200 #define transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \  argument
203 	vpunpckldq		x3, x2, t1; \
204 	vpunpckhdq		x3, x2, x3; \
208 	vpunpcklqdq		x3, t2, x2; \
209 	vpunpckhqdq		x3, t2, x3;
211 #define inpack_blocks(x0, x1, x2, x3, t0, t1, t2, rmask) \  argument
215 	vpshufb rmask, x3,	x3; \
217 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2)
219 #define outunpack_blocks(x0, x1, x2, x3, t0, t1, t2, rmask) \  argument
220 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
225 	vpshufb rmask,		x3, x3;