tract-linalg 0.23.4

Tiny, no-nonsense, self contained, TensorFlow and ONNX inference
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

use std::arch::aarch64::{float32x4_t, vdupq_n_f32, vgetq_lane_f32};

reduce_impl_wrap!(
    f32,
    arm64simd_min_f32_16n,
    16,
    4,
    (),
    f32::MAX,
    #[inline(never)]
    fn run(buf: &[f32], _: ()) -> f32 {
        assert!(buf.len() % 16 == 0);
        assert!(buf.len() > 0);
        unsafe {
            let len = buf.len();
            let ptr = buf.as_ptr();
            let mut out: float32x4_t = vdupq_n_f32(f32::MAX);
            std::arch::asm!("
            mov v1.16b, v0.16b
            mov v2.16b, v0.16b
            mov v3.16b, v0.16b
            2:
                ld1 {{v4.4s, v5.4s, v6.4s, v7.4s}}, [{ptr}], 64
                fmin v0.4s, v0.4s, v4.4s
                fmin v1.4s, v1.4s, v5.4s
                fmin v2.4s, v2.4s, v6.4s
                fmin v3.4s, v3.4s, v7.4s
                subs {len}, {len}, 16
                bne 2b
            fmin v0.4s, v0.4s, v1.4s
            fmin v2.4s, v2.4s, v3.4s
            fmin v0.4s, v0.4s, v2.4s
            fminv s0, v0.4s
            ",
            len = inout(reg) len => _,
            ptr = inout(reg) ptr => _,
            inout("v0") out, out("v1") _, out("v2") _, out("v3") _,
            out("v4") _, out("v5") _, out("v6") _, out("v7") _,);
            vgetq_lane_f32(out, 0)
        }
    },
    #[inline(never)]
    fn reduce_two(a: f32, b: f32) -> f32 {
        a.min(b)
    }
);

#[cfg(test)]
mod test_arm64simd_min_f32_16n {
    use super::*;
    crate::min_frame_tests!(true, f32, arm64simd_min_f32_16n);
}