    // All u32::MAX values in arguments will interpreted as -1i32.
    // u16x8_narrow_i32x4() converts all negative values into 0.
    let scaled_alpha_u16 = u16x8_narrow_i32x4(scaled_alpha_lo_u32, scaled_alpha_hi_u32);

    let luma_u16 = u16x8_shl(pixels, 8);
    let scaled_luma_u16 = u16x8_mul_add_shr16(luma_u16, scaled_alpha_u16, u32x4_splat(0x8000));
    let luma_max = u16x8_splat(0xff);
    let scaled_luma_u16 = u16x8_min(scaled_luma_u16, luma_max);

    // Blend scaled luma with original alpha channel.
    let alpha_mask = u16x8_splat(0xff00);
    let alpha = v128_and(pixels, alpha_mask);
    v128_or(scaled_luma_u16, alpha)
}