typedef int __attribute__((address_space(5))) *l0cptr;
typedef signed char __attribute__((address_space(3))) *l0aptr;
typedef signed char __attribute__((address_space(4))) *l0bptr;

void cube_mmad(l0cptr dst, l0aptr a, l0bptr b) {
    __builtin_cce___mad(dst, a, b, 16, 16, 16, 0, 0);
}