trueno 0.17.1

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
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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246

//! Golden Vector Tests: Q6K scalar reference vs dispatch/SIMD paths

use super::super::*;

/// Golden Test: Q6K scalar == dispatch for random input
#[test]
fn test_golden_q6k_scalar_vs_dispatch() {
    // Realistic LLM dimensions
    let in_dim = 512; // 2 super-blocks
    let out_dim = 8;

    let mut q6k_data = Vec::new();
    for row in 0..out_dim {
        for sb in 0..(in_dim / 256) {
            // ql: varied 4-bit low values
            for i in 0..128 {
                let low = ((row + sb + i) % 16) as u8;
                let high = ((row + sb + i + 3) % 16) as u8;
                q6k_data.push(low | (high << 4));
            }
            // qh: varied 2-bit high values
            for i in 0..64 {
                let vals = [
                    ((row + i) % 4) as u8,
                    ((row + i + 1) % 4) as u8,
                    ((row + i + 2) % 4) as u8,
                    ((row + i + 3) % 4) as u8,
                ];
                q6k_data.push(vals[0] | (vals[1] << 2) | (vals[2] << 4) | (vals[3] << 6));
            }
            // scales: varied signed 8-bit
            for i in 0..16 {
                q6k_data.push(((row * 7 + sb * 3 + i) % 64) as u8);
            }
            // d ~ 0.1
            q6k_data.extend_from_slice(&[0x66, 0x2E]);
        }
    }

    // Sinusoidal input
    let input: Vec<f32> = (0..in_dim).map(|i| ((i as f32) * 0.019).sin() * 0.4).collect();

    let scalar_output = matmul_q6k_f32_scalar(&q6k_data, &input, out_dim, in_dim);
    let dispatch_output = matmul_q6k_f32_dispatch(&q6k_data, &input, out_dim, in_dim);

    assert_eq!(scalar_output.len(), dispatch_output.len());
    let mut max_abs_error = 0.0f32;

    for (i, (scalar, dispatch)) in scalar_output.iter().zip(dispatch_output.iter()).enumerate() {
        let abs_error = (scalar - dispatch).abs();
        max_abs_error = max_abs_error.max(abs_error);

        // Scalar and dispatch should match closely (minor FMA ordering differences)
        assert!(
            abs_error < 2e-4,
            "Row {}: scalar={}, dispatch={}, diff={}",
            i,
            scalar,
            dispatch,
            abs_error
        );
    }

    eprintln!("[Golden Q6K Test] max_abs_error={:.6}", max_abs_error);
}

/// Golden Test: Q6K colmajor path consistency
#[test]
#[allow(deprecated)]
fn test_golden_q6k_colmajor_consistency() {
    let in_dim = 512;
    let out_dim = 4;

    let mut q6k_data = Vec::new();
    for row in 0..out_dim {
        for sb in 0..2 {
            // ql
            for i in 0..128 {
                q6k_data.push(((row * 5 + sb * 13 + i) % 256) as u8);
            }
            // qh
            for i in 0..64 {
                q6k_data.push(((row * 7 + sb * 11 + i * 2) % 256) as u8);
            }
            // scales
            for i in 0..16 {
                q6k_data.push(((row + sb + i) % 128) as u8);
            }
            // d ~ 0.5
            q6k_data.extend_from_slice(&[0x00, 0x38]);
        }
    }

    let input: Vec<f32> = (0..in_dim).map(|i| ((i as f32) * 0.011 + 0.3).cos() * 0.5).collect();

    let colmajor_output = matmul_q6k_f32_colmajor(&q6k_data, &input, out_dim, in_dim);
    let colmajor_dispatch = matmul_q6k_f32_colmajor_dispatch(&q6k_data, &input, out_dim, in_dim);

    assert_eq!(colmajor_output.len(), colmajor_dispatch.len());
    for (i, (base, dispatch)) in colmajor_output.iter().zip(colmajor_dispatch.iter()).enumerate() {
        let diff = (base - dispatch).abs();
        assert!(
            diff < 1e-4,
            "Row {}: colmajor base={}, dispatch={}, diff={}",
            i,
            base,
            dispatch,
            diff
        );
    }
}

