ORQ  R12, DI
	ADDQ R14, R10

	ADDQ R15, CX
	ADDQ R15, R10
	ADDQ DI, R10

	VPSRLQ $19, Y2, Y3
	VPSLLQ $(64-19), Y2, Y1
	VPOR   Y1, Y3, Y3
	VPXOR  Y3, Y8, Y8
	VPSRLQ $61, Y2, Y3
	VPSLLQ $(64-61), Y2, Y1
	VPOR   Y1, Y3, Y3
	VPXOR  Y3, Y8, Y8

	VPADDQ Y8, Y4, Y4

	VPSRLQ $6, Y4, Y8

	MOVQ  R10, DI
	RORXQ $41, CX, R13
	ADDQ  2*8+frame_YFER(SP), R9

	VPSRLQ 17(SP)(BP*2), Y13, K3, Y24                  // 6261952bd3846c11000000
	VPSRLQ -7(DI)(R8*4), Y13, K3, Y24                  // 6221952bd38487f9ffffff
	VPSRLQ X30, Z27, K3, Z23                           // 6281a543d3fe
	VPSRLQ 15(R8), Z27, K3, Z23                        // 62c1a543d3b80f000000
	VPSRLQ (BP), Z27, K3, Z23                          // 62e1a543d37d00
	VPSRLQ X30, Z14, K3, Z23                           // 62818d4bd3fe

	RORXL   $25, e, y0;                  \ // y0 = e >> 25						// S1A
	ADDL    (disp + 2*4)(SP)(SRND*1), h; \ // h = k + w + h			// --
	;                                    \
	VPSRLQ  $19, XTMP2, XTMP3;           \ // XTMP3 = W[-2] ror 19 {xBxA}
	RORXL   $11, e, y1;                  \ // y1 = e >> 11						// S1B
	ORL     c, y3;                       \ // y3 = a|c                         // MAJA

	"VPSLLQ",
	"VPSLLVD",
	"VPSLLVQ",
	"VPSLLVW",
	"VPSLLW",
	"VPSRAD",
	"VPSRAQ",
	"VPSRAVD",
	"VPSRAVQ",
	"VPSRAVW",
	"VPSRAW",
	"VPSRLD",
	"VPSRLDQ",
	"VPSRLQ",
	"VPSRLVD",
	"VPSRLVQ",
	"VPSRLVW",
	"VPSRLW",
	"VPSUBB",
	"VPSUBD",
	"VPSUBQ",
	"VPSUBSB",
	"VPSUBSW",
	"VPSUBUSB",
	"VPSUBUSW",
	"VPSUBW",
	"VPTERNLOGD",

	VPSRLD $-1, X1, X2           // c5e972d1ff
	VPSRLD $-1, Y1, Y2           // c5ed72d1ff
	VPSRLDQ $-1, X1, X2          // c5e973d9ff
	VPSRLDQ $-1, Y1, Y2          // c5ed73d9ff
	VPSRLQ $-1, X1, X2           // c5e973d1ff
	VPSRLQ $-1, Y1, Y2           // c5ed73d1ff
	VPEXTRW $-1, X1, (AX)        // c4e3791508ff
	VPEXTRW $-1, X1, AX          // c4e37915c8ff
	VEXTRACTF128 $-1, Y1, X2     // c4e37d19caff

	VPSRLDQ $7, Y11, Y15            // c4c10573db07
	VPSRLQ (BX), X9, X2             // c4e131d313 or c5b1d313
	VPSRLQ (R11), X9, X2            // c4c131d313
	VPSRLQ X2, X9, X2               // c4e131d3d2 or c5b1d3d2
	VPSRLQ X11, X9, X2              // c4c131d3d3
	VPSRLQ (BX), X9, X11            // c46131d31b or c531d31b
	VPSRLQ (R11), X9, X11           // c44131d31b
	VPSRLQ X2, X9, X11              // c46131d3da or c531d3da