// vim: ft=arm

.{{label}}:

{% capture mr_over_4 %}{{ mr | divided_by: 4}}{%endcapture%}
{% capture mr_over_4_min_1 %}{{ mr | divided_by: 4 | minus: 1}}{%endcapture%}

{%capture cols%}{{to | plus: 1| minus:from| divided_by:mr_over_4}}{%endcapture%}
{%capture cols_min_1%}{{to | plus: 1| minus:from| divided_by:mr_over_4|minus:1}}{%endcapture%}

{% if cols == "1" %}
    vld1.f32        d0[0], [ r3 ]
{% else %}
    {%capture cols_over_2_minus_1%}{{cols | divided_by:2 | minus:1}}{%endcapture%}
    {% for c in (0..cols_over_2_minus_1) %}
        vldmia      r3!, { d{{c}} }
    {% endfor %}
{% endif %}


{% for right in (0..cols_min_1) %}
    vdup.f32 q3, d{{right|divided_by:2}}[{{right| modulo:2}}]
    {% for down in (0..mr_over_4_min_1) %}
        {%capture acc%}{{mr_over_4|times:right|plus:from|plus:down}}{%endcapture%}
        {% if flipped %}
            {{op}} q{{acc}}, q{{acc}}, q3
        {% else %}
            {{op}} q{{acc}}, q3, q{{acc}}
        {% endif %}
    {% endfor %}
{% endfor %}

    b .non_linear_loop