// vim: set syntax=asm :

{{L}}{{label}}:
    mov             rax, [ rdi + 8 ]

{% capture mr_over_16 %}{{ mr | divided_by: 16}}{%endcapture%}
{% capture mr_over_16_min_1 %}{{ mr | divided_by: 16 | minus: 1}}{%endcapture%}

{% for ix in (0..mr_over_16_min_1) %}
    vmovups         zmm{{to | plus: 1 | plus: ix}},  [rax + {{ix | times: 64}}]
{% endfor %}

{% if flipped %}
    {% for acc in (from..to) %}
        {{op}} zmm{{acc}}, zmm{{acc}}, zmm{{ acc | modulo: mr_over_16 | plus: to | plus: 1 }}
    {% endfor %}
{% else %}
    {% for acc in (from..to) %}
        {{op}} zmm{{acc}}, zmm{{ acc | modulo: mr_over_16 | plus: to | plus: 1 }}, zmm{{acc}}
    {% endfor %}
{% endif %}

    jmp {{L}}non_linear_loop