half 2.7.1

Half-precision floating point f16 and bf16 types for Rust implementing the IEEE 754-2008 standard binary16 and bfloat16 types.
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

use core::{
    arch::{
        aarch64::{float32x4_t, float64x2_t, uint16x4_t},
        asm,
    },
    mem::MaybeUninit,
    ptr,
};
use zerocopy::transmute;

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn f16_to_f32_fp16(i: u16) -> f32 {
    let result: f32;
    asm!(
        "fcvt {0:s}, {1:h}",
        out(vreg) result,
        in(vreg) i,
        options(pure, nomem, nostack, preserves_flags));
    result
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn f16_to_f64_fp16(i: u16) -> f64 {
    let result: f64;
    asm!(
        "fcvt {0:d}, {1:h}",
        out(vreg) result,
        in(vreg) i,
        options(pure, nomem, nostack, preserves_flags));
    result
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn f32_to_f16_fp16(f: f32) -> u16 {
    let result: u16;
    asm!(
        "fcvt {0:h}, {1:s}",
        out(vreg) result,
        in(vreg) f,
        options(pure, nomem, nostack, preserves_flags));
    result
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn f64_to_f16_fp16(f: f64) -> u16 {
    let result: u16;
    asm!(
        "fcvt {0:h}, {1:d}",
        out(vreg) result,
        in(vreg) f,
        options(pure, nomem, nostack, preserves_flags));
    result
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn f16x4_to_f32x4_fp16(v: &[u16; 4]) -> [f32; 4] {
    let vec: uint16x4_t = transmute!(*v);
    let result: float32x4_t;
    asm!(
        "fcvtl {0:v}.4s, {1:v}.4h",
        out(vreg) result,
        in(vreg) vec,
        options(pure, nomem, nostack));
    transmute!(result)
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn f32x4_to_f16x4_fp16(v: &[f32; 4]) -> [u16; 4] {
    let vec: float32x4_t = transmute!(*v);
    let result: uint16x4_t;
    asm!(
        "fcvtn {0:v}.4h, {1:v}.4s",
        out(vreg) result,
        in(vreg) vec,
        options(pure, nomem, nostack));
    transmute!(result)
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn f16x4_to_f64x4_fp16(v: &[u16; 4]) -> [f64; 4] {
    let vec: uint16x4_t = transmute!(*v);
    let low: float64x2_t;
    let high: float64x2_t;
    asm!(
        "fcvtl {2:v}.4s, {3:v}.4h", // Convert to f32
        "fcvtl {0:v}.2d, {2:v}.2s", // Convert low part to f64
        "fcvtl2 {1:v}.2d, {2:v}.4s", // Convert high part to f64
        lateout(vreg) low,
        lateout(vreg) high,
        out(vreg) _,
        in(vreg) vec,
        options(pure, nomem, nostack));
    transmute!([low, high])
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn f64x4_to_f16x4_fp16(v: &[f64; 4]) -> [u16; 4] {
    let mut low = MaybeUninit::<float64x2_t>::uninit();
    let mut high = MaybeUninit::<float64x2_t>::uninit();
    ptr::copy_nonoverlapping(v.as_ptr(), low.as_mut_ptr().cast(), 2);
    ptr::copy_nonoverlapping(v[2..].as_ptr(), high.as_mut_ptr().cast(), 2);
    let result: uint16x4_t;
    asm!(
        "fcvtn {1:v}.2s, {2:v}.2d", // Convert low to f32
        "fcvtn2 {1:v}.4s, {3:v}.2d", // Convert high to f32
        "fcvtn {0:v}.4h, {1:v}.4s", // Convert to f16
        lateout(vreg) result,
        out(vreg) _,
        in(vreg) low.assume_init(),
        in(vreg) high.assume_init(),
        options(pure, nomem, nostack));
    transmute!(result)
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn add_f16_fp16(a: u16, b: u16) -> u16 {
    let result: u16;
    asm!(
        "fadd {0:h}, {1:h}, {2:h}",
        out(vreg) result,
        in(vreg) a,
        in(vreg) b,
        options(pure, nomem, nostack));
    result
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn subtract_f16_fp16(a: u16, b: u16) -> u16 {
    let result: u16;
    asm!(
        "fsub {0:h}, {1:h}, {2:h}",
        out(vreg) result,
        in(vreg) a,
        in(vreg) b,
        options(pure, nomem, nostack));
    result
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn multiply_f16_fp16(a: u16, b: u16) -> u16 {
    let result: u16;
    asm!(
        "fmul {0:h}, {1:h}, {2:h}",
        out(vreg) result,
        in(vreg) a,
        in(vreg) b,
        options(pure, nomem, nostack));
    result
}

#[target_feature(enable = "fp16")]
#[inline]
pub(super) unsafe fn divide_f16_fp16(a: u16, b: u16) -> u16 {
    let result: u16;
    asm!(
        "fdiv {0:h}, {1:h}, {2:h}",
        out(vreg) result,
        in(vreg) a,
        in(vreg) b,
        options(pure, nomem, nostack));
    result
}