/// Edge case: maximum 6-bit values (63)
#[test]
fn test_golden_q6k_max_quant_values() {
    let in_dim = 256;
    let out_dim = 2;

    let mut q6k_data = Vec::new();
    for _ in 0..out_dim {
        // ql: all 0xF (low nibble = 15)
        q6k_data.extend_from_slice(&[0xFFu8; 128]);
        // qh: all 0xFF (all 2-bit high = 3), so value = 15 + 3*16 = 63
        q6k_data.extend_from_slice(&[0xFFu8; 64]);
        // scales: positive
        q6k_data.extend_from_slice(&[0x3Fu8; 16]); // scale = 63
                                                   // d = 1.0
        q6k_data.extend_from_slice(&[0x00, 0x3C]);
    }

    let input = vec![1.0f32; in_dim];
    let scalar_output = matmul_q6k_f32_scalar(&q6k_data, &input, out_dim, in_dim);
    let dispatch_output = matmul_q6k_f32_dispatch(&q6k_data, &input, out_dim, in_dim);

    for (i, (scalar, dispatch)) in scalar_output.iter().zip(dispatch_output.iter()).enumerate() {
        assert!(
            scalar.is_finite() && dispatch.is_finite(),
            "Row {}: max values should produce finite output",
            i
        );
        let diff = (scalar - dispatch).abs();
        assert!(
            diff < 1e-4,
            "Row {}: max quant scalar={}, dispatch={}, diff={}",
            i,
            scalar,
            dispatch,
            diff
        );
    }
}

/// Edge case: alternating positive/negative scales
#[test]
fn test_golden_q6k_alternating_scales() {
    let in_dim = 256;
    let out_dim = 2;

    let mut q6k_data = Vec::new();
    for _ in 0..out_dim {
        // ql: mid-range values
        q6k_data.extend_from_slice(&[0x77u8; 128]); // 7, 7 repeated
                                                    // qh: zeros (full value = 7)
        q6k_data.extend_from_slice(&[0x00u8; 64]);
        // scales: alternating +32, -32
        for i in 0..16 {
            if i % 2 == 0 {
                q6k_data.push(0x20); // +32
            } else {
                q6k_data.push(0xE0); // -32 (as signed i8)
            }
        }
        // d = 0.5
        q6k_data.extend_from_slice(&[0x00, 0x38]);
    }

    let input = vec![1.0f32; in_dim];
    let scalar_output = matmul_q6k_f32_scalar(&q6k_data, &input, out_dim, in_dim);
    let dispatch_output = matmul_q6k_f32_dispatch(&q6k_data, &input, out_dim, in_dim);

    for (i, (scalar, dispatch)) in scalar_output.iter().zip(dispatch_output.iter()).enumerate() {
        let diff = (scalar - dispatch).abs();
        assert!(
            diff < 1e-4,
            "Row {}: alternating scales scalar={}, dispatch={}, diff={}",
            i,
            scalar,
            dispatch,
            diff
        );
    }
}

/// Large scale test for SIMD path coverage
#[cfg(target_arch = "x86_64")]
#[test]
fn test_golden_q6k_large_simd() {
    if !is_x86_feature_detected!("avx2") || !is_x86_feature_detected!("fma") {
        eprintln!("Skipping Q6K SIMD test - no AVX2+FMA");
        return;
    }

    let in_dim = 2048; // 8 super-blocks
    let out_dim = 32;

    let mut q6k_data = Vec::new();
    for row in 0..out_dim {
        for sb in 0..(in_dim / 256) {
            for i in 0..128 {
                let val = ((row * 3 + sb * 7 + i) % 256) as u8;
                q6k_data.push(val);
            }
            for i in 0..64 {
                let val = ((row * 5 + sb * 11 + i * 2) % 256) as u8;
                q6k_data.push(val);
            }
            for i in 0..16 {
                q6k_data.push(((row + sb + i) % 64) as u8);
            }
            q6k_data.extend_from_slice(&[0x66, 0x2E]);
        }
    }

    let input: Vec<f32> = (0..in_dim).map(|i| ((i as f32) * 0.007 - 1.0).tanh()).collect();

    let scalar_output = matmul_q6k_f32_scalar(&q6k_data, &input, out_dim, in_dim);
    let dispatch_output = matmul_q6k_f32_dispatch(&q6k_data, &input, out_dim, in_dim);

    let mut max_rel_error = 0.0f32;
    for (i, (scalar, dispatch)) in scalar_output.iter().zip(dispatch_output.iter()).enumerate() {
        let abs_error = (scalar - dispatch).abs();
        let rel_error = if scalar.abs() > 1e-6 { abs_error / scalar.abs() } else { abs_error };
        max_rel_error = max_rel_error.max(rel_error);

        assert!(
            rel_error < 1e-4 || abs_error < 1e-4,
            "Row {}: large SIMD scalar={}, dispatch={}, rel_err={:.6}",
            i,
            scalar,
            dispatch,
            rel_error
        );
    }

    eprintln!("[Golden Q6K Large SIMD] max_rel_error={:.6}", max_rel_error);
}