MOVOU yout(16*0), X2
	MOVOU yout(16*1), X3
	MOVOU zout(16*0), X4
	MOVOU zout(16*1), X5

	MOVL BX, X6
	MOVL CX, X7

	PXOR X8, X8
	PCMPEQL X9, X9

	PSHUFD $0, X6, X6
	PSHUFD $0, X7, X7

	PCMPEQL X8, X6
	PCMPEQL X8, X7

	MOVOU X6, X15
	PANDN X9, X15

	MOVOU x1in(16*0), X9
	MOVOU x1in(16*1), X10
	MOVOU y1in(16*0), X11
	MOVOU y1in(16*1), X12

	"PADDSB",
	"PADDSW",
	"PADDUSB",
	"PADDUSW",
	"PADDW",
	"PALIGNR",
	"PAND",
	"PANDN",
	"PAUSE",
	"PAVGB",
	"PAVGW",
	"PBLENDVB",
	"PBLENDW",
	"PCLMULQDQ",
	"PCMPEQB",
	"PCMPEQL",
	"PCMPEQQ",
	"PCMPEQW",
	"PCMPESTRI",
	"PCMPESTRM",
	"PCMPGTB",
	"PCMPGTL",
	"PCMPGTQ",
	"PCMPGTW",
	"PCMPISTRI",
	"PCMPISTRM",
	"PDEPL",
	"PDEPQ",
	"PEXTL",
	"PEXTQ",