// then folded (accumulated) with the next data chunk in V5 and
	// stored in V1.  Repeat this step for the register contents
	// in V2, V3, and V4 respectively.

	VGFMAG  CONST_R2R1, V1, V5, V1
	VGFMAG  CONST_R2R1, V2, V6, V2
	VGFMAG  CONST_R2R1, V3, V7, V3
	VGFMAG  CONST_R2R1, V4, V8 ,V4

	// Adjust buffer pointer and length for next loop
	ADD     $64, R3                  // BUF = BUF + 64

	"WFSQDB",
	"VFS",
	"VFSDB",
	"WFSDB",
	"VFTCI",
	"VFTCIDB",
	"WFTCIDB",
	"VGFM",
	"VGFMB",
	"VGFMH",
	"VGFMF",
	"VGFMG",
	"VGFMA",
	"VGFMAB",
	"VGFMAH",
	"VGFMAF",
	"VGFMAG",
	"VGEF",
	"VGEG",
	"VGBM",
	"VZERO",
	"VONE",
	"VGM",
	"VGMB",
	"VGMH",
	"VGMF",
	"VGMG",
	"VISTR",
	"VISTRB",
	"VISTRH",
	"VISTRF",
	"VISTRBS",
	"VISTRHS",