microgemm

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

The implementation is based on the BLIS microkernel approach.

Install

cargo add microgemm

Usage

The Kernel trait is the main abstraction of microgemm. You can implement it yourself or use kernels that are already provided out of the box.

gemm

use microgemm::{kernels::GenericKernel8x8, Kernel as _, MatMut, MatRef, PackSizes};

fn main() {
    let kernel = GenericKernel8x8::<f32>::new();
    assert_eq!(kernel.mr(), 8);
    assert_eq!(kernel.nr(), 8);

    let pack_sizes = PackSizes {
        mc: 5 * kernel.mr(), // MC must be divisible by MR
        kc: 190,
        nc: 9 * kernel.nr(), // NC must be divisible by NR
    };
    let mut packing_buf = vec![0.0; pack_sizes.buf_len()];

    let (alpha, beta) = (2.0, -3.0);
    let (m, k, n) = (100, 380, 250);

    let a = vec![2.0; m * k];
    let b = vec![3.0; k * n];
    let mut c = vec![4.0; m * n];

    let a = MatRef::row_major(m, k, &a);
    let b = MatRef::row_major(k, n, &b);
    let mut c = MatMut::row_major(m, n, &mut c);

    // c <- alpha a b + beta c
    kernel.gemm(alpha, &a, &b, beta, &mut c, &pack_sizes, &mut packing_buf);
    println!("{:?}", c.as_slice());
}

Also see no_alloc example for use without Vec.

Implemented Kernels

Custom Kernel Implementation

use microgemm::{typenum::U4, Kernel, MatMut, MatRef};

struct CustomKernel;

impl Kernel for CustomKernel {
    type Scalar = f64;
    type Mr = U4;
    type Nr = U4;

    // dst <- alpha lhs rhs + beta dst
    fn microkernel(
        &self,
        alpha: f64,
        lhs: &MatRef<f64>,
        rhs: &MatRef<f64>,
        beta: f64,
        dst: &mut MatMut<f64>,
    ) {
        // lhs is col-major
        assert_eq!(lhs.row_stride(), 1);
        assert_eq!(lhs.nrows(), Self::MR);

        // rhs is row-major
        assert_eq!(rhs.col_stride(), 1);
        assert_eq!(rhs.ncols(), Self::NR);

        // dst is col-major
        assert_eq!(dst.row_stride(), 1);
        assert_eq!(dst.nrows(), Self::MR);
        assert_eq!(dst.ncols(), Self::NR);

        // your microkernel implementation...
    }
}

Benchmarks

All benchmarks are performed in a single thread on square matrices of dimension n.

f32

PackSizes { mc: n, kc: n, nc: n }

aarch64 (M1)

   n  NeonKernel8x8           faer matrixmultiply
 128         75.5µs        242.6µs         46.2µs
 256        466.3µs          3.2ms        518.2µs
 512            3ms         15.9ms          2.7ms
1024         23.9ms        128.4ms           22ms
2048          191ms             1s        182.8ms

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.