lightweight-simd 0.1.0

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
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

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

//! Correctness tests for bitwise/shift operators, fused multiply-add, float
//! functions, comparisons/masks, and reductions.

#[cfg(not(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma")))]
use core::ops::{Add, Mul};
use core::ops::{BitAnd, BitAndAssign, BitOr, BitOrAssign, BitXor, BitXorAssign};
#[cfg(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma"))]
use num_traits::MulAdd;
use num_traits::{Float, Num};

use lightweight_simd::{Simd, f64x4, f64x8, mask8};

fn check_bitwise<T, const N: usize>(mk: &impl Fn(usize) -> T)
where
    T: BitAnd<Output = T>
        + BitOr<Output = T>
        + BitXor<Output = T>
        + BitAndAssign
        + BitOrAssign
        + BitXorAssign
        + Copy
        + PartialEq
        + std::fmt::Debug,
{
    let a = Simd::<T, N>::from_fn(mk);
    let b = Simd::<T, N>::from_fn(|i| mk(i + 1));
    assert_eq!((a & b).to_array(), core::array::from_fn(|i| mk(i) & mk(i + 1)));
    assert_eq!((a | b).to_array(), core::array::from_fn(|i| mk(i) | mk(i + 1)));
    assert_eq!((a ^ b).to_array(), core::array::from_fn(|i| mk(i) ^ mk(i + 1)));
    let mut c = a;
    c &= b;
    assert_eq!(c, a & b);
    c |= b;
    assert_eq!(c.to_array(), core::array::from_fn(|i| (mk(i) & mk(i + 1)) | mk(i + 1)));
    c ^= b;
    assert_eq!(c.to_array(), core::array::from_fn(|i| ((mk(i) & mk(i + 1)) | mk(i + 1)) ^ mk(i + 1)));
}

#[test]
fn bitwise_ints_and_bool() {
    check_bitwise::<i64, 8>(&|i| i as i64 * 5 + 1);
    check_bitwise::<i64, 3>(&|i| i as i64 * 5 + 1);
    check_bitwise::<i32, 5>(&|i| i as i32 + 12345);
    check_bitwise::<bool, 8>(&|i| i % 3 == 0);
}

#[test]
fn shifts_i64() {
    let v = Simd::<i64, 4>::from([1, 2, 3, 4]);
    assert_eq!((v << 2).to_array(), [4, 8, 12, 16]);
    assert_eq!(((v << 3) >> 1).to_array(), [4, 8, 12, 16]);
}

/// Reference and crate method must agree under whichever evaluation mode the
/// crate was built in. The predicate mirrors src/ops.rs: fused `MulAdd` when
/// the target has hardware FMA (or the `use_libm_fma` feature pins it on),
/// separated `a * b + c` otherwise.
#[cfg(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma"))]
fn check_fused<T, const N: usize>(a_mk: &impl Fn(usize) -> T, b_mk: &impl Fn(usize) -> T)
where
    T: MulAdd<Output = T> + Copy + PartialEq + std::fmt::Debug,
{
    let a = Simd::<T, N>::from_fn(a_mk);
    let b = Simd::<T, N>::from_fn(b_mk);
    let c = Simd::<T, N>::from_fn(|i| a_mk(i + 7));
    assert_eq!(a.mul_add(b, c).to_array(), core::array::from_fn(|i| MulAdd::mul_add(a_mk(i), b_mk(i), a_mk(i + 7))));
    let mut acc = c;
    acc.fma_from(b, a);
    assert_eq!(acc.to_array(), core::array::from_fn(|i| MulAdd::mul_add(b_mk(i), a_mk(i), a_mk(i + 7))));
}

#[cfg(not(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma")))]
fn check_fused<T, const N: usize>(a_mk: &impl Fn(usize) -> T, b_mk: &impl Fn(usize) -> T)
where
    T: Mul<Output = T> + Add<Output = T> + Copy + PartialEq + std::fmt::Debug,
{
    let a = Simd::<T, N>::from_fn(a_mk);
    let b = Simd::<T, N>::from_fn(b_mk);
    let c = Simd::<T, N>::from_fn(|i| a_mk(i + 7));
    assert_eq!(a.mul_add(b, c).to_array(), core::array::from_fn(|i| a_mk(i) * b_mk(i) + a_mk(i + 7)));
    let mut acc = c;
    acc.fma_from(b, a);
    assert_eq!(acc.to_array(), core::array::from_fn(|i| b_mk(i) * a_mk(i) + a_mk(i + 7)));
}

#[test]
fn fused_floats() {
    check_fused::<f64, 4>(&|i| i as f64 * 0.5 + 1.0, &|i| i as f64 * 0.25 - 2.0);
    check_fused::<f64, 3>(&|i| i as f64 * 0.5 + 1.0, &|i| i as f64 * 0.25 - 2.0);
    check_fused::<f32, 8>(&|i| i as f32 * 0.5 + 1.0, &|i| i as f32 * 0.25 - 2.0);
    check_fused::<f32, 5>(&|i| i as f32 * 0.5 + 1.0, &|i| i as f32 * 0.25 - 2.0);

