trueno 0.17.5

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
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

//! AVX-512 arithmetic operations (add, sub, mul, div).
//!
//! For large vectors (≥8192 elements), uses non-temporal stores to bypass
//! cache pollution and 4-way unrolling for instruction-level parallelism.
//! Based on: Drepper (2007) "What Every Programmer Should Know About Memory"
//! Section 6.1: non-temporal stores eliminate read-for-ownership traffic.

#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;

/// Threshold above which non-temporal stores are beneficial.
/// Below this, data fits in L2 and cache-through stores are faster.
const NT_THRESHOLD: usize = 8192;

/// AVX-512 vector addition.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn add(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        if len >= NT_THRESHOLD {
            add_nt(a, b, result);
        } else {
            add_cached(a, b, result);
        }
    }
}

/// Cached-store path for small vectors (fits in L2).
#[inline]
#[target_feature(enable = "avx512f")]
unsafe fn add_cached(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        let mut i = 0;
        while i + 16 <= len {
            let va = _mm512_loadu_ps(a.as_ptr().add(i));
            let vb = _mm512_loadu_ps(b.as_ptr().add(i));
            _mm512_storeu_ps(result.as_mut_ptr().add(i), _mm512_add_ps(va, vb));
            i += 16;
        }
        for j in i..len {
            result[j] = a[j] + b[j];
        }
    }
}

/// Non-temporal store path for large vectors.
/// 4-way unrolled (64 f32 = 256 bytes per iteration = 4 cache lines).
/// Prefetches 8 cache lines ahead (~512 bytes).
#[inline]
#[target_feature(enable = "avx512f")]
unsafe fn add_nt(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        let ap = a.as_ptr();
        let bp = b.as_ptr();
        let rp = result.as_mut_ptr();
        let mut i = 0;

        // 4-way unrolled non-temporal loop
        while i + 64 <= len {
            // Prefetch 8 cache lines ahead (512 bytes = 128 f32)
            _mm_prefetch(ap.add(i + 128).cast::<i8>(), _MM_HINT_T0);
            _mm_prefetch(bp.add(i + 128).cast::<i8>(), _MM_HINT_T0);

            let va0 = _mm512_loadu_ps(ap.add(i));
            let vb0 = _mm512_loadu_ps(bp.add(i));
            let va1 = _mm512_loadu_ps(ap.add(i + 16));
            let vb1 = _mm512_loadu_ps(bp.add(i + 16));
            let va2 = _mm512_loadu_ps(ap.add(i + 32));
            let vb2 = _mm512_loadu_ps(bp.add(i + 32));
            let va3 = _mm512_loadu_ps(ap.add(i + 48));
            let vb3 = _mm512_loadu_ps(bp.add(i + 48));

            _mm512_stream_ps(rp.add(i), _mm512_add_ps(va0, vb0));
            _mm512_stream_ps(rp.add(i + 16), _mm512_add_ps(va1, vb1));
            _mm512_stream_ps(rp.add(i + 32), _mm512_add_ps(va2, vb2));
            _mm512_stream_ps(rp.add(i + 48), _mm512_add_ps(va3, vb3));

            i += 64;
        }

        // Cleanup: remaining full SIMD widths
        while i + 16 <= len {
            let va = _mm512_loadu_ps(ap.add(i));
            let vb = _mm512_loadu_ps(bp.add(i));
            _mm512_stream_ps(rp.add(i), _mm512_add_ps(va, vb));
            i += 16;
        }

        // Memory fence after non-temporal stores
        _mm_sfence();

        // Scalar remainder
        for j in i..len {
            result[j] = a[j] + b[j];
        }
    }
}

/// AVX-512 vector subtraction.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn sub(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        if len >= NT_THRESHOLD {
            sub_nt(a, b, result);
        } else {
            sub_cached(a, b, result);
        }
    }
}

#[inline]
#[target_feature(enable = "avx512f")]
unsafe fn sub_cached(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        let mut i = 0;
        while i + 16 <= len {
            let va = _mm512_loadu_ps(a.as_ptr().add(i));
            let vb = _mm512_loadu_ps(b.as_ptr().add(i));
            _mm512_storeu_ps(result.as_mut_ptr().add(i), _mm512_sub_ps(va, vb));
            i += 16;
        }
        for j in i..len {
            result[j] = a[j] - b[j];
        }
    }
}

