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

#![cfg_attr(feature = "as_crate", no_std)] // We are std!
#![cfg_attr(feature = "as_crate", feature(platform_intrinsics), feature(portable_simd))]
#[cfg(not(feature = "as_crate"))]
use core::simd;
#[cfg(feature = "as_crate")]
use core_simd::simd;

use simd::{LaneCount, Simd, SupportedLaneCount};

#[cfg(feature = "as_crate")]
mod experimental {
    pub trait Sealed {}
}

#[cfg(feature = "as_crate")]
use experimental as sealed;

use crate::sealed::Sealed;

// "platform intrinsics" are essentially "codegen intrinsics"
// each of these may be scalarized and lowered to a libm call
extern "platform-intrinsic" {
    // ceil
    fn simd_ceil<T>(x: T) -> T;

    // floor
    fn simd_floor<T>(x: T) -> T;

    // round
    fn simd_round<T>(x: T) -> T;

    // trunc
    fn simd_trunc<T>(x: T) -> T;

    // fsqrt
    fn simd_fsqrt<T>(x: T) -> T;

    // fma
    fn simd_fma<T>(x: T, y: T, z: T) -> T;
}

/// This trait provides a possibly-temporary implementation of float functions
/// that may, in the absence of hardware support, canonicalize to calling an
/// operating system's `math.h` dynamically-loaded library (also known as a
/// shared object). As these conditionally require runtime support, they
/// should only appear in binaries built assuming OS support: `std`.
///
/// However, there is no reason SIMD types, in general, need OS support,
/// as for many architectures an embedded binary may simply configure that
/// support itself. This means these types must be visible in `core`
/// but have these functions available in `std`.
///
/// [`f32`] and [`f64`] achieve a similar trick by using "lang items", but
/// due to compiler limitations, it is harder to implement this approach for
/// abstract data types like [`Simd`]. From that need, this trait is born.
///
/// It is possible this trait will be replaced in some manner in the future,
/// when either the compiler or its supporting runtime functions are improved.
/// For now this trait is available to permit experimentation with SIMD float
/// operations that may lack hardware support, such as `mul_add`.
pub trait StdFloat: Sealed + Sized {
    /// Fused multiply-add.  Computes `(self * a) + b` with only one rounding error,
    /// yielding a more accurate result than an unfused multiply-add.
    ///
    /// Using `mul_add` *may* be more performant than an unfused multiply-add if the target
    /// architecture has a dedicated `fma` CPU instruction.  However, this is not always
    /// true, and will be heavily dependent on designing algorithms with specific target
    /// hardware in mind.
    #[inline]
    #[must_use = "method returns a new vector and does not mutate the original value"]
    fn mul_add(self, a: Self, b: Self) -> Self {
        unsafe { simd_fma(self, a, b) }
    }

    /// Produces a vector where every lane has the square root value
    /// of the equivalently-indexed lane in `self`
    #[inline]
    #[must_use = "method returns a new vector and does not mutate the original value"]
    fn sqrt(self) -> Self {
        unsafe { simd_fsqrt(self) }
    }

    /// Returns the smallest integer greater than or equal to each lane.
    #[must_use = "method returns a new vector and does not mutate the original value"]
    #[inline]
    fn ceil(self) -> Self {
        unsafe { simd_ceil(self) }
    }

    /// Returns the largest integer value less than or equal to each lane.
    #[must_use = "method returns a new vector and does not mutate the original value"]
    #[inline]
    fn floor(self) -> Self {
        unsafe { simd_floor(self) }
    }

    /// Rounds to the nearest integer value. Ties round toward zero.
    #[must_use = "method returns a new vector and does not mutate the original value"]
    #[inline]
    fn round(self) -> Self {
        unsafe { simd_round(self) }
    }

    /// Returns the floating point's integer value, with its fractional part removed.
    #[must_use = "method returns a new vector and does not mutate the original value"]
    #[inline]
    fn trunc(self) -> Self {
        unsafe { simd_trunc(self) }
    }

    /// Returns the floating point's fractional value, with its integer part removed.
    #[must_use = "method returns a new vector and does not mutate the original value"]
    fn fract(self) -> Self;
}

impl<const N: usize> Sealed for Simd<f32, N> where LaneCount<N>: SupportedLaneCount {}
impl<const N: usize> Sealed for Simd<f64, N> where LaneCount<N>: SupportedLaneCount {}

// We can safely just use all the defaults.
impl<const N: usize> StdFloat for Simd<f32, N>
where
    LaneCount<N>: SupportedLaneCount,
{
    /// Returns the floating point's fractional value, with its integer part removed.
    #[must_use = "method returns a new vector and does not mutate the original value"]
    #[inline]
    fn fract(self) -> Self {
        self - self.trunc()
    }
}

impl<const N: usize> StdFloat for Simd<f64, N>
where
    LaneCount<N>: SupportedLaneCount,
{
    /// Returns the floating point's fractional value, with its integer part removed.
    #[must_use = "method returns a new vector and does not mutate the original value"]
    #[inline]
    fn fract(self) -> Self {
        self - self.trunc()
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use simd::*;

    #[test]
    fn everything_works() {
        let x = f32x4::from_array([0.1, 0.5, 0.6, -1.5]);
        let x2 = x + x;
        let _xc = x.ceil();
        let _xf = x.floor();
        let _xr = x.round();
        let _xt = x.trunc();
        let _xfma = x.mul_add(x, x);
        let _xsqrt = x.sqrt();
        let _ = x2.abs() * x2;
    }
}