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

//! Generated GEMM microkernels via trueno-gemm-codegen proc macro.
//!
//! Contract: cgp-gemm-codegen-v1.yaml (C-CODEGEN-001 through C-CODEGEN-004)
//! Sovereign: all code generated at compile time from trueno's own proc macro.
//!
//! These kernels are shape-specialized at compile time, producing fully-unrolled
//! FMA code with optimal register allocation for each (MR, NR) combination.

use trueno_gemm_codegen::{avx512_microkernel, avx512_microkernel_broadcast_b};

// === Broadcast-A kernels (A scalar broadcast, B zmm vector load) ===

// Generate the same 8x32 shape as the hand-written kernel for validation.
// C-CODEGEN-001: must match hand-written output within 1e-5.
// C-CODEGEN-002: must not be slower than hand-written (within 5%).
avx512_microkernel!(mr = 8, nr = 32);

// Generate 8x16 for small-N path validation.
avx512_microkernel!(mr = 8, nr = 16);

// New shapes not previously hand-written — explore register space.
// 8x48: 24 accumulators (8*3 zmm) + 3 B loads = 27 zmm. Fits in 32.
avx512_microkernel!(mr = 8, nr = 48);

// === Broadcast-B kernels (faer-style: A zmm vector load, B scalar broadcast) ===
// Advantage: small NR → tiny B panel → large KC → less packing overhead.
// MR must be multiple of 16 (zmm width).

// 32×6: 2 A-loads × 6 B-broadcasts = 12 accumulators + 2 A + 4 headroom = 18 zmm.
avx512_microkernel_broadcast_b!(mr = 32, nr = 6);

// 48×6: 3 A-loads × 6 B-broadcasts = 18 accumulators + 3 A + 4 headroom = 25 zmm.
avx512_microkernel_broadcast_b!(mr = 48, nr = 6);

// 64×6: 4 A-loads × 6 B-broadcasts = 24 accumulators + 4 A + 4 headroom = 32 zmm.
// Matches faer's register utilization. Maximum tile that fits in zmm register file.
avx512_microkernel_broadcast_b!(mr = 64, nr = 6);

#[cfg(test)]
mod tests {
    use super::*;

    /// Helper: scalar reference GEMM for validation.
    fn gemm_reference(m: usize, n: usize, k: usize, a: &[f32], b: &[f32], c: &mut [f32]) {
        for i in 0..m {
            for j in 0..n {
                for p in 0..k {
                    c[i * n + j] += a[p * m + i] * b[p * n + j];
                }
            }
        }
    }

    /// FALSIFY-CODEGEN-001: Generated 8x32 matches scalar reference.
    #[test]
    fn test_codegen_8x32_correctness() {
        let mr = 8;
        let nr = 32;
        let k = 64;

        let a: Vec<f32> = (0..mr * k).map(|i| ((i * 7 + 3) % 100) as f32 / 100.0).collect();
        let b: Vec<f32> = (0..k * nr).map(|i| ((i * 11 + 5) % 100) as f32 / 100.0).collect();
        let mut c_gen = vec![0.0f32; mr * nr];
        let mut c_ref = vec![0.0f32; mr * nr];

        // Reference
        gemm_reference(mr, nr, k, &a, &b, &mut c_ref);

        // Generated kernel
        // SAFETY: AVX-512 is available (test runs on x86_64 with avx512f)
        if std::arch::is_x86_feature_detected!("avx512f") {
            unsafe {
                microkernel_8x32_avx512_gen(k, a.as_ptr(), b.as_ptr(), c_gen.as_mut_ptr(), nr);
            }

            let max_diff =
                c_gen.iter().zip(c_ref.iter()).map(|(g, r)| (g - r).abs()).fold(0.0f32, f32::max);

            assert!(max_diff < 1e-2, "C-CODEGEN-001: max diff {max_diff} >= 1e-2 for 8x32");
        }
    }

    /// FALSIFY-CODEGEN-001b: Generated 8x16 matches scalar reference.
    #[test]
    fn test_codegen_8x16_correctness() {
        let mr = 8;
        let nr = 16;
        let k = 64;

        let a: Vec<f32> = (0..mr * k).map(|i| ((i * 7 + 3) % 100) as f32 / 100.0).collect();
        let b: Vec<f32> = (0..k * nr).map(|i| ((i * 11 + 5) % 100) as f32 / 100.0).collect();
        let mut c_gen = vec![0.0f32; mr * nr];
        let mut c_ref = vec![0.0f32; mr * nr];

        gemm_reference(mr, nr, k, &a, &b, &mut c_ref);

        if std::arch::is_x86_feature_detected!("avx512f") {
            unsafe {
                microkernel_8x16_avx512_gen(k, a.as_ptr(), b.as_ptr(), c_gen.as_mut_ptr(), nr);
            }

            let max_diff =
                c_gen.iter().zip(c_ref.iter()).map(|(g, r)| (g - r).abs()).fold(0.0f32, f32::max);

            assert!(max_diff < 1e-2, "C-CODEGEN-001: max diff {max_diff} >= 1e-2 for 8x16");
        }
    }

