trueno 0.18.0

High-performance SIMD compute library with GPU support, LLM inference engine, and GGUF model loading
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
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360

//! FALSIFY tests for uninit allocation correctness (CGP-DBUF).
//!
//! These tests verify that the `Vec::with_capacity + set_len` pattern
//! produces identical results to the original `vec![0.0; n]` pattern.
//!
//! If any test fails, the uninit optimization is UNSOUND and must be reverted.

use crate::{Matrix, Vector};

// ========================================================================
// FALSIFY-UNINIT-001: Vector::sqrt uninit matches zero-init
// ========================================================================

#[test]
fn falsify_uninit_001_sqrt_correctness() {
    let data = vec![0.0, 1.0, 4.0, 9.0, 16.0, 0.25, 100.0, f32::INFINITY];
    let v = Vector::from_slice(&data);
    let result = v.sqrt().unwrap();

    let expected: Vec<f32> = data.iter().map(|x| x.sqrt()).collect();
    assert_eq!(result.as_slice(), &expected[..], "FALSIFY-UNINIT-001: sqrt output mismatch");
}

#[test]
fn falsify_uninit_001b_sqrt_large() {
    let n = 200_000;
    let data: Vec<f32> = (0..n).map(|i| (i as f32) * 0.01).collect();
    let v = Vector::from_slice(&data);
    let result = v.sqrt().unwrap();

    for (i, (&got, &x)) in result.as_slice().iter().zip(data.iter()).enumerate() {
        let expected = x.sqrt();
        assert!(
            (got - expected).abs() < 1e-6 || (got.is_nan() && expected.is_nan()),
            "FALSIFY-UNINIT-001b: sqrt[{i}] got {got}, expected {expected}"
        );
    }
}

// ========================================================================
// FALSIFY-UNINIT-002: Vector::recip uninit matches zero-init
// ========================================================================

#[test]
fn falsify_uninit_002_recip_correctness() {
    let data = vec![1.0, 2.0, 4.0, 0.5, 10.0, 0.1, -3.0];
    let v = Vector::from_slice(&data);
    let result = v.recip().unwrap();

    let expected: Vec<f32> = data.iter().map(|x| 1.0 / x).collect();
    for (i, (&got, &exp)) in result.as_slice().iter().zip(expected.iter()).enumerate() {
        assert!(
            (got - exp).abs() < 1e-6,
            "FALSIFY-UNINIT-002: recip[{i}] got {got}, expected {exp}"
        );
    }
}

// ========================================================================
// FALSIFY-UNINIT-003: blis::softmax uninit matches reference
// ========================================================================

#[test]
fn falsify_uninit_003_softmax_sums_to_one() {
    let logits = vec![1.0, 2.0, 3.0, 4.0, 5.0];
    let result = crate::blis::softmax::softmax_1d_alloc(&logits);

    let sum: f32 = result.iter().sum();
    assert!((sum - 1.0).abs() < 1e-5, "FALSIFY-UNINIT-003: softmax sum={sum}");

    for (i, &v) in result.iter().enumerate() {
        assert!(v > 0.0, "FALSIFY-UNINIT-003: softmax[{i}]={v} <= 0");
    }

    for i in 1..result.len() {
        assert!(result[i] >= result[i - 1], "FALSIFY-UNINIT-003: not monotone at {i}");
    }
}

#[test]
fn falsify_uninit_003b_softmax_large_avx2() {
    // n=1024 exercises AVX2 path (chunks of 32)
    let n = 1024;
    let logits: Vec<f32> = (0..n).map(|i| (i as f32) * 0.01 - 5.0).collect();
    let result = crate::blis::softmax::softmax_1d_alloc(&logits);

    let sum: f32 = result.iter().sum();
    assert!((sum - 1.0).abs() < 1e-4, "FALSIFY-UNINIT-003b: sum={sum}");
    assert!(result.iter().all(|v| *v >= 0.0 && v.is_finite()));
}

#[test]
fn falsify_uninit_003c_softmax_deterministic() {
    let logits: Vec<f32> = (0..256).map(|i| (i as f32) * 0.1 - 12.8).collect();
    let r1 = crate::blis::softmax::softmax_1d_alloc(&logits);
    let r2 = crate::blis::softmax::softmax_1d_alloc(&logits);
    assert_eq!(r1, r2, "FALSIFY-UNINIT-003c: softmax non-deterministic");
}

// ========================================================================
// FALSIFY-UNINIT-004: Matrix::matvec uninit matches reference
// ========================================================================

#[test]
fn falsify_uninit_004_matvec_small() {
    let m = Matrix::from_vec(3, 2, vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]).unwrap();
    let v = Vector::from_slice(&[7.0, 8.0]);
    let result = m.matvec(&v).unwrap();
    assert_eq!(result.as_slice(), &[23.0, 53.0, 83.0], "FALSIFY-UNINIT-004");
}

