inari 1.0.0

A Rust implementation of interval arithmetic
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

use std::{arch::x86_64::*, mem::transmute};

pub(crate) const HAS_MAXIMUM: bool = false;

pub(crate) type F64X2 = __m128d;

pub(crate) fn abs(x: F64X2) -> F64X2 {
    // Drop sign bits.
    unsafe { _mm_andnot_pd(splat(-0.0), x) }
}

/// `add_rn([x0, x1], [y0, y1]) = [x0 + y0, x1 + y1]` rounded to nearest.
pub(crate) fn add_rn(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_add_pd(x, y) }
}

pub(crate) fn all(x: F64X2) -> bool {
    bitmask(x) == 3
}

pub(crate) fn and(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_and_pd(x, y) }
}

pub(crate) fn any(x: F64X2) -> bool {
    bitmask(x) != 0
}

pub(crate) fn bitmask(x: F64X2) -> u32 {
    unsafe { _mm_movemask_pd(x) as u32 }
}

pub(crate) fn ceil(x: F64X2) -> F64X2 {
    unsafe { _mm_ceil_pd(x) }
}

pub(crate) fn constant(x: f64, y: f64) -> F64X2 {
    unsafe { transmute([x, y]) }
}

pub(crate) fn eq(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_cmpeq_pd(x, y) }
}

pub(crate) fn extract(x: F64X2) -> [f64; 2] {
    unsafe { transmute::<_, [f64; 2]>(x) }
}

pub(crate) fn extract0(x: F64X2) -> f64 {
    unsafe { transmute::<_, [f64; 2]>(x)[0] }
}

// This is slower than extract0, which usually turns into a no-op.
pub(crate) fn extract1(x: F64X2) -> f64 {
    unsafe { transmute::<_, [f64; 2]>(x)[1] }
}

pub(crate) fn floor(x: F64X2) -> F64X2 {
    unsafe { _mm_floor_pd(x) }
}

pub(crate) fn ge(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_cmpge_pd(x, y) }
}

pub(crate) fn gt(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_cmpgt_pd(x, y) }
}

/// `hadd_rn([x0, x1], [y0, y1]) = [x0 + x1, y0 + y1]` rounded to nearest.
pub(crate) fn hadd_rn(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_hadd_pd(x, y) }
}

pub(crate) fn le(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_cmple_pd(x, y) }
}

pub(crate) fn lt(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_cmplt_pd(x, y) }
}

// The vector version of the `maximumNumber` operation defined in IEEE 754-2019,
// which does not propagate NaNs.
pub(crate) fn max(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_max_pd(x, y) }
}

// The vector version of the `maximum` operation defined in IEEE 754-2019,
// which propagates NaNs.
pub(crate) fn maximum(_x: F64X2, _y: F64X2) -> F64X2 {
    unimplemented!()
}

pub(crate) fn min(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_min_pd(x, y) }
}

pub(crate) fn minimum(_x: F64X2, _y: F64X2) -> F64X2 {
    unimplemented!()
}

pub(crate) fn neg(x: F64X2) -> F64X2 {
    xor(splat(-0.0), x)
}

pub(crate) fn neg0(x: F64X2) -> F64X2 {
    unsafe { xor(x, _mm_set_sd(-0.0)) }
}

pub(crate) fn or(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_or_pd(x, y) }
}

// This one is hard to implement correctly.
// https://www.cockroachlabs.com/blog/rounding-implementations-in-go/
pub(crate) fn round(x: F64X2) -> F64X2 {
    constant(extract0(x).round(), extract1(x).round())
}

pub(crate) fn round_ties_to_even(x: F64X2) -> F64X2 {
    unsafe { _mm_round_pd(x, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC) }
}

/// `shuffle02([x0, x1], [x2, x3]) = [x0, x2]`
pub(crate) fn shuffle02(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_shuffle_pd(x, y, 0) }
}

/// `shuffle03([x0, x1], [x2, x3]) = [x0, x3]`
pub(crate) fn shuffle03(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_shuffle_pd(x, y, 2) }
}

/// `shuffle13([x0, x1], [x2, x3]) = [x1, x3]`
pub(crate) fn shuffle13(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_shuffle_pd(x, y, 3) }
}

pub(crate) fn splat(x: f64) -> F64X2 {
    unsafe { transmute([x, x]) }
}

/// `sub_rn([x0, x1], [y0, y1]) = [x0 - y0, x1 - y1]` rounded to nearest.
pub(crate) fn sub_rn(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_sub_pd(x, y) }
}

pub(crate) fn swap(x: F64X2) -> F64X2 {
    shuffle12(x, x)
}

pub(crate) fn trunc(x: F64X2) -> F64X2 {
    unsafe { _mm_round_pd(x, _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC) }
}

/// `shuffle13([x0, x1], [x2, x3]) = [x1, x2]`
fn shuffle12(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_shuffle_pd(x, y, 1) }
}

fn xor(x: F64X2, y: F64X2) -> F64X2 {
    unsafe { _mm_xor_pd(x, y) }
}

cfg_if::cfg_if! {
    if #[cfg(target_feature = "avx512f")] {
        mod avx512f;
        pub(crate) use avx512f::*;
    } else if #[cfg(all(target_feature = "avx", target_feature = "fma"))] {
        mod avx_fma;
        pub(crate) use avx_fma::*;
    } else {
        compile_error!("RUSTFLAGS='-Ctarget-cpu=haswell' or later is required. See https://doc.rust-lang.org/rustc/codegen-options/#target-cpu");
    }
}