    // concrete values
    let v = f64x4::from([2.0; 4]);
    assert_eq!(v.mul_add(f64x4::from([3.0; 4]), f64x4::from([1.0; 4])).to_array(), [7.0; 4]);
    let mut acc = f64x4::from([1.0; 4]);
    acc.fma_from(f64x4::from([2.0; 4]), f64x4::from([3.0; 4]));
    assert_eq!(acc.to_array(), [7.0; 4]);
}

/// The two evaluation modes genuinely differ in rounding: with
/// a = 1 + 2^-52, b = 1 - 2^-53, c = 2^-53, the exact a*b + c lies just
/// below 1 + 2^-52, so a single rounding keeps it there, while rounding
/// a*b first (to 1.0) and then adding c (tie to even) lands on 1.0.
/// Expected values are bit literals, not compiler-evaluated expressions.
#[test]
fn mul_add_rounding_mode() {
    let a = f64x8::splat(1.0 + f64::EPSILON);
    let b = f64x8::splat(1.0 - 0.5 * f64::EPSILON);
    let c = f64x8::splat(0.5 * f64::EPSILON);
    let got = a.mul_add(b, c)[0].to_bits();
    // predicate mirrors src/ops.rs (a plain `cargo test` builds for generic
    // x86-64, so the default run takes the separated path on this host)
    #[cfg(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma"))]
    assert_eq!(got, 0x3ff0000000000001); // fused: 1 + 2^-52
    #[cfg(not(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma")))]
    assert_eq!(got, 0x3ff0000000000000); // separated: rounds to 1.0
}

fn check_float<T, const N: usize>(mk: &impl Fn(usize) -> T)
where
    T: Float + Copy + PartialEq + std::fmt::Debug,
{
    let a = Simd::<T, N>::from_fn(mk);
    assert_eq!(a.abs().to_array(), core::array::from_fn(|i| mk(i).abs()));
    assert_eq!(a.sqrt().to_array(), core::array::from_fn(|i| mk(i).sqrt()));
    assert_eq!(a.recip().to_array(), core::array::from_fn(|i| mk(i).recip()));
    assert_eq!(a.floor().to_array(), core::array::from_fn(|i| mk(i).floor()));
    assert_eq!(a.ceil().to_array(), core::array::from_fn(|i| mk(i).ceil()));
    assert_eq!(a.round().to_array(), core::array::from_fn(|i| mk(i).round()));
    assert_eq!(a.trunc().to_array(), core::array::from_fn(|i| mk(i).trunc()));
    assert_eq!(a.fract().to_array(), core::array::from_fn(|i| mk(i).fract()));
    let b = Simd::<T, N>::from_fn(|i| mk(i + 1));
    assert_eq!(a.min(b).to_array(), core::array::from_fn(|i| mk(i).min(mk(i + 1))));
    assert_eq!(a.max(b).to_array(), core::array::from_fn(|i| mk(i).max(mk(i + 1))));
    assert_eq!(a.copysign(b).to_array(), core::array::from_fn(|i| mk(i).copysign(mk(i + 1))));
}

#[test]
fn float_generic() {
    check_float::<f64, 8>(&|i| i as f64 * 0.5 + 0.5);
    check_float::<f64, 3>(&|i| i as f64 * 0.5 + 0.5);
    check_float::<f32, 5>(&|i| i as f32 * 0.5 + 0.5);
    check_float::<f32, 4>(&|i| i as f32 * 0.5 + 0.5);
}

#[test]
fn float_concrete_f64() {
    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 t = f64x4::from([0.5, 1.5, 2.5, -0.5]);
    assert_eq!(t.floor().to_array(), [0.0, 1.0, 2.0, -1.0]);
    assert_eq!(t.ceil().to_array(), [1.0, 2.0, 3.0, -0.0]);
    assert_eq!(t.round().to_array(), [1.0, 2.0, 3.0, -1.0]);
    assert_eq!(t.trunc().to_array(), [0.0, 1.0, 2.0, -0.0]);
    assert_eq!(t.fract().to_array(), [0.5, 0.5, 0.5, -0.5]);
    assert_eq!(t.abs().to_array(), [0.5, 1.5, 2.5, 0.5]);
    assert_eq!(t.copysign(f64x4::splat(1.0)).to_array(), [0.5, 1.5, 2.5, 0.5]);
    assert_eq!(t.copysign(f64x4::splat(-1.0)).to_array(), [-0.5, -1.5, -2.5, -0.5]);
    assert_eq!(f64x4::splat(2.0).recip().to_array(), [0.5; 4]);
    let u = f64x4::from([1.5, 2.5, -1.5, -2.5]);
    assert_eq!(u.round_ties_even().to_array(), [2.0, 2.0, -2.0, -2.0]);
    // scalar right-hand side
    assert_eq!(f64x4::splat(1.0) * 2.0, f64x4::splat(2.0));
}

