LVX	(R3+off96),V6
	LVX	(R3+off112),V7

	ADD	$128,R3

	VPMSUMW	V16,V0,V0
	VPMSUMW	V17,V1,V1
	VPMSUMW	V18,V2,V2
	VPMSUMW	V19,V3,V3
	VPMSUMW	V20,V4,V4
	VPMSUMW	V21,V5,V5
	VPMSUMW	V22,V6,V6
	VPMSUMW	V23,V7,V7

	// now reduce the tail

	CMP	$0,R7
	BEQ	next1

	LVX	(R4),V16
	LVX	(R3),V17
	VPMSUMW	V16,V17,V16
	VXOR	V0,V16,V0
	BC	18,0,next1

	LVX	(R4+off16),V16

	"VMULEUB",
	"VMULOUB",
	"VMULESH",
	"VMULOSH",
	"VMULEUH",
	"VMULOUH",
	"VMULESW",
	"VMULOSW",
	"VMULEUW",
	"VMULOUW",
	"VMULUWM",
	"VPMSUM",
	"VPMSUMB",
	"VPMSUMH",
	"VPMSUMW",
	"VPMSUMD",
	"VMSUMUDM",
	"VR",
	"VRLB",
	"VRLH",
	"VRLW",
	"VRLD",
	"VS",
	"VSLB",
	"VSLH",
	"VSLW",
	"VSL",
	"VSLO",
	"VSRB",
	"VSRH",
	"VSRW",
	"VSR",

	VMULOUW V1, V2, V3              // 10611088
	VMULUWM V1, V2, V3              // 10611089
	VPMSUMB V1, V2, V3              // 10611408
	VPMSUMH V1, V2, V3              // 10611448
	VPMSUMW V1, V2, V3              // 10611488
	VPMSUMD V1, V2, V3              // 106114c8
	VMSUMUDM V1, V2, V3, V4         // 108110e3
	VRLB V1, V2, V3                 // 10611004
	VRLH V1, V2, V3                 // 10611044

	VPKSDUS:        "vpksdus",
	VPKUDUM:        "vpkudum",
	VPKUDUS:        "vpkudus",
	VPMSUMB:        "vpmsumb",
	VPMSUMD:        "vpmsumd",
	VPMSUMH:        "vpmsumh",
	VPMSUMW:        "vpmsumw",
	VPOPCNTB:       "vpopcntb",
	VPOPCNTD:       "vpopcntd",
	VPOPCNTH:       "vpopcnth",
	VPOPCNTW:       "vpopcntw",
	VRLD:           "vrld",
	VSBOX:          "vsbox",

			opset(AVMULEUH, r0)
			opset(AVMULOUH, r0)
			opset(AVMULESW, r0)
			opset(AVMULOSW, r0)
			opset(AVMULEUW, r0)
			opset(AVMULOUW, r0)
			opset(AVMULUWM, r0)
		case AVPMSUM: /* vpmsumb, vpmsumh, vpmsumw, vpmsumd */
			opset(AVPMSUMB, r0)
			opset(AVPMSUMH, r0)
			opset(AVPMSUMW, r0)
			opset(AVPMSUMD, r0)

		case AVR: /* vrlb, vrlh, vrlw, vrld */
			opset(AVRLB, r0)
			opset(AVRLH, r0)