vyre-primitives 0.6.3

Compositional primitives for vyre - marker types (always on) + Tier 2.5 LEGO substrate (feature-gated per domain).
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

use super::*;

#[test]
fn matvec_cpu_matches_dequantized_reference() {
    let weights = vec![
        vec![-8, -4, -1, 0, 1, 2, 6, 7, 5],
        vec![7, 5, 3, 1, -1, -3, -5, -7, 6],
        vec![0, 1, 0, -1, 2, -2, 3, -3, 4],
    ];
    let x = [0.5_f32, -1.0, 2.0, 0.25, -0.5, 1.5, -2.0, 3.0, 0.75];
    let scales = [0.125_f32, 0.25, 0.5];
    let packed = pack_i4_matrix_rows(&weights);
    let actual = i4x8_matvec_f32_scaled_cpu(&packed, &x, &scales, 3, 9);
    let expected = weights
        .iter()
        .zip(scales)
        .map(|(row, scale)| {
            row.iter()
                .zip(x)
                .fold(0.0_f32, |acc, (&w, x)| acc + w as f32 * x)
                * scale
        })
        .collect::<Vec<_>>();

    assert_eq!(
        actual.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
        expected.iter().map(|v| v.to_bits()).collect::<Vec<_>>()
    );
}

#[test]
fn generated_matvec_matches_dequantized_reference_across_pack_boundaries() {
    let pattern = [-8, -3, -1, 0, 1, 3, 7, 6, 5, 4, 2, -2, -4, -6, -7, -5];
    for rows in 1..=8_u32 {
        for cols in [1_u32, 7, 8, 9, 16, 17, 31, 32, 33, 65] {
            let weights = (0..rows as usize)
                .map(|row| {
                    pattern
                        .iter()
                        .copied()
                        .cycle()
                        .skip(row)
                        .take(cols as usize)
                        .collect::<Vec<_>>()
                })
                .collect::<Vec<_>>();
            let x = (0..cols)
                .map(|col| (col % 11) as f32 * 0.125 - 0.5)
                .collect::<Vec<_>>();
            let scales = (0..rows)
                .map(|row| 0.125_f32 + row as f32 * 0.0625)
                .collect::<Vec<_>>();
            let packed = pack_i4_matrix_rows(&weights);
            let actual = i4x8_matvec_f32_scaled_cpu(&packed, &x, &scales, rows, cols);
            let expected = weights
                .iter()
                .zip(scales.iter().copied())
                .map(|(row, scale)| {
                    row.iter()
                        .zip(x.iter().copied())
                        .fold(0.0_f32, |acc, (&w, x)| acc + w as f32 * x)
                        * scale
                })
                .collect::<Vec<_>>();

            assert_eq!(
                actual.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
                expected.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
                "rows={rows} cols={cols}"
            );
        }
    }
}

#[test]
fn batched_matvec_cpu_matches_repeated_matvec_reference() {
    let weights = vec![
        vec![1, 2, 3, 4, -1, -2, -3, -4, 5],
        vec![4, 3, 2, 1, -4, -3, -2, -1, -5],
        vec![-8, -7, -6, -5, -4, -3, -2, -1, 0],
    ];
    let x_batches = [
        1.0_f32, -0.5, 0.25, 2.0, -1.5, 0.75, 1.25, -2.0, 0.5, -1.0, 0.5, -0.25, -2.0, 1.5, -0.75,
        -1.25, 2.0, -0.5,
    ];
    let scales = [0.5_f32, 0.25, 0.125];
    let packed = pack_i4_matrix_rows(&weights);
    let actual = i4x8_batched_matvec_f32_scaled_cpu(&packed, &x_batches, &scales, 2, 3, 9);
    let mut expected = Vec::new();
    for x in x_batches.chunks_exact(9) {
        expected.extend(i4x8_matvec_f32_scaled_cpu(&packed, x, &scales, 3, 9));
    }

    assert_eq!(
        actual.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
        expected.iter().map(|v| v.to_bits()).collect::<Vec<_>>()
    );
}

#[test]
fn generated_batched_matvec_matches_repeated_matvec_across_pack_boundaries() {
    let pattern = [-8, -3, -1, 0, 1, 3, 7, 6, 5, 4, 2, -2, -4, -6, -7, -5];
    for batch in 1..=4_u32 {
        for rows in 1..=5_u32 {
            for cols in [1_u32, 7, 8, 9, 16, 17, 31, 32, 33] {
                let weights = (0..rows as usize)
                    .map(|row| {
                        pattern
                            .iter()
                            .copied()
                            .cycle()
                            .skip(row * 2)
                            .take(cols as usize)
                            .collect::<Vec<_>>()
                    })
                    .collect::<Vec<_>>();
                let x_batches = (0..batch * cols)
                    .map(|index| (index % 13) as f32 * 0.125 - 0.75)
                    .collect::<Vec<_>>();
                let scales = (0..rows)
                    .map(|row| 0.125_f32 + row as f32 * 0.0625)
                    .collect::<Vec<_>>();
                let packed = pack_i4_matrix_rows(&weights);
                let actual = i4x8_batched_matvec_f32_scaled_cpu(
                    &packed, &x_batches, &scales, batch, rows, cols,
                );
                let mut expected = Vec::new();
                for x in x_batches.chunks_exact(cols as usize) {
                    expected.extend(i4x8_matvec_f32_scaled_cpu(&packed, x, &scales, rows, cols));
                }

                assert_eq!(
                    actual.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
                    expected.iter().map(|v| v.to_bits()).collect::<Vec<_>>(),
                    "batch={batch} rows={rows} cols={cols}"
                );
            }
        }
    }
}