PRECALC_32_79(Y8,Y12,Y13,Y15,Y7,0x40,0x180) \
	PRECALC_32_79(Y7,Y8,Y12,Y14,Y5,0x40,0x1a0) \
	PRECALC_32_79(Y5,Y7,Y8,Y13,Y3,0x40,0x1c0) \
	PRECALC_32_79(Y3,Y5,Y7,Y12,Y15,0x60,0x1e0) \
	PRECALC_32_79(Y15,Y3,Y5,Y8,Y14,0x60,0x200) \
	PRECALC_32_79(Y14,Y15,Y3,Y7,Y13,0x60,0x220) \
	PRECALC_32_79(Y13,Y14,Y15,Y5,Y12,0x60,0x240) \
	PRECALC_32_79(Y12,Y13,Y14,Y3,Y8,0x60,0x260)

	VPGATHERQQ X12, (R13)(X14*2), X11       // c40299915c7500
	VPGATHERQQ Y12, (R13)(Y14*2), Y11       // c4029d915c7500
	VPGATHERQQ X2, (BP)(X7*2), X1           // c4e2e9914c7d00
	VPGATHERQQ Y2, (BP)(Y7*2), Y1           // c4e2ed914c7d00
	VPGATHERQQ X12, (R13)(X14*2), X11       // c40299915c7500
	VPGATHERQQ Y12, (R13)(Y14*2), Y11       // c4029d915c7500
	VGATHERDPD X2, (BP)(X7*2), X1           // c4e2e9924c7d00

	VPGATHERDQ Y5, 664(DX*8), Y6    // ERROR "invalid instruction"
	VPGATHERDQ Y5, (DX), Y0         // ERROR "invalid instruction"
	// VM/X rejects Y index register.
	VPGATHERDQ Y5, 664(Y14*8), Y6   // ERROR "invalid instruction"
	VPGATHERQQ X2, (BP)(Y7*2), X1   // ERROR "invalid instruction"
	// VM/Y rejects X index register.
	VPGATHERQQ Y2, (BP)(X7*2), Y1   // ERROR "invalid instruction"

#define ctr2StoreAVX2 (5*32)(BP)
#define ctr3StoreAVX2 (6*32)(BP)
#define tmpStoreAVX2 (7*32)(BP) // 256 bytes on stack
#define AA0 Y0
#define AA1 Y5
#define AA2 Y6
#define AA3 Y7
#define BB0 Y14
#define BB1 Y9
#define BB2 Y10
#define BB3 Y11
#define CC0 Y12
#define CC1 Y13
#define CC2 Y8
#define CC3 Y15
#define DD0 Y4
#define DD1 Y1
#define DD2 Y2
#define DD3 Y3
#define TT0 DD3