simdeez 3.0.1

SIMD library to abstract over different instruction sets and widths
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

use crate::prelude::*;

use super::shared::assert_f32_slices_close;

fn reference_quantize_f32(input: &[f32]) -> Vec<i32> {
    input
        .iter()
        .map(|&value| (value.abs() * 1.75 + 0.125).floor().clamp(0.0, 255.0) as i32)
        .collect()
}

fn reference_normalize_luma_f32(input: &[f32]) -> Vec<f32> {
    input
        .iter()
        .map(|&value| {
            let normalized = ((value - 16.0) * (1.0 / 219.0)).clamp(0.0, 1.0);
            let lifted = (normalized * 1.12).clamp(0.0, 1.0);
            let softened = normalized * 0.9;
            if normalized > 0.62 {
                lifted
            } else {
                softened
            }
        })
        .collect()
}

simd_unsafe_generate_all!(
    fn quantize_f32_to_i32(input: &[f32], output: &mut [i32]) {
        assert_eq!(input.len(), output.len());

        let mut input = input;
        let mut output = output;
        let scale = S::Vf32::set1(1.75);
        let bias = S::Vf32::set1(0.125);
        let lower = S::Vf32::set1(0.0);
        let upper = S::Vf32::set1(255.0);

        while input.len() >= S::Vf32::WIDTH {
            let values = S::Vf32::load_from_slice(input);
            let quantized = values
                .abs()
                .mul_add(scale, bias)
                .floor()
                .max(lower)
                .min(upper)
                .cast_i32();
            quantized.copy_to_slice(output);

            input = &input[S::Vf32::WIDTH..];
            output = &mut output[S::Vf32::WIDTH..];
        }

        for (value, slot) in input.iter().zip(output.iter_mut()) {
            *slot = (value.abs() * 1.75 + 0.125).floor().clamp(0.0, 255.0) as i32;
        }
    }
);

simd_unsafe_generate_all!(
    fn normalize_luma_f32(input: &[f32], output: &mut [f32]) {
        assert_eq!(input.len(), output.len());

        let mut input = input;
        let mut output = output;
        let black_level = S::Vf32::set1(16.0);
        let inv_range = S::Vf32::set1(1.0 / 219.0);
        let zero = S::Vf32::set1(0.0);
        let one = S::Vf32::set1(1.0);
        let threshold = S::Vf32::set1(0.62);
        let lift = S::Vf32::set1(1.12);
        let soften = S::Vf32::set1(0.9);

        while input.len() >= S::Vf32::WIDTH {
            let values = S::Vf32::load_from_slice(input);
            let normalized = (values - black_level).mul(inv_range).max(zero).min(one);
            let lifted = normalized.mul(lift).min(one);
            let softened = normalized.mul(soften);
            let output_chunk = normalized.cmp_gt(threshold).blendv(softened, lifted);
            output_chunk.copy_to_slice(output);

            input = &input[S::Vf32::WIDTH..];
            output = &mut output[S::Vf32::WIDTH..];
        }

        for (value, slot) in input.iter().zip(output.iter_mut()) {
            let normalized = ((value - 16.0) * (1.0 / 219.0)).clamp(0.0, 1.0);
            let lifted = (normalized * 1.12).clamp(0.0, 1.0);
            let softened = normalized * 0.9;
            *slot = if normalized > 0.62 { lifted } else { softened };
        }
    }
);

fn assert_quantize_matches_all_backends(input: &[f32]) {
    let expected = reference_quantize_f32(input);

    let mut portable = vec![0; input.len()];
    quantize_f32_to_i32(input, &mut portable);
    assert_eq!(portable, expected);

