bits 64
    
    vpdpbusd xmm2, xmm7, xmm0
    vpdpbusd xmm2, xmm7, [rbx]
    vpdpbusd xmm2, xmm7, [rbx]{1to4}
    vpdpbusd xmm2, xmm7, [rbx+r11*8+256]
    vpdpbusd xmm2, xmm7, [rbx+r11*8-256]
    vpdpbusd xmm2{k5}, xmm7, xmm0
    vpdpbusd xmm2{k5}, xmm7, [rbx]
    vpdpbusd xmm2{k5}, xmm7, [rbx]{1to4}
    vpdpbusd xmm2{k5}, xmm7, [rbx+r11*8+256]
    vpdpbusd xmm2{k5}, xmm7, [rbx+r11*8-256]
    vpdpbusd xmm2{k5}{z}, xmm7, xmm0
    vpdpbusd xmm2{k5}{z}, xmm7, [rbx]
    vpdpbusd xmm2{k5}{z}, xmm7, [rbx]{1to4}
    vpdpbusd xmm2{k5}{z}, xmm7, [rbx+r11*8+256]
    vpdpbusd xmm2{k5}{z}, xmm7, [rbx+r11*8-256]
    vpdpbusd ymm16, ymm13, ymm15
    vpdpbusd ymm16, ymm13, [rbx]
    vpdpbusd ymm16, ymm13, [rbx]{1to8}
    vpdpbusd ymm16, ymm13, [rbx+r11*8+256]
    vpdpbusd ymm16, ymm13, [rbx+r11*8-256]
    vpdpbusd ymm16{k5}, ymm13, ymm15
    vpdpbusd ymm16{k5}, ymm13, [rbx]
    vpdpbusd ymm16{k5}, ymm13, [rbx]{1to8}
    vpdpbusd ymm16{k5}, ymm13, [rbx+r11*8+256]
    vpdpbusd ymm16{k5}, ymm13, [rbx+r11*8-256]
    vpdpbusd ymm16{k5}{z}, ymm13, ymm15
    vpdpbusd ymm16{k5}{z}, ymm13, [rbx]
    vpdpbusd ymm16{k5}{z}, ymm13, [rbx]{1to8}
    vpdpbusd ymm16{k5}{z}, ymm13, [rbx+r11*8+256]
    vpdpbusd ymm16{k5}{z}, ymm13, [rbx+r11*8-256]
    vpdpbusd zmm24, zmm24, zmm31
    vpdpbusd zmm24, zmm24, [rbx]
    vpdpbusd zmm24, zmm24, [rbx]{1to16}
    vpdpbusd zmm24, zmm24, [rbx+r11*8+256]
    vpdpbusd zmm24, zmm24, [rbx+r11*8-256]
    vpdpbusd zmm24{k5}, zmm24, zmm31
    vpdpbusd zmm24{k5}, zmm24, [rbx]
    vpdpbusd zmm24{k5}, zmm24, [rbx]{1to16}
    vpdpbusd zmm24{k5}, zmm24, [rbx+r11*8+256]
    vpdpbusd zmm24{k5}, zmm24, [rbx+r11*8-256]
    vpdpbusd zmm24{k5}{z}, zmm24, zmm31
    vpdpbusd zmm24{k5}{z}, zmm24, [rbx]
    vpdpbusd zmm24{k5}{z}, zmm24, [rbx]{1to16}
    vpdpbusd zmm24{k5}{z}, zmm24, [rbx+r11*8+256]
    vpdpbusd zmm24{k5}{z}, zmm24, [rbx+r11*8-256]
    vpdpbusds xmm2, xmm7, xmm0
    vpdpbusds xmm2, xmm7, [rbx]
    vpdpbusds xmm2, xmm7, [rbx]{1to4}
    vpdpbusds xmm2, xmm7, [rbx+r11*8+256]
    vpdpbusds xmm2, xmm7, [rbx+r11*8-256]
    vpdpbusds xmm2{k5}, xmm7, xmm0
    vpdpbusds xmm2{k5}, xmm7, [rbx]
    vpdpbusds xmm2{k5}, xmm7, [rbx]{1to4}
    vpdpbusds xmm2{k5}, xmm7, [rbx+r11*8+256]
    vpdpbusds xmm2{k5}, xmm7, [rbx+r11*8-256]
    vpdpbusds xmm2{k5}{z}, xmm7, xmm0
    vpdpbusds xmm2{k5}{z}, xmm7, [rbx]
    vpdpbusds xmm2{k5}{z}, xmm7, [rbx]{1to4}
    vpdpbusds xmm2{k5}{z}, xmm7, [rbx+r11*8+256]
    vpdpbusds xmm2{k5}{z}, xmm7, [rbx+r11*8-256]
    vpdpbusds ymm16, ymm13, ymm15
