8x8 kernel after

.section .text.rav1e::asm::x86::dist::hbd::satd_kernel_8x8_hbd_avx2,"ax",@progbits
    .p2align    4, 0x90
    .type   rav1e::asm::x86::dist::hbd::satd_kernel_8x8_hbd_avx2,@function
rav1e::asm::x86::dist::hbd::satd_kernel_8x8_hbd_avx2:

    .cfi_startproc
    push rbp
    .cfi_def_cfa_offset 16
    .cfi_offset rbp, -16
    mov rbp, rsp
    .cfi_def_cfa_register rbp

    push r15
    push r14
    push r12
    push rbx
    and rsp, -64
    sub rsp, 320
    .cfi_offset rbx, -48
    .cfi_offset r12, -40
    .cfi_offset r14, -32
    .cfi_offset r15, -24

    vpmovsxwd ymm0, xmmword ptr [rdi]

    vpmovsxwd ymm1, xmmword ptr [rdx]

    vpmovsxwd ymm2, xmmword ptr [rdx + 2*rcx]

    vpmovsxwd ymm3, xmmword ptr [rdi + 4*rsi]

    vpmovsxwd ymm4, xmmword ptr [rdx + 4*rcx]

    lea r8, [rsi + 2*rsi]

    lea rbx, [rcx + 2*rcx]

    vpmovsxwd ymm6, xmmword ptr [rdx + 8*rcx]

    lea rax, [rsi + 4*rsi]

    vpmovsxwd ymm5, xmmword ptr [rdi + 2*r8]

    vpmovsxwd ymm7, xmmword ptr [rdi + 4*r8]
    vpmovsxwd ymm8, xmmword ptr [rdx + 4*rbx]

    vpsubd ymm0, ymm0, ymm1

    vpmovsxwd ymm1, xmmword ptr [rdi + 2*rsi]

    vmovd r12d, xmm0

    vpsubd ymm1, ymm1, ymm2
    vpsubd ymm2, ymm3, ymm4

    vpmovsxwd ymm3, xmmword ptr [rdx + 2*rbx]

    vpmovsxwd ymm4, xmmword ptr [rdi + 8*rsi]

    vmovd r9d, xmm1
    vmovd r8d, xmm2

    vpsubd ymm3, ymm5, ymm3

    vpmovsxwd ymm5, xmmword ptr [rdi + 2*rax]

    lea rax, [rcx + 4*rcx]

    vpsubd ymm4, ymm4, ymm6

    vpmovsxwd ymm6, xmmword ptr [rdx + 2*rax]

    mov rax, rsi

    vmovd r10d, xmm3
    vmovd r15d, xmm4

    shl rax, 4
    sub rax, rsi
    sub rax, rsi

    vpsubd ymm6, ymm5, ymm6
    vpsubd ymm5, ymm7, ymm8

    vpmovsxwd ymm7, xmmword ptr [rdi + rax]

    mov rax, rcx
    shl rax, 4

    vmovd r11d, xmm6
    vmovd r14d, xmm5

    sub rax, rcx
    sub rax, rcx
    mov rcx, -7

    vpmovsxwd ymm8, xmmword ptr [rdx + rax]

    vmovdqa ymmword ptr [rsp], ymm0

    vmovdqa ymmword ptr [rsp + 32], ymm1

    vmovdqa ymmword ptr [rsp + 64], ymm2

    vmovdqa ymmword ptr [rsp + 96], ymm3

    vmovdqa ymmword ptr [rsp + 128], ymm4

    vmovdqa ymmword ptr [rsp + 160], ymm6

    vmovdqa ymmword ptr [rsp + 192], ymm5

    vpsubd ymm7, ymm7, ymm8

    vmovd ebx, xmm7

    vmovdqa ymmword ptr [rsp + 224], ymm7

    .p2align    4, 0x90
.LBB666_1:
    lea esi, [r9 + r12]

    sub r12d, r9d

    lea edx, [r10 + r8]

    sub r8d, r10d

    lea edi, [r11 + r15]

    lea r9d, [rbx + r14]

    sub r14d, ebx

    sub r15d, r11d

    lea r10d, [rdx + rsi]

    sub esi, edx

    lea edx, [r8 + r12]

    lea ebx, [r9 + rdi]

    sub r12d, r8d

    lea eax, [r14 + r15]

    sub edi, r9d

    sub r15d, r14d

    lea r8d, [rbx + r10]

    sub r10d, ebx

    lea ebx, [rax + rdx]

    sub edx, eax

    lea eax, [rdi + rsi]

    sub esi, edi

    lea edi, [r15 + r12]

