lightweight-simd 0.1.1

Lightweight generic fixed-array vector types relying on compiler auto-vectorization
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

// Most or all contents generated by AI (model glm-5.3)

//! Tests for the enforced-alignment wrappers and aligned aliases: layout
//! guarantees, delegation equivalence against plain `Simd`, conversions, and
//! mask handling.

use num_traits::Zero as _;

use lightweight_simd::*;

#[test]
fn alignment_and_size() {
    use core::mem::{align_of, size_of};
    assert_eq!(align_of::<f64x8>(), 64);
    assert_eq!(size_of::<f64x8>(), 64);
    assert_eq!(align_of::<f64x4>(), 32);
    assert_eq!(align_of::<f64x2>(), 16);
    assert_eq!(align_of::<f64x16>(), 64); // capped at 64
    assert_eq!(size_of::<f64x16>(), 128);
    assert_eq!(align_of::<f32x16>(), 64);
    assert_eq!(align_of::<f32x4>(), 16);
    assert_eq!(align_of::<f32x2>(), 8);
    assert_eq!(align_of::<i64x8>(), 64);
    assert_eq!(align_of::<i32x8>(), 32);
    assert_eq!(align_of::<i16x4>(), 8);
    assert_eq!(align_of::<i8x64>(), 64);
    // masks stay plain
    assert_eq!(align_of::<mask8>(), 1);
    // generic wrappers usable for any (T, N); size padded to alignment
    assert_eq!(align_of::<Aligned64<f64, 2>>(), 64);
    assert_eq!(size_of::<Aligned64<f64, 2>>(), 64);
}

#[test]
fn array_alignment() {
    let v: Vec<f64x8> = vec![f64x8::zero(); 16];
    assert_eq!(v.as_ptr() as usize % 64, 0);
    let w: Vec<f32x4> = vec![f32x4::zero(); 16];
    assert_eq!(w.as_ptr() as usize % 16, 0);
}

#[test]
fn ops_equivalence() {
    let a = f64x8::from_fn(|i| i as f64 * 0.5 + 1.0);
    let b = f64x8::from_fn(|i| i as f64 * 0.25 - 2.0);
    let got = (a * b).mul_add(a, b) + b * 2.0 - a;
    let a2: Simd<f64, 8> = a.into();
    let b2 = b.to_simd();
    let want = (a2 * b2).mul_add(a2, b2) + b2 * 2.0 - a2;
    assert_eq!(got.to_array(), want.to_array());
    assert_eq!(got.reduce_sum(), want.reduce_sum());

    // assigns and negation
    let mut c = a;
    c += b;
    c -= b;
    c *= b;
    c /= b;
    let mut d = a2;
    d += b2;
    d -= b2;
    d *= b2;
    d /= b2;
    assert_eq!(c.to_simd(), d);
    assert_eq!((-c).reduce_sum(), (-d).reduce_sum());
}

#[test]
fn conversions() {
    let s: Simd<f32, 4> = Simd::from_fn(|i| i as f32);
    let a: f32x4 = s.into();
    let back: Simd<f32, 4> = a.into();
    assert_eq!(back, s);
    assert_eq!(f32x4::from([1.0, 2.0, 3.0, 4.0]).reduce_sum(), 10.0);
    assert_eq!(a.to_simd(), s);
}

#[test]
fn mask_select_on_aligned() {
    let a = f64x8::splat(1.0);
    let b = f64x8::splat(2.0);
    let m: Simd<bool, 8> = Simd::from_fn(|i| i % 2 == 0);
    // per lane: 1, 2, 1, 2, 1, 2, 1, 2
    assert_eq!(a.mask_select(m, b).reduce_sum(), 12.0);
    // comparisons on aligned vectors produce plain masks
    let m2 = a.simd_lt(b);
    assert_eq!(m2.to_array(), [true; 8]);
    assert!(m2.all_true());
    assert_eq!(a.simd_eq(a).to_array(), [true; 8]);
    // mask_select also exists on plain Simd
    assert_eq!(Simd::<f64, 8>::splat(1.0).mask_select(m, Simd::splat(2.0)).reduce_sum(), 12.0);
}

#[test]
fn float_and_memory_delegation() {
    let v = f64x4::from([4.0, 9.0, 0.25, 2.25]);
    assert_eq!(v.sqrt().to_array(), [2.0, 3.0, 0.5, 1.5]);
    let mut buf = [0.0f64; 4];
    v.sqrt().store_slice(&mut buf);
    assert_eq!(buf, [2.0, 3.0, 0.5, 1.5]);
    let padded = f64x4::from_slice_pad(&[1.0, 2.0]);
    assert_eq!(padded.reduce_sum(), 3.0);
    assert_eq!(f64x4::from([1.5, 2.5, -1.5, -2.5]).round_ties_even().to_array(), [2.0, 2.0, -2.0, -2.0]);
}

#[test]
fn common_traits() {
    assert!(f64x8::zero().is_zero());
    assert_eq!(i64x4::one().reduce_product(), 1);
    assert_eq!(f64x4::default(), f64x4::zero());
    let z: f64x4 = num_traits::Zero::zero();
    assert!(z.is_zero());
    // bitwise/shift delegation on integer aliases
    let v = i64x4::from([1, 2, 3, 4]);
    assert_eq!((v << 2).to_array(), [4, 8, 12, 16]);
    assert_eq!((v & v).to_simd(), v.to_simd());
}

#[test]
fn uninit_scratch_aligned() {
    let mut v = unsafe { f64x8::uninit() };
    v.fill(1.5);
    assert_eq!(v.reduce_sum(), 12.0);
}

#[cfg(feature = "complex")]
#[test]
fn complex_alignment() {
    assert_eq!(core::mem::align_of::<lightweight_simd::c64x2>(), 32);
    assert_eq!(core::mem::align_of::<lightweight_simd::c64x4>(), 64);
}

#[cfg(feature = "half")]
#[test]
fn half_alignment() {
    assert_eq!(core::mem::align_of::<lightweight_simd::f16x2>(), 4);
    assert_eq!(core::mem::align_of::<lightweight_simd::f16x32>(), 64);
    assert_eq!(core::mem::align_of::<lightweight_simd::bf16x8>(), 16);
}