// vim: ft=arm

{% include "armv7neon_mmm_q_scalar.tmpliq" label:"scalar_min", op:"vmin.s32", from:from, to:to%}
{% include "armv7neon_mmm_q_scalar.tmpliq" label:"scalar_max", op:"vmax.s32", from:from, to:to%}
{% include "armv7neon_mmm_q_scalar.tmpliq" label:"scalar_mul", op:"vmul.s32", from:from, to:to%}
{% include "armv7neon_mmm_q_scalar.tmpliq" label:"scalar_add", op:"vadd.s32", from:from, to:to%}
{% include "armv7neon_mmm_q_scalar.tmpliq" label:"scalar_sub", op:"vsub.s32", from:from, to:to%}
{% include "armv7neon_mmm_q_scalar.tmpliq" label:"scalar_sub_flipped", op:"vsub.s32", from:from, to:to, flipped:true%}

.leaky_relu:
    vmov            s0, r3
    vdup.32         q0, d0[0]
    {% for reg in (from..to) %}
        vmul.s32    q2, q{{reg}}, q0
        vcgt.s32    q1, q{{reg}}, 0
        vbsl        q1, q{{reg}}, q2
        vmov        q{{reg}}, q1
    {% endfor %}
    b .non_linear_loop