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

//! SSE2 reduction operations (dot, sum, max, min, argmax, argmin).

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

use crate::backends::VectorBackend;

/// SSE2 dot product.
#[inline]
#[target_feature(enable = "sse2")]
// 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 = _mm_setzero_ps();

        while i + 4 <= len {
            let va = _mm_loadu_ps(a.as_ptr().add(i));
            let vb = _mm_loadu_ps(b.as_ptr().add(i));
            acc = _mm_add_ps(acc, _mm_mul_ps(va, vb));
            i += 4;
        }

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

/// SSE2 vector sum.
#[inline]
#[target_feature(enable = "sse2")]
// 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 = _mm_setzero_ps();

        while i + 4 <= len {
            acc = _mm_add_ps(acc, _mm_loadu_ps(a.as_ptr().add(i)));
            i += 4;
        }

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

/// SSE2 vector max.
#[inline]
#[target_feature(enable = "sse2")]
// 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 = _mm_set1_ps(a[0]);

        while i + 4 <= len {
            vmax = _mm_max_ps(vmax, _mm_loadu_ps(a.as_ptr().add(i)));
            i += 4;
        }

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

/// SSE2 vector min.
#[inline]
#[target_feature(enable = "sse2")]
// 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 = _mm_set1_ps(a[0]);

        while i + 4 <= len {
            vmin = _mm_min_ps(vmin, _mm_loadu_ps(a.as_ptr().add(i)));
            i += 4;
        }

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

/// SSE2 argmax.
#[inline]
#[target_feature(enable = "sse2")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn argmax(a: &[f32]) -> usize {
    let _len = a.len();
    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
}

/// SSE2 argmin.
#[inline]
#[target_feature(enable = "sse2")]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn argmin(a: &[f32]) -> usize {
    let _len = a.len();
    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 (delegates to scalar).
#[inline]
// SAFETY: caller ensures preconditions are met for this unsafe function
pub(crate) unsafe fn sum_kahan(a: &[f32]) -> f32 {
    unsafe { crate::backends::scalar::ScalarBackend::sum_kahan(a) }
}

// Helper: horizontal sum of 4 floats
#[inline]
#[target_feature(enable = "sse2")]
// SAFETY: caller ensures preconditions are met for this unsafe function
unsafe fn horizontal_sum_ps(v: __m128) -> f32 {
    let temp = _mm_add_ps(v, _mm_movehl_ps(v, v));
    let temp = _mm_add_ss(temp, _mm_shuffle_ps(temp, temp, 1));
    _mm_cvtss_f32(temp)
}

// Helper: horizontal max of 4 floats
#[inline]
#[target_feature(enable = "sse2")]
// SAFETY: caller ensures preconditions are met for this unsafe function
unsafe fn horizontal_max_ps(v: __m128) -> f32 {
    let temp = _mm_max_ps(v, _mm_movehl_ps(v, v));
    let temp = _mm_max_ss(temp, _mm_shuffle_ps(temp, temp, 1));
    _mm_cvtss_f32(temp)
}

// Helper: horizontal min of 4 floats
#[inline]
#[target_feature(enable = "sse2")]
// SAFETY: caller ensures preconditions are met for this unsafe function
unsafe fn horizontal_min_ps(v: __m128) -> f32 {
    let temp = _mm_min_ps(v, _mm_movehl_ps(v, v));
    let temp = _mm_min_ss(temp, _mm_shuffle_ps(temp, temp, 1));
    _mm_cvtss_f32(temp)
}