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

//! AVX-512 reduction operations (dot, sum, max, min, argmax, argmin).

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

/// AVX-512 dot product.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn dot(a: &[f32], b: &[f32]) -> f32 {
    unsafe {
        let len = a.len();
        let mut i = 0;
        let mut acc = _mm512_setzero_ps();

        while i + 16 <= len {
            let va = _mm512_loadu_ps(a.as_ptr().add(i));
            let vb = _mm512_loadu_ps(b.as_ptr().add(i));
            acc = _mm512_fmadd_ps(va, vb, acc);
            i += 16;
        }

        let mut result = _mm512_reduce_add_ps(acc);
        result += a[i..].iter().zip(&b[i..]).map(|(x, y)| x * y).sum::<f32>();
        result
    }
}

/// AVX-512 vector sum.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn sum(a: &[f32]) -> f32 {
    unsafe {
        let len = a.len();
        let mut i = 0;
        let mut acc = _mm512_setzero_ps();

        while i + 16 <= len {
            acc = _mm512_add_ps(acc, _mm512_loadu_ps(a.as_ptr().add(i)));
            i += 16;
        }

        let mut result = _mm512_reduce_add_ps(acc);
        result += a[i..].iter().sum::<f32>();
        result
    }
}

/// AVX-512 vector max.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn max(a: &[f32]) -> f32 {
    unsafe {
        let len = a.len();
        let mut i = 0;
        let mut vmax = _mm512_set1_ps(a[0]);

        while i + 16 <= len {
            vmax = _mm512_max_ps(vmax, _mm512_loadu_ps(a.as_ptr().add(i)));
            i += 16;
        }

        let mut result = _mm512_reduce_max_ps(vmax);
        for &val in &a[i..] {
            if val > result {
                result = val;
            }
        }
        result
    }
}

/// AVX-512 vector min.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn min(a: &[f32]) -> f32 {
    unsafe {
        let len = a.len();
        let mut i = 0;
        let mut vmin = _mm512_set1_ps(a[0]);

        while i + 16 <= len {
            vmin = _mm512_min_ps(vmin, _mm512_loadu_ps(a.as_ptr().add(i)));
            i += 16;
        }

        let mut result = _mm512_reduce_min_ps(vmin);
        for &val in &a[i..] {
            if val < result {
                result = val;
            }
        }
        result
    }
}

/// AVX-512 argmax.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn argmax(a: &[f32]) -> usize {
    let mut max_idx: usize = 0;
    let mut max_val = a[0];
    for (i, &val) in a.iter().enumerate() {
        if val > max_val {
            max_val = val;
            max_idx = i;
        }
    }
    max_idx
}

/// AVX-512 argmin.
#[inline]
#[target_feature(enable = "avx512f")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn argmin(a: &[f32]) -> usize {
    let mut min_idx: usize = 0;
    let mut min_val = a[0];
    for (i, &val) in a.iter().enumerate() {
        if val < min_val {
            min_val = val;
            min_idx = i;
        }
    }
    min_idx
}

/// Kahan sum (scalar implementation).
#[inline]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn sum_kahan(a: &[f32]) -> f32 {
    let mut sum = 0.0;
    let mut c = 0.0;
    for &x in a {
        let y = x - c;
        let t = sum + y;
        c = (t - sum) - y;
        sum = t;
    }
    sum
}