#[inline]
#[target_feature(enable = "avx512f")]
unsafe fn sub_nt(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        let ap = a.as_ptr();
        let bp = b.as_ptr();
        let rp = result.as_mut_ptr();
        let mut i = 0;

        while i + 64 <= len {
            _mm_prefetch(ap.add(i + 128).cast::<i8>(), _MM_HINT_T0);
            _mm_prefetch(bp.add(i + 128).cast::<i8>(), _MM_HINT_T0);

            let va0 = _mm512_loadu_ps(ap.add(i));
            let vb0 = _mm512_loadu_ps(bp.add(i));
            let va1 = _mm512_loadu_ps(ap.add(i + 16));
            let vb1 = _mm512_loadu_ps(bp.add(i + 16));
            let va2 = _mm512_loadu_ps(ap.add(i + 32));
            let vb2 = _mm512_loadu_ps(bp.add(i + 32));
            let va3 = _mm512_loadu_ps(ap.add(i + 48));
            let vb3 = _mm512_loadu_ps(bp.add(i + 48));

            _mm512_stream_ps(rp.add(i), _mm512_sub_ps(va0, vb0));
            _mm512_stream_ps(rp.add(i + 16), _mm512_sub_ps(va1, vb1));
            _mm512_stream_ps(rp.add(i + 32), _mm512_sub_ps(va2, vb2));
            _mm512_stream_ps(rp.add(i + 48), _mm512_sub_ps(va3, vb3));

            i += 64;
        }

        while i + 16 <= len {
            let va = _mm512_loadu_ps(ap.add(i));
            let vb = _mm512_loadu_ps(bp.add(i));
            _mm512_stream_ps(rp.add(i), _mm512_sub_ps(va, vb));
            i += 16;
        }

        _mm_sfence();

        for j in i..len {
            result[j] = a[j] - b[j];
        }
    }
}

/// AVX-512 vector multiplication.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn mul(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        if len >= NT_THRESHOLD {
            mul_nt(a, b, result);
        } else {
            mul_cached(a, b, result);
        }
    }
}

#[inline]
#[target_feature(enable = "avx512f")]
unsafe fn mul_cached(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        let mut i = 0;
        while i + 16 <= len {
            let va = _mm512_loadu_ps(a.as_ptr().add(i));
            let vb = _mm512_loadu_ps(b.as_ptr().add(i));
            _mm512_storeu_ps(result.as_mut_ptr().add(i), _mm512_mul_ps(va, vb));
            i += 16;
        }
        for j in i..len {
            result[j] = a[j] * b[j];
        }
    }
}

#[inline]
#[target_feature(enable = "avx512f")]
unsafe fn mul_nt(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        let ap = a.as_ptr();
        let bp = b.as_ptr();
        let rp = result.as_mut_ptr();
        let mut i = 0;

        while i + 64 <= len {
            _mm_prefetch(ap.add(i + 128).cast::<i8>(), _MM_HINT_T0);
            _mm_prefetch(bp.add(i + 128).cast::<i8>(), _MM_HINT_T0);

            let va0 = _mm512_loadu_ps(ap.add(i));
            let vb0 = _mm512_loadu_ps(bp.add(i));
            let va1 = _mm512_loadu_ps(ap.add(i + 16));
            let vb1 = _mm512_loadu_ps(bp.add(i + 16));
            let va2 = _mm512_loadu_ps(ap.add(i + 32));
            let vb2 = _mm512_loadu_ps(bp.add(i + 32));
            let va3 = _mm512_loadu_ps(ap.add(i + 48));
            let vb3 = _mm512_loadu_ps(bp.add(i + 48));

            _mm512_stream_ps(rp.add(i), _mm512_mul_ps(va0, vb0));
            _mm512_stream_ps(rp.add(i + 16), _mm512_mul_ps(va1, vb1));
            _mm512_stream_ps(rp.add(i + 32), _mm512_mul_ps(va2, vb2));
            _mm512_stream_ps(rp.add(i + 48), _mm512_mul_ps(va3, vb3));

            i += 64;
        }

        while i + 16 <= len {
            let va = _mm512_loadu_ps(ap.add(i));
            let vb = _mm512_loadu_ps(bp.add(i));
            _mm512_stream_ps(rp.add(i), _mm512_mul_ps(va, vb));
            i += 16;
        }

        _mm_sfence();

        for j in i..len {
            result[j] = a[j] * b[j];
        }
    }
}

/// AVX-512 vector division.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn div(a: &[f32], b: &[f32], result: &mut [f32]) {
    unsafe {
        let len = a.len();
        let mut i = 0;
        while i + 16 <= len {
            let va = _mm512_loadu_ps(a.as_ptr().add(i));
            let vb = _mm512_loadu_ps(b.as_ptr().add(i));
            _mm512_storeu_ps(result.as_mut_ptr().add(i), _mm512_div_ps(va, vb));
            i += 16;
        }
        for j in i..len {
            result[j] = a[j] / b[j];
        }
    }
}