    vpdpbusds ymm16, ymm13, [rbx]
    vpdpbusds ymm16, ymm13, [rbx]{1to8}
    vpdpbusds ymm16, ymm13, [rbx+r11*8+256]
    vpdpbusds ymm16, ymm13, [rbx+r11*8-256]
    vpdpbusds ymm16{k5}, ymm13, ymm15
    vpdpbusds ymm16{k5}, ymm13, [rbx]
    vpdpbusds ymm16{k5}, ymm13, [rbx]{1to8}
    vpdpbusds ymm16{k5}, ymm13, [rbx+r11*8+256]
    vpdpbusds ymm16{k5}, ymm13, [rbx+r11*8-256]
    vpdpbusds ymm16{k5}{z}, ymm13, ymm15
    vpdpbusds ymm16{k5}{z}, ymm13, [rbx]
    vpdpbusds ymm16{k5}{z}, ymm13, [rbx]{1to8}
    vpdpbusds ymm16{k5}{z}, ymm13, [rbx+r11*8+256]
    vpdpbusds ymm16{k5}{z}, ymm13, [rbx+r11*8-256]
    vpdpbusds zmm24, zmm24, zmm31
    vpdpbusds zmm24, zmm24, [rbx]
    vpdpbusds zmm24, zmm24, [rbx]{1to16}
    vpdpbusds zmm24, zmm24, [rbx+r11*8+256]
    vpdpbusds zmm24, zmm24, [rbx+r11*8-256]
    vpdpbusds zmm24{k5}, zmm24, zmm31
    vpdpbusds zmm24{k5}, zmm24, [rbx]
    vpdpbusds zmm24{k5}, zmm24, [rbx]{1to16}
    vpdpbusds zmm24{k5}, zmm24, [rbx+r11*8+256]
    vpdpbusds zmm24{k5}, zmm24, [rbx+r11*8-256]
    vpdpbusds zmm24{k5}{z}, zmm24, zmm31
    vpdpbusds zmm24{k5}{z}, zmm24, [rbx]
    vpdpbusds zmm24{k5}{z}, zmm24, [rbx]{1to16}
    vpdpbusds zmm24{k5}{z}, zmm24, [rbx+r11*8+256]
    vpdpbusds zmm24{k5}{z}, zmm24, [rbx+r11*8-256]
    vpdpwssd xmm2, xmm7, xmm0
    vpdpwssd xmm2, xmm7, [rbx]
    vpdpwssd xmm2, xmm7, [rbx]{1to4}
    vpdpwssd xmm2, xmm7, [rbx+r11*8+256]
    vpdpwssd xmm2, xmm7, [rbx+r11*8-256]
    vpdpwssd xmm2{k5}, xmm7, xmm0
    vpdpwssd xmm2{k5}, xmm7, [rbx]
    vpdpwssd xmm2{k5}, xmm7, [rbx]{1to4}
    vpdpwssd xmm2{k5}, xmm7, [rbx+r11*8+256]
    vpdpwssd xmm2{k5}, xmm7, [rbx+r11*8-256]
    vpdpwssd xmm2{k5}{z}, xmm7, xmm0
    vpdpwssd xmm2{k5}{z}, xmm7, [rbx]
    vpdpwssd xmm2{k5}{z}, xmm7, [rbx]{1to4}
    vpdpwssd xmm2{k5}{z}, xmm7, [rbx+r11*8+256]
    vpdpwssd xmm2{k5}{z}, xmm7, [rbx+r11*8-256]
    vpdpwssd ymm16, ymm13, ymm15
    vpdpwssd ymm16, ymm13, [rbx]
    vpdpwssd ymm16, ymm13, [rbx]{1to8}
    vpdpwssd ymm16, ymm13, [rbx+r11*8+256]
    vpdpwssd ymm16, ymm13, [rbx+r11*8-256]
    vpdpwssd ymm16{k5}, ymm13, ymm15
    vpdpwssd ymm16{k5}, ymm13, [rbx]
    vpdpwssd ymm16{k5}, ymm13, [rbx]{1to8}
    vpdpwssd ymm16{k5}, ymm13, [rbx+r11*8+256]
    vpdpwssd ymm16{k5}, ymm13, [rbx+r11*8-256]
    vpdpwssd ymm16{k5}{z}, ymm13, ymm15
    vpdpwssd ymm16{k5}{z}, ymm13, [rbx]
    vpdpwssd ymm16{k5}{z}, ymm13, [rbx]{1to8}
    vpdpwssd ymm16{k5}{z}, ymm13, [rbx+r11*8+256]
    vpdpwssd ymm16{k5}{z}, ymm13, [rbx+r11*8-256]
    vpdpwssd zmm24, zmm24, zmm31