    sub r12d, r15d

    mov dword ptr [rsp + 4*rcx + 28], r8d

    mov dword ptr [rsp + 4*rcx + 60], ebx

    mov dword ptr [rsp + 4*rcx + 92], eax

    mov dword ptr [rsp + 4*rcx + 124], edi

    mov dword ptr [rsp + 4*rcx + 156], r10d

    mov dword ptr [rsp + 4*rcx + 188], edx

    mov dword ptr [rsp + 4*rcx + 220], esi

    mov dword ptr [rsp + 4*rcx + 252], r12d

    test rcx, rcx

    je .LBB666_2

    mov r12d, dword ptr [rsp + 4*rcx + 32]
    mov r9d, dword ptr [rsp + 4*rcx + 64]

    mov r8d, dword ptr [rsp + 4*rcx + 96]
    mov r10d, dword ptr [rsp + 4*rcx + 128]

    mov r15d, dword ptr [rsp + 4*rcx + 160]
    mov r11d, dword ptr [rsp + 4*rcx + 192]

    mov r14d, dword ptr [rsp + 4*rcx + 224]
    mov ebx, dword ptr [rsp + 4*rcx + 256]

    inc rcx

    jmp .LBB666_1

.LBB666_2:
    xor eax, eax

    .p2align    4, 0x90
.LBB666_3:
    mov ecx, dword ptr [rsp + rax]

    mov esi, dword ptr [rsp + rax + 4]

    mov ebx, dword ptr [rsp + rax + 8]

    mov edx, dword ptr [rsp + rax + 20]

    mov edi, dword ptr [rsp + rax + 28]

    lea r10d, [rsi + rcx]

    sub ecx, esi

    mov esi, dword ptr [rsp + rax + 12]

    lea r9d, [rsi + rbx]

    sub ebx, esi

    mov esi, dword ptr [rsp + rax + 16]

    lea r8d, [r9 + r10]

    sub r10d, r9d

    lea r9d, [rbx + rcx]

    sub ecx, ebx

    lea r11d, [rdx + rsi]

    sub esi, edx

    mov edx, dword ptr [rsp + rax + 24]

    lea r14d, [rdi + rdx]

    sub edx, edi

    lea edi, [r14 + r11]

    sub r11d, r14d

    lea ebx, [rdx + rsi]

    sub esi, edx

    lea r14d, [rdi + r8]

    sub r8d, edi

    lea edi, [rbx + r9]

    sub r9d, ebx

    lea ebx, [r11 + r10]

    lea edx, [rsi + rcx]

    sub r10d, r11d

    sub ecx, esi

    mov dword ptr [rsp + rax], r14d

    mov dword ptr [rsp + rax + 4], edi

    mov dword ptr [rsp + rax + 8], ebx

    mov dword ptr [rsp + rax + 12], edx

    mov dword ptr [rsp + rax + 16], r8d

    mov dword ptr [rsp + rax + 20], r9d

    mov dword ptr [rsp + rax + 24], r10d

    mov dword ptr [rsp + rax + 28], ecx

    add rax, 32

    cmp rax, 256

    jne .LBB666_3

    vpabsd ymm0, ymmword ptr [rsp]

    vpabsd ymm1, ymmword ptr [rsp + 32]

    vpabsd ymm2, ymmword ptr [rsp + 96]

    vpaddd ymm0, ymm1, ymm0

    vpabsd ymm1, ymmword ptr [rsp + 64]

    vpaddd ymm1, ymm1, ymm2

    vpabsd ymm2, ymmword ptr [rsp + 160]

    vpaddd ymm0, ymm0, ymm1

    vpabsd ymm1, ymmword ptr [rsp + 128]

    vpaddd ymm1, ymm1, ymm2

    vpabsd ymm2, ymmword ptr [rsp + 192]

    vpaddd ymm1, ymm1, ymm2

    vpabsd ymm2, ymmword ptr [rsp + 224]

    vpaddd ymm0, ymm0, ymm1

    vpaddd ymm0, ymm0, ymm2

    vmovd eax, xmm0

    vpextrd ecx, xmm0, 1

    add rcx, rax

    vpextrd eax, xmm0, 2

    add rax, rcx

    vpextrd ecx, xmm0, 3

    vextracti128 xmm0, ymm0, 1

    add rcx, rax

    vmovd eax, xmm0
    vpextrd edx, xmm0, 2

    add rax, rcx

    vpextrd ecx, xmm0, 1

    add rcx, rax

    vpextrd eax, xmm0, 3

    add rdx, rcx

    add rax, rdx

    lea rsp, [rbp - 32]

    pop rbx
    pop r12
    pop r14
    pop r15
    pop rbp
    .cfi_def_cfa rsp, 8
    vzeroupper

    ret

.Lfunc_end666:
    .size   rav1e::asm::x86::dist::hbd::satd_kernel_8x8_hbd_avx2, .Lfunc_end666-rav1e::asm::x86::dist::hbd::satd_kernel_8x8_hbd_avx2