// Now we have X0-X3, X12-X15 available for temporaries.
	// Add seed rows back to output. We left seed in DI, SI, R8..R13 above.
	REPLREG(DI, X0)
	REPLREG(SI, X1)
	REPLREG(R8, X2)
	REPLREG(R9, X3)
	REPLREG(R10, X12)
	REPLREG(R11, X13)
	REPLREG(R12, X14)
	REPLREG(R13, X15)
	PADDD X0, X4
	PADDD X1, X5
	PADDD X2, X6
	PADDD X3, X7
	PADDD X12, X8
	PADDD X13, X9
	PADDD X14, X10