#[test]
fn falsify_uninit_004b_matvec_large_parallel() {
    // >=4096 rows triggers parallel path
    let rows = 4096;
    let cols = 128;
    let data: Vec<f32> = (0..rows * cols).map(|i| ((i % 17) as f32) * 0.1).collect();
    let v_data: Vec<f32> = (0..cols).map(|i| ((i % 7) as f32) * 0.2).collect();

    let m = Matrix::from_vec(rows, cols, data.clone()).unwrap();
    let v = Vector::from_slice(&v_data);
    let result = m.matvec(&v).unwrap();

    for r in 0..rows {
        let mut expected = 0.0f32;
        for c in 0..cols {
            expected += data[r * cols + c] * v_data[c];
        }
        let got = result.as_slice()[r];
        assert!(
            (got - expected).abs() < expected.abs() * 1e-5 + 1e-6,
            "FALSIFY-UNINIT-004b: matvec[{r}] got={got}, expected={expected}"
        );
    }
}

// ========================================================================
// FALSIFY-UNINIT-005: Q4K GEMV uninit is deterministic
// ========================================================================

#[test]
fn falsify_uninit_005_q4k_gemv_deterministic() {
    use crate::backends::q4k::matmul_q4k_f32;

    let in_dim = 256;
    let out_dim = 4;
    let num_blocks_per_row = (in_dim + 255) / 256;
    let row_bytes = num_blocks_per_row * 144;
    let q4k_data = vec![0x55u8; out_dim * row_bytes];
    let input = vec![1.0f32; in_dim];

    let r1 = matmul_q4k_f32(&q4k_data, &input, out_dim, in_dim);
    let r2 = matmul_q4k_f32(&q4k_data, &input, out_dim, in_dim);

    assert_eq!(r1, r2, "FALSIFY-UNINIT-005: Q4K non-deterministic");
    for (i, &v) in r1.iter().enumerate() {
        assert!(v.is_finite(), "FALSIFY-UNINIT-005: Q4K output[{i}]={v}");
    }
}

// ========================================================================
// FALSIFY-UNINIT-006: AttentionOp output bounded by V range
// ========================================================================

#[test]
fn falsify_uninit_006_attention_bounded() {
    use crate::brick::{AttentionOp, ComputeBackend, ComputeOp};

    let seq_len = 4;
    let kv_seq_len = 4;
    let head_dim = 8;

    let op = AttentionOp::new(seq_len, kv_seq_len, head_dim);

    let q = vec![1.0f32; seq_len * head_dim];
    let k = vec![1.0f32; kv_seq_len * head_dim];
    let v: Vec<f32> = (0..kv_seq_len * head_dim).map(|i| (i as f32) * 0.1).collect();

    let backend = ComputeBackend::default();
    let output = op.execute((q, k, v.clone()), backend).unwrap();

    let v_min = v.iter().cloned().fold(f32::INFINITY, f32::min);
    let v_max = v.iter().cloned().fold(f32::NEG_INFINITY, f32::max);

    for (i, &o) in output.iter().enumerate() {
        assert!(
            o >= v_min - 1e-5 && o <= v_max + 1e-5,
            "FALSIFY-UNINIT-006: output[{i}]={o} outside [{v_min}, {v_max}]"
        );
    }
}

// ========================================================================
// FALSIFY-UNINIT-007: FusedQKV SIMD dot matches scalar reference
// ========================================================================

#[test]
fn falsify_uninit_007_fused_qkv_simd_vs_scalar() {
    use crate::brick::{ComputeBackend, ComputeOp, FusedQKVOp, FusedQKVWeights};

    let hidden = 8;
    let num_heads = 2;
    let kv_heads = 1;
    let kv_dim = hidden / num_heads * kv_heads;

    let op = FusedQKVOp::new(hidden, num_heads, kv_heads);

    let w_q: Vec<f32> = (0..hidden * hidden).map(|i| ((i % 11) as f32) * 0.1 - 0.5).collect();
    let w_k: Vec<f32> = (0..kv_dim * hidden).map(|i| ((i % 7) as f32) * 0.1 - 0.3).collect();
    let w_v: Vec<f32> = (0..kv_dim * hidden).map(|i| ((i % 13) as f32) * 0.1 - 0.6).collect();
    let x: Vec<f32> = (0..hidden).map(|i| ((i % 5) as f32) * 0.2).collect();

    let weights =
        FusedQKVWeights { q_weight: w_q.clone(), k_weight: w_k.clone(), v_weight: w_v.clone() };
    let backend = ComputeBackend::default();

    let (q, k, v) = op.execute((x.clone(), weights), backend).unwrap();

    // Scalar reference
    for i in 0..hidden {
        let expected: f32 = (0..hidden).map(|j| x[j] * w_q[i * hidden + j]).sum();
        assert!(
            (q[i] - expected).abs() < 1e-4,
            "FALSIFY-UNINIT-007: Q[{i}] got={}, expected={expected}",
            q[i]
        );
    }
    for i in 0..kv_dim {
        let exp_k: f32 = (0..hidden).map(|j| x[j] * w_k[i * hidden + j]).sum();
        let exp_v: f32 = (0..hidden).map(|j| x[j] * w_v[i * hidden + j]).sum();
        assert!((k[i] - exp_k).abs() < 1e-4, "FALSIFY-UNINIT-007: K[{i}]");
        assert!((v[i] - exp_v).abs() < 1e-4, "FALSIFY-UNINIT-007: V[{i}]");
    }
}