    #[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
    {
        if std::arch::is_x86_feature_detected!("sse2") {
            let mut output = vec![0; input.len()];
            unsafe { quantize_f32_to_i32_sse2(input, &mut output) };
            assert_eq!(output, expected);
        }
        if std::arch::is_x86_feature_detected!("sse4.1") {
            let mut output = vec![0; input.len()];
            unsafe { quantize_f32_to_i32_sse41(input, &mut output) };
            assert_eq!(output, expected);
        }
        if std::arch::is_x86_feature_detected!("avx2") && std::arch::is_x86_feature_detected!("fma")
        {
            let mut output = vec![0; input.len()];
            unsafe { quantize_f32_to_i32_avx2(input, &mut output) };
            assert_eq!(output, expected);
        }
    }
}

fn assert_normalize_luma_matches_all_backends(input: &[f32]) {
    let expected = reference_normalize_luma_f32(input);

    let mut portable = vec![0.0; input.len()];
    normalize_luma_f32(input, &mut portable);
    assert_f32_slices_close(&portable, &expected);

    #[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
    {
        if std::arch::is_x86_feature_detected!("sse2") {
            let mut output = vec![0.0; input.len()];
            unsafe { normalize_luma_f32_sse2(input, &mut output) };
            assert_f32_slices_close(&output, &expected);
        }
        if std::arch::is_x86_feature_detected!("sse4.1") {
            let mut output = vec![0.0; input.len()];
            unsafe { normalize_luma_f32_sse41(input, &mut output) };
            assert_f32_slices_close(&output, &expected);
        }
        if std::arch::is_x86_feature_detected!("avx2") && std::arch::is_x86_feature_detected!("fma")
        {
            let mut output = vec![0.0; input.len()];
            unsafe { normalize_luma_f32_avx2(input, &mut output) };
            assert_f32_slices_close(&output, &expected);
        }
    }
}

// This is the kind of magnitude-to-bucket stage used before writing image intensities or feature
// maps back to integer buffers, where clamp and floor must agree across portable and native paths.
#[test]
fn real_world_quantize_f32_to_i32_matches_reference() {
    let input = [
        -300.0, -255.9, -144.3, -33.2, -8.75, -1.1, -0.4, 0.0, 0.2, 0.7, 1.1, 3.9, 7.3, 12.8, 31.6,
        63.2, 91.7, 127.4, 144.6, 199.1, 255.3, 300.7,
    ]
    .into_iter()
    .cycle()
    .take(53)
    .enumerate()
    .map(|(i, value)| value + (i % 5) as f32 * 0.03125)
    .collect::<Vec<_>>();

    assert_quantize_matches_all_backends(&[]);
    assert_quantize_matches_all_backends(&input[..3]);
    assert_quantize_matches_all_backends(&input[..17]);
    assert_quantize_matches_all_backends(&input);
}

// This mirrors a luma-normalization pass from imaging/video pipelines, where incoming studio-range
// values are normalized, thresholded, and gently lifted before later stages consume them.
#[test]
fn real_world_normalize_luma_f32_matches_reference() {
    let input = [
        -20.0, 0.0, 8.0, 15.5, 16.0, 23.0, 48.0, 64.0, 96.0, 128.0, 160.0, 186.0, 210.0, 235.0,
        260.0,
    ]
    .into_iter()
    .cycle()
    .take(67)
    .enumerate()
    .map(|(i, value)| value + ((i % 9) as f32 - 4.0) * 0.375)
    .collect::<Vec<_>>();

    assert_normalize_luma_matches_all_backends(&[]);
    assert_normalize_luma_matches_all_backends(&input[..5]);
    assert_normalize_luma_matches_all_backends(&input[..19]);
    assert_normalize_luma_matches_all_backends(&input);
}

#[test]
fn real_world_quantize_f32_to_i32_threshold_boundaries_match_reference() {
    let input = vec![
        -0.125, -0.071_428, 0.0, 0.071_428, 0.499_999, 0.5, 0.500_001, 145.642_84, 145.642_85,
        145.642_86, 400.0,
    ];
    assert_quantize_matches_all_backends(&input);
}

#[test]
fn real_world_normalize_luma_f32_threshold_boundaries_match_reference() {
    let threshold_input = 16.0 + 219.0 * 0.62;
    let input = vec![
        15.999,
        16.0,
        16.001,
        threshold_input - 0.001,
        threshold_input,
        threshold_input + 0.001,
        234.999,
        235.0,
        235.001,
    ];
    assert_normalize_luma_matches_all_backends(&input);
}