    vpdpwssd zmm24, zmm24, [rbx]
    vpdpwssd zmm24, zmm24, [rbx]{1to16}
    vpdpwssd zmm24, zmm24, [rbx+r11*8+256]
    vpdpwssd zmm24, zmm24, [rbx+r11*8-256]
    vpdpwssd zmm24{k5}, zmm24, zmm31
    vpdpwssd zmm24{k5}, zmm24, [rbx]
    vpdpwssd zmm24{k5}, zmm24, [rbx]{1to16}
    vpdpwssd zmm24{k5}, zmm24, [rbx+r11*8+256]
    vpdpwssd zmm24{k5}, zmm24, [rbx+r11*8-256]
    vpdpwssd zmm24{k5}{z}, zmm24, zmm31
    vpdpwssd zmm24{k5}{z}, zmm24, [rbx]
    vpdpwssd zmm24{k5}{z}, zmm24, [rbx]{1to16}
    vpdpwssd zmm24{k5}{z}, zmm24, [rbx+r11*8+256]
    vpdpwssd zmm24{k5}{z}, zmm24, [rbx+r11*8-256]
    vpdpwssds xmm2, xmm7, xmm0
    vpdpwssds xmm2, xmm7, [rbx]
    vpdpwssds xmm2, xmm7, [rbx]{1to4}
    vpdpwssds xmm2, xmm7, [rbx+r11*8+256]
    vpdpwssds xmm2, xmm7, [rbx+r11*8-256]
    vpdpwssds xmm2{k5}, xmm7, xmm0
    vpdpwssds xmm2{k5}, xmm7, [rbx]
    vpdpwssds xmm2{k5}, xmm7, [rbx]{1to4}
    vpdpwssds xmm2{k5}, xmm7, [rbx+r11*8+256]
    vpdpwssds xmm2{k5}, xmm7, [rbx+r11*8-256]
    vpdpwssds xmm2{k5}{z}, xmm7, xmm0
    vpdpwssds xmm2{k5}{z}, xmm7, [rbx]
    vpdpwssds xmm2{k5}{z}, xmm7, [rbx]{1to4}
    vpdpwssds xmm2{k5}{z}, xmm7, [rbx+r11*8+256]
    vpdpwssds xmm2{k5}{z}, xmm7, [rbx+r11*8-256]
    vpdpwssds ymm16, ymm13, ymm15
    vpdpwssds ymm16, ymm13, [rbx]
    vpdpwssds ymm16, ymm13, [rbx]{1to8}
    vpdpwssds ymm16, ymm13, [rbx+r11*8+256]
    vpdpwssds ymm16, ymm13, [rbx+r11*8-256]
    vpdpwssds ymm16{k5}, ymm13, ymm15
    vpdpwssds ymm16{k5}, ymm13, [rbx]
    vpdpwssds ymm16{k5}, ymm13, [rbx]{1to8}
    vpdpwssds ymm16{k5}, ymm13, [rbx+r11*8+256]
    vpdpwssds ymm16{k5}, ymm13, [rbx+r11*8-256]
    vpdpwssds ymm16{k5}{z}, ymm13, ymm15
    vpdpwssds ymm16{k5}{z}, ymm13, [rbx]
    vpdpwssds ymm16{k5}{z}, ymm13, [rbx]{1to8}
    vpdpwssds ymm16{k5}{z}, ymm13, [rbx+r11*8+256]
    vpdpwssds ymm16{k5}{z}, ymm13, [rbx+r11*8-256]
    vpdpwssds zmm24, zmm24, zmm31
    vpdpwssds zmm24, zmm24, [rbx]
    vpdpwssds zmm24, zmm24, [rbx]{1to16}
    vpdpwssds zmm24, zmm24, [rbx+r11*8+256]
    vpdpwssds zmm24, zmm24, [rbx+r11*8-256]
    vpdpwssds zmm24{k5}, zmm24, zmm31
    vpdpwssds zmm24{k5}, zmm24, [rbx]
    vpdpwssds zmm24{k5}, zmm24, [rbx]{1to16}
    vpdpwssds zmm24{k5}, zmm24, [rbx+r11*8+256]
    vpdpwssds zmm24{k5}, zmm24, [rbx+r11*8-256]
    vpdpwssds zmm24{k5}{z}, zmm24, zmm31
    vpdpwssds zmm24{k5}{z}, zmm24, [rbx]
    vpdpwssds zmm24{k5}{z}, zmm24, [rbx]{1to16}
    vpdpwssds zmm24{k5}{z}, zmm24, [rbx+r11*8+256]
    vpdpwssds zmm24{k5}{z}, zmm24, [rbx+r11*8-256]