pub fn matmul(a: var, b: var) [a.len][b[0].len]@typeOf(a[0][0]) {
    const T = @typeOf(a[0][0]);
    const M = a.len;
    const N = a[0].len;
    const P = b[0].len;
    if (T != @typeOf(b[0][0])) @compileError("element type mismatch");
    if (N != b.len) @compileError("size mismatch");
    var result = [_][P]T{[_]T{0} ** P} ** M;
    var m: u64 = 0;
    while (m < M) : (m += 1) {
        var n: u64 = 0;
        while (n < N) : (n += 1) {
            var p: u64 = 0;
            while (p < P) : (p += 1) {
                result[m][p] += a[m][n] * b[n][p];
            }
        }
    }
    return result;
}