    /// FALSIFY-CODEGEN-001c: Generated 8x48 (new shape) matches scalar reference.
    #[test]
    fn test_codegen_8x48_correctness() {
        let mr = 8;
        let nr = 48;
        let k = 32;

        let a: Vec<f32> = (0..mr * k).map(|i| ((i * 7 + 3) % 100) as f32 / 100.0).collect();
        let b: Vec<f32> = (0..k * nr).map(|i| ((i * 11 + 5) % 100) as f32 / 100.0).collect();
        let mut c_gen = vec![0.0f32; mr * nr];
        let mut c_ref = vec![0.0f32; mr * nr];

        gemm_reference(mr, nr, k, &a, &b, &mut c_ref);

        if std::arch::is_x86_feature_detected!("avx512f") {
            unsafe {
                microkernel_8x48_avx512_gen(k, a.as_ptr(), b.as_ptr(), c_gen.as_mut_ptr(), nr);
            }

            let max_diff =
                c_gen.iter().zip(c_ref.iter()).map(|(g, r)| (g - r).abs()).fold(0.0f32, f32::max);

            assert!(max_diff < 1e-2, "C-CODEGEN-001: max diff {max_diff} >= 1e-2 for 8x48");
        }
    }

    // === Broadcast-B kernel tests ===

    /// FALSIFY-CODEGEN-002a: broadcast-B 32×6 matches scalar reference.
    #[test]
    fn test_codegen_bcast_b_32x6_correctness() {
        let mr = 32;
        let nr = 6;
        let k = 64;

        let a: Vec<f32> = (0..mr * k).map(|i| ((i * 7 + 3) % 100) as f32 / 100.0).collect();
        let b: Vec<f32> = (0..k * nr).map(|i| ((i * 11 + 5) % 100) as f32 / 100.0).collect();
        let mut c_gen = vec![0.0f32; mr * nr];
        let mut c_ref = vec![0.0f32; mr * nr];

        gemm_reference(mr, nr, k, &a, &b, &mut c_ref);

        if std::arch::is_x86_feature_detected!("avx512f") {
            unsafe {
                microkernel_32x6_avx512_bcast_b(k, a.as_ptr(), b.as_ptr(), c_gen.as_mut_ptr(), nr);
            }

            let max_diff =
                c_gen.iter().zip(c_ref.iter()).map(|(g, r)| (g - r).abs()).fold(0.0f32, f32::max);

            assert!(max_diff < 1e-2, "FALSIFY-CODEGEN-002a: max diff {max_diff} >= 1e-2");
        }
    }

    /// FALSIFY-CODEGEN-002b: broadcast-B 48×6 matches scalar reference.
    #[test]
    fn test_codegen_bcast_b_48x6_correctness() {
        let mr = 48;
        let nr = 6;
        let k = 64;

        let a: Vec<f32> = (0..mr * k).map(|i| ((i * 7 + 3) % 100) as f32 / 100.0).collect();
        let b: Vec<f32> = (0..k * nr).map(|i| ((i * 11 + 5) % 100) as f32 / 100.0).collect();
        let mut c_gen = vec![0.0f32; mr * nr];
        let mut c_ref = vec![0.0f32; mr * nr];

        gemm_reference(mr, nr, k, &a, &b, &mut c_ref);

        if std::arch::is_x86_feature_detected!("avx512f") {
            unsafe {
                microkernel_48x6_avx512_bcast_b(k, a.as_ptr(), b.as_ptr(), c_gen.as_mut_ptr(), nr);
            }

            let max_diff =
                c_gen.iter().zip(c_ref.iter()).map(|(g, r)| (g - r).abs()).fold(0.0f32, f32::max);

            assert!(max_diff < 1e-2, "FALSIFY-CODEGEN-002b: max diff {max_diff} >= 1e-2");
        }
    }

    /// FALSIFY-CODEGEN-002c: broadcast-B 64×6 matches scalar reference.
    #[test]
    fn test_codegen_bcast_b_64x6_correctness() {
        let mr = 64;
        let nr = 6;
        let k = 32;

        let a: Vec<f32> = (0..mr * k).map(|i| ((i * 7 + 3) % 100) as f32 / 100.0).collect();
        let b: Vec<f32> = (0..k * nr).map(|i| ((i * 11 + 5) % 100) as f32 / 100.0).collect();
        let mut c_gen = vec![0.0f32; mr * nr];
        let mut c_ref = vec![0.0f32; mr * nr];

        gemm_reference(mr, nr, k, &a, &b, &mut c_ref);

        if std::arch::is_x86_feature_detected!("avx512f") {
            unsafe {
                microkernel_64x6_avx512_bcast_b(k, a.as_ptr(), b.as_ptr(), c_gen.as_mut_ptr(), nr);
            }

            let max_diff =
                c_gen.iter().zip(c_ref.iter()).map(|(g, r)| (g - r).abs()).fold(0.0f32, f32::max);

            assert!(max_diff < 1e-2, "FALSIFY-CODEGEN-002c: max diff {max_diff} >= 1e-2");
        }
    }
}