fn check_compare<T, const N: usize>(mk: &impl Fn(usize) -> T)
where
    T: PartialOrd + Copy + std::fmt::Debug,
{
    let a = Simd::<T, N>::from_fn(mk);
    let b = Simd::<T, N>::from_fn(|i| mk(i + 1));
    assert_eq!(a.simd_lt(b).to_array(), core::array::from_fn(|i| a[i] < b[i]));
    assert_eq!(a.simd_le(b).to_array(), core::array::from_fn(|i| a[i] <= b[i]));
    assert_eq!(a.simd_gt(b).to_array(), core::array::from_fn(|i| a[i] > b[i]));
    assert_eq!(a.simd_ge(b).to_array(), core::array::from_fn(|i| a[i] >= b[i]));
    assert_eq!(a.simd_eq(a).to_array(), [true; N]);
    assert_eq!(a.simd_ne(a).to_array(), [false; N]);
    // select
    let m = a.simd_ge(b);
    assert_eq!(m.select(a, b).to_array(), core::array::from_fn(|i| if a[i] >= b[i] { a[i] } else { b[i] }));
    // mask composition via bitwise operators
    let m1 = a.simd_lt(b);
    let m2 = a.simd_le(b);
    assert_eq!((m1 & m2).to_array(), core::array::from_fn(|i| (a[i] < b[i]) & (a[i] <= b[i])));
    assert_eq!((m1 | m2).to_array(), core::array::from_fn(|i| (a[i] < b[i]) | (a[i] <= b[i])));
    assert_eq!((m1 ^ m2).to_array(), core::array::from_fn(|i| (a[i] < b[i]) ^ (a[i] <= b[i])));
    assert_eq!((!m1).to_array(), core::array::from_fn(|i| a[i] >= b[i]));
}

#[test]
fn compare_and_masks() {
    check_compare::<f64, 8>(&|i| i as f64 + 0.5);
    check_compare::<f64, 3>(&|i| i as f64 + 0.5);
    check_compare::<i64, 5>(&|i| i as i64 * 3);
    check_compare::<f32, 4>(&|i| i as f32 * 1.5);
}

#[test]
fn mask_concrete() {
    let m = mask8::from([true, false, true, true, false, false, true, false]);
    assert!(m.any_true());
    assert!(!m.all_true());
    assert_eq!(f64x8::splat(1.0).mask_select(m, f64x8::splat(2.0)).reduce_sum(), 12.0);
    assert!(!(m & mask8::splat(false)).any_true());
    assert!((m | mask8::splat(true)).all_true());
}

fn check_reduce<T, const N: usize>(mk: &impl Fn(usize) -> T)
where
    T: Num + Copy + PartialEq + std::fmt::Debug,
{
    let v = Simd::<T, N>::from_fn(mk);
    let mut sum = T::zero();
    let mut prod = T::one();
    for i in 0..N {
        sum = sum + mk(i);
        prod = prod * mk(i);
    }
    assert_eq!(v.reduce_sum(), sum);
    assert_eq!(v.reduce_product(), prod);
    for n in 0..=N {
        let mut s = T::zero();
        let mut p = T::one();
        for i in 0..n {
            s = s + mk(i);
            p = p * mk(i);
        }
        assert_eq!(v.reduce_sum_n(n), s);
        assert_eq!(v.reduce_product_n(n), p);
    }
}

#[test]
fn reduce_floats_and_ints() {
    check_reduce::<f64, 1>(&|i| i as f64 * 0.5 + 1.0);
    check_reduce::<f64, 3>(&|i| i as f64 * 0.5 + 1.0);
    check_reduce::<f64, 8>(&|i| i as f64 * 0.5 + 1.0);
    check_reduce::<f64, 32>(&|i| i as f64 * 0.5 + 1.0);
    check_reduce::<f32, 5>(&|i| i as f32 * 0.5 + 1.0);
    check_reduce::<i64, 8>(&|i| i as i64 % 2 + 1);
    check_reduce::<i64, 32>(&|i| i as i64 % 2 + 1);
}

#[cfg(feature = "complex")]
#[test]
fn complex_aliases() {
    use num_complex::Complex;
    let v: lightweight_simd::c64x8 = lightweight_simd::c64x8::splat(Complex::new(1.0, -1.0));
    assert_eq!(v.reduce_sum(), Complex::new(8.0, -8.0));
}

#[cfg(feature = "half")]
#[test]
fn half_aliases() {
    use half::{bf16, f16};
    let v = lightweight_simd::f16x8::splat(f16::from_f32(1.0));
    assert_eq!(v.reduce_sum_n(4), f16::from_f32(4.0));
    let w = lightweight_simd::bf16x16::splat(bf16::from_f32(0.5));
    assert_eq!(w.reduce_sum(), bf16::from_f32(8.0));
}