47 #define G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
48   row1l = _mm_add_epi64(_mm_add_epi64(row1l, b0), row2l); \
60   row2l = _mm_xor_si128(row2l, row3l); \
63   row2l = _mm_roti_epi64(row2l, -24); \
66 #define G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
67   row1l = _mm_add_epi64(_mm_add_epi64(row1l, b0), row2l); \
79   row2l = _mm_xor_si128(row2l, row3l); \
82   row2l = _mm_roti_epi64(row2l, -63); \
86 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
87   t0 = _mm_alignr_epi8(row2h, row2l, 8); \
88   t1 = _mm_alignr_epi8(row2l, row2h, 8); \
89   row2l = t0; \
101 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
102   t0 = _mm_alignr_epi8(row2l, row2h, 8); \
103   t1 = _mm_alignr_epi8(row2h, row2l, 8); \
104   row2l = t0; \
117 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
119   t1 = row2l;\
125   row2l = _mm_unpackhi_epi64(row2l, _mm_unpacklo_epi64(row2h, row2h)); \
128 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
132   t0 = row2l;\
134   row2l = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(row2l, row2l)); \
149   G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
151   G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
152   DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
154   G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
156   G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
157   UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h);