tract_linalg/generic/
silu.rs1#![allow(clippy::excessive_precision)]
2use crate::frame::element_wise::ElementWiseKer;
3use tract_data::internal::*;
4
5#[derive(Clone, Debug)]
6pub struct SSiLU4;
7
8impl ElementWiseKer<f32> for SSiLU4 {
9 fn name() -> &'static str {
10 "generic"
11 }
12
13 fn alignment_bytes() -> usize {
14 16
15 }
16
17 fn alignment_items() -> usize {
18 4
19 }
20
21 fn nr() -> usize {
22 4
23 }
24
25 fn run(x: &mut [f32], _: ()) {
26 debug_assert!(x.len() % Self::nr() == 0);
27 debug_assert!(x.as_ptr() as usize % Self::alignment_bytes() == 0);
28 x.iter_mut().for_each(|px| {
29 let sigmoid = 1.0 / (1.0 + (-*px).exp());
30 *px = *px * sigmoid;
31 });
32 }
33}
34
35#[derive(Clone, Debug)]
36pub struct HSiLU8;
37
38impl ElementWiseKer<f16> for HSiLU8 {
39 fn name() -> &'static str {
40 "generic"
41 }
42
43 fn alignment_bytes() -> usize {
44 16
45 }
46
47 fn alignment_items() -> usize {
48 4
49 }
50
51 fn nr() -> usize {
52 8
53 }
54
55 fn run(x: &mut [f16], _: ()) {
56 debug_assert!(x.len() % Self::nr() == 0);
57 debug_assert!(x.as_ptr() as usize % Self::alignment_bytes() == 0);
58 x.iter_mut().for_each(|px| {
59 let x_f32 = px.to_f32();
60 let sigmoid = 1.0 / (1.0 + (-x_f32).exp());
61 *px = f16::from_f32(x_f32 * sigmoid);
62 });
63 }
64}
65
66#[cfg(test)]
67#[macro_use]
68pub mod s {
69 silu_frame_tests!(true, f32, crate::generic::silu::SSiLU4);
70}
71
72#[cfg(test)]
73#[macro_use]
74pub mod h {
75 silu_frame_tests!(true, tract_data::internal::f16, crate::generic::silu::HSiLU8);
76}