microgemm 0.3.1

General matrix multiplication with custom configuration in Rust. Supports no_std and no_alloc environments.
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

#![cfg(target_arch = "aarch64")]

use microgemm::{
    kernels::{NeonKernel4x4, NeonKernel8x8},
    Kernel, MatMut, MatRef, PackSizes,
};

#[test]
fn test_neon8x8() {
    let kernel = if cfg!(target_feature = "neon") {
        unsafe { NeonKernel8x8::<f32>::new() }
    } else {
        println!("neon feature is not supported");
        return;
    };
    test_kernel_f32(kernel);
}

#[test]
fn test_neon4x4() {
    let kernel = if cfg!(target_feature = "neon") {
        unsafe { NeonKernel4x4::<f32>::new() }
    } else {
        println!("neon feature is not supported");
        return;
    };
    test_kernel_f32(kernel);
}

fn test_kernel_f32(kernel: impl Kernel<Scalar = f32>) {
    let a = [1., 2., 3., 4., 5., 6.];
    let b = [10., 11., 20., 21., 30., 31.];
    let mut c = vec![0f32; 2 * 2];
    let a = MatRef::row_major(2, 3, a.as_ref());
    let b = MatRef::row_major(3, 2, b.as_ref());
    let mut c = MatMut::row_major(2, 2, c.as_mut());
    let pack_sizes = PackSizes {
        mc: kernel.mr(),
        kc: 2,
        nc: kernel.nr(),
    };
    let mut packing_buf = vec![0f32; pack_sizes.buf_len()];
    kernel.gemm(
        1f32,
        a,
        b,
        0f32,
        c.as_mut(),
        pack_sizes,
        packing_buf.as_mut(),
    );
    assert_eq!(c.as_slice(), [140., 146., 320., 335.]);
}