// vim: ft=arm

.{{label}}:
    ldr         x2, [x0, #8]

{% capture mr_over_4 %}{{ mr | divided_by: 4}}{%endcapture%}
{% capture cols%}{{to | plus: 1| minus:from| divided_by:mr_over_4}}{%endcapture%}

{% capture loads %}{{cols | divided_by:4}}{% endcapture %}

{%if cols == "1" %}
        ld1         {v0.s}[0], [ x2 ]
{% else %}
    {% for reg in (1..loads) %}
        ldr         q{{reg |minus:1}}, [ x2 ], #16
    {% endfor %}
{% endif %}

// {{mr}} {{cols}}

{% for col in (1..cols) %}
    dup v3.4s, v{{col| minus: 1|divided_by:4}}.s[{{col| minus: 1|modulo:4}}]
    {% for row in (1..mr_over_4) %}
        {% capture acc %}{{ col|minus:1|times:mr_over_4|plus:row|minus:1|plus:from }}{% endcapture %}
        {% if flipped %}
            {{op}} v{{acc}}.4s, v{{acc}}.4s, v3.4s
        {% else %}
            {{op}} v{{acc}}.4s, v3.4s, v{{acc}}.4s
        {% endif %}
    {% endfor %}
{% endfor %}

b           .non_linear_loop