// vim: ft=arm

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

{% capture mr_over_8 %}{{ mr | divided_by: 8}}{%endcapture%}
{% capture cols%}{{to | plus: 1| minus:from| divided_by:mr_over_8}}{%endcapture%}

{% capture loads %}{{cols | divided_by: 8}}{% endcapture %}

{%if cols == "1" %}
        ld1         {v0.h}[0], [ x2 ]
{% elsif cols == "3" %}
        ld1         {v0.s}[0], [ x2 ], #4
        ld1         {v0.h}[2], [ x2 ]
{% elsif cols == "4" %}
        ldr         d0, [ x2 ]
{% elsif cols == "6" %}
        ld1         {v0.d}[0], [ x2 ], #8
        ld1         {v0.s}[2], [ x2 ]
{% else %}
    {% for reg in (1..loads) %}
        ldr         q{{reg |minus:1}}, [ x2 ], #16
    {% endfor %}
{% endif %}

// mr:{{mr}} {{ loads }} {{cols}}

{% for col in (1..cols) %}
    dup v3.8h, v{{col| minus: 1|divided_by:8}}.h[{{col| minus: 1|modulo:8}}]
    {% for row in (1..mr_over_8) %}
        {% capture acc %}{{ col|minus:1|times:mr_over_8|plus:row|minus:1|plus:from }}{% endcapture %}
        {% if flipped %}
            {{op}} v{{acc}}.8h, v{{acc}}.8h, v3.8h
        {% else %}
            {{op}} v{{acc}}.8h, v3.8h, v{{acc}}.8h
        {% endif %}
    {% endfor %}
{% endfor %}

b           .non_linear_loop