// ========================================================================
// FALSIFY-SIMD-001: AVX2 axpy in attention matches scalar reference
// ========================================================================

#[test]
fn falsify_simd_001_axpy_correctness() {
    use crate::brick::AttentionOp;

    // head_dim=128 (typical LLM): exercises full AVX2 path (128/8=16 iterations)
    let n = 128;
    let x: Vec<f32> = (0..n).map(|i| (i as f32) * 0.1).collect();
    let alpha = 0.75f32;

    let mut out_simd = vec![1.0f32; n]; // non-zero initial values
    AttentionOp::simd_axpy(alpha, &x, &mut out_simd);

    // Scalar reference
    let mut out_ref = vec![1.0f32; n];
    for i in 0..n {
        out_ref[i] += alpha * x[i];
    }

    for i in 0..n {
        assert!(
            (out_simd[i] - out_ref[i]).abs() < 1e-5,
            "FALSIFY-SIMD-001: axpy[{i}] got={}, expected={}",
            out_simd[i],
            out_ref[i]
        );
    }
}

#[test]
fn falsify_simd_001b_axpy_remainder() {
    use crate::brick::AttentionOp;

    // head_dim=17: tests remainder path (17 % 8 = 1)
    let n = 17;
    let x: Vec<f32> = (0..n).map(|i| ((i * 3 + 1) as f32) * 0.1).collect();
    let alpha = -0.5f32;

    let mut out = vec![2.0f32; n];
    AttentionOp::simd_axpy(alpha, &x, &mut out);

    for i in 0..n {
        let expected = 2.0 + alpha * x[i];
        assert!(
            (out[i] - expected).abs() < 1e-5,
            "FALSIFY-SIMD-001b: axpy[{i}] got={}, expected={expected}",
            out[i]
        );
    }
}

// ========================================================================
// FALSIFY-PARALLEL-001: Parallel transpose matches serial at boundary
// ========================================================================

#[test]
fn falsify_parallel_001_transpose_boundary() {
    // 1000×1000 = 1M elements: right at PARALLEL_THRESHOLD=1M.
    // Both parallel and serial paths must produce identical results.
    let rows = 1000;
    let cols = 1000;
    let data: Vec<f32> = (0..rows * cols).map(|i| (i as f32) * 0.001).collect();

    let mut result = vec![0.0f32; rows * cols];
    crate::blis::transpose::transpose(rows, cols, &data, &mut result).unwrap();

    // Verify: result[j*rows+i] == data[i*cols+j]
    for i in 0..rows.min(100) {
        for j in 0..cols.min(100) {
            let expected = data[i * cols + j];
            let got = result[j * rows + i];
            assert!(
                (got - expected).abs() < 1e-6,
                "FALSIFY-PARALLEL-001: transpose[{j},{i}] got={got}, expected={expected}"
            );
        }
    }
}

// ========================================================================
// FALSIFY-PARALLEL-002: Parallel matvec matches serial at boundary
// ========================================================================

#[test]
fn falsify_parallel_002_matvec_boundary() {
    // 2048 rows: right at PARALLEL_THRESHOLD=2048 for matvec.
    let rows = 2048;
    let cols = 64;
    let data: Vec<f32> = (0..rows * cols).map(|i| ((i % 31) as f32) * 0.1).collect();
    let v_data: Vec<f32> = (0..cols).map(|i| ((i % 11) as f32) * 0.2).collect();

    let m = Matrix::from_vec(rows, cols, data.clone()).unwrap();
    let v = Vector::from_slice(&v_data);
    let result = m.matvec(&v).unwrap();

    // Scalar reference
    for r in 0..rows {
        let expected: f32 = (0..cols).map(|c| data[r * cols + c] * v_data[c]).sum();
        let got = result.as_slice()[r];
        assert!(
            (got - expected).abs() < expected.abs() * 1e-5 + 1e-5,
            "FALSIFY-PARALLEL-002: matvec[{r}] got={got}, expected={expected}"
        );
    }
}

// ========================================================================
// FALSIFY-PARALLEL-003: Parallel transpose deterministic
// ========================================================================

#[test]
fn falsify_parallel_003_transpose_deterministic() {
    let rows = 1024;
    let cols = 1024;
    let data: Vec<f32> = (0..rows * cols).map(|i| (i as f32) * 0.001).collect();

    let mut r1 = vec![0.0f32; rows * cols];
    let mut r2 = vec![0.0f32; rows * cols];
    crate::blis::transpose::transpose(rows, cols, &data, &mut r1).unwrap();
    crate::blis::transpose::transpose(rows, cols, &data, &mut r2).unwrap();
    assert_eq!(r1, r2, "FALSIFY-PARALLEL-003: transpose non-deterministic");
}