trueno 0.17.4

High-performance SIMD compute library with GPU support for matrix operations
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

use super::super::*;

// ================================================================
// Q4K AVX2 GEMV coverage - dispatch with various sizes
// ================================================================

/// Test dispatch with multiple super-blocks, exercising AVX2 inner loops
#[test]
fn test_q4k_dispatch_multi_superblock() {
    if !is_x86_feature_detected!("avx2") || !is_x86_feature_detected!("fma") {
        return;
    }

    // 4 super-blocks = 1024 elements
    let in_dim = 1024;
    let out_dim = 8;
    let num_blocks = in_dim / SUPER_BLOCK_SIZE;

    let mut q4k_data = Vec::new();
    for row in 0..out_dim {
        for sb in 0..num_blocks {
            q4k_data.extend_from_slice(&[0x66, 0x2E]); // d
            q4k_data.extend_from_slice(&[0x66, 0x2A]); // dmin
            let sv = ((row + sb) as u8 + 1) | (((row + sb) as u8 + 2) << 4);
            q4k_data.extend_from_slice(&[sv; 12]);
            for i in 0..128 {
                let low = ((row + sb + i) % 16) as u8;
                let high = ((row + sb + i + 3) % 16) as u8;
                q4k_data.push(low | (high << 4));
            }
        }
    }

    let input: Vec<f32> = (0..in_dim).map(|i| (i as f32) * 0.001 - 0.5).collect();

    let scalar = matmul_q4k_f32_scalar(&q4k_data, &input, out_dim, in_dim);
    let dispatch = matmul_q4k_f32_dispatch(&q4k_data, &input, out_dim, in_dim);

    for (i, (s, d)) in scalar.iter().zip(dispatch.iter()).enumerate() {
        let diff = (s - d).abs();
        let rel_diff = if s.abs() > 1e-6 { diff / s.abs() } else { diff };
        assert!(
            rel_diff < 1e-4 || diff < 1e-4,
            "Row {}: scalar={}, dispatch={}, diff={}, rel={}",
            i,
            s,
            d,
            diff,
            rel_diff
        );
    }
}

/// Test dispatch with 3 super-blocks (768 elements), exercising more inner loop iterations
#[test]
fn test_q4k_dispatch_three_superblocks() {
    if !is_x86_feature_detected!("avx2") || !is_x86_feature_detected!("fma") {
        return;
    }

    let in_dim = 768; // 3 super-blocks
    let out_dim = 2;
    let num_blocks = in_dim / SUPER_BLOCK_SIZE;

    let mut q4k_data = Vec::new();
    for row in 0..out_dim {
        for sb in 0..num_blocks {
            q4k_data.extend_from_slice(&[0x00, 0x3C]); // d ~ 1.0
            q4k_data.extend_from_slice(&[0x00, 0x38]); // dmin ~ 0.5
            let sv = ((row + sb) as u8 + 1) & 0x3F;
            q4k_data.extend_from_slice(&[sv; 12]);
            q4k_data.extend_from_slice(&[0x77u8; 128]);
        }
    }

    let input: Vec<f32> = (0..in_dim).map(|i| (i as f32) * 0.005).collect();

    let scalar = matmul_q4k_f32_scalar(&q4k_data, &input, out_dim, in_dim);
    let dispatch = matmul_q4k_f32_dispatch(&q4k_data, &input, out_dim, in_dim);

    for (i, (s, d)) in scalar.iter().zip(dispatch.iter()).enumerate() {
        let diff = (s - d).abs();
        let rel_diff = if s.abs() > 1e-6 { diff / s.abs() } else { diff };
        assert!(
            rel_diff < 1e-4 || diff < 1e-4,
            "Row {}: scalar={}, dispatch={}, diff={}",
            i,
            s,
            d,
            diff
        );
    }
}

/// Test dispatch with negative input values (exercises subtract branches)
#[test]
fn test_q4k_dispatch_negative_inputs() {
    if !is_x86_feature_detected!("avx2") || !is_x86_feature_detected!("fma") {
        return;
    }

    let in_dim = 256;
    let out_dim = 4;

    let mut q4k_data = Vec::new();
    for row in 0..out_dim {
        q4k_data.extend_from_slice(&[0x66, 0x2E]); // d ~ 0.1
        q4k_data.extend_from_slice(&[0x66, 0x2A]); // dmin ~ 0.05
        q4k_data.extend_from_slice(&[((row + 1) as u8); 12]);
        for i in 0..128 {
            q4k_data.push(((i * 3 + row) % 256) as u8);
        }
    }

    let input: Vec<f32> = (0..in_dim).map(|i| -1.0 + (i as f32) * 0.008).collect();

    let scalar = matmul_q4k_f32_scalar(&q4k_data, &input, out_dim, in_dim);
    let dispatch = matmul_q4k_f32_dispatch(&q4k_data, &input, out_dim, in_dim);

    for (i, (s, d)) in scalar.iter().zip(dispatch.iter()).enumerate() {
        let diff = (s - d).abs();
        let rel_diff = if s.abs() > 1e-6 { diff / s.abs() } else { diff };
        assert!(
            rel_diff < 1e-4 || diff < 1e-4,
            "Row {}: scalar={}, dispatch={}, diff={}",
            i,
            s,
            d,
            diff
        );
    }
}