{% comment %}
// vim: set syntax=asm :

/* mmm 16 x 8

    zmm0 zmm1 ... zmm8

System V ABI:
    args: rdi, rsi, rdx, rcx, r8, r9
    preserve: rbx, rsp, rbp, r12, r13, r14, r15
    scratch: rax, rdi, rsi, rdx, rcx, r8, r9, r10, r11
    return: rax (+rdx)

Windows ABI:
    args: RCX, RDX, R8, R9
    preserve: RBX, RBP, RDI, RSI, RSP, R12, R13, R14, R15, and XMM6-15
    scratch: RAX, RCX, RDX, R8, R9, R10, R11, XMM0-5, and the upper portions of ZMM0-15 and ZMM0-15
    return: rax (+rdx)
*/
{% endcomment %}


{% include "preamble.tmpliq" size:"16x8", suffix:suffix, G:G, arch:"avx512" %}

{{L}}clear:
    vzeroall
    jmp     {{L}}non_linear_loop

{{L}}add_mat_mul:
    mov     rcx,    [rdi + 24]   // B
    mov     rax,    [rdi + 16]   // A

    mov     rbx,    [rdi + 8]    // k
    test    rbx,    rbx
    jz      {{L}}non_linear_loop

	cmp rbx, 2
	jl {{L}}main_loop_packed_packed_tail

{{align}} 16
{{L}}main_loop_packed_packed:
	{% include "8x8/packed_packed_loop1/avx-512-unroll.tmpli" %}

    sub             rbx, 2
	cmp rbx,        2
	jge              {{L}}main_loop_packed_packed

    test    rbx, rbx
    jz      {{L}}non_linear_loop

{{align}} 16
{{L}}main_loop_packed_packed_tail:
	{% include "8x8/packed_packed_loop1/avx-512.tmpli" %}

	sub             rbx, 1
    jnz				{{L}}main_loop_packed_packed_tail

    jmp      {{L}}non_linear_loop

{% include "f32_scalars.tmpliq" from:0, to:7 %}
{% include "f32_per_rows.tmpliq" mr:16, from:0, to:7 %}
{% include "f32_per_cols.tmpliq" mr:16, from:0, to:7 %}

{{L}}add_unicast:

    mov     r10,    [rdi + 8]           // c ptr
    mov     rsi,    [rdi + 16]          // row stride
    mov     rbx,    [rdi + 24]          // col stride

    mov     eax,    0

{% for i in (0..3) %}
    pinsrd  xmm14, eax, {{i}}
    add     eax,    esi
{% endfor %}
{% for i in (0..3) %}
    pinsrd  xmm15, eax, {{i}}
    add     eax,    esi
{% endfor %}
{% for i in (0..3) %}
    pinsrd  xmm12, eax, {{i}}
    add     eax,    esi
{% endfor %}
{% for i in (0..3) %}
    pinsrd  xmm13, eax, {{i}}
    add     eax,    esi
{% endfor %}

    vperm2f128      ymm14,  ymm14, ymm15,         32 // ymm14 <- xmm14::xmm15
    vperm2f128      ymm13,  ymm12, ymm13,         32 // ymm12 <- xmm12::xmm13
    vinsertf32x8    zmm14, zmm14, ymm13, 1

{% for i in (0..7) %}
    kxnorw k1,k1,k1
    vgatherdps      zmm12{k1},  [ r10 + zmm14 ]
    add     r10, rbx
    vaddps          zmm{{i}},   zmm{{i}},   zmm12
{% endfor %}

    jmp    {{L}}non_linear_loop

{{L}}add_row_col_products:
    mov             rax, [ rdi + 8 ]
    mov             rbx, [ rdi + 16 ]

    vmovups         zmm12, zmmword ptr [rax]

{% for i in (0..7) %}
    vbroadcastss    zmm14, dword ptr [rbx + {{i|times:4}} ]
    vfmadd231ps     zmm{{i}},   zmm12, zmm14
{% endfor %}
    jmp    {{L}}non_linear_loop

{{L}}store:
    mov     r8,     [rdi + 8]           // c ptr
    mov     rsi,    [rdi + 16]          // row stride
    mov     rbx,    [rdi + 24]          // col stride

    // tops of cols
    lea     r9,     [ r8 + rbx ]
    lea     r10,    [ r8 + 2 * rbx ]
    lea     r12,    [ r8 + 4 * rbx ]
    lea     r11,    [ r10 + rbx ]
    lea     r13,    [ r12 + rbx ]
    lea     r14,    [ r12 + 2 * rbx ]
    lea     r15,    [ r13 + 2 * rbx ]
    
    {% for quarter in (0..3) %}
        {% for r in (0..7) %}
            vextractf32x4 xmm{{r | plus: 8}}, zmm{{r}}, {{quarter}}
        {% endfor %}
        {% for row in (0..3) %}
            {% for i in (0..7) %}
                vextractps  dword ptr [r{{i | plus: 8}}], xmm{{i | plus: 8}}, {{row}}
                add         r{{i | plus: 8}}, rsi
            {% endfor %}
        {% endfor %}
    {% endfor %}

    jmp     {{L}}non_linear_loop

{% include "postamble.tmpliq" size:"16x8", suffix:suffix, G:G, L:L, arch:"avx512" %}